Resource adaptive joint management and control method and system in the integrated scenario of synergy and computing

Abstract

The present invention discloses a resource adaptive joint management and control method and system in a tele-sensing and computing integration scenario, which relates to the field of communications and is used to achieve a balance between the stability and optimality of multiple experience delays and the amount of effective information with low system overhead. In the present invention, in a scenario of communication, perception, and computing integration, each perception node obtains real-time network status parameters and obtains the target state and perception data that each perception node is concerned about; based on the obtained network state parameters, the network central control node constructs a long-term optimization model for network asset management, converts it into a time-discrete resource management model using the Lyapunov method, and divides it into an initial stage, an iterative stage, and an end stage, and then solves the stage to which the time slot query belongs, sends the resource optimization decision to the corresponding perception node to schedule resources, and aggregates and analyzes the target state and perception data. The present invention achieves a balance between the stability and optimality of multiple experience delays and the amount of effective information.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a resource adaptive joint management method in a telecom-computing integrated scenario, and a resource adaptive joint management system in a telecom-computing integrated scenario. Background Art

[0002] With the development of B5G / 6G communication technology, the integrated perception, communication and computing technology has been widely used in artificial intelligence, perception, computing and other service scenarios due to its ability to integrate communication and intelligence, as well as perception and communication. It aims to achieve ubiquitous intelligent connection between people, machines and things, and empower new digital information infrastructure for new applications, which has attracted widespread attention from scholars in the industry.

[0003] The integrated communication, perception, and computing system samples the physical world through perception services, provides connectivity to the physical world through communication services, and optimizes the network and processes perception data through computing services. The integration of communication, perception, and computing enables the deep integration and mutually beneficial enhancement of multi-dimensional perception, collaborative communication, and intelligent computing capabilities, thereby opening a channel for the integration of the physical and digital worlds. It provides diverse capabilities such as positioning, ranging, and imaging, greatly meeting the needs of ultra-high resolution and precision applications. These include location perception, unmanned monitoring, and environmental reconstruction in the field of industrial upgrading; environmental monitoring and hazardous material detection in the field of social governance; and gesture and motion recognition, security monitoring, behavior monitoring, and health monitoring in the field of smart living.

[0004] Existing research on resource optimization for the integrated communication, perception, and computing scenarios involving multi-dimensional and even multi-system resource scheduling focuses solely on optimizing communication, computing, or both. This separates network data perception, communication, and computing, making it difficult to meet the demands of real-time interaction and analysis of massive amounts of perception information. Therefore, there is an urgent need to explore an integrated communication, perception, and computing resource scheduling architecture for efficient delivery and analysis of perception data.

[0005] Considering the differentiated task response requirements and the impact of wireless channel fluctuations caused by fast fading, the constructed resource joint scheduling architecture needs to ensure the stability of perception data delivery and processing on a large time scale, avoiding the problems of queue data backlog and latency jitter degradation caused by traditional instantaneous optimization theoretical frameworks. It must achieve a stable and optimal balance between multiple experienced latency and effective information volume, effectively solving the problem of excessive resource allocation latency in existing integrated communication, perception, and computing scenarios. However, no effective solution to this problem has been found. Summary of the Invention

[0006] The purpose of the present invention is to provide a resource adaptive joint management method and system in an integrated telecomputing scenario to address the above-mentioned problems, so as to achieve a balance between multi-event delay, stability and optimality of effective information volume with low system overhead, thereby taking into account both device response speed and decision quality.

[0007] The technical solution adopted in the present invention is as follows:

[0008] A resource adaptive joint management method in a synergistic computing integrated scenario, comprising:

[0009] S1. In the integrated communication, perception, and computing scenario, each perception node acquires real-time network status parameters, including channel status parameters and network asset management parameters; acquires the target status and perception data of interest to each perception node; and based on the acquired network status parameters, the network central control node constructs a long-term network asset management optimization model.

[0010] S2. The network central control node uses the Lyapunov method to convert the long-term network asset management optimization model into a time-discrete resource optimization model, and divides the resource optimization model into an initial stage, an iterative stage, and an end stage in chronological order. The network central control node queries the stage according to the time slot, solves the resource optimization model corresponding to the stage, obtains the resource optimization strategy for the stage, and sends the resource optimization decision to the corresponding sensing node.

[0011] S3. Each sensing node performs resource scheduling according to resource optimization decisions, and aggregates and analyzes target status and sensing data.

