Time differentiation-oriented common inductance computing resource allocation method

Abstract

The invention belongs to the technical field of communication computing resource allocation, and particularly relates to a time-differentiation-oriented communication computing resource allocation method, which comprises the following steps of: constructing a time-differentiation ISCC system model which comprises a base station, a mobile user and a sensing target; establishing a time-varying request model, a service cache model, a communication model, a calculation model and a perception model; with maximization of a network utility function as a target, a joint optimization problem including service caching, task unloading, spectrum resources, power resources and computing power resources is established under delay tolerance constraint and storage capacity constraint; converting a joint optimization problem into a partially observable Markov decision process, taking a base station as an intelligent agent, and defining a triple; solving by adopting a depth deterministic strategy gradient algorithm based on LSTM; according to the invention, the LSTM network is utilized to capture historical information, and the intelligent agent makes an optimal decision.

Description

Technical Field

[0001] The present invention belongs to the technical field of synaesthesia resource allocation, and in particular relates to a synaesthesia resource allocation method oriented to time differentiation. Background Art

[0002] With the rapid development of technologies such as 5G, Internet of Things (IoT), artificial intelligence (AI), edge computing, and autonomous driving, the challenges faced by modern networks are increasing. Future networks require not only high-speed and low-latency communications, but also the ability to perceive the environment in real time and perform intelligent computing and decision-making. The core goal of the ISCC network is to organically combine perception, communication, and computing to form a collaborative and intelligent network system. This network system can collect and process environmental information in real time, while efficiently communicating and computing data to support smarter application scenarios. With limited communication and computing resources, how to dynamically process perception, communication, and computing tasks to achieve optimal network resource allocation.

[0003] However, in ISCC network scenarios, current research often focuses on simple network structures and consistent task characteristics, without fully considering time-differentiated, deeply uncertain multi-dimensional network states. The rapid growth of edge nodes and task traffic, the huge amount of equipment and data make the network state difficult to predict or even unpredictable, and bring obvious uncertainty. Applications such as autonomous driving, industrial Internet, and smart cities have different task requirements and time distribution characteristics, resulting in strong time dependence of communication systems in communication and computing. Therefore, it is of great significance to deeply study the uncertain network state and user task time dependence and design an efficient resource allocation solution for ISCC networks.

[0004] Deep reinforcement learning (DRL) combines the decision-making ability of reinforcement learning with the feature extraction ability of deep learning, and can cope with the multi-dimensional and complex environment of ISCC. However, in the highly dynamic ISCC of the future, the time dependency of user task requests leads to poor results of traditional deep learning in extracting network features, and thus low resource allocation efficiency of the entire network. Therefore, how to design a more flexible resource allocation method in the time-differentiated ISCC network has become an urgent problem to be solved in current research. Summary of the invention

[0005] In order to solve the above problems and process time differentiation more effectively, the present invention provides a time-differentiated synaesthesia resource allocation method, comprising the following steps:

[0006] S1. Construct a time-differentiated ISCC system model, which includes base stations, mobile users and sensing targets;

[0007] S2. Based on the time-differentiated ISCC system model, establish the time-varying request model, service cache model, communication model, calculation model and perception model;

[0008] S3. Based on the various models constructed in steps S1-S2, with the goal of maximizing the network utility function, under the constraints of delay tolerance and storage capacity, a joint optimization problem including service cache, task offloading, spectrum resources, power resources and computing resources is established;

[0009] S4. Convert the joint optimization problem into a partially observable Markov decision process, take the base station as the intelligent agent, and define the triple {s, a, r}, where s represents the global state space, a represents the global action space set, and r represents the reward function;

[0010] S5. Use LSTM-based deep deterministic policy gradient algorithm to solve, including:

[0011] S51. Network Modeling:

[0012] Construct actor network μ(s|θ μ ) and the critic network Q(s,a|θ Q ), θ μ represents the actor network parameters, θ Q represents the critic network parameters; add LSTM network to the actor network and use h a (t) represents the network status information of the actor network at time slot t; an LSTM network is added to the critic network, and h c (t) represents the network status information of the critic network at time slot t; the actor network corresponds to an actor target network μ' with the same structure and parameter θ μ '; The critic network corresponds to a critic target network Q' with the same structure, and the parameter is θ Q ';

[0013] S52. Network Update:

[0014] According to the noise ξ and the current strategy μ(s(t),h a (t)|θ μ ) selects action a(t), where s(t) is the state of the current time slot t; the agent performs action a(t) and obtains reward r(t) and the next state s(t+1);

[0015] Minimize the loss function L(θ Q ) Update the critic network parameters to maximize the policy gradient Update actor network parameters;

[0016] Update the actor target network parameters and the critic target network parameters through soft update;

[0017] Repeat the above content and update continuously until the optimal solution is finally obtained.

