Content collaborative caching method and system based on digital twinning

Abstract

The invention discloses a content collaborative caching method and system based on digital twinning, and belongs to the technical field of industrial Internet of Things. According to the system, an efficient content collaborative caching strategy is designed according to dynamic content request characteristics of the mobile terminal, calculation and storage resources of edge nodes can be fully utilized through global scheduling of a digital twin layer, the time delay and transmission cost of content acquisition of the mobile terminal are reduced, and the overall service quality and income of the system are improved. According to the method, a real-time interactive digital twinborn layer is constructed based on multi-dimensional data of a terminal and an edge, and real-time sharing of virtual data between different base stations and updating and maintenance of a digital twinborn model are realized; through organic combination of content popularity prediction and global collaborative caching decision based on digital twinning, a global optimal caching strategy is obtained. According to the method, the perception capability of global information can be enhanced, the reliability of system state synchronization is improved, the precision of the global digital twinborn model is improved, and the global optimality of a cache decision is ensured.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) technology, and in particular to a content collaborative caching method and system based on digital twins. Background Technology

[0002] In the Industrial Internet of Things (IIoT) scenario, terminal devices typically need to retrieve large amounts of instructions or business data from the cloud to maintain normal production operations. However, traditional cloud-based data retrieval models are limited by long-distance transmission, often accompanied by high backhaul latency. In recent years, edge caching technology has been widely used. By pre-storing the content required by the terminal on edge nodes, the data transmission path can be effectively shortened, improving system response efficiency. However, due to the limited caching resources of edge nodes, it is difficult to support the distributed storage of all content. Furthermore, the high mobility and dynamic content requests of mobile devices in industrial scenarios make it difficult for static caching strategies to maintain a high hit rate, ultimately requiring requests to the cloud and generating additional latency. Although collaborative caching technology expands the ways terminals retrieve content through cloud, edge, and terminal resource sharing and collaborative scheduling, the additional communication overhead and transmission latency generated during cross-node collaboration remain bottlenecks restricting system real-time performance. Therefore, how to design an efficient caching strategy to achieve seamless collaboration between mobile terminals, edge nodes, and the cloud has become an urgent problem to be solved.

[0003] Content popularity prediction provides key insights for optimizing collaborative caching. By analyzing differentiated request preferences on device sides and capturing demand trends, it's possible to rationally allocate limited cache resources. This not only significantly reduces the overhead of long-distance transmission but also effectively improves edge cache hit rate and reduces content retrieval latency. However, existing technologies often focus on caching globally frequently requested popular content. While this meets most general needs, in the IIoT vertical domain, the operation of all mobile terminals serves the overall factory's efficiency improvement and cost control. Due to differences in production processes, devices exhibit significant priority differences: high-priority devices have stricter real-time requirements for task processing and need to acquire the required content first. Since there is no positive correlation between device priority and content popularity, relying solely on traditional popularity prediction mechanisms often overlooks the specific needs of critical tasks. Therefore, when making edge caching decisions, it's crucial to consider not only content popularity but also device priority. Device priority should be incorporated as a constraint or key factor into the training and decision-making models of collaborative caching to ensure that resources required for core business operations receive priority response and protection.

[0004] Furthermore, existing collaborative caching strategies based on multi-agent reinforcement learning are mostly executed directly in the physical environment. Edge nodes are limited by their perception range and can only train strategies based on local information, resulting in slow model convergence and difficulty in reaching the global optimum. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a content collaborative caching method and system based on digital twins. By organically combining content popularity prediction based on digital twins with global collaborative caching decision-making, the cache hit rate and system benefits are improved, and the efficient operation of the content collaborative caching system in the industrial Internet of Things scenario is realized.

[0006] Technical Solution: A content collaborative caching system based on digital twins includes terminal devices deployed at the terminal layer, edge servers deployed at the edge layer, and a digital twin layer. The terminal devices send real-time content requests and related operational data to the edge servers and receive specific content from the edge servers based on the requests. Upon receiving the requests from the terminal devices, the edge servers perform two actions: first, they obtain the optimal caching strategy using a locally deployed strategy model; second, they continue to upload data to the digital twin layer. Simultaneously, the edge servers receive the optimal strategy model from the digital twin layer. After receiving various types of data from the edge servers, the digital twin layer updates and maintains the global digital twin model, extracts latent features using the AE model to obtain predicted content, and inputs the predicted content as prior knowledge into the virtual agent to train the strategy model, thereby obtaining the globally optimal caching strategy model.

[0007] Furthermore, it also includes:

[0008] The data collection module, deployed at the terminal layer, senses and collects various operational data and content request data during the device's operation through a gateway installed on the mobile device.

[0009] The priority module, deployed at the terminal layer, divides mobile devices into multiple priority levels according to preset rules and ensures that each device has a fixed priority attribute.

[0010] The edge data module, deployed at the edge layer, receives basic device data and content request information from the data collection module, forwards the basic device data and content request information to the content popularity prediction module of the digital twin layer for predictive analysis, and forwards the basic device data and content request information to the global collaborative cache training module for training to obtain the optimal cache strategy model.

[0011] The cost design module, deployed at the edge layer, builds an edge base station content delivery cost model and a priority-based total system cost model, and expresses the system optimization objectives in a formulaic way.