[0012] Furthermore, each sensing node adopts a pilot-assisted channel estimation method to decompose the transmitted signal into two parts: pilot and data, and completes channel estimation based on the reception of pilot information to obtain channel state parameters.

[0013] Furthermore, the network asset management parameters include:

[0014] The upper limit of the time and frequency resources that each access point n can call for during short packet communication is l Smax , and the upper limit of the time and frequency resources that can be called during long packet communication Lmax ; The total computing resources available to the server associated with each access point n The maximum experienced delay limit of the perception node in the initial stage, iteration stage, and end stage As well as the delay thresholds for reinforcement learning queues, model training queues, and sensor data sinking queues. The accuracy loss threshold e of the neural network model of the perception node and server at the end of the iteration phase.

[0015] Furthermore, the long-term optimization model for network asset management constructed by the network central control node is:

[0016]

[0017] st for have:

[0018] C1: Limited network time and frequency resources

[0019] C2: Server CPU resources are limited

[0020] C3: Average experienced latency limit of each sensing node at each stage:

[0021]

[0022] C4: Average delay limit of each queue during the iteration phase:

[0023]

[0024]

[0025] C5: Limit on model accuracy loss at the end of the iteration phase:

[0026] Where n∈N represents any access point n in the access point set N, u∈U represents any sensing node u in the sensing node set U, the number of sensing nodes in the sensing node set U is represented by |U|, and the set of sensing nodes within the coverage of access point n is U n , there is U n∈N U n =U; t represents the time step, t = 0 represents the initial stage, t = 1, 2, 3, ..., T represents the iteration stage, t = T + 1 represents the end stage, and T represents the number of iteration cycles; I init , I ite , I end They represent the effective information that the sensing node u can obtain in the initial stage, iterative stage and final stage respectively; E represents the mathematical expectation; a u,n ∈{1,0} is the access status of the sensing node u, 1 means access is allowed, 0 means access is not allowed; are the time-frequency resources occupied by the sensing node u during the short packet communication and long packet communication processes respectively; represents the total time-frequency resources of each access point n; h u,n (t) represents the amount of CPU computing resources that can be called by the server associated with the access point n at time t by the sensing node u; T init 、 T endThey represent the experienced delays in the initial stage, iterative stage, and final stage respectively. The server integrated by access point n configures three queues for each sensing node u within its service range: Sink queue for sensor data, For reinforcement learning queue, Model training queue, and the data leaving the three queues at time t are Each sensing node u is equipped with two queues: reinforcement learning queue and model training queue The data leaving the two queues at time t are x u is the splitting of sensor data by sensing node u, D u (t) is the size of the associated data called by the sensing node u from the sensing network, x u D u and (1-x u )D u The amount of sensor data analyzed locally and analyzed with the help of server; v u represents the model parameters of the main evaluation network in the long-term optimization model of network asset management, L(v u ) represents the loss function of the main evaluation network, TD T represents the temporal difference error of the tuple sampled by the main evaluation network at the end of iteration T; d u (t) represents the data size perceived by the sensing node u at time t.

[0027] Furthermore, the network central control node converts the network asset management long-term optimization model into a discrete-time resource optimization model as follows:

[0028]

[0029] st for have:

[0030] C1: Limited network time and frequency resources

[0031] C2: Server CPU resources are limited

[0032] C3: Average experienced latency limit of each sensing node at each stage:

[0033]

[0034] C4: Limit on model accuracy loss at the end of the iteration phase:

[0035] in: Vite (t) is the auxiliary parameter for queue update,

[0036] The queue Virtual queue; qv(t) represents each item defined in QV(t), and They are the sensor data sink queues of the server Reinforcement Learning Queue Model training queue and reinforcement learning queues of perception nodes and model training queue The input and output of each parameter involved; ξ is a constant and ξ>0; Represents the average amount of information obtained by the sensing node.

