Caching Strategies and Resource Allocation Methods in Recommendation Mechanism-Driven UAV Edge Caching Networks

By constructing an ARIS-assisted edge caching network and combining user preferences and RSMA technology to optimize the caching strategy and resource allocation of the UAV edge caching network, the coordination problems of cache placement, content recommendation and wireless resource allocation in the UAV edge caching network are solved, improving cache hit rate and transmission efficiency, and enhancing the stability of decision-making.

CN123123944APending Publication Date: 2026-11-17GUILIN UNIV OF ELECTRONIC TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611252172.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In UAV edge caching networks, cache placement, content recommendation, and wireless resource allocation are difficult to coordinate as a whole through independent optimization methods. User preferences and cache hit requirements are difficult to balance, and insufficient UAV flight boundary constraints affect the stability of joint decision-making and the effectiveness of resource allocation.

Method used

An ARIS-assisted edge caching network is constructed, and cache-oriented recommendation results are formed by combining user preference distortion and personalization tolerance threshold. User messages are split using RSMA and interference is suppressed by SIC. A joint optimization problem of caching strategy and resource allocation is constructed, and soft and hard penalties are set in the Markov decision process. The CAD-PPO algorithm and constraint-aware discriminant network are used for value correction.

Benefits of technology

It improves cache hit rate and transmission efficiency, enhances the stability of joint decision-making in dynamic wireless environments, and optimizes the collaborative decision-making between caching strategies and communication resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123944A_ABST
    Figure CN123123944A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of wireless communication and edge computing technology, specifically to a caching strategy and resource allocation method in a recommendation mechanism-driven UAV edge caching network, comprising the following steps: S1, constructing an ARIS-assisted edge caching network and generating network state information; S2, determining the recommendation window and recommendation decision, and generating cache-guided recommendation results; S3, using RSMA transmission to construct a joint optimization problem of caching and resources; S4, setting soft and hard penalties at the UAV boundary and constructing a joint optimization decision model; S5, using CAD-PPO for value correction and policy update, and outputting the caching and resource allocation results. This invention achieves coordinated adjustment of user requests, caching strategies, and communication resources, improving cache hit rate, transmission efficiency, and the stability of joint decision-making in dynamic wireless environments while satisfying UAV flight boundary constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and edge computing technology, and in particular to caching strategies and resource allocation methods in a recommendation mechanism-driven UAV edge caching network. Background Technology

[0002] With the development of UAV communication, edge caching, and reconfigurable smart surface technologies, utilizing UAVs equipped with aerial active reconfigurable smart surfaces (ARIS) to assist ground user communication, and deploying caching resources at edge nodes, can reduce content retrieval latency by flexibly adjusting airborne links and cached content. Meanwhile, RSMA can improve multi-user transmission performance through rate segmentation of public and private streams and continuous interference cancellation, while content recommendation mechanisms can guide request behavior based on user preferences, providing a technical foundation for the coordinated configuration of caching, recommendation, and wireless communication resources.

[0003] Currently, cache placement, content recommendation, and wireless resource allocation in UAV edge caching networks typically involve multiple coupled factors such as user requests, cache status, UAV trajectories, ARIS reflection, beamforming, and multi-user transmission, making it difficult to achieve overall coordination through independent optimization methods. Furthermore, there is a lack of sufficient linkage between recommendation results and cached content, making it difficult to balance user preferences and cache hit requirements. In dynamic wireless environments, traditional value estimation is also susceptible to biases due to environmental randomness, and if UAV flight boundary constraints only employ simple boundary crossing restrictions, it is difficult to simultaneously achieve advance constraints on near boundaries and effective penalties after actual boundary crossings, thus affecting the stability of joint decision-making and the effectiveness of resource allocation. Summary of the Invention

[0004] This invention provides a caching strategy and resource allocation method in a recommendation mechanism-driven UAV edge caching network. By constructing an ARIS-assisted edge caching network, it utilizes user preference distortion and personalized tolerance thresholds to form cache-oriented recommendation results. Furthermore, it combines RSMA to jointly optimize UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement, recommendation decision, and RSMA common rate allocation. It further introduces UAV boundary soft and hard penalties and CAD-PPO distributed value generation and constraint-aware value correction to achieve collaborative decision-making between caching strategy and communication resources, thereby improving cache hit rate, transmission efficiency, and the stability of joint decision-making in dynamic wireless environments.

[0005] The caching strategy and resource allocation method in a recommendation mechanism-driven drone edge caching network include the following steps: S1. Construct an airborne active reconfigurable smart surface ARIS-assisted edge caching network to obtain the initial user request distribution, user preferences, cache content status and wireless communication status, and generate network status information by combining UAV trajectory, ARIS reflection coefficient matrix and BS beamforming status. S2, based on the network state information, calculate the user preference distortion degree, combine it with the personalized tolerance threshold to determine the maximum feasible recommendation window, generate recommendation decisions from the cached content, reshape the initial user request distribution into a cache-oriented user request distribution, and obtain cache-oriented recommendation results; S3. Based on the cache-guided recommendation results and network status information, RSMA is used to split user messages into public and private streams, and SIC is used to suppress interference between users. UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement, recommendation decision and RSMA common rate allocation are used as joint optimization variables. With the dynamic weighted sum of latency efficiency and cache hit rate as the objective, a joint optimization problem of cache strategy and resource allocation is constructed. S4. Based on the caching strategy and resource allocation joint optimization problem, a Markov decision process is constructed, and soft and hard penalties are set for the UAV flight boundary to generate a joint optimization decision model with UAV boundary constraints. S5. The CAD-PPO algorithm is used to solve the joint optimization decision model. The return distribution is modeled and the value is estimated by a distributed value generator. The constraint violation signal is constructed based on the normalized numerical deviation between the generated value and the TD target. The constraint-aware discrimination network is used for adversarial training and value correction. The policy is updated and the caching policy and resource allocation results are output.

[0006] Optionally, S1 includes: S11. Construct an airborne active reconfigurable smart surface ARIS-assisted edge buffer network. The ARIS equipped with active load impedance elements is set on the UAV. The UAV provides an airborne reflection link to the ground mobile user according to the current trajectory position. The UAV spatial position, ARIS reflection coefficient matrix and BS beamforming state at the current moment are determined to form the airborne ARIS communication state parameters. Each reflection unit of the ARIS is subjected to amplitude and phase control to characterize the active reflection adjustment state of the ARIS to the incident signal. S12: Obtain content request records and user preference information of users in various locations, obtain the content storage status in the edge cache and the current wireless communication status, calculate the initial request probability of each content according to the Zipf distribution parameters corresponding to each user and the content sorting, form the initial user request distribution from the initial request probability of each content, form the user preference representation according to the number of historical requests of users, and mark the status of each content as cached, so as to obtain the initial user request distribution, user preference, cached content status and wireless communication status. S13. Based on the airborne ARIS communication status parameters, initial user request distribution, user preferences, cached content status, and wireless communication status, the UAV trajectory status, ARIS reflection coefficient matrix, BS beamforming status, and user-side, cache-side, and communication-side status at the same time are uniformly correlated to generate network status information. The network status information is then used as the status input for cache-oriented recommendation and joint optimization of cache strategy and resource allocation.