[0012] Markov Decision Design Module: Deployed at the edge layer, it models the agent's cache replacement process as a Markov decision process, defines the agent's state space and action space, and designs a reward function based on a formulaic optimization objective to guide the agent's training direction;

[0013] The content caching module, deployed at the edge layer, receives the optimal caching strategy model obtained from the global collaborative caching training module, obtains the optimal caching strategy based on Markov decision design and state data, and performs caching operations on the corresponding content.

[0014] The content popularity prediction module is deployed in the digital twin layer. It uses the AE model to build a content popularity evaluation matrix based on the device's historical request information, and selects high-evaluation content as predicted popular content. The predicted popular content is then used as prior knowledge and input into the global collaborative cache training module.

[0015] The global collaborative caching training module is deployed in the digital twin layer. It trains the global collaborative caching strategy model based on the content popularity prediction results and the received global information. During the global collaborative caching training process, each agent cooperates by observing each other's states and finally solves the global optimal caching strategy model.

[0016] A content collaborative caching method based on digital twins is applied to any of the aforementioned content collaborative caching systems based on digital twins. By organically combining content popularity prediction based on digital twins with global collaborative caching decisions, a globally optimal caching strategy is obtained. The specific steps are as follows:

[0017] S1, each mobile device sends its location, remaining power, and content caching requests to the edge base stations under its coverage area;

[0018] S2, after receiving data sent by the mobile device, the edge base station constructs a digital twin layer. Different edge base stations upload data from mobile devices within their coverage area to the digital twin layer, and virtual data is shared between different base stations;

[0019] S3, the digital twin layer, monitors the system's operating status in each round, predicts the content request preferences of mobile devices, and stores the predicted content in the training data of the global collaborative cache;

[0020] S4, the edge base station designs the total system cost model according to priority, formulates the optimization objective, and establishes the system problem formula;

[0021] S5. Based on the total cost model and the system problem formula, design a Markov decision process;

[0022] S6, based on the given problem and Markov decision process, train the global collaborative caching strategy model in the digital twin layer to obtain the globally optimal caching strategy model;

[0023] S7: The digital twin layer distributes the globally optimal caching strategy model to the edge layer. The edge layer, based on the state data and Markov decision design, derives the optimal strategy and caches the corresponding content.

[0024] Furthermore, the digital twin layer predicts the content request preferences of mobile devices and stores the predicted content in the training data of a global collaborative cache, specifically including:

[0025] S31, Design the backtracking time window for the r-th round. The system predicts future content demand based on data within the digital twin of the window, and the window size is adaptively adjusted based on the cache hit rate.

[0026] S32, in the backtracking time window Within, popular content is predicted based on historical request information from each mobile device;

[0027] S33, the predicted popular content is used as prior knowledge and input into the training data of the global collaborative cache.

[0028] Furthermore, the backtracking time window for the r-th round is obtained. The steps include:

[0029] S311, Design the backtracking time ratio coefficient for the r-th round. :

[0030] ,

[0031] in, , These represent the lower and upper bounds of the proportionality coefficient, respectively. Let represent the cache hit rate in round r-1, and ;

[0032] S312, based on the backtracking time scaling factor Design the final backtracking time window for:

[0033] ,

[0034] in, This indicates the current moment, which is the moment when round (r-1) ends and round r is about to begin; This represents the total duration of the actual training in the (r-1)th round.

[0035] Furthermore, the specific steps for predicting popular content based on historical request information from each mobile device include:

[0036] S321, determine whether a certain request time is within the backtracking time window. If the request time... satisfy:

[0037] ,

[0038] This indicates that the request occurred during the last segment of the previous round, and is a high-value recent feature, which should be retained and input into the AE model.

[0039] S322, the AE model deployed in the digital twin layer extracts historical request information of each mobile device within the jurisdiction of each base station within the retrospective time window, and constructs a content popularity evaluation matrix for each device. ;in The first dimension is the identity characteristics of the mobile device, and the second dimension is the evaluation level of the content by the mobile device.

[0040] S323: Sort the content of mobile devices under each base station in descending order of level, and select the top N content with the highest evaluation level as the predicted popular content, where the size of N is determined by the cache capacity of the base station.

[0041] Furthermore, the specific steps for designing the total system cost model for edge base stations based on priorities include:

[0042] S41, Design the overall system model. Mobile devices send content requests to the local base station, and there are three ways to provide content. If the local base station stores the content, it transmits it directly; if a neighboring base station stores the content, it transmits it to the local base station; if neither the local base station nor a neighboring base station stores the content, the cloud transmits it to the local base station.