[0037] Furthermore, sensor data sinks into the queue Reinforcement Learning Queue Model training queue Reinforcement Learning Queue and model training queue The update method is:

[0038]

[0039]

[0040] queue The update method is:

[0041]

[0042] Furthermore, the network central control node divides the resource optimization model into an initial phase, an iterative phase, and an end phase in chronological order, including:

[0043] The resource optimization model for the initial phase of network central control node division is:

[0044]

[0045] st for have:

[0046] C1: Limited network time and frequency resources

[0047] C2: Server CPU resources are limited

[0048] C3: Average experienced latency limit in the initial stage:

[0049] The resource optimization model for the iterative phase of network central control node division is:

[0050]

[0051] st for have:

[0052] C1: Limited network time and frequency resources

[0053] C2: Server CPU resources are limited

[0054] C3: Average experienced delay limit for each time slot during the iteration phase:

[0055] The resource optimization model at the end stage of the network central control node division is:

[0056]

[0057] st for have:

[0058] C1: Limited network time and frequency resources

[0059] C2: Server CPU resources are limited

[0060] C3: Average experienced latency limit at the end stage

[0061] C4: Limit on model accuracy loss at the end of the iteration phase

[0062] Furthermore, after the sensing node performs resource scheduling according to the resource optimization decision, the result of aggregating and analyzing the target state and sensing data is:

[0063]

[0064] in, and is the local processing result of the perception node u and the processing result of the associated server.

[0065] The present invention also provides a resource adaptive joint management and control system in a telecomputing integrated scenario, which executes the above method.

[0066] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0067] The present invention utilizes a multi-dimensional resource adaptive joint management method to achieve dual-end performance optimization of devices and nodes in an intelligent network, while solving the resource adaptive scheduling efficiency problem to balance the resource decision response speed and quality of the perception node. The specific innovations include the following two aspects: 1) Multi-dimensional resource adaptive joint management problem, that is, in the scenario of integrated communication and perception, with the goal of optimizing the long-term average effective information volume of the perception node, under the main constraints such as the effective delay of each working stage of the perception node, a multi-dimensional resource adaptive joint management and model training method is implemented, a balanced design of delay stability-information volume optimization, while achieving lower delay and solving the data backlog problem of massive data access. 2) Resource adaptive scheduling efficiency problem, because when solving problem 1), the constructed optimization scheme models the resource management in the iterative stage as a long-term average optimization problem, the computer solution complexity is very high, and it will consume too much computing power resources. Therefore, the present invention introduces the Lyapunov optimization method to convert the designed network asset management long-term optimization model into a time-discrete resource optimization model, thereby reducing the complexity of the model solution and alleviating the system's resource overhead. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0069] Figure 1 This is a flow chart of a resource adaptive joint management method in a synergistic computing integrated scenario according to an embodiment of the present invention;

[0070] Figure 2 This is a diagram of a data queue model architecture according to an embodiment of the present invention;

[0071] Figure 3 This is a graph showing the length of each resource training queue for adaptive joint management of resources in a synergistic computing integrated scenario according to an embodiment of the present invention;

[0072] Figure 4 This is a diagram illustrating adaptive joint resource management and control of the delay of each resource training queue in a synergistic computing integration scenario according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] All features disclosed in this specification, or all steps in the disclosed methods or processes, except mutually exclusive features and / or steps, can be combined in any manner.

[0074] Any feature disclosed in this specification (including any appended claims and abstract), unless otherwise stated, may be replaced by other equivalent or similar features. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0075] like Figures 1 to 2 As shown, the first embodiment of the present invention proposes a resource adaptive joint management method in a telecomputing integrated scenario, and the method specifically includes the following steps:

[0076] S1. In the scenario of integrated communication, perception and computing, each perception node obtains real-time network status parameters, which include channel status parameters and network asset management parameters; obtains the target status and perception data that each perception node is concerned about; based on the obtained network status parameters, the network central control node builds a long-term network asset management optimization model.

[0077] S11. In the scenario of integrated communication, perception and computing, each perception node obtains real-time channel state parameters, network asset management parameters and other network state parameters, and at the same time, obtains the target state and perception data that each perception node is concerned about.

[0078] The set of sensing nodes and the set of access points are represented by U and N respectively, n∈N represents any access point n in the access point set N, u∈U represents any sensing node u in the sensing node set U, the number of sensing nodes in the sensing node set U is represented by |U|, and the set of sensing nodes within the coverage of access point n is U n , there is U n∈N U n =U.

[0079] 1) Channel state parameters: Each sensing node uses a pilot-assisted channel estimation method to split the transmitted signal into two parts: pilot and data, and completes channel estimation based on the reception of the pilot information. represents the channel gain between sensing node u and access point n on channel m, where is the small-scale fading channel coefficient, Indicates road loss. Represents shadow fading, and its variance is The lognormal distribution of .