[0018] Beneficial effects of the present invention:

[0019] This paper proposes a time-differentiated ISCC resource allocation method. Under the constraints of multi-dimensional resources such as delay tolerance and storage capacity, considering factors such as user time-varying task requests and resource constraints, a joint optimization problem of service cache, task offloading, spectrum resource allocation, power resource allocation, and computing power resource allocation is established to maximize the utility function of the ISCC network. Finally, the optimization problem is transformed into a Markov decision process, and a deep deterministic policy gradient algorithm based on LSTM is proposed. The LSTM network is used to capture historical information, and the intelligent agent makes the best decision. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the joint optimization method of service caching, task offloading, spectrum resources, power resources and computing power resource allocation in the time-differentiated ISCC network of the present invention;

[0021] Figure 2 It is a model diagram of the ISCC network system oriented to time differentiation in the present invention;

[0022] Figure 3 This is a framework diagram of the deep deterministic policy gradient algorithm of LSTM in the present invention;

[0023] Figure 4 This is a performance comparison chart of the proposed algorithm, the DDPG algorithm, and the DDPG-LFU algorithm under different BS cache capacities in the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] The present invention provides a time-differentiated synaesthesia computing resource allocation method, such as Figure 1 As shown, the following steps are included:

[0026] S1. Construct a time-differentiated ISCC system model, which includes base stations, mobile users and sensing targets.

[0027] Specifically, Figure 2 As shown in the figure, the present invention considers a time-differentiated ISCC system model, in which there is a base station (BS) equipped with ISAC (Integrated Sensing And Communications) technology and MEC server, K mobile users (UEs) and Q sensing targets, and the user set is recorded as Perception target set The ISCC system supports communication, perception and computing functions. The communication function is concentrated on the wireless communication link, the perception function is concentrated on the BS estimating the perception information related to the target by transmitting the echo signal to the perception target, and the computing function is concentrated on processing the computing tasks of the user and processing the perception information obtained by the BS. The entire ISCC system runs in discrete time slots, and the time set T represents the total number of time slots in a time period; defining time slots The computationally intensive task generated by mobile user k is d = {d1, d2, d K}; where d k ∈d represents a mobile user The task generated in time slot t is defined as d k ={D k (t),C k (t),X k ,F k}, D k (t) represents the data size of the task generated by mobile user k in time slot t, in bits; C k (t) represents the maximum tolerable delay of the task generated by mobile user k in time slot t, X k represents the number of CPU cycles required to process k-bit tasks of mobile users, F k represents the service type required to process the task generated by mobile user k.

[0028] S2. Based on the time-differentiated ISCC system model, a time-varying request model, service cache model, communication model, calculation model and perception model are established.

[0029] Preferably, the present invention assumes that task requests cover multiple service types and have periodic changes. To improve user experience, tasks need to be processed in a timely manner, so popular content is cached at the edge of the network rather than in the cloud. The BS has limited cache capacity, and the cached content needs to be dynamically updated according to time changes. Assume there are I services, which are composed of a set If collected, the corresponding service The task can only be processed by the i-th service placed on the BS or cloud server. Represents a service library. It contains multiple large time periods T. The service ranking changes over time within period T, and the change interval is set to There are V service popularity in period T, and the popularity of each service remains unchanged within time τ.

[0030] Specifically, for the time-varying request model, the ranking change set of all service types under the base station within the time period T is defined as g = {g1, g2, ..., g V}, g v represents the popularity of v=1,2,…,V services, where V represents the number of service popularity; definition Indicates that the base station is down The ranking in the popularity of the vth service; the popularity of each service changes dynamically and follows the Zipf distribution; define the probability P of mobile user k requesting service i under the base station i for:

[0031]

[0032] in, represents the ranking of service i in the popularity of the vth service, δ k is the skewness coefficient, as δ k As the number of hotspots increases, the likelihood that users will request hotspot services also increases.