[0043] S42, after the mobile device sends a content request to the base station, the cost of the base station fulfilling the content transmission request is calculated as follows:

[0044] If the local base station to which the mobile device is connected has cached the content in advance, the content will be transmitted directly through the wireless link. The transmission cost is denoted as the first transmission cost. , is represented as:

[0045] ,

[0046] in, This represents the unit cost of transmitting content using a wireless link. Content Data size, This indicates the wireless link transmission rate between the base station and the mobile device;

[0047] If the local base station connected to the mobile device cannot provide content. However, if a neighboring base station caches this content, it is transmitted to the local base station via fiber optic cable, and then distributed via wireless link. This method of content delivery... The transmission cost is denoted as the second transmission cost. , is represented as:

[0048] ,

[0049] in, This represents the unit cost of transmitting content using a wireless link. This indicates the wired link transmission rate between the local base station and its neighboring base stations;

[0050] If content If the content is neither cached at the local base station nor at a neighboring base station, the local base station needs to request the content from the cloud and transmit it to the corresponding mobile device. This method of content... The transmission cost is denoted as the third transmission cost. , is represented as:

[0051] ,

[0052] in, This indicates the wired link transmission rate between the local base station and the cloud;

[0053] Finally, the total content transmission cost per base station Represented as:

[0054] ,

[0055] in, This represents the set of devices within the coverage area of ​​the base station. A collection of content. Indicates whether device u requests this content from the base station. ; , respectively, indicate that the content request of device u is provided by the local base station, neighboring base stations, and the cloud, and the selection of the service method satisfies the constraints. ;

[0056] S43, Design a priority-based service cost model, including:

[0057] S431 divides all mobile devices into eight priority levels based on their importance. Indicates the priority of device u;

[0058] S432, designing the priority index for a single base station. for:

[0059] ,

[0060] in, This represents the cache hit rate of device u at this base station. This represents the total number of mobile devices under a single base station. Indicates the highest priority; The larger the value, the more priorities the strategy considers; conversely, the smaller the value, the fewer priorities it considers.

[0061] S433, Calculate the total system cost for:

[0062] ;

[0063] S44, Design the revenue model, including:

[0064] Calculate the waiting time for the mobile device to receive content after sending a request. Represented as:

[0065] ;

[0066] The revenue from base station design is inversely proportional to the content transmission latency; the revenue function... Represented as:

[0067] ,

[0068] in, Indicates the attenuation factor. Content Value;

[0069] Define an index to determine whether a cached request is satisfied within a specified time. for:

[0070] ,

[0071] in, This is the time limit for waiting for content requests; once the content is successfully transmitted, the index... If it fails, then index ;

[0072] Calculate the revenue of the base station , is represented as:

[0073] ;

[0074] S45, by maximizing the total revenue of a single base station Find the optimal caching strategy The optimization problem can then be expressed as:

[0075] ,

[0076] in, It is the total bandwidth of the base station. It is the bandwidth currently allocated to mobile device u by the edge base station. It is the total bandwidth of the base station. Indicates whether the base station caches content. For example, "yes" takes a value of 1, and "no" takes a value of 0; Represents a single content Size, This refers to the base station's cache capacity; constraint C1 represents the base station's total bandwidth limit; constraint C2 indicates that the base station needs to consider storage capacity limitations when caching content; and constraint C3 defines that the content requested by the mobile terminal can only be retrieved in one place.

[0077] Furthermore, each edge base station is modeled as an intelligent agent, and the Markov decision process is defined as follows:

[0078] Define the state of each base station as ,in This indicates the local cache status of the base station. The state space represents the received content request status. for: ,and , Indicates the total number of base stations;

[0079] Define the action in each base station as ,in For base station caching actions, Indicates the caching service method; then the action. for: ;

[0080] Define the reward function for each base station. The mathematical expression is: Then reward for: .

[0081] Furthermore, the specific steps to obtain the globally optimal caching strategy model are as follows:

[0082] S61, the digital twin layer receives prediction prior knowledge from the content popularity prediction module;

[0083] S62, the digital twin layer trains a global collaborative caching strategy model based on the stored node information and the discretized DDPG algorithm using Markov decision process, and finally obtains the globally optimal caching strategy model.

[0084] Compared with the prior art, the significant advantages of this invention are as follows:

[0085] 1. This invention proposes a content collaborative caching system based on digital twins. This system can design an efficient content collaborative caching strategy according to the dynamic content request characteristics of mobile terminals. Through global scheduling of the digital twin layer, it can make full use of the computing and storage resources of edge nodes, significantly reduce the latency and transmission cost of mobile terminals to obtain content, effectively improve the overall service quality and revenue of the system, and solve the problems of low edge resource utilization and difficulty in coping with dynamic requests in the prior art.

[0086] 2. This invention proposes a dynamic content preference prediction algorithm, which designs an adaptive backtracking time window mechanism that can dynamically adjust the window size based on the feedback of cache hit rate. It also combines the inherent logic of extracting historical request data of terminals within the window with the AE model to predict popular content with high request frequency and many preferences, which significantly improves the accuracy of content prediction and the effectiveness of caching strategy.

[0087] 3. This invention proposes a priority-based content cost model and problem formula design method. It addresses the problem that traditional popularity prediction mechanisms may neglect device priority due to excessive focus on content activity in industrial IoT scenarios. It pre-sets priority rules for different devices and calculates the priority index of base stations, incorporating priority weights into system cost considerations. This maximizes system benefits while maximizing the base station priority index, effectively solving the defect of existing technologies that tend to prioritize high-popularity content while neglecting the priority of key devices, and achieving a key balance between content activity and device priority.