[0080] 2) Network resource management parameters: the upper limit of the time and frequency resources that each access point n can call during short packet communication l Smax , and the upper limit of the time and frequency resources that can be called during long packet communication Lmax ; The total computing resources available to the server associated with each access point n The maximum experienced delay limit of the perception node in the initial stage, iteration stage, and end stage As well as the delay thresholds for reinforcement learning queues, model training queues, and sensor data sinking queues. The accuracy loss threshold e of the neural network model of the perception node and server at the end of the iteration phase.

[0081] S12. Based on the acquired network status data, the network central control node establishes long-term optimization goals for network asset management.

[0082] The constructed optimization objective is shown below. For the convenience of expression, let t represent the time step, t = 0, t = 1, 2, 3, ..., T, t = T + 1 represent the initial stage, the iteration stage and the end stage respectively, and T represents the number of iteration cycles.

[0083]

[0084] The network central control node evaluates the amount of effective information that can be obtained at each stage based on the established long-term optimization goals, as follows:

[0085] (1) Initial stage and final stage:

[0086] The sensing node is in a delay-sensitive communication mode in the initial and final stages. Therefore, the short packet communication business mode is used in these two stages, where the data size is shorter, the occupied time and frequency resources are less, and the uplink / downlink transmission spectrum efficiency is higher. and local / server computing efficiency For short packet services, the Shannon formula can no longer approximate its communication rate, and it is necessary to calculate the coding rate under a finite code length.

[0087]

[0088] Among them, the signal-to-noise ratio is the short packet communication power of the sensing node And the short packet communication power of the access point function; and Indicates the time-frequency resource blocks occupied by the short packet service in the uplink / downlink communication process; ò is the packet loss rate and also, and is the amount of short packets waiting to be calculated in the local queue and server queue. and is the corresponding computational delay.

[0089] At the end of the initial stage, the effective information that the sensing node u can obtain is:

[0090]

[0091] Among them, ε1 is the packet loss rate in the initial stage, a u,n To sense the node access situation, is the amount of time-frequency resources occupied by the sensing node u during the communication process, in units of s·Hz, is the spectrum efficiency of uplink short packet data transmission between sensing node u and access point n, in bit / s / Hz.

[0092] The effective amount of information that the sensing node u can obtain in the final stage is:

[0093]

[0094] Among them, ε3 is the packet loss rate at the end stage, is the spectrum efficiency of downlink short packet data transmission between sensing node u and access point n, in bit / s / Hz.

[0095] (2) Iteration phase:

[0096] The sensing node uses long packet transmission services for transmission in the iteration phase. Long packet services can also be understood as delay-tolerant services. Their data packets are long, occupying large time-frequency resource blocks and computing resources. Their uplink / downlink transmission spectrum efficiency is low. and local / server computing efficiency The spectral efficiency of transmission is approximately obtained by the Shannon formula.

[0097]

[0098] In the above formula, and is the uplink / downlink communication signal-to-noise ratio of the sensing node, which is the long packet communication power of the sensing node Long packet communication power with access point Function: and are the amount of long packet data to be calculated in the local queue and the server queue respectively, and is the corresponding calculation delay.

[0099] The effective amount of information that the perception node u can obtain in each iteration is:

[0100]

[0101] Among them, ε2 is the packet loss rate in the iteration phase, is the amount of time-frequency resources occupied by the sensing node u during the communication process, in units of s·Hz, is the spectrum efficiency of uplink / downlink long packet data transmission between sensing node u and access point n, in bit / s / Hz.

[0102] S13. Based on the long-term optimization goal of network resource management, the network central control node constructs a multi-dimensional resource adaptive network asset management long-term optimization model training problem for the integrated telecomputing scenario.

[0103] The long-term optimization model of network asset management includes the main network and the target network. Both the main network and the target network are "evaluation network-behavior network" models; the main network includes the main evaluation network and main behavior network The target network includes the target evaluation network and target behavior network For the evaluation network, s u (t) and a u (t) is the input, s u (t) represents the environmental parameter, v u and v u ′ are the model parameters of the two evaluation networks, and the evaluation Q value is output; for the behavior network, s u (t) is the input, η u and η u ′ are the model parameters of the two behavioral networks, a u (t) is the output, which represents the generated decision behavior.