[0007] Optionally, S2 includes: S21, based on the user preferences, initial user request distribution, and cached content status in the network state information, for the first... Each user generates candidate content based on their preference for different content types, and different recommendation window sizes are used for each user. Calculate user preference distortion The user preference distortion degree is compared with the first Personalized tolerance threshold for each user When comparing, When the corresponding recommendation window is determined to be a feasible recommendation window, When this happens, the corresponding recommendation window is determined to be an infeasible recommendation window, and the set of feasible recommendation windows consists of all recommendation windows that satisfy the user preference distortion degree not exceeding the personalization tolerance threshold. ; S22, based on the user preference distortion, select the feasible recommendation window with the largest recommendation window size as the first... The maximum feasible recommendation window for each user In conjunction with the cached content status, recommended candidate content is determined from the cached content according to user preferences, ensuring that the number of recommended content does not exceed the maximum feasible recommendation window, and a recommendation decision is generated for each user. ; S23, based on the recommendation decision and the initial user request distribution, the initial request probability corresponding to each user is reshaped using the recommendation decision, so that the request probability corresponding to cached and recommended content is redistributed towards the cached content, and the reshaped request probability is normalized to obtain the cache-oriented request probability of each user. The cache-oriented user request distribution is formed by the cache-oriented request probabilities. The recommendation decision and cache-oriented user request distribution are used as the cache-oriented recommendation result.

[0008] Optionally, S3 includes: S31, based on the recommendation decision and cache-oriented user request distribution in the cache-oriented recommendation result, and combined with the network status information, determine the current request content and corresponding wireless communication status of each user, and use Rate Division Multiple Access (RSMA) to split each user message into a common stream and a private stream, encode and transmit the common stream and the private stream respectively, so that each user first decodes the common stream and eliminates the corresponding interference through Continuous Interference Cancellation (SIC), and then decodes its own private stream to obtain the RSMA transmission status; S32, based on the RSMA transmission status and network status information, the UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement, recommendation decision, and RSMA common rate allocation are determined as joint optimization variables. The cache-guided user request distribution is used to characterize the user request status after the recommendation mechanism takes effect. This allows UAV spatial location adjustment, ARIS active reflection adjustment, BS beamforming, cache content configuration, recommended content selection, and RSMA common rate allocation to be coordinated and adjusted in the same joint optimization process, resulting in a set of joint optimization variables. S33. Based on the set of joint optimization variables, determine the latency efficiency and cache hit rate under the corresponding caching strategy and resource allocation state, and set dynamic weights for latency efficiency and cache hit rate according to the business scenario. Take the dynamic weighted sum of latency efficiency and cache hit rate as the joint optimization objective, and construct the joint optimization problem of caching strategy and resource allocation under the common constraints of the network state information, cache-oriented recommendation results and RSMA transmission state.

[0009] Optionally, S31 includes: S311, based on the recommendation decisions and cache-oriented user request distribution in the cache-oriented recommendation results, determine the time of each user. The cache-directed request probability for each content is determined, and combined with the wireless communication status in the network status information, the current requested content and corresponding wireless communication status of each user are determined, forming a correspondence between user requests and wireless communication status. S312, based on the current request content of each user and the corresponding wireless communication status, the message corresponding to each user is divided into a common part and a private part by using rate division multiple access (RSMA). The common parts of each user are combined to form a common stream and the private streams of the corresponding users are formed respectively. The common stream and each private stream are beamformed and superimposed for transmission by using BS beamforming state to obtain the RSMA transmission signal of the common stream and the private stream. S313, based on the RSMA transmission signal of the public stream and the private stream, each user decodes the public stream according to the corresponding wireless communication state, and after the public stream is successfully decoded, continuous interference cancellation (SIC) is used to eliminate the interference corresponding to the public stream, and the user decodes their own private stream. The transmission rate of each user is determined according to the decodeable rate of the public stream, the decodeable rate of the private stream, and the RSMA public rate allocation, and an RSMA transmission state is generated.

[0010] Optionally, S32 includes: S321, based on the RSMA transmission status and network status information, extract the current spatial position of the UAV, the ARIS reflection coefficient matrix and the BS beamforming status, form the UAV trajectory by combining the UAV spatial positions at consecutive times, and use the UAV trajectory, the ARIS reflection coefficient matrix and the BS beamforming as joint optimization variables on the communication side, so that the UAV spatial position adjustment, ARIS amplitude and phase adjustment and the BS beamforming of the common stream and the private stream change collaboratively under the same communication status; S322, based on the cache-oriented user request distribution and RSMA transmission status, the cache placement corresponding to the cache content status, the generated recommendation decision, and the RSMA common rate allocation for each user are determined as joint optimization variables for cache recommendation and transmission resources. The recommendation decision reflects the user request status after the recommendation mechanism takes effect through the cache-oriented user request distribution, and makes the recommended content subject to the cache placement status, and the RSMA common rate allocation subject to the common stream transmission rate. S323, based on the joint optimization variable set of the communication side and the joint optimization variable set of cache recommendation and transmission resources, the UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement status, recommendation decision and RSMA common rate allocation are uniformly associated, and combined with the cache-guided user request distribution and RSMA transmission status, a joint optimization variable set for joint optimization of latency efficiency and cache hit rate is formed.

[0011] Optionally, S33 includes: S331, based on the joint optimization variable set, cache-oriented user request distribution and RSMA transmission status, obtain the data volume and RSMA transmission rate corresponding to the current request content of each user, determine the content transmission delay of each user according to the request content data volume and RSMA transmission rate, and determine the delay efficiency corresponding to the current caching strategy and resource allocation status according to the content transmission delay of all users. S332, based on the cache-oriented user request distribution and the cache placement status in the joint optimization variable set, calculate the probability of each user requesting cached content, and statistically analyze the cache hit probability for all users to determine the cache hit rate under the current caching strategy and resource allocation status. S333, based on the latency efficiency and cache hit rate, set latency efficiency weight and cache hit rate weight according to the current business scenario, and determine the dynamic weighted sum of latency efficiency and cache hit rate as the joint optimization objective. Take the set of joint optimization variables as the variables to be optimized, and combine network status information, cache-oriented recommendation results and RSMA transmission status to determine the feasible range of joint optimization variables, and construct the joint optimization problem of cache strategy and resource allocation.

