Cellular-free site intelligent optimization method for communication and inductance fusion

Abstract

The invention relates to an intelligent optimization method for a non-cellular station address for sensing fusion, and the method comprises the steps: S1, building a mathematical relation model between communication performance and an access point station address based on a logarithmic distance path loss model; s2, on the basis of a bistatic system path loss model, constructing a mathematical relationship model between the sensing performance and the access point site; and S3, modeling an access point site optimization problem in the cellular-free sensing fusion system as a Markov decision process, constructing a state space and an action space, and calculating a state of the access point site according to a mathematical relationship model between communication performance and the access point site and a mathematical relationship model between sensing performance and the access point site. Fusing to obtain a reward function of a Markov decision process; and S4, solving the Markov decision process by adopting a flexible action-evaluation deep reinforcement learning algorithm to obtain an optimal access point deployment strategy. Compared with the prior art, the method has the advantages of improving the site selection accuracy, widening the application range, improving the site selection effect stability and the like.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and in particular to a synaesthesia fusion-oriented cellular-free site intelligent optimization method. Background Art

[0002] The cell-free massive MIMO architecture transcends the geographical limitations of traditional cellular networks. By deeply integrating distributed antenna systems with massive MIMO technology, it creates a completely new communications system. Access points are no longer confined to fixed base station locations; instead, they are distributed throughout the service area. User devices, regardless of their location, can receive coordinated services from multiple access points. The central processing unit, at the heart of the system, coordinates channel estimation, power allocation, and data transmission among access points in real time over the backhaul link, establishing a dynamically adaptive communications network.

[0003] Future mobile communications will build a new network where humans, machines, and objects are intelligently interconnected, and intelligent entities efficiently communicate with each other. The deep integration of communication and perception has become a key research direction. Synaesthesia integration pursues tighter functional integration, moving away from treating communication and perception as separate goals and instead integrating them into a unified design for mutual benefit. The ultimate goal is to fuse sensing and communication functions, achieving dynamic tradeoffs and mutual enhancement between the two, thereby maximizing overall system performance.

[0004] The deployment location of access points directly impacts the communication coverage, channel quality, target perception and positioning accuracy, and resource utilization efficiency of a cellular-free telemetry fusion system. AP site optimization requires careful consideration of the spatial distribution of communication users and perceived targets. Regarding communication, AP deployment location must consider communication signal path loss, system spectral efficiency, and throughput. Regarding perception, AP location determines the geometric structure between the AP and the target, thus affecting perceived positioning accuracy and multi-target resolution.

[0005] In this regard, some existing technologies provide site selection methods for synaesthesia fusion. For example, Chinese patent CN118555576A discloses a non-cellular synaesthesia fusion site planning method and system based on a flexible strategy evaluation network. Considering the problem of maximizing the communication perception fusion performance index in the non-cellular system, the Euclidean distance criterion is used to construct the expression of the relationship between communication performance and site topology under zero-forcing detection, and the relationship between the determinant of the Fisher information matrix and the site topology in the joint ranging parameter estimation problem is studied to improve the perception performance, and the communication perception performance is combined into a multi-objective optimization problem. Then, the flexible strategy evaluation reinforcement learning with the introduction of the maximum entropy term is used to solve this complex non-convex problem. This method has good optimization performance. The site topology obtained by the method of the present invention can better improve the communication performance and perception performance of the non-cellular system. However, the above-mentioned existing technologies have the following defects:

[0006] 1. When deriving the Fisher Information Matrix (FIM), the distance between each transmitting and receiving node and the target is not explicitly considered. Only the directional angle difference term is reflected in the expression, and the distance-related accuracy degradation mechanism is not reflected. As a result, the FIM no longer has the physical meaning of distinguishing "near is better, far is worse", resulting in distorted perception performance estimation.

[0007] 2. The product optimization of communication and perception performance was performed without normalizing their different scales and convergence rates. This posed a risk of training instability, ultimately leading to large fluctuations in site selection results and failing to meet stability requirements.

