Full-duplex radar communication system joint transmission beam forming method based on deep learning

Abstract

The invention discloses a full-duplex radar communication system joint transmission beam forming method based on deep learning, and the method comprises the steps: firstly carrying out the mathematical modeling of the performances of transmitting and receiving signals, radar perception and communication rate of an FD-ISAC system; in the constructed model, the base station receives a signal and obtains an estimated angle of the user through a traditional channel estimation method. Secondly, uplink historical channel state information and downlink historical channel state information are calculated based on the angle; and then designing a cross attention-gated cyclic unit network (CAttn-GRU Net) which takes the uplink historical channel state information and the downlink historical channel state information as input to predict a radar signal covariance matrix and a communication signal beam forming matrix, so as to respectively assist in constructing a sensing signal and a beam forming design at the current moment. In addition, in the training process, a multi-objective optimization loss function is adopted.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless transmission, and specifically relates to a joint transmit beamforming method for a full-duplex radar communication system based on deep learning. Background Art

[0002] Effective transmit beamforming design is the key to unlock the potential of multiple-input multiple-output (MIMO) communication systems and MIMO radar systems. Many existing works study the transmit design in multi-antenna ISAC systems by focusing on joint beamforming optimization. Specifically, beamforming is optimized by minimizing the beam pattern matching error while taking into account the individual signal-to-interference-plus-noise ratio (SINR) requirements of the communication users. However, these works only design transmit beamforming without considering the reception of radar echoes.

[0003] The main function of a radar system is to estimate the channel parameters of a target, such as delay and Doppler frequency, from the received radar echo signal. Considering radar echo reception in an ISAC system, the relevant scenarios are divided into two cases. The first case corresponds to downlink ISAC, where radar sensing reuses the resources of downlink transmission, and the BS acts as a radar transceiver and a communication transmitter. The transmitted downlink ISAC signal is known to the BS and can be used for receive processing for sensing. By applying a linear receive beamformer to the echo signal, the radar signal-to-interference-and-noise ratio (SINR) for target detection is explicitly obtained. The second case considers integrating sensing with uplink communication, where the BS can be regarded as a radar transceiver and a communication receiver. The radar receiver operates while transmitting, i.e., in full-duplex (FD) mode. Self-interference (SI) is a key issue in FD operation. In a FD ISAC system, SIC should be performed only for direct signal coupling between transceiver antennas, while target reflections should be retained. However, for FD radar, the integrated communication function only occurs in the downlink or uplink, operating in half-duplex (HD) mode. In order to achieve higher spectrum efficiency, people also consider the FD capability of communication, that is, let the BS be used as a radar transceiver and a communication transceiver at the same time. In this setting, not only is there interference between the sensing and communication functions, but there is also coupling between the uplink and downlink transmissions, which makes the ISAC design very complicated. The algorithms designed in the past did not consider the impact of uplink communication, only a pure downlink sensing signal was sent, the uplink transmission power was fixed, the possibility of downlink communication was not considered, and the uplink transmission was not designed.

[0004] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgement or any form of suggestion that the information constitutes the prior art already known to ordinary technicians in this field. Summary of the invention

[0005] This paper proposes a joint transmit beamforming method for a full-duplex radar communication system based on deep learning. This method significantly reduces the complexity of signal detection and improves the signal perception accuracy and communication rate of the system in complex environments through innovative algorithm design, neural network architecture optimization and effective combination with traditional methods.

[0006] A deep learning-based joint transmit beamforming method for a full-duplex radar communication system of the present invention comprises the following steps:

[0007] Step 1: The full-duplex base station receives the echo signal and estimates the angle of the user in each time slot through traditional channel estimation methods, such as multiple signal classification (MUSIC);

[0008] Step 2, constructing a sending and receiving signal model of the FD-ISAC system, and calculating the uplink historical channel state information and the downlink historical channel state information according to the angle of the user in each time slot;

[0009] Step 3, construct the CAttn-GRU network, including the computational attention module, the gated recurrent unit module, and the fully connected network;

[0010] The estimated uplink historical channel state information and downlink historical channel state information are input into the computational attention module in the CAttn-GRU network to obtain the attention modeling channel state information; the attention modeling channel state information is input into the gated recurrent unit module to obtain the global historical channel state information; the global historical channel state information is sent to the fully connected network to predict the radar signal covariance matrix and the communication signal beamforming matrix;

[0011] The radar signal covariance matrix and the communication signal beamforming matrix are used to assist in the current moment perception signal and beamforming design.