[0012] Optionally, S4 includes: S41, Based on the joint optimization problem of caching strategy and resource allocation, network state information is used as the state of Markov decision process, the adjustment result corresponding to the set of joint optimization variables is used as action, the joint optimization objective value is used as basic immediate benefit, and a state transition relationship is established according to the network state at the next moment formed after the current state performs the action, thus constructing a Markov decision process to describe the continuous decision process of caching strategy and resource allocation. S42, based on the UAV's spatial position, set boundary buffer areas for the X-axis and Y-axis boundaries of the UAV's flight area respectively. When the UAV is within the flight boundary and enters the boundary buffer area, a soft penalty is calculated. When the UAV goes beyond the flight boundary, a hard penalty is calculated. The penalty values ​​corresponding to the X-axis and Y-axis are summarized respectively to obtain the UAV boundary penalty. S43. Based on the Markov decision process, the joint optimization objective, the total soft penalty of the UAV, and the total hard penalty of the UAV, the joint optimization objective is taken as the positive reward, and the soft penalty and hard penalty are taken as UAV flight boundary constraints to construct an immediate reward. The immediate reward, state space, action space, and state transition relationship are uniformly associated to generate a joint optimization decision model with UAV boundary constraints.

[0013] Optionally, S5 includes: S51, Based on the joint optimization decision model with UAV boundary constraints, obtain the state, action, immediate reward and next time state formed in the Markov decision process, input the state and multiple independently sampled noise vectors into the distributed value generator, generate multiple value samples through noise-driven random forward propagation, use the multiple value samples to represent the reward distribution corresponding to the state, and obtain the value estimate of the state through Monte Carlo averaging. S52, based on the generated value and Monte Carlo average value estimation, the TD target is determined according to the value estimation of the immediate reward and the state at the next time step. The numerical deviation between each generated value and the TD target is calculated, and the numerical deviation is normalized to obtain the constraint violation signal. The generated value and the corresponding constraint violation signal are input into the constraint perception discrimination network. The distribution difference and numerical deviation between the generated value and the empirical TD target are judged through adversarial training, and the discrimination result is used to correct the value of the distributed value generator. S53. Based on the value estimation after correction by the constrained perception discriminant network, the advantage estimation corresponding to the joint optimization action is calculated. The ratio of the selection probability of the same joint optimization action by the current policy and the policy before the update is used as the policy probability ratio. The pruning policy objective of the PPO is optimized by the near-end policy to iteratively update the joint decision policy. Based on the updated joint decision policy, the UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement status, recommendation decision and RSMA common rate allocation are determined, and the cache policy and resource allocation results are output.

[0014] Optionally, S53 includes: S531, based on the TD target and the value estimate after correction by the constrained perception discrimination network, calculate the advantage estimate corresponding to the joint optimization action in the current state, and calculate the policy probability ratio based on the selection probability of the same joint optimization action by the current joint decision policy and the joint decision policy before the policy update. S532, based on the advantage estimation and policy probability ratio, the PPO pruning policy objective is used to limit the single update range of the joint decision policy, and the joint decision policy is iteratively updated by maximizing the PPO policy optimization objective to obtain the updated joint decision policy; S533, based on the updated joint decision-making strategy, determine the optimal set of joint optimization variables, and obtain the optimized UAV spatial location, ARIS reflection coefficient matrix, BS beamforming, buffer placement status, recommendation decision and RSMA common rate allocation respectively. The optimized UAV spatial location at continuous time intervals forms the UAV trajectory, and the buffer strategy and resource allocation results are output.

[0015] The beneficial effects of this invention are: This invention constructs an airborne active reconfigurable smart surface ARIS-assisted edge caching network, which unifies the initial user request distribution, user preferences, cached content status, wireless communication status, UAV trajectory, ARIS reflection coefficient matrix, and BS beamforming status. Based on user preference distortion and personalization tolerance thresholds, it determines the maximum feasible recommendation window and generates recommendation decisions from cached content. This reshapes the initial user request distribution into a cache-oriented user request distribution, enabling the selection of recommended content to simultaneously consider user preferences and cached content status. This improves the matching degree between recommendation results and cached resources, providing a unified network status and cache-oriented recommendation results for subsequent collaborative optimization of caching strategies and communication resources.

[0016] This invention employs RSMA to split user messages into public and private streams, and combines it with SIC to suppress inter-user interference. It integrates UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, buffer placement, recommendation decision-making, and RSMA common rate allocation into a single joint optimization process. Using a dynamic weighted sum of latency efficiency and buffer hit rate as the joint optimization objective, it enables coordinated adjustments to UAV spatial location, ARIS active reflection adjustment, communication resource allocation, buffer content configuration, and recommendation content selection based on the business scenario. This avoids resource mismatch caused by independent optimization of buffering, recommendation, and wireless communication resources, thereby improving buffer hit rate and transmission latency performance.

[0017] This invention constructs the joint optimization problem of caching strategy and resource allocation as a Markov decision process, and sets soft and hard penalties for UAV flight boundaries, so that UAVs are subject to early constraints when approaching the flight boundary and stronger constraints when crossing the boundary. Furthermore, the CAD-PPO algorithm is adopted, which models the complete reward distribution through a distributed value generator and obtains the value estimate by Monte Carlo averaging. The constraint violation signal is constructed based on the normalized numerical deviation between the generated value and the TD target. Then, the constraint-aware discrimination network is used for adversarial training and value correction, thereby reducing the value estimation bias in random wireless environments, improving the stability of policy updates, and finally obtaining the caching strategy and resource allocation results that satisfy the UAV boundary constraints. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the allocation method according to an embodiment of the present invention; Figure 2This is a schematic diagram illustrating the joint optimization problem of caching strategy and resource allocation in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] like Figures 1-2 As shown, the caching strategy and resource allocation method in a recommendation mechanism-driven UAV edge caching network includes the following steps: S1. Construct an airborne active reconfigurable smart surface ARIS-assisted edge caching network to obtain the initial user request distribution, user preferences, cache content status and wireless communication status, and generate network status information by combining UAV trajectory, ARIS reflection coefficient matrix and BS beamforming status. S2, based on network state information, calculate user preference distortion, combine with personalized tolerance threshold to determine the maximum feasible recommendation window, generate recommendation decisions from cached content, reshape the initial user request distribution into a cache-oriented user request distribution, and obtain cache-oriented recommendation results; S3, based on cache-guided recommendation results and network status information, uses RSMA to split user messages into public and private streams, and suppresses interference between users through SIC. It takes UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement, recommendation decision and RSMA common rate allocation as joint optimization variables, and constructs a joint optimization problem of cache strategy and resource allocation with the dynamic weighted sum of latency efficiency and cache hit rate as the objective. S4. Based on the joint optimization problem of caching strategy and resource allocation, a Markov decision process is constructed, and soft and hard penalties are set for the UAV flight boundary to generate a joint optimization decision model with UAV boundary constraints. S5 uses the CAD-PPO algorithm to solve the joint optimization decision model. It models the return distribution and obtains the value estimate through a distributed value generator. It constructs a constraint violation signal based on the normalized numerical deviation between the generated value and the TD target. It uses a constraint-aware discrimination network for adversarial training and value correction, updates the policy, and outputs the caching policy and resource allocation results.