[0088] 4. The content collaborative caching method proposed in this invention constructs a real-time interactive digital twin layer based on multi-dimensional data from terminals and the edge, realizing real-time sharing of virtual data between different base stations and updating and maintaining the digital twin model; it enhances the perception of global information, improves the reliability of system state synchronization, increases the accuracy of the global digital twin model, and ensures the global optimality of caching decisions. Attached Figure Description

[0089] Figure 1 This is the overall architecture diagram of the content collaborative caching system based on digital twins of this invention;

[0090] Figure 2 This is a detailed framework diagram of the content collaborative caching system based on digital twins according to the present invention;

[0091] Figure 3 This is a flowchart of the content collaborative caching method based on digital twins according to the present invention;

[0092] Figure 4This is a comparison chart of the changes in system cache hit rate before and after the introduction of the dynamic content preference prediction mechanism in this embodiment of the invention;

[0093] Figure 5 This is a comparison chart of the changes in system revenue before and after the introduction of the dynamic content preference prediction mechanism in this embodiment of the invention;

[0094] Figure 6 This is a comparison chart of the changes in system cache hit rate before and after the introduction of the digital twin mechanism in this embodiment of the invention;

[0095] Figure 7 This is a comparison chart of the changes in system revenue before and after the introduction of the digital twin mechanism in this embodiment of the invention. Detailed Implementation

[0096] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0097] Digital twins, acting as a link between physical entities and virtual space, can construct real-time mappings of physical entities in virtual space through perception and high-fidelity modeling technologies, thereby enabling predictive analysis and decision optimization. By constructing a global digital twin layer and training multiple agents based on it, agents can acquire global perception information for simulation and policy evolution, significantly improving the effectiveness of training and the accuracy of decision-making.

[0098] In a content collaborative caching solution based on digital twins, the comprehensive integration of technologies such as content prediction and digital twins, along with consideration of device priority design, is of great significance for optimizing edge network resource allocation, improving system training efficiency, and ensuring the quality of service (QoS) in industrial scenarios.

[0099] This embodiment addresses the caching challenge of mobile devices in Industrial Internet of Things (IIoT) scenarios by introducing a dynamic content prediction mechanism based on digital twins and a global collaborative caching decision-making method. Specifically, in the digital twin layer, an autoencoder (AE model) is used to mine data features of mobile devices within an adaptive backtracking time window, constructing a content popularity evaluation matrix to accurately identify popular content. In the edge base station (edge ​​layer) where edge servers are deployed, a Markov decision process and a priority-based cost modeling method are designed for global collaborative caching training of the digital twin layer. Based on prior predicted content, the digital twin layer utilizes a multi-agent collaborative mechanism to optimize the best caching strategy. The introduction of the digital twin layer successfully transforms the local observation information of multiple agents into global state information, significantly improving the accuracy of strategy training. The content popularity prediction design is as follows: after the edge server aggregates mobile device data, it sends it to the digital twin layer, which uses the AE model to construct a content evaluation matrix for devices within the base station's coverage area, selecting high-scoring content as predicted popular content. Priority-based problem-solving formulas are implemented as follows: mobile devices within the base station coverage area are divided into eight priority levels based on their importance; the higher the priority of a device, the greater the response weight for its content requests. A priority index is also defined and incorporated into the constraint model of the total system cost. Markov decision design is implemented as follows: each edge base station is modeled as an agent, and its internal cache is replaced with a triplet-based system. The Markov decision process described Representing the state space, Indicates an action, This represents a reward. The global collaborative caching decision design is embodied in: global collaborative training based on the full data of the digital twin layer, enabling each agent to overcome the limitations of the local environment and collaboratively explore the optimal strategy in the global state space. Through the organic combination of content popularity prediction and global collaborative caching decision-making, the efficient operation of the content collaborative caching system in the industrial IoT scenario is effectively guaranteed.

[0100] like Figure 1The diagram shows the overall architecture of the content collaborative caching system based on digital twins according to the present invention. This architecture includes a terminal device, an edge server, and a digital twin layer. The terminal device sends real-time content requests and related operational data to the edge server and receives specific content from the edge server based on the requests. After receiving the requests from the terminal device, the edge server performs two functions: first, it obtains the corresponding optimal caching strategy through a locally deployed strategy model; second, it continues to upload data to the digital twin layer (including various types of data from the edge server's runtime, used to construct a global digital twin model). Furthermore, the edge server receives the optimal strategy model from the digital twin layer. After receiving the various types of data from the edge server, the digital twin layer updates and maintains the global digital twin model, and uses an AE model (Auto-Encoder) to extract latent features to obtain predicted content. This predicted content is then used as prior knowledge input to the virtual agent for training, resulting in a globally optimal caching strategy model.

[0101] like Figure 2 The diagram shown is a detailed framework of the content collaborative caching system based on digital twins according to the present invention. The system includes:

[0102] The data collection module is configured to be deployed at the terminal layer and, through a gateway installed on the mobile device, sense and collect various operational data and content request data during the device's operation.

[0103] The priority module is configured to be deployed at the terminal layer, divide mobile devices into 8 priority levels according to preset rules, and ensure that each device has a fixed priority attribute.