[0008] 3. A single target is set to move on a fixed trajectory, which lacks consideration of the uncertainty of the target's spatial distribution. It is difficult to apply to random or multi-target environments. The complexity of the target environment in the synaesthesia scenario is not fully incorporated into the modeling. It is only applicable to orbital motion application scenarios and has limited generalization capabilities. Summary of the Invention

[0009] The purpose of the present invention is to provide a synaesthesia fusion-oriented non-cellular site intelligent optimization method in order to solve the defects of the prior art.

[0010] The purpose of the present invention can be achieved by the following technical solutions:

[0011] A method for intelligent optimization of non-cellular sites for synaesthesia fusion, comprising:

[0012] Step S1: Based on the logarithmic distance path loss model, a mathematical relationship model between communication performance and access point location is constructed;

[0013] Step S2: Based on the bistatic system path loss model, a mathematical relationship model between sensing performance and access point locations is constructed;

[0014] Step S3: Modeling the access point site optimization problem in the cellular-free synaesthesia fusion system as a Markov decision process, constructing a state space and an action space, and fusing the mathematical relationship model between communication performance and access point site and the mathematical relationship model between perception performance and access point site to obtain a reward function of the Markov decision process;

[0015] Step S4: Use the flexible action-evaluation deep reinforcement learning algorithm to solve the Markov decision process and obtain the optimal access point deployment strategy.

[0016] The mathematical relationship model between the communication performance and the access point site is:

[0017]

[0018] Where: R sumis the total communication system rate, R i is the communication rate of the i-th user, K u is the number of communication users, ρ0 is the uplink transmission power of all communication users, SNR i is the signal-to-noise ratio of the i-th user, U is the number of transmitting APs, V is the number of receiving APs, d mi is the distance between the i-th user and the m-th AP, β(d mi ) is the large-scale fading coefficient between the i-th user and the m-th AP.

[0019] The mathematical relationship model between the sensing performance and the access point location is:

[0020]

[0021] Where: |F k | is the Fisher information matrix of the k-th perception target, is the distance between the uth transmitting access point and the sensing target, is the distance between the vth receiving access point and the sensing target, is the polar angle between the u-th transmitting access point and the sensing target, is the polar angle between the vth receiving access point and the sensing target.

[0022] The reward function of the Markov decision process is:

[0023]

[0024] Among them: K s is the number of perceived targets, is an adjustable parameter related to communication performance, is the x-coordinate of the target AP site, is the y coordinate of the target AP site.

[0025] The state space is

[0026]

[0027] Where: s t is the state space, q1 is the position of the i-th user, For K u The position of the user, p1 is the position of the first perceived target, For K s The position of the sensing target, t1 is the position of the first transmitting access point, t U is the location of the Uth transmitting access point, r1 is the location of the first receiving access point, r V is the position of the Vth receiving access point;

[0028] The action space is:

[0029] a t ={Δt1,…,Δt U ,Δr1,…Δr V}

[0030] Among them: a t is the action space at time t, Δt1 is the movement of the first transmitting access point, Δt U is the movement of the Uth transmitting access point, Δr1 is the movement of the first receiving access point, and Δr V is the movement amount of the position of the Vth receiving access point.

[0031] Each position information in the state space is normalized.

[0032] The range of the single-step movement amount at each position in the action space is:

[0033] [-d max ,d max ]

[0034] Where: d max The maximum moving distance of a single step.