[0022] S1 includes: S11, construct an airborne active reconfigurable smart surface (ARIS) assisted edge buffer network. An ARIS equipped with active load impedance elements is placed on a UAV. The UAV provides an airborne reflection link to ground mobile users according to its current trajectory position. The current spatial position of the UAV, the ARIS reflection coefficient matrix, and the BS beamforming state are determined to form airborne ARIS communication state parameters. Each reflection unit of the ARIS undergoes amplitude and phase control to characterize the ARIS's active reflection adjustment state of the incident signal, expressed as: ; in, For UAV at time Spatial location, , , These represent the UAV's horizontal coordinates, vertical coordinates, and flight altitude, respectively. ; in, For a moment The ARIS reflection coefficient matrix, This refers to the number of ARIS reflector units. For the first Amplitude coefficient of each ARIS reflector unit, For the first The phase of each ARIS reflector unit, The imaginary unit; ; in, For a moment Base station BS for the first Beamforming vectors set by the user. For the number of users; S12: Obtain content request records and user preference information from users in various locations, and obtain the content storage status in the edge cache and the current wireless communication status. Calculate the initial request probability of each content based on the Zipf distribution parameters corresponding to each user and the content sorting. Form the initial user request distribution from the initial request probabilities of each content. Form the user preference representation based on the number of historical requests. Simultaneously, mark the status of each content as cached. Obtain the initial user request distribution, user preferences, cached content status, and wireless communication status, represented as follows: ; ; ; ; ; in, For the first The user on the first The initial request probability of each piece of content. The total number of contents in the content library. For the first Zipf distribution parameters for each user Sort by content numbering. For the first Initial user request distribution for each user The initial user request distribution set for all users. For cached content status, For the first The content is at a certain moment. Cached This indicates that the cache is not cached. The cached content status for all content; S13, based on the airborne ARIS communication state parameters, initial user request distribution, user preferences, cached content status, and wireless communication status, uniformly correlates the UAV trajectory status, ARIS reflection coefficient matrix, BS beamforming status, and user-side, cache-side, and communication-side status at the same time to generate network state information. This network state information is then used as the state input for cache-oriented recommendation and joint optimization of cache strategy and resource allocation, enabling recommendation decisions, cache decisions, and communication resource decisions to be processed collaboratively under the same network state. This is expressed as: ; in, For a moment Network status information, For user preferences, For cached content status, In wireless communication mode. This represents the current spatial position corresponding to the UAV trajectory. This is the ARIS reflection coefficient matrix. This is the BS beamforming state.

[0023] S2 includes: S21, based on user preferences, initial user request distribution, and cached content status in the network state information, for the first... Each user generates candidate content based on their preference for different content types, and different recommendation window sizes are used for each user. Calculate user preference distortion The distortion of user preferences is compared with the first Personalized tolerance threshold for each user When comparing, When the corresponding recommendation window is determined to be a feasible recommendation window, When this happens, the corresponding recommendation window is determined to be an infeasible recommendation window, and the set of feasible recommendation windows consists of all recommendation windows that satisfy the user preference distortion degree not exceeding the personalization tolerance threshold. , represented as: ; ; ; ; ; in, For a moment Recommendation decision, Indicates time Should to the first The first user recommendation For each piece of content, a value of 1 is assigned when recommending it, and 0 otherwise. Indicates the first The user on the first Normalized preference values ​​for each content This refers to the size of the recommendation window, which is the number of recommended content items. Deadline No. The user on the first The number of historical requests for each piece of content. Deadline No. The user on the first The number of historical requests for each piece of content. For use in determining the first The number of historical effective recommendations for each user's personalized tolerance threshold For the first The user in the first User preference distortion corresponding to the second historical effective recommendation; S22, based on user preference distortion, select the feasible recommendation window with the largest recommendation window size as the first... The maximum feasible recommendation window for each user In conjunction with the cached content status, the system determines recommended candidate content from the cached content according to user preferences, ensuring that the number of recommended content does not exceed the maximum feasible recommendation window, and generates recommendation decisions for each user. , represented as: ; The recommended content satisfies: as well as ; ; S23, based on the recommendation decision and the initial user request distribution, the initial request probability corresponding to each user is reshaped using the recommendation decision. This redistributes the request probability corresponding to cached and recommended content towards the cached content, and the reshaped request probabilities are normalized to obtain the cache-oriented request probability for each user. The cache-oriented user request distribution is formed from these cache-oriented request probabilities. The recommendation decision and cache-oriented user request distribution are taken as the cache-oriented recommendation result, represented as: ; ; ; in, For the purpose of recommendation and decision-making, the [number]th individual users at any time For the The probability of cache-directed requests for each piece of content. For the first The cache for each user guides the distribution of user requests. The cache directs user request distribution sets for all users.

[0024] S3 includes: S31, based on the recommendation decision and cache-oriented user request distribution in the cache-oriented recommendation results, combined with network status information, determine the current request content and corresponding wireless communication status of each user, and use Rate Division Multiple Access (RSMA) to split each user message into a common stream and a private stream, encode and transmit the common stream and the private stream respectively, so that each user first decodes the common stream and eliminates the corresponding interference through continuous interference cancellation (SIC), and then decodes its own private stream to obtain the RSMA transmission status. S32, based on RSMA transmission status and network status information, determines UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement, recommendation decision, and RSMA common rate allocation as joint optimization variables. Among them, the recommendation decision adopts the recommendation decision generated in S2, and uses cache-guided user request distribution to represent the user request status after the recommendation mechanism takes effect. This allows UAV spatial position adjustment, ARIS active reflection adjustment, BS beamforming, cache content configuration, recommendation content selection, and RSMA common rate allocation to be coordinated and adjusted in the same joint optimization process, resulting in a set of joint optimization variables. S33. Based on the joint optimization variable set, the latency efficiency and cache hit rate under the corresponding caching strategy and resource allocation state are determined respectively. Dynamic weights are set for latency efficiency and cache hit rate according to business scenarios. The dynamic weighted sum of latency efficiency and cache hit rate is used as the joint optimization objective. Under the common constraints of network state information, cache-oriented recommendation results and RSMA transmission state, a joint optimization problem of caching strategy and resource allocation is constructed.

