A Modal Endogenous Cross-Sensory Array Antenna Beamforming Method

Abstract

The present invention provides a method for modal endogenous cross-sensory array antenna beamforming, belonging to the field of wireless communication technology. The method includes: establishing a communication and sensing integrated network for radar communication duplex; based on the established communication and sensing integrated network for radar communication duplex, converting the information rate of the duplex base station into an ergodic optimization problem with a total beam pattern error constraint; adopting a model-free learning method, introducing primal-dual learning to solve the ergodic optimization problem by controlling power, and obtaining an optimal power control strategy; using transfer learning to optimize the array antenna beamforming through iteration on the basis of power control. By adopting the present invention, a beamforming strategy with the optimal information rate can be achieved, and the detection accuracy can be ensured.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly to a beamforming method for a communication and sensing integrated array antenna with endogenous modality. Background Art

[0002] With the rapid development of communication and sensing integrated networks, ubiquitous sensing in future wireless communication networks is urgently needed. Beamforming in 5G networks can concentrate the energy of signals in the direction where the receiving end is located, thereby improving the spectrum utilization efficiency. Transmit beamforming enables each function to use its individual waveform, and by utilizing the traditional dedicated signals of each function, it may support higher data rates and guaranteed radar performance. Especially in new high-frequency bands such as millimeter waves and terahertz, it is more necessary to combine beamforming technology to better perform detection and communication.

[0003] In recent years, beamforming in wireless communication networks has begun to be equipped with artificial intelligence (AI). However, these works are established in a pre-designed training set. The learning performance depends on the work of the training set derived from traditional management methods, which results in beamforming being restricted in some special cases. Although the current work on dataset collection has been underway, it requires a large amount of preparation and accumulation and is not yet sufficient to cope with the flexible and changeable communication and sensing integrated network of radar communication duplex. Summary of the Invention

[0004] Embodiments of the present invention provide a beamforming method for a communication and sensing integrated array antenna with endogenous modality, which can achieve a beamforming strategy with the optimal information rate and ensure the detection accuracy. The method includes:

[0005] Establish a communication and sensing integrated network of radar communication duplex;

[0006] Based on the established communication and sensing integrated network of radar communication duplex, convert the information rate of the duplex base station into an ergodic optimization problem with a total beam pattern error constraint;

[0007] Adopt a model-free learning method, introduce primal-dual learning, and solve the ergodic optimization problem by controlling the power to obtain the optimal power control strategy;

[0008] Utilize transfer learning to optimize the beamforming of the array antenna through iteration on the basis of power control.

[0009] Further, the communication and sensing integrated network of radar communication duplex includes: 1 access point equipped with a radar communication duplex antenna and K single-antenna users; where

[0010] Let N C and NR represent the number of communication antennas and the number of radar antennas of the access point respectively, then N C +N R = M;

[0011] At time slot t, d k [t] and n k [t] represent the communication signal and the corresponding noise signal respectively. The noise signal n k [t] follows a complex normal distribution with a mean of 0 and a standard deviation of N0 Let s[t] be the radar signal and its corresponding covariance matrix be:

[0012]

[0013] where represents the complex domain, T represents the total time slot length, (·) H represents the conjugate transpose operation. The signal-to-interference-plus-noise ratio γ of the k-th user k is expressed as:

[0014]

[0015] where (·) T represents the matrix transpose operation; and represent the channel vectors of the communication signal and the radar signal respectively; represents the beamforming vector channel of the k-th user;

[0016] radar communication dual-functional channel matrix where the k-th channel state in the channel matrix H

[0017] Furthermore, for the communication and sensing integrated network based on the established radar communication duplex, converting the information rate of the duplex base station into an ergodic optimization problem with a total beam pattern error constraint includes:

[0018] Determine the beam pattern error L r，1 (R):

[0019]

[0020] where L represents the total number of paths; d(θ l ) represents the ideal beam pattern, θ l represents the azimuth angle of the θ l -th path; P(θ l ) = a H (θ l )Ra(θ l ) represents θ lPower consumption in the direction; a(·) represents the steering vector;