[0104] The set of optimization parameters related to the settings is defined as:

[0105]

[0106] Among them, x u (t)∈[0,1] is the splitting of the sensor data by the sensing node, x u D u and (1-x u )D u The amount of sensor data analyzed locally and analyzed with the help of server; a u,n ∈{0,1} is the network access status of the node. Its value is 1, which means that the target is allowed to access the network when the uplink / downlink communication request is made, and the value is 0, which is the opposite. The amount of computing resources that can be called by each sensing node is h u , the amount of CPU computing resources that the sensing node u can call on the server associated with access point n at time t is h u,n ; is the amount of time-frequency resources occupied by the sensing node u during the initial and final phases (short packet) of communication, in s·Hz; is the amount of time-frequency resources occupied by the sensing node u during the iterative phase (long packet) communication process, in s·Hz;

[0107] The constructed optimization problem is:

[0108]

[0109] stC1: Network time and frequency resources are limited

[0110] C2: Server CPU resources are limited

[0111] C3: Average experienced latency limit of each sensing node at each stage

[0112]

[0113] C4: Average delay limit of each queue during the iteration phase

[0114]

[0115]

[0116] C5: Limit on model accuracy loss at the end of the iteration phase

[0117] Among them, the effective information amount that can be obtained in the initial stage, iterative stage and final stage is I init , and I end The server integrated in the access point configures three queues for each sensing node within its service range: Sink queue for sensor data, For reinforcement learning queue, Model training queue, and the data leaving the three queues at time t are Each perception node is equipped with two queues: reinforcement learning queue and model training queue The data leaving the two queues at time t are d u (t) Local sensing data size of sensing node u; D u (t) is the size of the associated data called by the perception node u from the perception network. Each perception node u updates the main evaluation network to minimize the loss function, and the model parameter is ν u , and its loss function is defined as TD t Represents the time series difference error of the sampled tuple at time t.

[0118] S2. For the constructed long-term optimization model of network asset management, the network central control node adopts the Lyapunov method to split the long-term optimization model into multiple time-discrete resource optimization models, and divides each resource optimization model into the initial stage, iterative stage, and end stage in chronological order; the network central control node queries the stage according to the time slot, solves the sub-optimization problem corresponding to the stage, and sends the resource optimization decision to the corresponding perception node.

[0119] S21. In view of the constructed long-term optimization goal, the network central control node adopts the Lyapunov method to convert the long-term optimization model of network asset management into a time-discrete resource optimization model, and divides each resource optimization model into the initial stage, iterative stage, and end stage in chronological order.

[0120] The specific steps are as follows:

[0121] Step 1: The server adopts a parallel queue architecture and equips each sensing node within its service range with three queues: sensor data sinking queue Reinforcement Learning Queue Model training queue The data leaving the three queues at time t are The update rules for each queue are as follows:

[0122]

[0123] Step 2: Each sensing node is equipped with two queues: reinforcement learning queue and model training queue The data leaving the two queues at time t are The update rules for each queue are as follows:

[0124]

[0125] Step 3: Define virtual queues for the data queues involved in steps 1 and 2 The rules are updated as follows:

[0126]

[0127] Step 4: Based on each queue in steps 1 to 3, calculate the Lyapunov function:

[0128]

[0129] in, V ite (t) is the auxiliary parameter for queue update.

[0130] Step 5: Define the Lyapunov drift term based on the Lyapunov function to describe the stability of the above delay indicators. The Lyapunov drift term can be expressed as:

[0131] Δ(QV(t))=E[F L (QV(t+1))-F L (QV(t))|QV(t)].

[0132] Step 6: Based on the Lyapunov drift term and the optimization objective of the optimization problem OP1, calculate the Lyapunov drift penalty term:

[0133]

[0134] The drift penalty term satisfies for any constant ξ>0 and any QV(t)

[0135]

[0136] Where C>0 is a constant, qv(t) represents each term defined in QV(t), and They are sensor data sink queues Reinforcement Learning Queue Model training queue Reinforcement Learning Queue and model training queue The input and output quantities of each parameter involved.

[0137] Step 7: Based on the upper bound in step 6, the time-discrete optimization problem OP2 is given:

[0138]

[0139] stC1: Network time and frequency resources are limited

[0140] C2: Server CPU resources are limited

[0141] C3: Average experienced latency limit of each sensing node at each stage

[0142]

[0143] C4: Limit on model accuracy loss at the end of the iteration phase

[0144] S22. The network central control node queries the phase according to the time slot, solves the resource optimization model corresponding to the phase to obtain the resource optimization strategy for the phase, and sends the resource optimization decision to the corresponding perception node.