[0025] S31 includes: S311, Based on the recommendation decisions and cache-oriented user request distribution in the cache-oriented recommendation results, determine the time of each user. The probability of cache-directed requests for each content is determined, and combined with the wireless communication status in the network status information, the current requested content and corresponding wireless communication status of each user are determined, forming a correspondence between user requests and wireless communication status, expressed as: ; ; ; ; ; ; in, For the first individual users at any time Seek the first The probability of each content. Indicates the first individual users at any time For the The request indicator variable for the content, when the first content The current request from the user is number [number]. When there is content, ,otherwise, , For the first individual users at any time The index of the currently requested content. In wireless communication mode. For the first individual users at any time The current wireless communication status, In accordance with the first The operation of retrieving the corresponding wireless communication status from the wireless communication status of all users using a single user identifier. For the first individual users at any time Request—Wireless communication status correspondence item, For a moment The correspondence between user requests and wireless communication status; S312, based on the current request content and corresponding wireless communication status of each user, Rate Division Multiple Access (RSMA) is used to divide the messages corresponding to each user into a common part and a private part. The common parts of each user are combined to form a common stream, and the private streams of each user are formed separately. Beamforming is performed on the common stream and each private stream through the BS beamforming state, and they are superimposed and transmitted to obtain the RSMA transmission signal of the common stream and the private stream, represented as: ; ; ; in, For a moment RSMA transmit signal, For public flow, For the first Private streams for individual users For the common stream beamforming vector, For the first Beamforming vectors for individual user private streams For the first individual users at any time Signal-to-interference-plus-noise ratio (SIR) when decoding a common stream. For the first The conjugate transpose of the channel vector corresponding to each user's wireless communication state For the first Beamforming vectors for individual user private streams For the first Noise power of each user receiver, For the first individual users at any time Decoding rate of the common stream; Since the common stream needs to be decoded by all users, the common stream transmission rate satisfies... ,in, For a moment The common stream transmission rate that all users can complete decoding for; S313, based on RSMA transmission signals for the common and private flows, enables each user to decode the common flow according to its corresponding wireless communication state. After successful decoding of the common flow, continuous interference cancellation (SIC) is used to eliminate interference corresponding to the common flow, and the user then decodes its own private flow. The transmission rate of each user is determined based on the decodeable rate of the common flow, the decodeable rate of the private flow, and the RSMA common rate allocation, generating the RSMA transmission state, represented as: ; ; ; ; in, For the first individual users at any time Signal-to-interference-plus-noise ratio when decoding its own private stream For the first individual users at any time The private stream's decodeable rate, For a moment Assigned to the RSMA common rate for each user For the first individual users at any time The RSMA transmission rate obtained after transmission through both public and private streams.

[0026] S32 includes: S321, based on RSMA transmission status and network status information, extracts the current spatial position of the UAV, the ARIS reflection coefficient matrix, and the BS beamforming status. The UAV spatial positions at consecutive moments are used to form the UAV trajectory. The UAV trajectory, ARIS reflection coefficient matrix, and BS beamforming are used as joint optimization variables on the communication side, enabling coordinated changes in UAV spatial position adjustment, ARIS amplitude and phase adjustment, and BS beamforming for both public and private flows under the same communication state. This is expressed as: ; ; ; ; in, For UAV trajectory, Indicates time UAV spatial location, The total number of times for joint optimization. For a moment The ARIS reflection coefficient matrix, , The first The amplitude coefficient and phase of each ARIS reflector unit. For the first Beamforming vectors for individual user private streams For a moment BS beamforming state For a moment The set of joint optimization variables on the communication side; S322, based on cache-oriented user request distribution and RSMA transmission status, determines the cache placement corresponding to the cache content status, the generated recommendation decision, and the RSMA common rate allocation for each user as joint optimization variables for cache recommendation and transmission resources. The recommendation decision reflects the user request status after the recommendation mechanism takes effect through the cache-oriented user request distribution, and constrains the recommended content by the cache placement status. The RSMA common rate allocation is constrained by the common stream transmission rate, expressed as: ; ; in, The cache placement status for all content. For a moment The set of variables for joint optimization of cache recommendation and transmission resources; S323, based on the joint optimization variable set of the communication side and the joint optimization variable set of cache recommendation and transmission resources, unifies and correlates UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement status, recommendation decision, and RSMA common rate allocation. Combined with cache-guided user request distribution and RSMA transmission status, this allows six types of variables to participate in adjustment under the same network state information, forming a joint optimization variable set for joint optimization of latency efficiency and cache hit rate, expressed as: ; in, For a moment The set of joint optimization variables.

[0027] S33 includes: S331, based on the joint optimization variable set, cache-oriented user request distribution, and RSMA transmission status, obtains the data volume and RSMA transmission rate corresponding to the current request content of each user. It determines the content transmission latency of each user based on the request content data volume and RSMA transmission rate, and determines the latency efficiency corresponding to the current caching strategy and resource allocation status based on the content transmission latency of all users, expressed as: ; ; ; in, For the first individual users at any time Content transmission latency, For the first The content currently requested by the user The amount of data, For a moment Average content transmission latency, For a moment Latency efficiency corresponding to caching strategies and resource allocation states; S332, based on the cache-oriented user request distribution and the cache placement status in the joint optimization variable set, calculates the probability that each user requests cached content, and statistically analyzes the cache hit probability for all users to determine the cache hit rate under the current caching strategy and resource allocation status, expressed as: ; in, For a moment cache hit rate For the first individual users at any time Request No. The probability of a cache hit for a given piece of content that has already been cached; S333, based on latency efficiency and cache hit rate, sets latency efficiency weight and cache hit rate weight according to the current business scenario, and determines the joint optimization objective by dynamically weighting the latency efficiency and cache hit rate. The set of joint optimization variables is used as the variables to be optimized, and the feasible range of the joint optimization variables is determined by combining network state information, cache-oriented recommendation results, and RSMA transmission state. This constructs a joint optimization problem of caching strategy and resource allocation, expressed as: ; ; in, For a moment The joint optimization objective value, For a moment Dynamic weights for latency efficiency determined based on business scenarios. Dynamic weights for cache hit rate To obtain the optimal set of joint optimization variables, The joint optimization feasible region is determined by network state information, cache-guided recommendation results, and RSMA transmission state.