[0033] Specifically, for the service cache model, the service cache decision set of the base station in time slot t is defined as α(t) = {α1(t), α2(t), ..., α i (t),...,α I (t)}, where the binary variable αi(t)∈{0,1} represents whether service i in time slot t is cached to the caching strategy of the base station. i (t) = 1, it means that service i is cached at the base station in time slot t; if α i (t) = 0, it means that service i is not cached at the base station in time slot t. However, due to the limited storage capacity of the MEC server, the total capacity of the cached services cannot exceed the MEC server storage limit, and the following constraints must be met:

[0034]

[0035] Among them, BS Indicates the maximum cache capacity of BS; i represents the cache space required to cache service i at the BS.

[0036] Specifically, for the communication model, if the service corresponding to the task generated by mobile user k is i, when the base station has service i cached, the task is directly transmitted to the base station for processing, otherwise the task is forwarded by the base station to the cloud (i.e., the cloud end) for processing. The specific definition is as follows:

[0037] Define the spectrum resource allocation strategy x(t) as

[0038]

[0039] in represents the spectrum resource allocation coefficient of the base station to mobile user k, represents the spectrum resource allocation coefficient of the base station to the sensing target q;

[0040] Define the power resource allocation strategy y(t) as

[0041]

[0042] in represents the power resource allocation coefficient of the base station to mobile user k, represents the power resource allocation coefficient of the base station to the sensing target q;

[0043] Define the computing resource allocation strategy z(t) as

[0044]

[0045] in represents the computing resource allocation coefficient of the base station to mobile user k, represents the computing resource allocation coefficient of the base station to the sensing target q;

[0046] During the communication transmission process, the communication rate R of mobile user k k (t) is expressed as

[0047]

[0048] Where P represents the total power available to the base station, B represents the total spectrum resources available to the base station, represents the noise-normalized communication channel gain, which is calculated as

[0049]

[0050] represents the array gain factor, N tran Indicates the number of transmitting antennas, N recv Indicates the number of receiving antennas; represents the channel gain of mobile user k, represents a column of the beamforming matrix received by the base station, represents the transpose of a column of the beamforming matrix received by the base station, represents a column of the transmit precoding matrix of mobile user k, represents the transmission steering vector of mobile user k, a recv (θ k ) represents the receiving steering vector of mobile user k, which is specifically expressed as

[0051]

[0052]

[0053] θ k represents the angle of mobile user k under BS;

[0054] The communication delay of the task from mobile user k to the base station is

[0055]

[0056] The communication delay of the task from the base station to the cloud is

[0057]

[0058] Indicates the transmission rate from the base station to the cloud;

[0059] Therefore, the communication delay of mobile user k's transmission task is Expressed as

[0060]

[0061] The binary variable α i (t)∈{0,1} represents whether service i in time slot t is cached to the base station cache strategy. If α i (t) = 1, it means that service i is cached at the base station in time slot t; if α i (t)=0, which means that service i is not cached at the base station in time slot t.

[0062] Specifically, for the computational model, there are two processing methods for any task generated by a mobile user in the present invention. The first is the base station processing task, and the second is the cloud processing task. Define the offloading strategy set β(t) for time slot t = {β1(t), β2(t), ..., β K (t)}, represents the task offloading strategy of mobile user k in time slot t, represents the strategy of mobile user k transferring tasks to the base station for processing. It means that mobile user k transmits the task to the base station for processing. It means that mobile user k does not transmit the task to the base station for processing; ψ k (t)∈{0,1} represents the strategy of mobile user k transferring tasks to the cloud for processing, if ψ k (t) = 1, it means that mobile user k transfers the task to the cloud for processing. k (t) = 0, which means that mobile user k does not transfer the task to the cloud for processing.

[0063]

[0064] 1) Tasks are transmitted to the base station for processing

[0065] When the base station caches the service i corresponding to the task generated by mobile user k, the task generated by mobile user k can be directly transmitted to the base station through the wireless link for processing. Base station calculation delay It is expressed as:

[0066]

[0067] Where D k (t)X k Indicates the number of CPU cycles to complete the task, f BS Indicates the CPU frequency of the base station. represents the computing resource allocation coefficient of the base station to mobile user k;

[0068] 2) Tasks are transferred to the cloud for processing

[0069] The base station does not cache the service i corresponding to the task generated by mobile user k. Consider transmitting the task generated by mobile user k to the cloud for processing. At this time, ψ k (t) = 1, cloud computing latency It is expressed as:

[0070]

[0071] Among them, f CS Indicates the CPU frequency of the cloud,

[0072] Therefore, the computational delay of the task of mobile user k is for

[0073]

[0074] The binary variable α i (t)∈{0,1} represents whether service i in time slot t is cached to the base station cache strategy. If α i (t) = 1, it means that service i is cached at the base station in time slot t; if α i(t)=0, which means that service i is not cached at the base station in time slot t.