[0035] In step S4, a flexible action-evaluation algorithm is used to maximize the historical state-action chain ρ of the strategy under strategy π while maximizing the cumulative expected reward function. π The expected entropy under the condition of , in order to perform actions as randomly as possible, is conducive to finding the optimal solution. The objective function of the flexible strategy evaluation is:

[0036]

[0037] in: is the objective function of flexible strategy evaluation, T is the total number of steps in a reinforcement learning training round, r(s t ,a t ) is the reward for the current action, ω is the entropy adjustment coefficient, is the policy entropy,

[0038] The loss function of the Actor network is expressed as:

[0039]

[0040] in: is the experience replay buffer, Q φ (s t ,a t ) represents the reward value of the current state-action predicted by the Critic network, and the entropy regularization term -ωlogπθ (a t |s t ) increases the randomness of the strategy and prevents the strategy from converging to the local optimum too early. ω is the temperature coefficient, which is achieved by minimizing the following loss function:

[0041]

[0042] in: is the preset entropy lower limit, and ω is updated through back propagation so that the policy entropy gradually approaches the target value during training;

[0043] The loss function of the Critic network is defined as:

[0044]

[0045] in: is the target value function, For the parameters of the target Critic network, through the soft update strategy to update.

[0046] A cellular-free site intelligent optimization device for synaesthesia fusion includes a memory, a processor, and a program stored in the memory. When the processor executes the program, the method described above is implemented.

[0047] A storage medium stores a program, which implements the above method when executed.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. By optimizing the FIM determinant, the combined influence of the distance and angle between each transmitting and receiving AP and the target is considered. This directly reflects the impact of the geometric conditions of the perception link on the lower bound of the error, avoids overestimation in long-distance or angle-degraded scenarios, and improves the perception accuracy of the deployment strategy.

[0050] 2. Target distribution modeling uses a Gaussian distribution: The two sensing targets follow an independent and identically distributed Gaussian model. A Fisher information matrix for multiple target location parameters is constructed based on a dual-base system loss model. Target locations are sampled multiple times during simulation, ensuring that the optimized AP deployment plan is statistically optimal. This ensures stable and high perception accuracy even when the target's specific location changes during actual deployment. This site selection method does not rely on a fixed target trajectory or known path and is applicable to unknown targets, randomly moving targets, or multiple targets. This improves the generalization and robustness of the synaesthesia fusion system in complex dynamic environments.

[0051] 3. The communication rate and multiple objective FIM determinants are normalized and multiplied to improve the scale consistency of the objective functions. This can avoid the dominance of one of the communication and perception metrics during the optimization process due to differences in scale and numerical level. This makes the optimization process more balanced and effectively balances the conflict between communication link reliability and perception accuracy. The resulting site selection results can achieve the optimal synergy between perception and communication performance with a limited number of AP resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the overall framework of a synaesthesia-integrated, non-cellular site intelligent optimization method provided by an embodiment of the present invention;

[0053] Figure 2 A diagram of the Actor neural network structure;

[0054] Figure 3 Schematic diagram of the Critic neural network structure;

[0055] Figure 4 This is a schematic diagram of the flexible strategy evaluation network structure;

[0056] Figure 5 A cumulative reward curve diagram in one embodiment;

[0057] Figure 6 is a graph showing a communication sum rate-transmit power curve in one embodiment;

[0058] Figure 7 is a cumulative distribution diagram of positioning error in one embodiment;

[0059] Figure 8 A schematic diagram of an access point deployment solution in one embodiment;

[0060] Figure 9 A graph showing how communication and rate vary with the number of access nodes in one embodiment;

[0061] Figure 10 A diagram showing the cumulative distribution of positioning errors as a function of access nodes in one embodiment;

[0062] Figure 11 FIG. 1 is a schematic diagram of an access point deployment solution with different numbers of access points in one embodiment. DETAILED DESCRIPTION

[0063] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0064] A cellular-free site intelligent optimization method for synaesthesia fusion, such as Figure 1 As shown, including:

[0065] Step S1: Based on the logarithmic distance path loss model, a mathematical relationship model between communication performance and access point location is constructed;

[0066] The mathematical relationship model between communication performance and access point location is:

[0067]

[0068] Where: R sum is the total communication system rate, R i is the communication rate of the i-th user, K u is the number of communication users, ρ0 is the uplink transmission power of all communication users, SNR i is the signal-to-noise ratio of the i-th user, U is the number of transmitting APs, V is the number of receiving APs, d mi is the distance between the i-th user and the m-th AP, β(d mi ) is the large-scale fading coefficient between the i-th user and the m-th AP.