[0028] S4 includes: S41, based on the joint optimization problem of caching strategy and resource allocation, uses network state information as the state of a Markov decision process, the adjustment results corresponding to the set of joint optimization variables as actions, and the joint optimization objective value as the basic immediate reward. It then establishes state transition relationships based on the network state at the next time step after executing actions from the current state, constructing a Markov decision process to describe the continuous decision-making process of caching strategy and resource allocation, expressed as: ; ; ; ; in, For Markov decision-making processes, The state space is composed of network state information. For the action space, This is a state transition relation. For the reward function, For future reward discount factors, For a moment Joint optimization actions, The network state information at the next time step after executing the current joint optimization action; S42, based on the UAV's spatial position, set boundary buffer zones for the X-axis and Y-axis boundaries of the UAV's flight area respectively. When the UAV is within the flight boundary and enters the boundary buffer zone, a soft penalty is calculated. When the UAV goes beyond the flight boundary, a hard penalty is calculated. The penalty values ​​corresponding to the X-axis and Y-axis are summarized to obtain the UAV boundary penalty. Let the boundaries of the UAV's permitted flight area on the X and Y axes be: ; ; in, , These represent the lower and upper boundaries of the permissible flight range along the X-axis, respectively. , These represent the lower and upper boundaries of the allowable flight range along the Y-axis, respectively. Set the boundary buffer distances for the X and Y axes respectively. and The soft penalty on the X-axis is then expressed as: ; The Y-axis soft penalty is represented as follows: ; in, , They are time points The X-axis soft penalty and Y-axis soft penalty are 0 when the UAV is far away from the boundary buffer area, and the soft penalty increases when the UAV enters the boundary buffer area and gradually approaches the flight boundary. The X-axis hard penalty is represented as: ; The Y-axis hard penalty is represented as follows: ; in, , These are the X-axis hard penalty and the Y-axis hard penalty, used to identify whether the UAV has exceeded the corresponding flight boundary; Therefore, soft punishment and hard punishment are represented as follows: ; ; in, For UAV total soft penalty, Total hard penalty for UAV; S43, based on the Markov decision process, joint optimization objective, total soft penalty of the UAV, and total hard penalty of the UAV, the joint optimization objective is taken as the positive reward, and the soft penalty and hard penalty are used as UAV flight boundary constraints to construct an immediate reward. This ensures that the UAV is subject to early constraints when approaching the boundary and stronger constraints when crossing the boundary. The immediate reward, state space, action space, and state transition relationship are uniformly correlated to generate a joint optimization decision model with UAV boundary constraints, expressed as: ; in, For a moment Instant rewards As a soft penalty weight, Hard penalty weight; ; in, For the optimal joint decision-making strategy, This is a joint decision-making strategy for caching and resource allocation. In accordance with the strategy Expected cumulative reward when performing joint optimization actions This is the discount factor.

[0029] S5 includes: S51, based on a joint optimization decision-making model with UAV boundary constraints, obtains the state, action, immediate reward, and next-time state formed during the Markov decision-making process. The state and multiple independently sampled noise vectors are input into a distributed value generator. Multiple value samples are generated through noise-driven random forward propagation. These multiple value samples represent the reward distribution corresponding to the state, and the value estimate of the state is obtained through Monte Carlo averaging, expressed as: ; ; ; in, For the first A noise vector, The mean is zero and the covariance matrix is Gaussian distribution, For the noise vector dimension, For parameters Distributed value generator, For distributed value generators based on state and the Value samples generated from noise vectors For state The corresponding Monte Carlo average value estimate, The number of noise samples; S52, based on generated value and Monte Carlo average value estimation, the TD objective is determined according to the immediate reward and the value estimation of the next time step state. The numerical deviation between each generated value and the TD objective is calculated, and the numerical deviation is normalized to obtain the constraint violation signal. The generated value and the corresponding constraint violation signal are input into the constraint-aware discriminant network. Through adversarial training, the distribution difference and numerical deviation between the generated value and the empirical TD objective are discriminated, and the discriminant results are used to correct the value of the distributed value generator, as shown in: ; ; ; ; ; ; in, For a moment The TD target, For the value estimate corresponding to the state at the next moment, To generate the absolute numerical deviation between the value and the TD target, For the first Constraint violation signals corresponding to each generated value A constant used to prevent the denominator from being zero and to perform numerical normalization. For parameters The constraint-aware discrimination network simultaneously receives generated value and constraint violation signals , For parameters The composite loss function of the distributed value generator, For distributed value generator Network parameters, To counteract the loss, the adversarial feedback of the constraint-aware discriminative network is used to align the reward distribution corresponding to the generated value with the empirical TD target distribution. This is the value correction loss, used to constrain the generated value to numerically approach the TD target. Weighting coefficients to counteract losses, The weighting coefficient for value correction loss; S53, based on the value estimation after correction by the constrained perception discriminant network, calculates the advantage estimation corresponding to the joint optimization action. The ratio of the selection probability of the current policy and the previous policy for the same joint optimization action is used as the policy probability ratio. The pruning policy objective of the PPO optimized by the near-end policy is used to iteratively update the joint decision policy, so that the policy can improve the cumulative reward in the joint optimization decision model with UAV boundary constraints. Based on the updated joint decision policy, the UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement status, recommendation decision and RSMA common rate allocation are determined, and the cache policy and resource allocation results are output.

[0030] S53 includes: S531, based on the TD objective and the value estimate after correction by the constrained perception discrimination network, calculate the advantage estimate corresponding to the joint optimization action in the current state, and calculate the policy probability ratio based on the selection probability of the same joint optimization action by the current joint decision policy and the joint decision policy before the policy update, expressed as: ; ; in, For a moment The advantage estimate, For a moment The probability ratio of the strategies For parameters The current joint decision-making strategy This refers to the joint decision-making strategy before the strategy update. For a moment Joint optimization actions; S532, based on the advantage estimate and policy probability ratio, the PPO pruning policy objective is used to limit the single update magnitude of the joint decision policy, and the joint decision policy is iteratively updated by maximizing the PPO policy optimization objective to obtain the updated joint decision policy, expressed as: ; in, To optimize the PPO strategy, Update the pruning parameters for the PPO strategy. To limit the policy probability ratio to Within the range, To calculate the expectation of the training samples; S533, based on the updated joint decision-making strategy, determines the optimal set of joint optimization variables, obtaining the optimized UAV spatial location, ARIS reflection coefficient matrix, BS beamforming, buffer placement status, recommendation decision, and RSMA common rate allocation, respectively. The optimized UAV spatial locations at consecutive time steps form the UAV trajectory, and the buffering strategy and resource allocation results are output, as follows: ; in, For the optimal joint optimization variable set, For the optimized UAV spatial location, This is the optimized ARIS reflection coefficient matrix. For optimized BS beamforming, The optimized cache placement state, For the optimized recommendation decision, For the optimized RSMA common rate allocation.