[0104] The edge data module is configured to: be deployed at the edge layer, receive device basic data and content request information from the terminal layer data collection module, forward the device basic data and content request information to the content popularity prediction module of the digital twin layer for predictive analysis, and forward the device basic data and content request information to the global collaborative cache training module for training to obtain the optimal cache strategy model;

[0105] The cost design module is configured to be deployed at the edge layer to build an edge base station content delivery cost model and a priority-based total system cost model, and to express the system optimization objectives in a formulaic way.

[0106] Markov Decision Design Module: Deployed at the edge layer, it models the agent's cache replacement process as a Markov decision process, defines the agent's state space and action space, and designs a reward function based on a formulaic optimization objective to guide the agent's training direction;

[0107] The content caching module is configured to: be deployed at the edge layer, receive the optimal caching strategy model obtained by the global collaborative caching training module, obtain the optimal caching strategy based on Markov decision design and state data, and perform caching operations on the corresponding content;

[0108] The content popularity prediction module is configured to: be deployed in the digital twin layer, use the AE model to construct a content popularity evaluation matrix based on the device's historical request information, select high-evaluation content as predicted popular content, and input the predicted popular content as prior knowledge into the global collaborative cache training module;

[0109] The global collaborative caching training module is configured to be deployed in the digital twin layer and train a global collaborative caching strategy model based on the content popularity prediction results and the received global information. In this process, each agent achieves collaboration by observing each other's states and finally solves the global optimal caching strategy model.

[0110] like Figure 3 The diagram shows the flowchart of the content collaborative caching method based on digital twins according to the present invention. The present invention aims to improve cache hit rate and system benefits by organically combining content popularity prediction based on digital twins with global collaborative caching decision-making, thereby achieving efficient operation of the content collaborative caching system in the industrial Internet of Things scenario. The specific steps are as follows:

[0111] Step 1: Each mobile device sends data such as location and remaining energy, as well as content caching requests, to the edge base stations under its coverage area.

[0112] Step 2: After receiving data sent by the mobile device, the edge base station constructs a digital twin layer. Different edge base stations upload data from mobile devices within their coverage area to the digital twin layer, and virtual data is shared between different base stations.

[0113] Step 3: The digital twin layer monitors the system's operating status in each round, predicts the content request preferences of mobile devices, and stores the predicted content in the training data of the global collaborative cache;

[0114] Step 31: Design the backtracking time window for the r-th round The system predicts future content demand based on data within the digital twin of the window, and the window size is adaptively adjusted based on the cache hit rate.

[0115] Step 311: Design the backtracking time ratio coefficient for the r-th round:

[0116] ,

[0117] in, and These represent the lower and upper bounds of the proportionality coefficient, respectively. from This indicates that the backtracking time window lasts for a maximum of the entire duration of the previous training round, or 0.2 times the last training time. Let represent the cache hit rate in round r-1, and , The initial value is 0.2;

[0118] Step 312: Design the final backtracking time window based on the backtracking time ratio coefficient. for:

[0119] ,

[0120] in, This indicates the current moment, that is, the moment when round (r-1) ends and round r is about to begin. This represents the total duration of the actual training in the (r-1)th round.

[0121] Step 32: In the backtracking time window Within, popular content is predicted based on historical request information from each mobile device;

[0122] Step 321: Determine if a request time is within the backtracking time window. If the request time... satisfy:

[0123] ,

[0124] This indicates that the request occurred during the last segment of the previous round, and is a high-value recent feature, which should be retained and input into the AE model.

[0125] Step 322: The AE model deployed in the digital twin layer extracts the historical request information of each mobile device within the jurisdiction of each base station within the retrospective time window, and constructs a content popularity evaluation matrix for each device. ,in The first dimension is the identity characteristics of the mobile device, and the second dimension is the evaluation level of the content by the mobile device.

[0126] Step 323: Sort the content of mobile devices under each base station in descending order of rating, and select the top N content items with the highest rating as predicted popular content. The size of N is determined by the cache capacity of the base station.

[0127] Step 33: Input the popular content predicted in Step 32 as prior knowledge into the data for training the global collaborative cache.

[0128] Step S4: The edge base station designs the total system cost model according to priority, formulates the optimization objective, and establishes the system problem formula;

[0129] Step 41: Design the overall system model. Mobile devices send content requests to the local base station, and there are three ways to provide content. These are: the local base station stores the content and transmits it directly; the neighboring base station stores the content and transmits it to the local base station; and neither the local base station nor the neighboring base station stores the content, so the cloud transmits it to the local base station.

[0130] Step 42: Calculate the content transmission cost for the base station to fulfill the content request after the mobile device sends the content request to the base station;

[0131] If the local base station to which the mobile device is connected has pre-cached the content, it can be transmitted directly via the wireless link. The transmission cost is denoted as the first transmission cost. , is represented as:

[0132] ,

[0133] in, This represents the unit cost of transmitting content using a wireless link. Content Data size, This indicates the wireless link transmission rate between the base station and the mobile device;

[0134] If the local base station connected to the mobile device cannot provide content. However, if a neighboring base station has cached this content, it can be transmitted to the local base station via fiber optic cable and then distributed via wireless link. In this case, the content obtained through this method... The transmission cost is denoted as the second transmission cost. , is represented as:

[0135] ,

[0136] in, This represents the unit cost of transmitting content using a wireless link. This indicates the wired link transmission rate between the local base station and its neighboring base stations;