[0075] Specifically, for the perception model, the perception model of the present invention mainly focuses on the positioning accuracy of the perception target.

[0076] Define the perceptual accuracy of the perceptual target q for

[0077]

[0078] Where γ1 and γ2 represent normalization factors, CRB(d q ) represents the distance estimation of the perceived target q, CRB(θ q ) represents the direction estimation of the perceived target q, and the Cramr-Rao Bound (CRB) of the qth perceived target is evaluated in a generalized way. The specific calculation formula is:

[0079]

[0080] Where χ1 and χ2 represent constants related to system configuration, signal design, beam gain, matched filter gain, and specific signal processing algorithms, P represents the total power available to the base station, B represents the total spectrum resources available to the base station, and G q represents the perceptual channel gain normalized by noise, represents the spectrum resource allocation coefficient of the base station to the sensing target q, represents the power resource allocation coefficient of the base station to the sensing target q;

[0081] After the base station realizes the perception and positioning service of the perceived target q, the computational delay of computing and processing the radar raw data collected by the radar front end is for

[0082]

[0083] Among them, S k (t) is the amount of original radar data perceiving target q, X q Indicates the number of CPU cycles required to process a unit bit of radar raw data. Represents the computing resource allocation coefficient of the base station to the perception target q.

[0084] S3. Based on the various models constructed in steps S1-S2, with the goal of maximizing the network utility function, under the constraints of delay tolerance and storage capacity, a joint optimization problem including service cache, task offloading, spectrum resources, power resources and computing resources is established.

[0085] Specifically, the network utility function u(t) of the time-differentiated ISCC system model is defined as

[0086]

[0087] Among them, λ1, λ2, and λ3 represent the weight coefficients of computing utility, communication utility, and perception utility, respectively. Indicates the maximum tolerable delay of task calculation, Indicates the maximum tolerable communication delay of the task, Indicates the perceived maximum tolerable delay.

[0088] Since the resources in the ISCC system are limited, in order to maximize the network utility of the entire ISCC system, the joint optimization problem of service cache, task offloading, spectrum resources, power resources and computing resources is defined as

[0089]

[0090] Among them, the first constraint is the base station service cache capacity limit, which means that the service cached by the base station cannot exceed the storage capacity of the base station; the second constraint is the spectrum resource allocation limit, which means that the sum of the spectrum resources of mobile users and perception targets cannot exceed the total spectrum resources of the base station; the third constraint is the power resource allocation limit, which means that the sum of the power resources of mobile users and perception targets cannot exceed the total power resources of the base station; the fourth constraint is the computing power resource allocation limit, which means that the sum of the computing power resources of mobile users and perception targets cannot exceed the computing network resources of the base station; the fifth constraint is the communication tolerance delay limit, which means that the communication transmission delay cannot exceed the maximum tolerance value; the sixth constraint is the calculation tolerance delay limit, which means that the calculation processing delay cannot exceed the maximum tolerance value; the seventh constraint is the perception tolerance delay limit, which means that the perception processing delay cannot exceed the maximum tolerance value.

[0091] S4. Convert the joint optimization problem into a partially observable Markov decision process, take the base station as the intelligent agent, and define the triple {s, a, r}, where s represents the global state space, a represents the global action space set, and r represents the reward function.

[0092] Specifically, the resource allocation problem in the ISCC system is affected by the highly dynamic changes in the network environment, the time-differentiated requests of user tasks, and the changes in the popularity of cached content. Traditional convex optimization methods are difficult to solve. Deep reinforcement learning has the ability to balance multi-objective optimization in complex state spaces. Therefore, the joint optimization problem is transformed into a partially observable Markov decision process (DEC-POMDP), with the base station acting as an intelligent agent, and the triple {s, a, r} is defined, where:

[0093] 1) Environmental status

[0094] s = {s(1), s(2), ..., s(T)} represents the global state space; s(t)∈s represents the total information of the time-varying request tasks and the perceived target requests received by the agent in the coverage area at time slot t, which can be expressed as

[0095] s(t)={d1,d2,...,d K ,P,B}

[0096] d k represents the task generated by mobile user k, P represents the total power available to the base station, and B represents the total spectrum resources available to the base station.