[0021] Determine the root mean square cross-correlation pattern L r,2 (R):

[0022]

[0023] where P c (θ l , θ r ; R) = a H (θ l )Ra(θ r ) represents the root mean square cross-correlation beam pattern in the directions of θ l and θ r ;

[0024] Determine the total error of the beam pattern L r (R) and its constraint:

[0025] L r (R) = L r,1 (R) + L r,2 (R)

[0026] L r (R) ≤ ε

[0027] where ε represents the threshold value;

[0028] Determine the spectral efficiency r of user k k as:

[0029] r k = log(1 + γ k )

[0030] where the information rate

[0031] Combined with the constraints, the ergodic optimization problem of the multi-user multiple-input single-output scenario in the communication-sensing integrated network is expressed as:

[0032]

[0033] s.t. τ k |w k | 2 ≤ P max ,

[0034] L r (R) ≤ ε.

[0035] where represents the mathematical expectation, f(h, p(h)) is the instantaneous performance, h is the channel state, and p(h) is the power control strategy, Denote utility, s.t. denote constraint conditions, τ k denotes the scheduling factor of user k, P max denotes the power budget.

[0036] Furthermore, the model-free learning method is adopted, and the primal-dual learning is introduced to solve the ergodic optimization problem by controlling the power. The obtained optimal power control strategy includes:

[0037] Characterize the power control strategy p(h) through the parameter π(h; ω) in the deep neural network, that is: p(h) = π(h; ω), and introduce the primal-dual learning to determine the Lagrangian function of the ergodic optimization problem

[0038]

[0039] where is composed of the weighted sum of the utility and the constraint function multiplier; f(h, π(h; ω)) is the loss function, g k (h, π(h; ω)) is the constraint function, including: τ k |w k | 2 ≤P max and L r (R) ≤ ε; non-negative dual variable λ k denotes the Lagrangian multiplier of the dual variable for the k-th user; the primal variable ω r denotes the weight w r and bias b r of the r-th layer of the deep neural network, R represents the number of layers of the deep neural network; π(h; ω) is the power control strategy output by the deep neural network, π(h; ω) = ξ R (ξ R-1 (...ξ1(ξ0(h; ω0); ω1)...; ω R-1 ) ; ω R )), ξ r (·) is the activation function of the r-th layer of the deep neural network; G k denotes the threshold value of the power control strategy π(h; ω) output by the deep neural network;

[0040] Determine the dual function of the ergodic optimization problem as:

[0041]

[0042] Write the dual problem as:

[0043]

[0044] such that λ k ≥0, k = 1, ..., K

[0045] Adopt the mini - batch gradient descent method to iteratively optimize the original variable ω and the dual variable λ until the loss function converges or reaches the maximum number of iterations, and obtain the optimal power control strategy π*(h; ω*) to achieve power control; among them, the iterative formulas for the original variable ω and the dual variable λ are:

[0046]

[0047]

[0048] Among them, the superscript t represents the t - th iteration, η represents the iteration step size, denotes the gradient with respect to ω, denotes the gradient with respect to λ k with respect to, the form (x) + = max[0, x] means taking the non - negative number, taking zero when less than zero.

[0049] Furthermore, the scheduling factor τ of user k k will be scheduled for access by seeking the local maximum, that is, satisfying:

[0050]

[0051] Among them, represents the optimized scheduling factor of user k, and k* represents k when f(h, p(h)) takes the maximum value.