[0069] The specific process is as follows:

[0070] The system model of a non-cellular massive MIMO system has M = 8 single-antenna APs, of which U = 4 are transmitting APs and V = 4 are receiving APs, with K u = 3 communication users and K s = 2 sensing targets with unknown locations. Each user is served by all APs at the same time, and all APs are connected to the central processing unit through a fault-free backhaul network. k =[x k ,y k ],1≤k≤K s is the position vector of the perceived target, q i =[x i ,y i ],1≤i≤K u Table is the communication user location, are the locations of the transmitting AP and the receiving AP respectively. In addition, are the polar angles between the u-th transmitting AP and the v-th receiving AP and the target, respectively.

[0071] Communication users are generated by independent sampling based on the Gaussian mixture model, where H = 3 is the number of Gaussian sub-distributions. μ1=[10,15]m, μ2=[25,45]m, μ3=[40,10]m, variance σ j =5m. For uplink transmission, the data symbol received by the mth AP from the user is,

[0072]

[0073] Among them, ρ i is the uplink transmission power, and the channel coefficient between the mth AP and the i-th user is γ mi is the large-scale fading coefficient, is the small-scale fading coefficient, and n m represents additive Gaussian noise. The data symbols of each user are independent of each other, and all users transmit with equal power ρ0 = 30 dB, so perfect channel state information can be obtained at the CPU.

[0074] According to the logarithmic distance path loss model, the transmission loss is,

[0075]

[0076] Where d is the path length of signal transmission, PL(d0) is the reference path loss, d0 = 1m, n is the path loss coefficient, n = 2 in free space, and shadow fading X is ignored. σ Influence, the large-scale fading coefficient is,

[0077]

[0078] Referring to the asymptotic expression of user communication SNR, The signal-to-noise ratio of the i-th user is,

[0079]

[0080] The system sum rate is the sum of all user rates.

[0081]

[0082] The communication site optimization problem is expressed as:

[0083]

[0084] Step S2: Based on the bistatic system path loss model, a mathematical relationship model between sensing performance and access point locations is constructed;

[0085] The mathematical relationship model between perception performance and access point location is:

[0086]

[0087] Where: |F k | is the Fisher information matrix of the k-th perception target, is the distance between the uth transmitting access point and the sensing target, is the distance between the vth receiving access point and the sensing target, is the polar angle between the u-th transmitting access point and the sensing target, is the polar angle between the vth receiving access point and the sensing target.

[0088] The specific process is as follows:

[0089] The two sensing targets follow an independent and identically distributed Gaussian model, μ0 = [25, 25]m, σ0 = 10m, and the Fisher information matrix of multiple target position parameters is constructed based on the dual-base system loss model. For the k-th sensing target, the FIM is,

[0090]

[0091] t=[t1,…,t U ],r=[r1,…,r V ], In the case of Gaussian noise, the elements of FIM are simplified to,

[0092]

[0093] The time delay of signal transmission τ uv The partial derivative with respect to the k-th sensory target position is an important component in building the FIM:

[0094]

[0095] have are the polar angles between the u-th transmitting AP and the v-th receiving AP and the target respectively:

[0096]

[0097] Simplify the constant term, let The Fisher information matrix is expressed as:

[0098]

[0099] Step S3: Modeling the access point site optimization problem in the cellular-free synaesthesia fusion system as a Markov decision process, constructing a state space and an action space, and fusing the mathematical relationship model between communication performance and access point site and the mathematical relationship model between perception performance and access point site to obtain a reward function of the Markov decision process;

[0100] The optimization objective of the synaesthesia-integrated non-cellular site optimization problem is the product of the communication user sum rate and the determinant of the average normalized Fisher information matrix. The reward function of the Markov decision process is:

[0101]

[0102] Among them: K s is the number of perceived targets, is an adjustable parameter related to communication performance, is the x-coordinate of the target AP site, is the y coordinate of the target AP site.