[0031] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0032] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A caching strategy and resource allocation method in a recommendation mechanism-driven UAV edge caching network, characterized in that, Includes the following steps: S1. Construct an airborne active reconfigurable smart surface ARIS-assisted edge caching network to obtain the initial user request distribution, user preferences, cache content status and wireless communication status, and generate network status information by combining UAV trajectory, ARIS reflection coefficient matrix and BS beamforming status. S2, based on the network state information, calculate the user preference distortion degree, combine it with the personalized tolerance threshold to determine the maximum feasible recommendation window, generate recommendation decisions from the cached content, reshape the initial user request distribution into a cache-oriented user request distribution, and obtain cache-oriented recommendation results; S3. Based on the cache-guided recommendation results and network status information, RSMA is used to split user messages into public and private streams, and SIC is used to suppress interference between users. UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement, recommendation decision and RSMA common rate allocation are used as joint optimization variables. With the dynamic weighted sum of latency efficiency and cache hit rate as the objective, a joint optimization problem of cache strategy and resource allocation is constructed. S4. Based on the caching strategy and resource allocation joint optimization problem, a Markov decision process is constructed, and soft and hard penalties are set for the UAV flight boundary to generate a joint optimization decision model with UAV boundary constraints. S5. The CAD-PPO algorithm is used to solve the joint optimization decision model. The return distribution is modeled and the value is estimated by a distributed value generator. The constraint violation signal is constructed based on the normalized numerical deviation between the generated value and the TD target. The constraint-aware discrimination network is used for adversarial training and value correction. The policy is updated and the caching policy and resource allocation results are output.

2. The caching strategy and resource allocation method in the UAV edge caching network driven by the recommendation mechanism according to claim 1, characterized in that, S1 includes: S11. Construct an airborne active reconfigurable smart surface ARIS-assisted edge buffer network. The ARIS equipped with active load impedance elements is set on the UAV. The UAV provides an airborne reflection link to the ground mobile user according to the current trajectory position. The UAV spatial position, ARIS reflection coefficient matrix and BS beamforming state at the current moment are determined to form the airborne ARIS communication state parameters. Each reflection unit of the ARIS is subjected to amplitude and phase control to characterize the active reflection adjustment state of the ARIS to the incident signal. S12: Obtain content request records and user preference information of users in various locations, obtain the content storage status in the edge cache and the current wireless communication status, calculate the initial request probability of each content according to the Zipf distribution parameters corresponding to each user and the content sorting, form the initial user request distribution from the initial request probability of each content, form the user preference representation according to the number of historical requests of users, and mark the status of each content as cached, so as to obtain the initial user request distribution, user preference, cached content status and wireless communication status. S13. Based on the airborne ARIS communication status parameters, initial user request distribution, user preferences, cached content status, and wireless communication status, the UAV trajectory status, ARIS reflection coefficient matrix, BS beamforming status, and user-side, cache-side, and communication-side status at the same time are uniformly correlated to generate network status information. The network status information is then used as the status input for cache-oriented recommendation and joint optimization of cache strategy and resource allocation.

3. The caching strategy and resource allocation method in the UAV edge caching network driven by the recommendation mechanism according to claim 2, characterized in that, S2 includes: S21, based on the user preferences, initial user request distribution, and cached content status in the network state information, for the first... Each user generates candidate content based on their preference for different content types, and different recommendation window sizes are used for each user. Calculate user preference distortion The user preference distortion is compared with the first Personalized tolerance threshold for each user When comparing, When, the corresponding recommendation window is determined to be a feasible recommendation window, when When this happens, the corresponding recommendation window is determined to be an infeasible recommendation window, and the set of feasible recommendation windows consists of all recommendation windows that satisfy the user preference distortion degree not exceeding the personalization tolerance threshold. ; S22, based on the user preference distortion, select the feasible recommendation window with the largest recommendation window size as the first... The maximum feasible recommendation window for each user In conjunction with the cached content status, recommended candidate content is determined from the cached content according to user preferences, ensuring that the number of recommended content does not exceed the maximum feasible recommendation window, and a recommendation decision is generated for each user. ; S23, based on the recommendation decision and the initial user request distribution, the initial request probability corresponding to each user is reshaped using the recommendation decision, so that the request probability corresponding to cached and recommended content is redistributed towards the cached content, and the reshaped request probability is normalized to obtain the cache-oriented request probability of each user. The cache-oriented user request distribution is formed by the cache-oriented request probabilities. The recommendation decision and cache-oriented user request distribution are used as the cache-oriented recommendation result.

4. The caching strategy and resource allocation method in the UAV edge caching network driven by the recommendation mechanism according to claim 3, characterized in that, S3 includes: S31, based on the recommendation decision and cache-oriented user request distribution in the cache-oriented recommendation result, and combined with the network status information, determine the current request content and corresponding wireless communication status of each user, and use Rate Division Multiple Access (RSMA) to split each user message into a common stream and a private stream, encode and transmit the common stream and the private stream respectively, so that each user first decodes the common stream and eliminates the corresponding interference through Continuous Interference Cancellation (SIC), and then decodes its own private stream to obtain the RSMA transmission status; S32, based on the RSMA transmission status and network status information, the UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement, recommendation decision, and RSMA common rate allocation are determined as joint optimization variables. The cache-guided user request distribution is used to characterize the user request status after the recommendation mechanism takes effect. This allows UAV spatial location adjustment, ARIS active reflection adjustment, BS beamforming, cache content configuration, recommended content selection, and RSMA common rate allocation to be coordinated and adjusted in the same joint optimization process, resulting in a set of joint optimization variables. S33. Based on the set of joint optimization variables, determine the latency efficiency and cache hit rate under the corresponding caching strategy and resource allocation state, and set dynamic weights for latency efficiency and cache hit rate according to the business scenario. Take the dynamic weighted sum of latency efficiency and cache hit rate as the joint optimization objective, and construct the joint optimization problem of caching strategy and resource allocation under the common constraints of the network state information, cache-oriented recommendation results and RSMA transmission state.

5. The caching strategy and resource allocation method in the UAV edge caching network driven by the recommendation mechanism according to claim 4, characterized in that, S31 includes: S311, based on the recommendation decisions and cache-oriented user request distribution in the cache-oriented recommendation results, determine the time of each user. The cache-directed request probability for each content is determined, and combined with the wireless communication status in the network status information, the current requested content and corresponding wireless communication status of each user are determined, forming a correspondence between user requests and wireless communication status. S312, based on the current request content of each user and the corresponding wireless communication status, the message corresponding to each user is divided into a common part and a private part by using rate division multiple access (RSMA). The common parts of each user are combined to form a common stream and the private streams of the corresponding users are formed respectively. The common stream and each private stream are beamformed and superimposed for transmission by using BS beamforming state to obtain the RSMA transmission signal of the common stream and the private stream. S313, based on the RSMA transmission signal of the public stream and the private stream, each user decodes the public stream according to the corresponding wireless communication state, and after the public stream is successfully decoded, continuous interference cancellation (SIC) is used to eliminate the interference corresponding to the public stream, and the user decodes their own private stream. The transmission rate of each user is determined according to the decodeable rate of the public stream, the decodeable rate of the private stream, and the RSMA public rate allocation, and an RSMA transmission state is generated.