[0097] 2) Action Space

[0098] a={a(1),a(2),…,a(T)} represents the global action space set, a(t)∈a represents the time slot t when the agent selects the corresponding action from the action space according to the environment state, observation information and current strategy, which is expressed as

[0099] a(t)={α(t),β(t),x(t),y(t),z(t)}

[0100] 3) Reward Function

[0101] The reward function measures the effect of the agent's actions in a specific state. The reward in this invention consists of two parts: the first part is the ISCC network utility, and the second part is the cache capacity limit. The optimization goal is to maximize the network utility function of the ISCC system, so the calculation formula for the cumulative reward r(t) obtained by the agent performing an action in time slot t is

[0102]

[0103] Among them, λ1, λ2, λ3 are weight coefficients, O(t) represents the cache capacity limit, which can be expressed as

[0104]

[0105] H(·) represents the Heaviside step function, ε represents the weight coefficient, O i represents the buffer space required by the base station to buffer service type i, O BS Indicates the maximum cache capacity of the base station.

[0106] S5. Use LSTM-based deep deterministic policy gradient algorithm to solve.

[0107] The LSTM network is a special recursive neural network that can effectively solve the gradient vanishing problem in long-distance sequences. Its structure includes a forget gate, an input gate, and an output gate, which are responsible for discarding, updating memory, and generating output, respectively. In view of the time dependency of task requests, the present invention considers introducing an LSTM network to extract time-related network state information, and therefore proposes a deep deterministic policy gradient algorithm (LDPGA) based on LSTM, such as Figure 3 shown.

[0108] Specifically, step S5 includes:

[0109] S51. Network Modeling:

[0110] The LDPGA framework is shown in Figure 3. It uses two different DNNs to approximate the actor network μ(s|θ μ ) and the critic network Q(s,a|θ Q ), θ μ represents the actor network parameters, θ Q represents the critic network parameters; add LSTM network to the actor network and use h a (t) represents the network status information of the actor network at time slot t; an LSTM network is added to the critic network, and h c (t) represents the network status information of the critic network at time slot t; the actor network corresponds to an actor target network μ' with the same structure and parameter θ μ '; The critic network corresponds to a critic target network Q' with the same structure, and the parameter is θ Q ';

[0111] S52. Network Update:

[0112] According to the noise ξ and the current strategy μ(s(t),h a (t)|θ μ ) selects action a(t), where s(t) is the state of the current time slot t; the agent performs action a(t) and obtains reward r(t) and the next state s(t+1).

[0113] Minimize the loss function L(θ Q ) Update the critic network parameters to maximize the policy gradient Update actor network parameters.

[0114] Specifically, the loss function L(θ Q ) is expressed as

[0115]

[0116] η(t)=r(t)+νQ(s(t+1),h c (t+1),μ(s(t+1))|θ Q )

[0117] in, represents the expected value, and v represents the discount factor;

[0118] Policy Gradient Expressed as

[0119]

[0120] represents the gradient operation, and J represents the strategy to be optimized;

[0121] In order to make the loss function L(θ Q ) is minimized, and the critic network Q will be updated by the given Adam optimizer. Then, the actor network μ sends a small batch of actions a = μ(s(t+1)) to the critic network to achieve the gradient of action a parameter can be derived by its own optimizer. Using these two gradients, the network can be updated with the following approximation:

[0122]

[0123] Update the actor target network parameters and the critic target network parameters by soft updating, denoted as θ μ '←τ a θ μ +(1-τ a )θ μ '

[0124] θ Q '←τ c θ Q +(1-τ c )θ Q '

[0125] τ a represents the actor network update coefficient, τ c Represents the critic network update coefficient.

[0126] Repeat the above content and update continuously until the optimal solution is finally obtained.

[0127] The implementation process of the resource allocation strategy based on the LDPGA algorithm is as follows

[0128]

[0129]

[0130] The present invention simulates and verifies the LDPGA algorithm on the Pycharm platform and compares it with the DDPG and DDPG-LFU algorithms. The LFU cache algorithm eliminates the least used services and only retains the commonly used services.