[0103] The state space contains the location information of the user equipment, the location information of the mobile target and the current location of the AP. The state space can be expressed as:

[0104]

[0105] Where: s t is the state space, q1 is the position of the i-th user, For K u The position of the user, p1 is the position of the first perceived target, For K s The position of the sensing target, t1 is the position of the first transmitting access point, t U is the location of the Uth transmitting access point, r1 is the location of the first receiving access point, r V is the position of the Vth receiving access point. The position information in the state space is normalized to improve the training stability.

[0106] The action space is:

[0107] a t ={Δt1,…,Δt U ,Δr1,…Δr V}

[0108] Among them: a t is the action space at time t, Δt1 is the movement of the first transmitting access point, Δt U is the movement of the Uth transmitting access point, Δr1 is the movement of the first receiving access point, and Δr V is the movement amount of the position of the Vth receiving access point.

[0109] The range of single-step movement at each position in the action space is:

[0110] [-d max ,d max ]

[0111] Where: d max The maximum moving distance of a single step.

[0112] Step S4: Use the flexible action-evaluation deep reinforcement learning algorithm to solve the Markov decision process and obtain the optimal access point deployment strategy.

[0113] In step S4, a flexible action-evaluation algorithm is used to maximize the historical state-action chain ρ of the strategy under strategy π while maximizing the cumulative expected reward function. π The expected entropy under the condition of , in order to perform actions as randomly as possible, is conducive to finding the optimal solution. The objective function of the flexible strategy evaluation is:

[0114]

[0115] in: is the objective function of flexible strategy evaluation, T is the total number of steps in a reinforcement learning training round, r(s t ,a t ) is the reward for the current action, ω is the entropy adjustment coefficient, is the policy entropy,

[0116] The input of the Actor network is the current state s t , using a two-layer multi-layer perceptron structure to output the probability distribution of actions. The goal of the Actor network is to maximize the expected return of entropy regularization, the gradient of which can be obtained by differentiating the policy parameter θ. The loss function of the Actor network is expressed as:

[0117]

[0118] in: is the experience replay buffer, Q φ (s t ,a t ) represents the reward value of the current state-action predicted by the Critic network, and the entropy regularization term -ωlogπ θ (a t |s t ) increases the randomness of the strategy and prevents the strategy from converging to the local optimum too early. ω is the temperature coefficient, which is achieved by minimizing the following loss function:

[0119]

[0120] in: The entropy lower limit is preset, and ω is updated through backpropagation, so that the policy entropy gradually approaches the target value during the training process. The design of the adaptive temperature coefficient makes the policy tend to try different site combinations in the early stage of site optimization. In the later stage of training, when the system tends to be stable, the policy gradually converges to a more precise deployment with better performance, thereby improving the utilization rate of the optimal policy.

[0121] The Critic network optimizes the Q-value function by minimizing the Bellman error. Its core is to recursively express the long-term value of the current state-action pair. These estimated values provide strategic guidance to the Actor network during training, establish a mapping relationship between the movement of synaesthesia AP sites and performance, and use Q-values to evaluate the pros and cons of various deployment strategies. To alleviate the problem of Q-value overestimation, SAC uses two independent Critic networks. and And take the minimum value as the final Q value estimate. The loss function of the Critic network is defined as:

[0122]

[0123] in: is the target value function, For the parameters of the target Critic network, through the soft update strategy This mechanism updates the target network parameters slowly to prevent the rapid changes of the Critic network from interfering with the Actor network, thereby reducing policy oscillation.