[0052] Furthermore, the use of transfer learning to iteratively optimize the array antenna beamforming on the basis of power control includes:

[0053] In the transfer learning algorithm, there are domain D = {F, P(X)} and task T = {Y, f(·)}, where F is the feature, P(X) is the distribution, Y is the label, f(·) is the prediction function, and X is the input data;

[0054] The power control in the real number domain in the communication - sensing integrated network is represented as D p , and the beamforming in the complex number domain is represented as D w , the training set of power control is T p , and when using transfer learning to solve beamforming, it is represented by the following formula:

[0055]

[0056]

[0057] Among them, l(·) represents the loss function of beamforming under transfer learning; t represents the t-th cycle; represents using the existing domain of power control and the task to perform calculations to obtain the beamforming prediction value; represents the t-th iteration of the input data set of beamforming, represents the true value of beamforming, represents the task executed in the t-th iteration under beamforming;

[0058] Based on the training results of power control, using transfer learning, transform the output layer of the trained deep neural network of power control to achieve the control of beamforming:

[0059]

[0060] s.t. τ k |w k | 2 ≤P max ,

[0061] L r (R)≤ε.

[0062] Among them, f w (t) is the beamforming prediction function, characterized by the information rate, that is

[0063] Using the primal-dual learning optimization, obtain the beamforming strategy π w (h; ω):

[0064] π w (h; ω) = ψ R (ξ R-1 (... ξ1(ξ0(h; ω0); ω1)...; ω R-1 ) ; φ R )

[0065] Among them, ψ R (·) represents the activation function of the R-th layer of the deep neural network, and φ R represents the output strategy of beamforming; among them, the output layer outputs a K×2 beamforming strategy π w (h; ω).

[0066] The beneficial effects brought by the technical solution provided by the embodiments of the present invention at least include:

[0067] In an embodiment of the present invention, a communication and sensing integrated network with radar communication duplex is established; based on the established communication and sensing integrated network with radar communication duplex, the information rate of the duplex base station is transformed into an ergodic optimization problem with a total beam pattern error constraint; a model-free learning method is adopted, and the primal-dual learning is introduced to solve the ergodic optimization problem by controlling the power, and an optimal power control strategy is obtained; by using transfer learning, the optimization of the array antenna beamforming is realized through iteration on the basis of power control; in this way, the beamforming strategy with the optimal information rate can be realized, and the detection accuracy can be guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0069] Figure 1 It is a schematic flow chart of a method for beamforming of a modal endogenous communication and sensing array antenna provided by an embodiment of the present invention;

[0070] Figure 2 It is a schematic structural diagram of a device corresponding to the method for beamforming of a communication and sensing array antenna provided by an embodiment of the present invention;

[0071] Figure 3 It is a schematic diagram of the modal endogenous principle provided by an embodiment of the present invention;

[0072] Figure 4 It is a detailed schematic flow chart of a method for beamforming of a modal endogenous communication and sensing array antenna provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0073] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the drawings.

[0074] As Figures 1 - 3 shown, an embodiment of the present invention provides a method for beamforming of a modal endogenous communication and sensing array antenna, including:

[0075] S101, establishing a communication and sensing integrated network with radar communication duplex;

[0076] In this embodiment, the communication and sensing integrated network with radar communication duplex includes: 1 access point equipped with a radar communication duplex antenna and K single-antenna users; where Figure 2 the target in

[0077] Let N Cand N R represent the number of communication antennas and the number of radar antennas of the access point respectively, then N C +N R = M; This antenna is a transmitting antenna, which means that the transmitting antenna of the access point can be divided into two exclusive functions.

[0078] At time slot t, d k [t] and n k [t] represent the communication signal and the corresponding noise signal respectively. The noise signal n k [t] follows a complex normal distribution with a mean of 0 and a standard deviation of N0 Let s[t] be the radar signal and its corresponding covariance matrix is:

[0079]

[0080] where represents the complex domain, T represents the total time slot length, (·) H represents the conjugate transpose operation. The signal-to-interference-to-noise ratio γ k (signal-to-interference-to-noise ratio, SINR) of the k-th user is expressed as:

[0081]

[0082] where, (·) T represents the matrix transpose operation; and represent the channel vectors of the communication signal and the radar signal respectively; represents the beamforming vector channel of the k-th user, assuming it follows a flat Rayleigh distribution (exponential distribution with an exponent of 2); The in the denominator is the multi-user interference, and is the interference generated by the sensing signal. The subscript k represents the communication user, and the subscript l represents the interfering user;