[0145] Step 1: The network central control node schedules the optimization problem OP2.1 in the initial stage and solves the resource optimization problem to optimize the amount of effective information in the initial stage. The resource optimization decision in the initial stage is sent to the network according to the time slot.

[0146] The optimization at this stage can be broken down into:

[0147]

[0148] stC1: Network time and frequency resources are limited

[0149] C2: Server CPU resources are limited

[0150] C3: Average experienced latency limit in the initial stage:

[0151] Step 2: The network's central control node schedules the optimization problem OP2.2 for each time slot during the iteration phase, solves the resource optimization problem, optimizes the amount of effective information during the iteration phase, and updates the queue in real time. Resource optimization decisions during the iteration phase are distributed to the network according to the time slot.

[0152] The optimization problem at this stage can be decomposed into:

[0153]

[0154] stC1: Network time and frequency resources are limited

[0155] C2: Server CPU resources are limited

[0156] C3: Average experienced delay limit for each time slot during the iteration phase:

[0157] Step 3: The network central control node schedules the optimization problem OP2.3 in the end phase and solves the resource optimization problem to achieve the optimal resolution of the effective information volume in the end phase. The resource optimization decision for the end phase is then sent to the network according to the time slot.

[0158] The optimization problem at this stage can be decomposed into:

[0159]

[0160] stC1: Network time and frequency resources are limited

[0161] C2: Server CPU resources are limited

[0162] C3: Average experienced latency limit at the end stage

[0163] C4: Limit on model accuracy loss at the end of the iteration phase

[0164] S3. Each sensing node performs resource scheduling (communication, computing, etc.) according to resource optimization decisions, and aggregates and analyzes target status and sensing data.

[0165] S31: Each sensing node schedules communication and computing resources according to the resource optimization decision and aggregates and analyzes the target state and sensing data. Through adaptive scheduling of resources in each time slot, the sensing node u outputs the sensing data analysis results at the end of the phase.

[0166]

[0167] in, and It is the local processing result of the perception node u and the processing result of the associated server.

[0168] Figure 2 A data queue model of the server queue and the perception node queue in the integrated scenario of synergy and computing is given. The parameters of the model correspond to the working mode of each data queue in the optimization model.

[0169] Figure 3 A data queue length curve is presented, showing the training message queue length of the resource allocation algorithm in the current communication, perception, and computing integration scenario. The graph shows that the training data quality of terminals in the communication system is effectively improved, and this improvement is more significant when the number of training data terminal time slots increases.

[0170] Figure 4 This diagram shows the data queue latency, representing the training message queue latency of the resource allocation algorithm in the current communication, perception, and computing integration scenario. The diagram shows that the latency of the perception node queue, data offload queue, and server queue is optimal, meeting the requirements of the communication, perception, and computing integration scenario.

[0171] The above can be applied to a resource adaptive joint management and control system or a wireless communication system in an integrated tele-sensing and computing scenario. The system executes the method in the above embodiment to achieve dynamic resource collaboration for device data training in the system.