[0124] Figure 1 The overall framework of the proposed method for synaesthesia-integrated cellular-free site intelligent optimization is shown in the following figure. The superiority of the algorithm's computational performance is verified through a specific implementation.

[0125] 1. Experimental parameter settings

[0126] The non-cellular interawareness fusion system environment is 50m×50m, with U, V = 2, 3, 4, 5, 6, and 7 transmitting and receiving APs, 3 communication users, and 2 sensing targets. For the flexible action-evaluation framework reinforcement learning training parameter setting, the learning rate is 10 -5 , the experience replay pool capacity is 2 21 , the network gradient descent batch size is 512, the number of hidden layer neurons is 64×32, the discount factor size is 0.98, and the soft update parameter is 0.005.

[0127] 2. Synaesthesia fusion convergence performance

[0128] The optimization results of the synaesthesia fusion target of the non-cellular system under different strategy optimization algorithms are compared. Figure 5 As shown in the figure, the proposed SAC-based AP site optimization algorithm can achieve higher cumulative rewards in a shorter period of time and maintain relatively consistent performance in multiple experiments. Its strategic advantages and stability in optimization problems are significantly better than other strategy optimization methods.

[0129] 3. Communication, speed, and positioning performance

[0130] Figure 6 、 Figure 7 The communication sum rate and positioning error performance curves of the synaesthesia-fusion site optimization results for a non-cellular system were compared using different optimization strategies. The proposed algorithm achieved a communication sum rate of 92.5% of that achieved by the communication-only algorithm, a significant improvement over other algorithms. While the proposed algorithm's perception performance fell short of the optimal perception solution, it still maintained a high level of accuracy, reducing the positioning error to within 0.22m at a 95% confidence level. This result fully demonstrates the effectiveness of the proposed synaesthesia AP site optimization algorithm in improving target positioning accuracy.

[0131] 4. Station layout results

[0132] Figure 8 Figure 2 shows the resulting deployment results for U = V = 4, three communication users, and two sensing metrics. The proposed SAC-based AP layout is relatively rational, dispersed across various areas, ensuring effective coverage for communication users. Furthermore, the geometric distribution between APs and sensing targets is relatively reasonable, ensuring both high communication rates and a small error ellipse in positioning tasks, thus optimizing overall performance.

[0133] 5. Communication, rate, and positioning performance with different numbers of access points

[0134] Figure 9 、 10 Comparisons were made of the communication and rate metrics and perception performance indicators of a cellular-free interawareness fusion system with varying numbers of transmitting nodes. The results demonstrate that the proposed algorithm's communication rate and perception accuracy improve significantly with varying numbers of APs, demonstrating that increasing the number of APs significantly improves ISAC performance. Furthermore, the proposed algorithm maintains stable operation and outputs high-quality deployment solutions across diverse network scales, demonstrating excellent scalability and adaptability for diverse deployment requirements.

[0135] 6. Site deployment results for different numbers of access points

[0136] Figure 11 Comparing the deployment results of a cellular-free telemetry fusion system with different numbers of transmitting nodes, we can observe that as the number of transmitting and receiving APs increases, the system's spatial coverage expands and the AP layout becomes more uniform. This not only improves the system's overall communication performance but also enables high-precision target perception over a larger area.

[0137] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

Claims

1. A method for intelligent optimization of non-cellular sites for synaesthesia fusion, characterized in that: include: Step S1: Based on the logarithmic distance path loss model, a mathematical relationship model between communication performance and access point location is constructed; Step S2: Based on the bistatic system path loss model, a mathematical relationship model between sensing performance and access point locations is constructed; Step S3: Modeling the access point site optimization problem in the cellular-free synaesthesia fusion system as a Markov decision process, constructing a state space and an action space, and fusing the mathematical relationship model between communication performance and access point site and the mathematical relationship model between perception performance and access point site to obtain a reward function of the Markov decision process; Step S4: Use the flexible action-evaluation deep reinforcement learning algorithm to solve the Markov decision process and obtain the optimal access point deployment strategy.