[0083] Radar communication dual-functional channel matrix where, the k-th channel state in the channel matrix H

[0084] In this embodiment, during communication, the transmit power p k of user k is limited by the power budget P max :

[0085] τ k ·||w k || 2 ≤ P max

[0086] where |w k | 2 = p k , τ k ∈ {0, 1} is the scheduling factor of user k, and τ k = 1 indicates that user k successfully accesses, otherwise it means τ k = 0. The rule followed by the scheduling factor is that each user occupies one channel at a time. Therefore, the user scheduling factor can be restricted as follows:

[0087]

[0088] S102. Based on the established integrated communication and sensing network for radar communication duplex, convert the information rate of the duplex base station into an ergodic optimization problem with the total beam pattern error constraint;

[0089] In this embodiment, for radar signals, the beam pattern error is considered in the integrated communication and sensing network. The multiple input multiple output (MIMO) radar transmit beamforming design aims to optimize the transmit power in a given direction or generally match the required beam pattern. The beam pattern error L r，1 (R) is the mean square error between the obtained beam pattern and the ideal beam pattern, which can be calculated by the following formula:

[0090]

[0091] where L represents the total number of paths; d(θ l ) represents the ideal beam pattern, θ l represents the azimuth angle of the θ l -th path; P(θ l ) = a H (θ l )Ra(θ l ) represents the power consumption in the direction of θ l ; a(·) represents the steering vector; the uniform linear antenna is the antenna structure considered in the system, so the steering vector can be expressed as:

[0092]

[0093] where d represents the antenna spacing and λ represents the wavelength;

[0094] The ideal beam pattern d(θ l ) is expressed as:

[0095]

[0096] where θ pDenote the ideal azimuth angle, △ denote the ideal azimuth angle beamwidth, p denote the number of ideal azimuth angles, and the azimuth angle varies from -90° to 90° at a resolution of 1°;

[0097] Determine the root mean square cross-correlation pattern L r,2 (R):

[0098]

[0099] where, P c (θ l , θ r ; R) = a H (θ l ) Ra(θ r ) represents the root mean square cross-correlation beam pattern in the directions of θ l and θ r , then calculate the loss function from the root mean square cross-correlation pattern;

[0100] Determine the total beam pattern error L r (R) and its constraint conditions:

[0101] L r (R) = L r,1 (R) + L r,2 (R)

[0102] L r (R) ≤ ε

[0103] where, ε represents the threshold value; this constraint indicates that the total beam pattern error is not allowed to exceed the threshold value ε, and this threshold value symbolizes the accuracy of the sensing performance.

[0104] In this embodiment, the considered beamforming strategy is the optimization of the information rate. Define the spectral efficiency r of user k k as:

[0105] r k = log(1 + γ k )

[0106] where, the information rate

[0107] In this embodiment, the resource management in the wireless communication system can be evaluated by the instantaneous performance f(h, p(h)), for example, rate, power consumption, energy efficiency, etc. h represents the channel state, and p(h) represents the corresponding instantaneous resource allocation strategy, which refers to the power control strategy in this embodiment. However, from the perspective of the end user, the instantaneous system performance often changes too fast. The long-term mathematical expectation is a more valuable metric. It can be calculated in the following way:

[0108]

[0109]

[0110] k = 1, ..., K.

[0111] Among them, G k represents the threshold value;

[0112] For the constraint conditions, they can be reflected by supplementing the objective function f(h, p(h)), so as to simplify the constraint conditions:

[0113]

[0114] Combined with the constraint conditions, the ergodic optimization problem of the multi-user multiple-input single-output scenario in the communication-sensing integrated network is expressed as:

[0115]

[0116] s.t. τ k |w k | 2 ≤ P max ,

[0117] L r (R) ≤ ε.