6. The caching strategy and resource allocation method in the UAV edge caching network driven by the recommendation mechanism according to claim 5, characterized in that, S32 includes: S321, based on the RSMA transmission status and network status information, extract the current spatial position of the UAV, the ARIS reflection coefficient matrix and the BS beamforming status, form the UAV trajectory by combining the UAV spatial positions at consecutive times, and use the UAV trajectory, the ARIS reflection coefficient matrix and the BS beamforming as joint optimization variables on the communication side, so that the UAV spatial position adjustment, ARIS amplitude and phase adjustment and the BS beamforming of the common stream and the private stream change collaboratively under the same communication status; S322, based on the cache-oriented user request distribution and RSMA transmission status, the cache placement corresponding to the cache content status, the generated recommendation decision, and the RSMA common rate allocation for each user are determined as joint optimization variables for cache recommendation and transmission resources. The recommendation decision reflects the user request status after the recommendation mechanism takes effect through the cache-oriented user request distribution, and makes the recommended content subject to the cache placement status, and the RSMA common rate allocation subject to the common stream transmission rate. S323, based on the joint optimization variable set of the communication side and the joint optimization variable set of cache recommendation and transmission resources, the UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement status, recommendation decision and RSMA common rate allocation are uniformly associated, and combined with the cache-guided user request distribution and RSMA transmission status, a joint optimization variable set for joint optimization of latency efficiency and cache hit rate is formed.

7. The caching strategy and resource allocation method in the UAV edge caching network driven by the recommendation mechanism according to claim 6, characterized in that, S33 includes: S331, based on the joint optimization variable set, cache-oriented user request distribution and RSMA transmission status, obtain the data volume and RSMA transmission rate corresponding to the current request content of each user, determine the content transmission delay of each user according to the request content data volume and RSMA transmission rate, and determine the delay efficiency corresponding to the current caching strategy and resource allocation status according to the content transmission delay of all users. S332, based on the cache-oriented user request distribution and the cache placement status in the joint optimization variable set, calculate the probability of each user requesting cached content, and statistically analyze the cache hit probability for all users to determine the cache hit rate under the current caching strategy and resource allocation status. S333, based on the latency efficiency and cache hit rate, set latency efficiency weight and cache hit rate weight according to the current business scenario, and determine the dynamic weighted sum of latency efficiency and cache hit rate as the joint optimization objective. Take the set of joint optimization variables as the variables to be optimized, and combine network status information, cache-oriented recommendation results and RSMA transmission status to determine the feasible range of joint optimization variables, and construct the joint optimization problem of cache strategy and resource allocation.

8. The caching strategy and resource allocation method in the UAV edge caching network driven by the recommendation mechanism according to claim 7, characterized in that, S4 includes: S41, Based on the joint optimization problem of caching strategy and resource allocation, network state information is used as the state of Markov decision process, the adjustment result corresponding to the set of joint optimization variables is used as action, the joint optimization objective value is used as basic immediate benefit, and a state transition relationship is established according to the network state at the next moment formed after the current state performs the action, thus constructing a Markov decision process to describe the continuous decision process of caching strategy and resource allocation. S42, based on the UAV's spatial position, set boundary buffer areas for the X-axis and Y-axis boundaries of the UAV's flight area respectively. When the UAV is within the flight boundary and enters the boundary buffer area, a soft penalty is calculated. When the UAV goes beyond the flight boundary, a hard penalty is calculated. The penalty values ​​corresponding to the X-axis and Y-axis are summarized respectively to obtain the UAV boundary penalty. S43. Based on the Markov decision process, the joint optimization objective, the total soft penalty of the UAV, and the total hard penalty of the UAV, the joint optimization objective is taken as the positive reward, and the soft penalty and hard penalty are taken as UAV flight boundary constraints to construct an immediate reward. The immediate reward, state space, action space, and state transition relationship are uniformly associated to generate a joint optimization decision model with UAV boundary constraints.

9. The caching strategy and resource allocation method in a recommendation mechanism-driven UAV edge caching network according to claim 8, characterized in that, S5 includes: S51, Based on the joint optimization decision model with UAV boundary constraints, obtain the state, action, immediate reward and next time state formed in the Markov decision process, input the state and multiple independently sampled noise vectors into the distributed value generator, generate multiple value samples through noise-driven random forward propagation, use the multiple value samples to represent the reward distribution corresponding to the state, and obtain the value estimate of the state through Monte Carlo averaging. S52, based on the generated value and Monte Carlo average value estimation, the TD target is determined according to the value estimation of the immediate reward and the state at the next time step. The numerical deviation between each generated value and the TD target is calculated, and the numerical deviation is normalized to obtain the constraint violation signal. The generated value and the corresponding constraint violation signal are input into the constraint perception discrimination network. The distribution difference and numerical deviation between the generated value and the empirical TD target are judged through adversarial training, and the discrimination result is used to correct the value of the distributed value generator. S53. Based on the value estimation after correction by the constrained perception discriminant network, the advantage estimation corresponding to the joint optimization action is calculated. The ratio of the selection probability of the same joint optimization action by the current policy and the policy before the update is used as the policy probability ratio. The pruning policy objective of the PPO is optimized by the near-end policy to iteratively update the joint decision policy. Based on the updated joint decision policy, the UAV trajectory, ARIS reflection coefficient matrix, BS beamforming, cache placement status, recommendation decision and RSMA common rate allocation are determined, and the cache policy and resource allocation results are output.

10. The caching strategy and resource allocation method in a recommendation mechanism-driven UAV edge caching network according to claim 9, characterized in that, S53 includes: S531, based on the TD target and the value estimate after correction by the constrained perception discrimination network, calculate the advantage estimate corresponding to the joint optimization action in the current state, and calculate the policy probability ratio based on the selection probability of the same joint optimization action by the current joint decision policy and the joint decision policy before the policy update. S532, based on the advantage estimation and policy probability ratio, the PPO pruning policy objective is used to limit the single update range of the joint decision policy, and the joint decision policy is iteratively updated by maximizing the PPO policy optimization objective to obtain the updated joint decision policy; S533, based on the updated joint decision-making strategy, determine the optimal set of joint optimization variables, and obtain the optimized UAV spatial location, ARIS reflection coefficient matrix, BS beamforming, buffer placement status, recommendation decision and RSMA common rate allocation respectively. The optimized UAV spatial location at continuous time intervals forms the UAV trajectory, and the buffer strategy and resource allocation results are output.