2. The method for intelligent optimization of non-cellular sites for synaesthesia fusion according to claim 1, characterized in that: The mathematical relationship model between the communication performance and the access point site is: Where: R sum is the total communication system rate, R i is the communication rate of the i-th user, K u is the number of communication users, ρ0 is the uplink transmission power of all communication users, SNR i is the signal-to-noise ratio of the i-th user, U is the number of transmitting APs, V is the number of receiving APs, d mi is the distance between the i-th user and the m-th AP, β(d mi ) is the large-scale fading coefficient between the i-th user and the m-th AP.

3. The method for intelligent optimization of non-cellular sites for synaesthesia fusion according to claim 2, characterized in that: The mathematical relationship model between the sensing performance and the access point location is: Where: |F k | is the Fisher information matrix of the k-th perception target, is the distance between the uth transmitting access point and the sensing target, is the distance between the vth receiving access point and the sensing target, is the polar angle between the u-th transmitting access point and the sensing target, is the polar angle between the vth receiving access point and the sensing target.

4. The method for intelligent optimization of non-cellular sites for synaesthesia fusion according to claim 3, characterized in that: The reward function of the Markov decision process is: Among them: K s is the number of perceived targets, is an adjustable parameter related to communication performance, is the x-coordinate of the target AP site, is the y coordinate of the target AP site.

5. The method for intelligent optimization of non-cellular sites for synaesthesia fusion according to claim 4, characterized in that: The state space is Where: s t is the state space, q1 is the position of the i-th user, For K u The position of the user, p1 is the position of the first perceived target, For K s The position of the sensing target, t1 is the position of the first transmitting access point, t U is the location of the Uth transmitting access point, r1 is the location of the first receiving access point, r V is the position of the Vth receiving access point; The action space is: a t ={Δt1,…,Δt U ,Δr1,…Δr V } Among them: a t is the action space at time t, Δt1 is the movement of the first transmitting access point, Δt U is the movement of the Uth transmitting access point, Δr1 is the movement of the first receiving access point, and Δr V is the movement amount of the position of the Vth receiving access point.

6. The method for intelligent optimization of non-cellular sites for synaesthesia fusion according to claim 5, characterized in that: Each position information in the state space is normalized.

7. The method for intelligent optimization of non-cellular sites for synaesthesia fusion according to claim 5, characterized in that: The range of the single-step movement amount at each position in the action space is: [-d max ,d max ] Where: d max The maximum moving distance of a single step.

8. The method for intelligent optimization of non-cellular sites for synaesthesia fusion according to claim 5, characterized in that: In step S4, a flexible action-evaluation algorithm is used to maximize the historical state-action chain ρ of the strategy under strategy π while maximizing the cumulative expected reward function. π The expected entropy under the condition of , in order to perform actions as randomly as possible, is conducive to finding the optimal solution. The objective function of the flexible strategy evaluation is: in: is the objective function of flexible strategy evaluation, T is the total number of steps in a reinforcement learning training round, r(s t ,a t ) is the reward for the current action, ω is the entropy adjustment coefficient, is the policy entropy, The loss function of the Actor network is expressed as: in: is the experience replay buffer, Q φ (s t ,a t ) represents the reward value of the current state-action predicted by the Critic network, and the entropy regularization term -ωlogπ θ (a t |s t ) increases the randomness of the strategy and prevents the strategy from converging to the local optimum too early. ω is the temperature coefficient, which is achieved by minimizing the following loss function: in: is the preset entropy lower limit, and ω is updated through back propagation so that the policy entropy gradually approaches the target value during training; The loss function of the Critic network is defined as: in: is the target value function, For the parameters of the target Critic network, through the soft update strategy to update.

9. A cellular-free site intelligent optimization device for synaesthesia fusion, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 8 is implemented.

Images