[0118] Among them, represents the mathematical expectation, f(h, p(h)) is the instantaneous performance, h is the channel state, p(h) is the power control strategy, represents the utility, s.t. represents the constraint conditions, τ k represents the scheduling factor of user k, P max represents the power budget.

[0119] S103. Adopt a model-free learning method, introduce the primal-dual learning to solve the ergodic optimization problem by controlling the power, and obtain the optimal power control strategy;

[0120] In this embodiment, the above ergodic optimization problem is a non-convex problem. To solve the ergodic optimization problem, the power control strategy p(h) can be characterized by the parameters π(h; ω) of a deep neural network (DNN):

[0121] p(h) = π(h; ω)

[0122] Among them, each ω in r represents the weight w r and bias b of the r-th layer of the deep neural networkr Since the DNN is fully connected and forward - propagating, the final output can be expressed as:

[0123] u r = ξ r (ω r u r-1 + b r ), r = 1,..., R

[0124] where u r represents the output of the current r - th layer, ξ r (·) is the activation function of the r - th layer, and R represents the number of layers of the deep neural network;

[0125] Assume that the power control strategy p(h) can be approximated by the parameters π(h; ω) of the DNN.

[0126] π(h; ω) = ξ R (ξ R-1 (...ξ1(ξ0(h; ω0); ω1)...; ω R-1 ) ; ω R )

[0127] Using sufficient linear and non - linear operations, it can be accurately approximated as p(h). For any small positive error σ > 0, π(h; ω) can obtain an approximation with sufficient R layers and activation functions:

[0128]

[0129] where sup represents the least upper bound;

[0130] Based on the universal approximation theorem, this means that for a given H, there exist parameters ω such that the policy π(h; ω) output by the DNN approximates the original power control strategy p(h) under the condition of ensuring the existence of any small positive error σ.

[0131] Introduce primal - dual learning to solve the ergodic optimization problem. The Lagrangian function of the ergodic optimization problem is composed of the weighted sum of the utility and the constraint function and its multipliers. For convenience, the Lagrangian function of its ergodic optimization problem is given by the following formula:

[0132]

[0133] where f(h, π(h; ω)) is the loss function, g k (h, π(h; ω)) is the constraint function (in this embodiment, it includes: τ k |w k | 2 ≤ P max and Lr (R) ≤ ε); G k Denotes the threshold value of the power control strategy π(h; ω) output by the deep neural network; non - negative dual variable λ k Denotes the Lagrange multiplier of the dual variable for the k - th user;

[0134] Correspondingly, the dual function of the ergodic optimization problem can be obtained It can be expressed as:

[0135]

[0136] The dual problem can be written as:

[0137]

[0138] s.t. λ k ≥ 0, k = 1,..., K

[0139] Adopt the mini - batch gradient descent method to iteratively optimize the original variable ω and the dual variable λ until the loss function converges or reaches the maximum number of iterations, and obtain the optimal power control strategy π*(h; ω*) to achieve power control; where the iterative formulas for the original variable ω and the dual variable λ are:

[0140]

[0141]

[0142] Among them, the form (x) + = max[0, x] means taking the non - negative number, taking zero when less than zero;

[0143] For the gradient value of the loss function to be solved and the gradient value of the constraint function They are calculated through the chain rule. Through the iteration of the primal - dual variables, an optimal power control strategy π*(h; ω*) can be finally obtained to achieve power control. At the same time, the scheduling factor τ k Will be scheduled for access by seeking the local maximum, that is, satisfying:

[0144]

[0145] Among them, Denotes the scheduling factor of the optimized user k, and k* denotes the k when f(h, p(h)) takes the maximum value;

[0146] Finally, based on the obtained optimized user scheduling factor τ* and power control strategy π*(h; ω*), the power control native module for the integrated communication and sensing network with optimal information rate is obtained. Among them, S101 - 103 are implemented by the power control native module. In the power control derivative module, the mini - batch gradient descent method is adopted, with as the loss function, and g k (h, π(h; ω)) as the constraint function. The original variable ω and the dual variable λ are iteratively optimized using the primal - dual learning method until the loss function converges (essentially making the information rate converge) or the maximum number of iterations is reached.