[0137] If content If the content is neither cached at the local base station nor at a neighboring base station, the local base station needs to request the content from the cloud and transmit it to the corresponding mobile device. This method of content delivery... The transmission cost is denoted as the third transmission cost. , is represented as:

[0138] ,

[0139] in, This indicates the wired link transmission rate between the local base station and the cloud;

[0140] Finally, the total content transmission cost per base station Represented as

[0141] ,

[0142] in, This represents the set of devices within the coverage area of ​​the base station. A collection of content. Indicates whether device u requests this content from the base station. ; , respectively, indicate that the content request of device u is provided by the local base station, neighboring base stations, and the cloud, and the selection of the service method satisfies the constraints. ;

[0143] Step 43: Design a priority-based service cost model;

[0144] Step 431: Divide all mobile devices into 8 priority levels according to their importance, using... Indicates the priority of device u;

[0145] Step 432: Design the priority index for a single base station for:

[0146] ,

[0147] in, This represents the cache hit rate of device u at this base station. This represents the total number of mobile devices under a single base station. This indicates the highest priority, which is 8 in this embodiment; The larger the value, the more priorities the strategy considers; conversely, the smaller the value, the fewer priorities it considers.

[0148] Step 433: Calculate the total system cost for:

[0149] ;

[0150] Step 44: Design the revenue model;

[0151] Calculate the waiting time for the mobile device to receive content after sending a request. Represented as:

[0152] ;

[0153] The revenue from base station design is inversely proportional to the content transmission latency; the revenue function... Represented as:

[0154] ,

[0155] in, Indicates the attenuation factor. Content Value;

[0156] Define an index to determine whether a cached request is satisfied within a specified time. for:

[0157] ,

[0158] in, This is the time limit for waiting for content requests; once the content is successfully transmitted, the index... If it fails, then index ;

[0159] Calculate the revenue of the base station , is represented as:

[0160] ;

[0161] Step 45: Establish the system problem formula and find the optimal caching strategy by maximizing the total revenue of a single base station. The optimization problem can then be expressed as:

[0162] ,

[0163] in, This represents the total revenue of a single base station; It is the bandwidth currently allocated to mobile device u by the edge base station. This is the total bandwidth of the base station; Indicates whether the base station caches content. For example, "yes" takes a value of 1, and "no" takes a value of 0; Represents a single content Size, This represents the base station's cache capacity. Constraint C1 indicates the base station's total bandwidth limit, constraint C2 indicates that the base station needs to consider storage capacity limitations when caching content, and constraint C3 defines that the content requested by the mobile terminal can only be retrieved in one place.

[0164] Step 5: Based on the total cost model and the system problem formula, design a Markov decision process; the details are as follows:

[0165] Define the state of each base station as ,in This indicates the local cache status of the base station. The state space represents the received content request status. : ,and , This indicates the total number of base stations.

[0166] Define the action in each base station as ,in For base station caching actions, Indicates the caching service method; action : ,and .

[0167] Since the reward function is directly related to the difference between system revenue and cost, a reward function is defined for each base station. The mathematical expression is: Then there is a defined reward. : ,and .

[0168] Step 6: Based on the given problem and Markov decision process, train the global collaborative caching strategy model in the digital twin layer to obtain the globally optimal caching strategy model;

[0169] Step 61: The digital twin layer receives prediction prior knowledge from the content popularity prediction module;

[0170] Step 62: Based on the stored node information, the digital twin layer trains a global collaborative caching strategy model using a discretized DDPG (Deep Deterministic Policy Gradient) algorithm based on Markov decision processes. Through logical mapping within the environment, a discrete caching strategy is ultimately generated. It is worth noting that in the digital twin layer, the environmental information perceived by the base station is global, not simply local, increasing the cache hit rate. Through global collaborative caching training, a globally optimal caching strategy model is finally obtained.

[0171] Step 7: The digital twin layer distributes the globally optimal caching strategy model to the edge layer. The edge layer, based on data such as state, designs the optimal strategy using Markov decision-making and caches the corresponding content.

[0172] To verify the effectiveness of the content collaborative caching method based on digital twins in this invention, the system's cache hit rate and total system revenue were selected as evaluation indicators. The number of mobile terminals was set to 50, the number of edge servers to 4, and the communication bandwidth conditions were set to 200MHz for wired transmission and 20MHz for wireless transmission.

[0173] like Figure 4 and Figure 5 As shown, to verify the effectiveness of the content popularity prediction algorithm in this invention, the system performance changes of the method under conditions of introducing a prediction mechanism and not introducing a prediction mechanism were compared. Simulation results show that as the training rounds progress, both the cache hit rate and the system reward exhibit a trend of first briefly decreasing and then continuously increasing, which is consistent with the convergence characteristics of deep reinforcement learning in the early stages of policy exploration. The comparative results show that, compared with the baseline algorithm without a prediction mechanism, the algorithm designed in this invention achieves significant improvements in both the cache hit rate and the system reward, two key indicators.