[0131] from Figure 4 It can be seen that as the BS cache capacity increases, the average rewards of the three algorithms increase, and LDPGA performs better than other algorithms. Since LDPGA introduces the LSTM network, it can process time-related network status information, and the agent makes better decisions and obtains higher rewards. However, DDPG can only use the current status information, and DDPG-LFU cannot learn the cache strategy, and the performance is poor. The effectiveness of LDPGA in ISCC network resource allocation is verified.

[0132] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0133] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A time-differentiated synaesthesia resource allocation method, characterized in that: The following steps are involved: S1. Construct a time-differentiated ISCC system model, which includes base stations, mobile users and sensing targets; S2. Based on the time-differentiated ISCC system model, establish the time-varying request model, service cache model, communication model, calculation model and perception model; S3. Based on the various models constructed in steps S1-S2, with the goal of maximizing the network utility function, under the constraints of delay tolerance and storage capacity, a joint optimization problem including service cache, task offloading, spectrum resources, power resources and computing resources is established; S4. Convert the joint optimization problem into a partially observable Markov decision process, take the base station as the intelligent agent, and define the triple {s, a, r}, where s represents the global state space, a represents the global action space set, and r represents the reward function; S5. Use LSTM-based deep deterministic policy gradient algorithm to solve, including: S51. Network Modeling: Construct actor network μ(s|θ μ ) and the critic network Q(s,a|θ Q ), θ μ represents the actor network parameters, θ Q represents the critic network parameters; add LSTM network to the actor network and use h a (t) represents the network status information of the actor network at time slot t; an LSTM network is added to the critic network, and h c (t) represents the network status information of the critic network at time slot t; the actor network corresponds to an actor target network μ' with the same structure and parameter θ μ '; The critic network corresponds to a critic target network Q' with the same structure, and the parameter is θ Q '; S52. Network Update: According to the noise ξ and the current strategy μ(s(t),h a (t)|θ μ ) selects action a(t), where s(t) is the state of the current time slot t; the agent performs action a(t) and obtains reward r(t) and the next state s(t+1); Minimize the loss function L(θ Q ) Update the critic network parameters to maximize the policy gradient Update actor network parameters; Update the actor target network parameters and the critic target network parameters through soft update; Repeat the above content and update continuously until the optimal solution is finally obtained.

2. The method for allocating time-differentiated synaesthesia computing resources according to claim 1, characterized in that: In the time-differentiated ISCC system model, there is a base station equipped with ISAC technology and MEC server, K mobile users and Q sensing targets. Perception target set The time-differentiated ISCC system model operates in discrete time slots, and the time set T represents the total number of time slots in a time period; defining time slots The task generated by mobile user k under the base station is d = {d1, d2, d K }; where d k ∈d represents a mobile user The task generated in time slot t is defined as d k ={D k (t),C k (t),X k ,F k }, D k (t) represents the data size of the task generated by mobile user k in time slot t, C k (t) represents the maximum tolerable delay of the task generated by mobile user k in time slot t, X k represents the number of CPU cycles required to process k-bit tasks of mobile users, F k Represents the services required to process tasks generated by mobile user k; defines the service library I represents the number of services.

3. The time-differentiated synaesthesia resource allocation method according to claim 2, characterized in that: Based on the time-differentiated ISCC system model, a time-varying request model is established. In the period T, there are multiple large time periods T. The service ranking changes with time in the period T, and the change interval is set to τ. There are V types of service popularity in the period T, and the popularity of each service remains unchanged in the time τ. The ranking change set of all service types of the base station in the period T is g = {g1, g2, ..., g V }, g v represents the popularity of v=1,2,…,V services, where V represents the number of service popularity; definition Indicates that the base station is down The ranking in the popularity of the vth service; the popularity of each service changes dynamically and follows the Zipf distribution; define the probability P of mobile user k requesting service i under the base station i,k for: in, represents the ranking of service i in the popularity of the vth service, δ k is the service skewness coefficient.

4. The method for allocating time-differentiated synaesthesia computing resources according to claim 2, characterized in that: Based on the time-differentiated ISCC system model, a service cache model is established. In the service cache model, the service cache decision set of the base station in time slot t is defined as α(t) = {α1(t), α2(t), ..., α i (t),...,α I (t)}, where the binary variable α i (t)∈{0,1} represents whether service i in time slot t is cached to the base station cache strategy. If α i (t) = 1, it means that service i is cached at the base station in time slot t; if α i (t)=0, which means that service i is not cached at the base station in time slot t.