[0147] S104, using transfer learning, iteratively optimizes the array antenna beamforming based on power control, and finally realizes the optimization of the information rate and beam pattern of the duplex base station in the integrated communication and sensing network.

[0148] Since in communication and sensing, more attention is paid to beamforming extended to the complex domain. And power control and beamforming are related. By combining the existing DNN with the idea of transfer learning, the beamforming strategy with optimal information rate can be further realized. Therefore, in this embodiment, the beamforming derivative module is generated by the power control native module. In the beamforming derivative module, the deep neural network in the power control native module is transformed, and the iterative optimization of beamforming is realized by learning and training faster in the form of primal - dual, and finally the beamforming optimization based on power control is realized. In this way, the detection accuracy can be guaranteed while meeting the information rate.

[0149] In the transfer learning algorithm, there are domain D = {F, P(X)} and task T = {Y, f(·)}, where F is the feature, P(X) is the distribution, Y is the label, f(·) is the prediction function, and X is the input data;

[0150] The power control in the real number domain in the integrated communication and sensing network is denoted as D p , and the beamforming in the complex domain is denoted as D w . The training set of power control is T p . Using transfer learning to solve beamforming can be expressed by the following formula:

[0151]

[0152]

[0153] Among them, l(·) represents the loss function of beamforming under transfer learning; t represents the t - th iteration; represents using the existing domain of power control and task Perform calculations to obtain the beamforming prediction value; Denote the t-th iteration of the input data set for beamforming, Denote the beamforming true value, Denote the task executed at the t-th iteration under beamforming;

[0154] Going back to the original problem, the local output layer can be changed through the DNN network of power control to achieve further control of beamforming:

[0155]

[0156] s.t. τ k |w k | 2 ≤ P max ,

[0157] L r (R) ≤ ε.

[0158] Among them, f w (t) is the beamforming prediction function, characterized by the information rate, that is

[0159] For the implementation of the corresponding algorithm, the original-dual learning optimization is still used, and the same allocation strategy will have corresponding changes on the original basis π(h; ω):

[0160] π w (h; ω) = ψ R (ξ R-1 (... ξ1(ξ0(h; ω0); ω1)...; ω R-1 ) ; φ R )

[0161] Among them, π w (h; ω) represents the beamforming strategy, ψ R (·) represents the activation function of the R-th layer of the deep neural network, and φ R represents the output strategy of beamforming;

[0162] In this embodiment, the output layer is no longer the K×1 power control strategy π(h; ω), but a K×2 beamforming strategy π w (h; ω) that includes the real part and the imaginary part. In this embodiment, only the parameters φ R of the last layer and the activation function ψ R (·) need to be changed because there is an equality constraint |w k | 2 = p k , which means that the energy of beamforming is provided by the transmission power, and this can reduce the number of training times.

[0163] In this embodiment, S104 is implemented by a beamforming derivative module.

[0164] As Figure 4 shown, the method for beamforming of a modal endogenetic communication and sensing array antenna according to an embodiment of the present invention may specifically include the following steps:

[0165] A1. Initialize the maximum number of iterations, the parameters of the deep neural network and the dual variable λ, and generate a channel data set H = [h1, h2,..., h K through a Rayleigh distribution;

[0166] A2. Perform mini-batch sampling on the generated channel data set

[0167] A3. Take f(h, π(h; ω)) as the loss function and g k (h, π(h; ω)) as the constraint function, and calculate the gradient values of the loss function and the constraint function respectively;

[0168] A4. Use the backpropagation of the deep neural network to perform iterative updates of the neural network parameters ;

[0169] A5. Update the dual variable to accelerate the training of the native module for power control;

[0170] A6. When the power control native module reaches the maximum number of iterations or the loss function converges, execute step A7; otherwise, return to step A2 and continue to execute;

[0171] A7. Based on the power control native module, use transfer learning to generate a beamforming derivative module;

[0172] A8. When the beamforming derivative module reaches the maximum number of iterations or the loss function converges, end the iteration; otherwise, return to step A2 and continue to execute.