[0174] like Figure 6 and Figure 7 As shown, to verify the effectiveness of the global collaborative caching decision algorithm based on digital twins in this invention, the algorithm's performance was compared under conditions of constructing a digital twin layer and without a digital twin layer. Simulation results show that the global collaborative caching decision algorithm proposed in this invention consistently outperforms the algorithm without a digital twin design in terms of cache hit rate and system benefits. Furthermore, the performance degradation caused by the lack of a digital twin collaboration mechanism is significantly greater than the impact of the lack of a content prediction mechanism. This result fully demonstrates that in the architecture of this invention, although content prediction can effectively improve the prior accuracy of the content to be cached, the global information perception and collaboration capabilities provided by the digital twin layer are the key to determining whether the caching strategy can achieve global optimization and thus maximize system benefits.

Claims

1. A content collaborative caching system based on digital twins, characterized in that, This includes terminal devices deployed at the terminal layer, edge servers deployed at the edge layer, and a digital twin layer. Terminal devices send real-time content requests and related operational data to the edge servers and receive specific content from the edge servers based on these requests. Upon receiving a request from a terminal device, the edge server performs two actions: first, it uses its locally deployed strategy model to obtain the optimal caching strategy; second, it continues to upload data to the digital twin layer. Simultaneously, the edge server receives the optimal strategy model from the digital twin layer. After receiving various types of data from the edge server, the digital twin layer updates and maintains the global digital twin model, extracts latent features using the AE model to obtain predicted content, and inputs this predicted content as prior knowledge into the virtual agent to train the strategy model, ultimately obtaining the globally optimal caching strategy model.

2. The content collaborative caching system based on digital twins according to claim 1, characterized in that, Also includes: The data collection module, deployed at the terminal layer, senses and collects various operational data and content request data during the device's operation through a gateway installed on the mobile device. The priority module, deployed at the terminal layer, divides mobile devices into multiple priority levels according to preset rules and ensures that each device has a fixed priority attribute. The edge data module, deployed at the edge layer, receives basic device data and content request information from the data collection module, forwards the basic device data and content request information to the content popularity prediction module of the digital twin layer for predictive analysis, and forwards the basic device data and content request information to the global collaborative cache training module for training to obtain the optimal cache strategy model. The cost design module, deployed at the edge layer, builds an edge base station content delivery cost model and a priority-based total system cost model, and expresses the system optimization objectives in a formulaic way. Markov Decision Design Module: Deployed at the edge layer, it models the agent's cache replacement process as a Markov decision process, defines the agent's state space and action space, and designs a reward function based on a formulaic optimization objective to guide the agent's training direction; The content caching module, deployed at the edge layer, receives the optimal caching strategy model obtained from the global collaborative caching training module, obtains the optimal caching strategy based on Markov decision design and state data, and performs caching operations on the corresponding content. The content popularity prediction module is deployed in the digital twin layer. It uses the AE model to build a content popularity evaluation matrix based on the device's historical request information, and selects high-evaluation content as predicted popular content. The predicted popular content is then used as prior knowledge and input into the global collaborative cache training module. The global collaborative caching training module is deployed in the digital twin layer. It trains the global collaborative caching strategy model based on the content popularity prediction results and the received global information. During the global collaborative caching training process, each agent cooperates by observing each other's states and finally solves the global optimal caching strategy model.

3. A content collaborative caching method based on digital twins, characterized in that, Applied to the content collaborative caching system based on digital twins as described in any one of claims 1-2, a globally optimal caching strategy is obtained through the organic combination of content popularity prediction based on digital twins and global collaborative caching decision-making. The specific steps are as follows: S1, each mobile device sends its location, remaining power, and content caching requests to the edge base stations under its coverage area; S2, after receiving data sent by the mobile device, the edge base station constructs a digital twin layer. Different edge base stations upload data from mobile devices within their coverage area to the digital twin layer, and virtual data is shared between different base stations; S3, the digital twin layer, monitors the system's operating status in each round, predicts the content request preferences of mobile devices, and stores the predicted content in the training data of the global collaborative cache; S4, the edge base station designs the total system cost model according to priority, formulates the optimization objective, and establishes the system problem formula; S5. Based on the total cost model and the system problem formula, design a Markov decision process; S6, based on the given problem and Markov decision process, train the global collaborative caching strategy model in the digital twin layer to obtain the globally optimal caching strategy model; S7: The digital twin layer distributes the globally optimal caching strategy model to the edge layer. The edge layer, based on the state data and Markov decision design, derives the optimal strategy and caches the corresponding content.

4. The content collaborative caching method based on digital twins according to claim 3, characterized in that, The digital twin layer predicts the content request preferences of mobile devices and stores the predicted content in the training data of a global collaborative cache, specifically including: S31, Design the backtracking time window for the r-th round. The system predicts future content demand based on data within the digital twin of the window, and the window size is adaptively adjusted based on the cache hit rate. S32, in the backtracking time window Within, popular content is predicted based on historical request information from each mobile device; S33, the predicted popular content is used as prior knowledge and input into the training data of the global collaborative cache.

5. The content collaborative caching method based on digital twins according to claim 4, characterized in that, Obtain the backtracking time window for round r. The steps include: S311, Design the backtracking time ratio coefficient for the r-th round. : ， in, , These represent the lower and upper bounds of the proportionality coefficient, respectively. Let represent the cache hit rate in round r-1, and ; S312, based on the backtracking time scaling factor Design the final backtracking time window for: ， in, This indicates the current moment, which is the moment when round (r-1) ends and round r is about to begin; This represents the total duration of the actual training in the (r-1)th round.