5. The time-differentiated synaesthesia resource allocation method according to claim 2, characterized in that: A communication model is established based on the time-differentiated ISCC system model. In the communication model, if the task generated by mobile user k corresponds to service i, when service i is cached at the base station, the task is directly transmitted to the base station for processing, otherwise the task is forwarded by the base station to the cloud for processing. The specific definition is as follows: Define the spectrum resource allocation strategy x(t) as in represents the spectrum resource allocation coefficient of the base station to mobile user k, represents the spectrum resource allocation coefficient of the base station to the sensing target q; Define the power resource allocation strategy y(t) as in represents the power resource allocation coefficient of the base station to mobile user k, represents the power resource allocation coefficient of the base station to the sensing target q; Define the computing resource allocation strategy z(t) as in represents the computing resource allocation coefficient of the base station to mobile user k, represents the computing resource allocation coefficient of the base station to the sensing target q; During the communication transmission process, the communication rate R of mobile user k k (t) is expressed as Where P represents the total power available to the base station, B represents the total spectrum resources available to the base station, represents the noise-normalized communication channel gain, which is calculated as represents the array gain factor, N tran Indicates the number of transmitting antennas, N recv Indicates the number of receiving antennas; represents the channel gain of mobile user k, represents a column of the beamforming matrix received by the base station, represents a column of the transmit precoding matrix of mobile user k, represents the transmission steering vector of mobile user k, a recv (θ k ) represents the receiving steering vector of mobile user k, which is specifically expressed as θ k represents the angle of mobile user k under BS; The communication delay of the task from mobile user k to the base station is The communication delay of the task from the base station to the cloud is Indicates the transmission rate from the base station to the cloud; Therefore, the communication delay of mobile user k's transmission task is Expressed as The binary variable α i (t)∈{0,1} represents whether service i in time slot t is cached to the base station cache strategy. If α i (t) = 1, it means that service i is cached at the base station in time slot t; if α i (t)=0, which means that service i is not cached at the base station in time slot t.

6. The method for allocating time-differentiated synaesthesia resources according to claim 2, characterized in that: A computational model is established based on the time-differentiated ISCC system model, and a set of unloading strategies for time slot t is defined in the computational model. represents the task offloading strategy of mobile user k in time slot t, represents the strategy of mobile user k transferring tasks to the base station for processing. It means that mobile user k transmits the task to the base station for processing. It means that mobile user k does not transmit the task to the base station for processing; ψ k (t)∈{0,1} represents the strategy of mobile user k transferring tasks to the cloud for processing, if ψ k (t) = 1, it means that mobile user k transfers the task to the cloud for processing. k (t) = 0, which means that mobile user k does not transfer the task to the cloud for processing. When mobile user k transmits the task to the base station for processing, the calculation delay of the base station is It is expressed as: Where D k (t)X k Indicates the number of CPU cycles to complete the task, f BS Indicates the CPU frequency of the base station. represents the computing resource allocation coefficient of the base station to mobile user k; When mobile user k transfers the task to the cloud for processing, the computing delay of the cloud It is expressed as: Among them, f CS Indicates the CPU frequency of the cloud, Therefore, the computational delay of the task of mobile user k is for The binary variable α i (t)∈{0,1} represents the cache strategy of whether service i in time slot t is cached to the base station. If α i (t) = 1, it means that service i is cached at the base station in time slot t; if α i (t)=0, which means that service i is not cached at the base station in time slot t.

7. The method for allocating time-differentiated synaesthesia computing resources according to claim 2, characterized in that: A perception model is established based on the time-differentiated ISCC system model, and the perception accuracy of the perception target q is defined in the perception model. for Where γ1 and γ2 represent normalization factors, CRB(d q ) represents the distance estimation of the perceived target q, CRB(θ q ) represents the direction estimation of the perceived target q, and the specific calculation formula is: Where χ1 and χ2 represent constants, P represents the total power available to the base station, B represents the total spectrum resources available to the base station, and G q represents the perceptual channel gain normalized by noise, represents the spectrum resource allocation coefficient of the base station to the sensing target q, represents the power resource allocation coefficient of the base station to the sensing target q; After the base station realizes the perception and positioning service of the perceived target q, the computational delay of computing and processing the radar raw data collected by the radar front end is for Among them, S k (t) is the amount of radar raw data perceiving target q, X q Indicates the number of CPU cycles required to process a unit bit of radar raw data. Represents the computing resource allocation coefficient of the base station to the perception target q.