[0173] As Figure 3 shown, an embodiment of the present invention further provides a device for beamforming of a modal endogenetic communication and sensing array antenna, which is used to implement the above-mentioned method for beamforming of a modal endogenetic communication and sensing array antenna. The device includes: a power control native module and a beamforming derivative module; wherein,

[0174] In the power control derivative module, use the primal-dual learning method to iteratively optimize this loss function until it converges;

[0175] In the beamforming derivative module, transform the deep neural network in the native module to achieve iterative optimization of beamforming, and finally achieve beamforming optimization based on power control.

[0176] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for modal endogenous cross-sensory array antenna beamforming, characterized in that Including: Establish a communication and sensing integrated network with radar communication duplex; Based on the established communication and sensing integrated network with radar communication duplex, convert the information rate of the duplex base station into an ergodic optimization problem with beam pattern total error constraint; Adopt a model-free learning method, introduce primal-dual learning, and solve the ergodic optimization problem by controlling power to obtain an optimal power control strategy; Utilize transfer learning to iteratively optimize the beamforming of the array antenna based on power control; Among them, the process of converting the information rate of the duplex base station into an ergodic optimization problem with beam pattern total error constraint based on the established communication and sensing integrated network with radar communication duplex includes: Determine the beam pattern error L r，1 (R): Among them, L represents the total number of paths; d(θ l ) represents the ideal beam pattern, and θ l represents the azimuth angle of the l-th path; P(θ l ) = a H (θ l )Ra(θ l ) represents the power consumption in the direction of θ l ; a(·) represents the steering vector; R represents the covariance matrix of the radar signal; (·) H represents the conjugate transpose operation; Determine the root mean square cross-correlation pattern L r,2 (R): where P c (θ l , θ r ; R) = a H (θ l ) Ra(θ r ) represents the root mean square cross-correlation beam pattern in the directions of θ l and θ r ; Determine the total error L of the beam pattern r (R) and its constraints: L r (R) = L r,1 (R) + L r,2 (R) L r (R) ≤ ε Where ε represents the threshold value; Determine the spectral effect r of user k k It is: r k = log(1 + γ k ) where γ k represents the signal-to-interference-plus-noise ratio (SINR) of the k-th user, and the information rate K represents the number of single-antenna users; Combined with the constraint conditions, represent the ergodic optimization problem in the multi-user multiple-input single-output scenario in the communication and sensing integrated network as: s.t.τ k |w k | 2 ≤P max , L r (R) ≤ ε. Among them, represents the mathematical expectation, f(h, p(h)) is the instantaneous performance, h is the channel state, and p(h) is the power control strategy. represents the utility, s.t. represents the constraint condition, and τ k represents the scheduling factor of user k, and P max represents the power budget, and w k represents the beamforming vector channel of the k-th user.