6. The content collaborative caching method based on digital twins according to claim 4, characterized in that, The specific steps for predicting popular content based on historical request information from various mobile devices include: S321, determine whether a certain request time is within the backtracking time window. If the request time... satisfy: ， This indicates that the request occurred during the last segment of the previous round, and is a high-value recent feature, which should be retained and input into the AE model. S322, the AE model deployed in the digital twin layer extracts historical request information of each mobile device within the jurisdiction of each base station within the retrospective time window, and constructs a content popularity evaluation matrix for each device. ;in The first dimension is the identity characteristics of the mobile device, and the second dimension is the evaluation level of the content by the mobile device. S323: Sort the content of mobile devices under each base station in descending order of level, and select the top N content with the highest evaluation level as the predicted popular content, where the size of N is determined by the cache capacity of the base station.

7. The content collaborative caching method based on digital twins according to claim 3, characterized in that, The specific steps for designing a total system cost model for edge base stations based on priority include: S41, Design the overall system model. Mobile devices send content requests to the local base station, and there are three ways to provide content. If the local base station stores the content, it transmits it directly; if a neighboring base station stores the content, it transmits it to the local base station; if neither the local base station nor a neighboring base station stores the content, the cloud transmits it to the local base station. S42, after the mobile device sends a content request to the base station, the cost of the base station fulfilling the content transmission request is calculated as follows: If the local base station to which the mobile device is connected has cached the content in advance, the content will be transmitted directly through the wireless link. The transmission cost is denoted as the first transmission cost. , is represented as: ， in, This represents the unit cost of transmitting content using a wireless link. Content Data size, This indicates the wireless link transmission rate between the base station and the mobile device; If the local base station connected to the mobile device cannot provide content. However, if a neighboring base station caches this content, it is transmitted to the local base station via fiber optic cable, and then distributed via wireless link. This method of content delivery... The transmission cost is denoted as the second transmission cost. , is represented as: ， in, This represents the unit cost of transmitting content using a wireless link. This indicates the wired link transmission rate between the local base station and its neighboring base stations; If content If the content is neither cached at the local base station nor at a neighboring base station, the local base station needs to request the content from the cloud and transmit it to the corresponding mobile device. This method of content... The transmission cost is denoted as the third transmission cost. , is represented as: ， in, This indicates the wired link transmission rate between the local base station and the cloud; Finally, the total content transmission cost per base station Represented as: ， in, This represents the set of devices within the coverage area of ​​the base station. A collection of content. Indicates whether device u requests this content from the base station. ; , respectively, indicate that the content request of device u is provided by the local base station, neighboring base stations, and the cloud, and the selection of the service method satisfies the constraints. ; S43, Design a priority-based service cost model, including: S431 divides all mobile devices into eight priority levels based on their importance. Indicates the priority of device u; S432, designing the priority index for a single base station. for: ， in, This represents the cache hit rate of device u at this base station. This represents the total number of mobile devices under a single base station. Indicates the highest priority; The larger the value, the more priorities the strategy considers; conversely, the smaller the value, the fewer priorities it considers. S433, Calculate the total system cost for: ； S44, Design the revenue model, including: Calculate the waiting time for the mobile device to receive content after sending a request. Represented as: ； The revenue from base station design is inversely proportional to the content transmission latency; the revenue function... Represented as: ， in, Indicates the attenuation factor. Content Value; Define an index to determine whether a cached request is satisfied within a specified time. for: ， in, This is the time limit for waiting for content requests; once the content is successfully transmitted, the index... If it fails, then index ; Calculate the revenue of the base station , is represented as: ； S45, by maximizing the total revenue of a single base station Find the optimal caching strategy The optimization problem can then be expressed as: ， in, It is the total bandwidth of the base station. It is the bandwidth currently allocated to mobile device u by the edge base station. It is the total bandwidth of the base station. Indicates whether the base station caches content. For example, "yes" takes a value of 1, and "no" takes a value of 0; Represents a single content Size, This refers to the base station's cache capacity; constraint C1 represents the base station's total bandwidth limit; constraint C2 indicates that the base station needs to consider storage capacity limitations when caching content; and constraint C3 defines that the content requested by the mobile terminal can only be retrieved in one place.

8. The content collaborative caching method based on digital twins according to claim 3, characterized in that, Each edge base station is modeled as an intelligent agent, and the Markov decision process is defined as follows: Define the state of each base station as ,in This indicates the local cache status of the base station. The state space represents the received content request status. for: ,and , Indicates the total number of base stations; Define the action in each base station as ,in For base station caching actions, Indicates the caching service method; then the action. for: ; Define the reward function for each base station. The mathematical expression is: Then reward for: .

9. The content collaborative caching method based on digital twins according to claim 3, characterized in that, The specific steps to obtain the globally optimal caching strategy model are as follows: S61, the digital twin layer receives prediction prior knowledge from the content popularity prediction module; S62, the digital twin layer trains a global collaborative caching strategy model based on the stored node information and the discretized DDPG algorithm using Markov decision process, and finally obtains the globally optimal caching strategy model.

Images