[0172] The present invention is not limited to the aforementioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A resource adaptive joint management method in a synergistic computing integrated scenario, characterized in that: include: S1. In the scenario of integrated communication, perception and computing, each perception node obtains real-time network status parameters, which include channel status parameters and network resource management parameters. The network resource management parameters include: The upper limit of the time and frequency resources that can be used during short packet communication , and the upper limit of the time and frequency resources that can be called during long packet communication ; Each access point The total computing resources available to the associated server ; The maximum experienced delay limit of the perception node in the initial stage, iteration stage, and end stage , as well as the delay thresholds of the reinforcement learning queue, model training queue, and sensor data sinking queue; obtain the target state and perception data of each sensing node; based on the obtained network state parameters, the network central control node constructs a long-term optimization model for network resource management: for ,have: C1: Limited network time and frequency resources ; C2: Server CPU resources are limited ; C3: Average experienced latency limit of each sensing node at each stage: ； C4: Average delay limit of each queue during the iteration phase: ， ， ； C5: Limit on model accuracy loss at the end of the iteration phase: ; in, Indicates the average amount of information obtained by the sensing node; Represents a set of access points Any access point in , Represents a set of perception nodes Any sensor node in , perception node set The number of sensing nodes in is expressed as , access point The set of sensing nodes within the coverage area is , ; represents the time step, Indicates the initial stage, , represents the iteration phase, Indicates the end stage, Indicates the number of iteration cycles; Represents the perception nodes in the initial stage, iterative stage and final stage respectively The amount of effective information available; It means finding the mathematical expectation; For the perception node Access status, 1 means access is allowed, 0 means access is not allowed; Perception nodes exist The amount of time and frequency resources occupied during short packet communication and long packet communication; Indicates each access point The total amount of time-frequency resources; Represents a perception node Access Point The associated server is The amount of CPU computing resources available at any time; Respectively represent the experience delay in the initial stage, iterative stage and final stage; access point The integrated server serves the sensing nodes within its service range Each is equipped with three queues: Sink queue for sensor data, For reinforcement learning queue, Model training queue, and The data leaving the three queues at the moment are ; Each sensing node Equipped with two queues: reinforcement learning queue and model training queue , The data leaving the two queues at time ; For the perception node exist The splitting of sensor data at all times, For the perception node exist The size of the associated data called from the perception network at any moment, and Respectively in The amount of sensor data analyzed locally and with server assistance at all times; represents the model parameters of the main evaluation network in the long-term optimization model of network resource management, represents the loss function of the main evaluation network, Indicates the end time of the iteration Temporal difference error of the tuples sampled by the main evaluation network; express Perception nodes at all times Perceived data size; S2. The network central control node uses the Lyapunov method to convert the long-term network resource management optimization model into a time-discrete resource optimization model, and divides the resource optimization model into an initial stage, an iterative stage, and an end stage in chronological order. The network central control node queries the stage according to the time slot, solves the resource optimization model corresponding to the stage, obtains the resource optimization strategy for the stage, and sends the resource optimization decision to the corresponding sensing node. S3. Each sensing node performs resource scheduling according to resource optimization decisions, and aggregates and analyzes target status and sensing data.

2. The resource adaptive joint management method in the integrated telepathy and computing scenario according to claim 1 is characterized in that: Each sensing node uses a pilot-assisted channel estimation method to split the transmitted signal into two parts: pilot and data, and completes channel estimation based on the reception of pilot information to obtain channel state parameters.

3. The resource adaptive joint management method in the integrated telepathy and computing scenario according to claim 1 is characterized in that: The network central control node converts the long-term optimization model of network resource management into a discrete-time resource optimization model: for ,have: C1: Limited network time and frequency resources ; C2: Server CPU resources are limited ; C3: Average experienced latency limit of each sensing node at each stage: ； C4: Limit on model accuracy loss at the end of the iteration phase: ; in: , Auxiliary parameters for queue updates, The queue 、 、 、 、 Virtual queue; express For each item defined in and Separate queues 、 、 、 、 The inputs and outputs involved; is a constant and .

4. The resource adaptive joint management method in the integrated telepathy and computing scenario according to claim 3 is characterized in that: Sensor data sink queue , reinforcement learning queue , model training queue , reinforcement learning queue and model training queue The update method is: queue The update method is: 。 5. The resource adaptive joint management method in the integrated telepathy and computing scenario according to claim 4 is characterized in that: The network central control node divides the resource optimization model into the initial stage, iterative stage, and final stage in chronological order, including: The resource optimization model for the initial phase of network central control node division is: for ,have: C1: Limited network time and frequency resources ; C2: Server CPU resources are limited ; C3: Average experienced latency limit in the initial stage: ; The resource optimization model for the iterative phase of network central control node division is: for ,have: C1: Limited network time and frequency resources ; C2: Server CPU resources are limited ; C3: Average experienced delay limit for each time slot during the iteration phase: ; The resource optimization model at the end stage of the network central control node division is: for ,have: C1: Limited network time and frequency resources ; C2: Server CPU resources are limited ; C3: Average experienced latency limit at the end stage ; C4: Limit on model accuracy loss at the end of the iteration phase .

6. The resource adaptive joint management method in the integrated telepathy and computing scenario according to claim 5 is characterized in that: After the sensing node schedules resources according to the resource optimization decision, it aggregates and analyzes the target state and sensing data. for: in, and For the perception node The local processing results and the processing results of the associated server.

Images