2. The method for modal endogenous synaesthetic array antenna beamforming according to claim 1, characterized in that, The communication and sensing integrated network with radar communication duplex includes: 1 access point equipped with a radar communication duplex antenna and K single-antenna users; where Let N C and N R respectively represent the number of communication antennas and the number of radar antennas of the access point, then N C + N R = M; At time slot t for d k [t] and n k [t] represent the communication signal and the corresponding noise signal respectively, and the noise signal n k [t] follows a complex normal distribution with a mean of 0 and a standard deviation of N0 Let s[t] be the radar signal and its corresponding covariance matrix be: Among them, represents the complex number field, T represents the total time slot length, (·) H represents the conjugate transpose operation, and the signal-to-interference-plus-noise ratio γ of the k-th user k is expressed as: Among them, (·) T represents the matrix transpose operation; and respectively represent the channel vectors of communication signals and radar signals; represents the beamforming vector channel of the k-th user; Radar communication dual-functional channel matrix Among them, the k-th channel state in the channel matrix H 3. The method for modal endogenetic synesthetic array antenna beamforming according to claim 1, characterized in that The process of adopting a model-free learning method, introducing primal-dual learning, and solving the ergodic optimization problem by controlling power to obtain an optimal power control strategy includes: Characterize the power control strategy p(h) by the parameter π(h; ω) in the deep neural network, that is: p(h) = π(h; ω), and introduce the primal-dual learning to determine the Lagrangian function of the ergodic optimization problem wherein, is composed of the weighted sum of the utility and the constraint function multiplier; f(h, π(h; ω)) is the loss function, g k (h, π(h; ω)) is the constraint function, including: τ k |w k | 2 ≤P max and L r (R) ≤ ε; the non - negative dual variable λ k represents the Lagrange multiplier of the dual variable for the k - th user; the original variables r = 1,..., R, ω r represents the weight w r and bias b r of the r - th layer of the deep neural network, R represents the number of layers of the deep neural network; π(h; ω) is the power control strategy output by the deep neural network, π(h; ω) = ξ R (ξ R-1 (... ξ1(ξ0(h; ω0); ω1)...; ω R-1 ); ω R ), ξ r (·) is the activation function of the r - th layer of the deep neural network; G k represents the threshold value of the power control strategy π(h; ω) output by the deep neural network; Determine the dual function for the ergodic optimization problem which is Write the dual problem as: s.t. λ k ≥ 0, k = 1, ..., K Adopt the mini-batch gradient descent method to iteratively optimize the primal variable ω and the dual variable λ until the loss function converges or reaches the maximum number of iterations, and obtain the optimal power control strategy π*(h; ω*) to achieve power control; where the iterative formulas for the primal variable ω and the dual variable λ are: where the superscript t represents the t-th iteration, and η represents the iteration step size. denotes taking the gradient with respect to ω. denotes taking the gradient with respect to λ k where the form (x) + = max[0, x] means taking the non-negative number, i.e., taking zero when it is less than zero.

4. The method for modal endogenetic synaesthetic array antenna beamforming according to claim 3, characterized in that, The scheduling factor τ of user k k Scheduling access will be performed by seeking local maxima, i.e., satisfying: Among them, represents the scheduling factor of the optimized user k, where k * represents the k when f(h, p(h)) reaches the maximum value.

5. The method for modal endogenous synesthesia array antenna beamforming according to claim 3, wherein The process of utilizing transfer learning to iteratively optimize the beamforming of the array antenna based on power control includes: In the transfer learning algorithm, there are domain D = {F, P(X)} and task T = {Y, f(·)}, where F is the feature, P(X) is the distribution, Y is the label, f(·) is the prediction function, and X is the input data; In the integrated communication and sensing network, the power control in the real number domain is denoted as D p , and the beamforming in the complex number domain is denoted as D w , the training set of power control is T p , when using transfer learning to solve beamforming, it is represented by the following formula: Among them, l(·) represents the loss function of beamforming under transfer learning; t represents the t-th cycle; represents using the existing domain of power control and the task to perform calculations to obtain the beamforming prediction value; represents the t-th iteration of the input data set of beamforming, represents the beamforming true value, represents the task executed in the t-th iteration under beamforming; Based on the training results of power control, utilize transfer learning to transform the output layer of the trained deep neural network for power control to achieve beamforming control: s.t.τ k |w k | 2 ≤P max , L r (R) ≤ ε. Among them, f w (t) is the beamforming prediction function, characterized by the information rate, that is Using primal-dual learning optimization, the beamforming policy π is obtained w (h; ω): π w (h; ω) = ψ R (ξ R-1 (...ξ1(ξ0(h; ω0); ω1)...; ω R-1 ); φ R ) Among them, ψ R (·) represents the activation function of the R-th layer of the deep neural network, and φ R represents the output strategy of beamforming; among them, the output layer outputs a K×2 beamforming strategy π w (h; ω).

Images