8. The time-differentiated synaesthesia resource allocation method according to claim 1, characterized in that: The network utility function u(t) of the time-differentiated ISCC system model is defined as Among them, λ1, λ2, λ3 are weight coefficients, Represents a collection of users, represents the set of sensing targets, K represents the number of mobile users, and Q represents the number of sensing targets; Indicates the maximum tolerable delay of task calculation, represents the computational delay of the task of mobile user k, Indicates the maximum tolerable communication delay of the task, represents the communication delay of the transmission task of mobile user k, represents the perception accuracy of the perceived target q, represents the maximum tolerable delay of perception, represents the computational delay of the radar raw data of the perceived target q; thus, the joint optimization problem is defined as Among them, α(t) represents the service cache decision set of the base station, x(t) represents the spectrum resource allocation strategy, y(t) represents the power resource allocation strategy, z(t) represents the computing resource allocation strategy, and β(t) represents the offloading strategy set; the binary variable α i (t)∈{0,1} represents whether service type i in time slot t is cached to the base station cache strategy. If α i (t) = 1, it means that service i is cached at the base station in time slot t; if α i (t) = 0, which means that service i is not cached at the base station in time slot t; i represents the cache space required by the base station to cache service i, O BS Indicates the maximum cache capacity of the base station; represents the spectrum resource allocation coefficient of the base station to mobile user k, represents the spectrum resource allocation coefficient of the base station to the sensing target q; represents the power resource allocation coefficient of the base station to mobile user k, represents the power resource allocation coefficient of the base station to the sensing target q; represents the computing resource allocation coefficient of the base station to mobile user k, represents the computing resource allocation coefficient of the base station to the sensing target q; D k (t) represents the data volume of the task generated by mobile user k in time slot t, P represents the total power available to the base station, B represents the total spectrum resources available to the base station, represents the noise-normalized communication channel gain, R k (t) represents the communication rate of mobile user k, f BS Indicates the CPU frequency of the base station, X k S represents the number of CPU cycles required to process k-bit tasks of mobile users. k (t) is the amount of radar raw data perceiving target q, X q Indicates the number of CPU cycles required to process a unit bit of radar raw data.

9. The method for allocating time-differentiated synaesthesia resources according to claim 1, characterized in that: In step S4, a triple {s, a, r} is defined, where: s = {s(1), s(2), ..., s(T)} represents the global state space; s(t)∈s represents the total information of the time-varying request tasks and the perceived target requests received by the agent in the coverage area at time slot t, which can be expressed as s(t)={d1,d2,...,d K ,P,B} d k represents the task generated by mobile user k, P represents the total power available to the base station, and B represents the total spectrum resources available to the base station; a={a(1),a(2),…,a(T)} represents the global action space set, a(t)∈a represents the action selected by the agent in time slot t from the action space, expressed as a(t)={α(t),β(t),x(t),y(t),z(t)} r represents the reward function, and the calculation formula for the reward r(t) obtained by the agent performing an action in time slot t is Among them, λ1, λ2, λ3 are weight coefficients, Represents a collection of users, represents the set of sensing targets, K represents the number of mobile users, and Q represents the number of sensing targets; Indicates the maximum tolerable delay of task calculation, represents the computational delay of the task of mobile user k, Indicates the maximum tolerable communication delay of the task, represents the communication delay of the transmission task of mobile user k, represents the perception accuracy of the perceived target q, represents the maximum tolerable delay of perception, represents the computational delay of the radar raw data of the perceived target q; O(t) represents the cache capacity limit, which can be expressed as H(·) represents the Heaviside step function, ε represents the weight coefficient, O i represents the buffer space required by the base station to buffer service type i, O BS Indicates the maximum cache capacity of the base station.

10. The time-differentiated synaesthesia resource allocation method according to claim 1, characterized in that: The loss function L(θ Q ) is expressed as η(t)=r(t)+νQ(s(t+1),h c (t+1),μ(s(t+1))∣θ Q ) in, represents the expected value, and v represents the discount factor; Policy Gradient Expressed as Represents the gradient operation; Update the actor target network parameters and the critic target network parameters by soft updating, denoted as θ μ '←τ a θ μ +(1-τ a )θ μ ' i Q '←t c i Q +(1-t c )i Q ' τ a represents the actor network update coefficient, τ c Represents the critic network update coefficient.

Images