[0012] Furthermore, in step 2, the receiving signal model of the FD-ISAC system is constructed, and the calculation formula of the base station receiving signal is expressed as follows:

[0013]

[0014] in, represents the antenna gain, γ k,n represents the uplink communication channel coefficient, γ l,n represents the downlink communication channel coefficient, γ l＝0,n represents the downlink communication channel coefficient of the sensing target, τ 0,n and ν 0,n Respectively represent the time delay and Doppler frequency shift of the perceived target, τ l,n and ν l,n They represent the time delay and Doppler shift of the downlink communication user respectively; represents the beamforming vector associated with downlink user l;

[0015] l∈{1,…,L}, Represents the unit power data symbol of user l, that is,

[0016] Represents the covariance matrix dedicated radar signal; Assuming the unit power signal of user l and dedicated radar signals Independent of each other, Indicates the unit power data symbol of uplink communication user k, a t (θ l,n ) represents the steering vector of the transmitting array in the direction of the downlink communication user l, a r (θ l,n ) represents the steering vector of the receiving array in the direction of downlink user l, a r (θ 0,n ) represents the steering vector of the receiving array to the perceived target, z represents undesired interference, and n represents noise;

[0017] In the above formula, the downlink beamforming is designed by and F0; once F0 is determined, a dedicated radar signal is generated where a t (θ k,n ) represents the guiding vector of the transmitting array in the θ direction, which can be written as:

[0018]

[0019] a r (θ k,n ) represents the steering vector of the receiving array in the θ direction, which can be written as:

[0020]

[0021] Among them, N t Indicates the number of base station transmitting antennas, N r Indicates the number of base station receiving antennas;

[0022] Assume that there are I signal-independent interference sources, located at angles and θ i,n ≠θ 0,n , these interferences will also reflect the sensor signal to the BS, generating undesired interference z, the calculation formula of z is as follows:

[0023]

[0024] In addition, considering the total transmit power constraint Where P max represents the maximum available power budget of the BS;

[0025] definition represents the uplink channel state information between the kth user and the BS;

[0026] Applying the arrival angle AOA estimation technique at the base station, we can get k,n The estimated results Substitute it into the uplink channel state information expression to obtain the uplink channel estimation result Thereby, the uplink historical channel state information described in step 2 is obtained.

[0027] Furthermore, in step 2, a signal transmission model of the FD-ISAC system is constructed, and the downlink signal received by the user is expressed as:

[0028]

[0029] in, represents the downlink channel state information between the lth user and the BS, where represents the transmit antenna gain, γ l,n represents the downlink communication channel coefficient, represents other downlink user signals other than l, f l′ represents the downlink beamforming vector related to other users other than l;

[0030] Apply the arrival angle AOA estimation technology at the base station to obtain the l,n The estimated results Substitute it into the downlink channel state information expression to obtain the downlink channel estimation result Thereby, the downlink historical channel state information described in step 2 is obtained.

[0031] Furthermore, the non-perceiving radar signal echo is filtered to estimate the angle of the perceived target. Get the echo θ of the perceived radar signal in the nth time slot from the received signal 0,n (t), expressed as

[0032]

[0033] in represents the receive beamforming vector for spatial filtering, The existing angle of arrival (AoA) estimation technology is used to estimate θ 0,n Estimate of; Since the receiving beamforming vector of spatial filtering is orthogonal, means the mean is zero and the variance is Then we use the classic matched filtering method to obtain the estimated downlink communication user delay and Doppler shift

[0034]

[0035] Where, ΔT d is the length of time to receive the signal; according to the estimated value and Assume that the remaining interference is ideally eliminated; thus derive the perceived radar signal echo for

[0036]

[0037] Among them G m is the matched filter gain, Show obedience The noise of the distribution has a variance of f 0,n represents the beamforming vector associated with the sensing target;

[0038] definition and Respectively represent τ 0,n and ν 0,n The estimation errors are related to the signal-to-noise ratio (SNR), specifically:

[0039]

[0040] where ρ τ and ρ υ is a constant determined by a specific system, represents the variance of the noise in the perceived radar signal echo, μ 0,n Represents the Doppler parameter in the perceived radar signal echo.

[0041] Furthermore, the uplink user rate, downlink user rate and Cramer-Rao lower bound (CRB) are used to evaluate the CAttn-GRU network performance, and the SINR corresponding to the uplink user k is expressed as:

[0042]

[0043] Q represents the covariance matrix of the downlink ISAC signal, CQC H I represents the interference caused to the uplink communication user k due to the presence of the downlink signal; Nr Indicates size N r ×N r The identity matrix of represents the variance of the noise in the received uplink signal; at this time, the uplink user rate is

[0044] The signal to interference and noise ratio SINR of downlink user l is expressed as:

[0045]

[0046] F0 represents the radar signal covariance, Represents the variance of the noise in the received downlink signal;

[0047] At this time, the downlink user rate is

[0048] Calculate the distance d of the perceived target based on the Fish information matrix 0,n and angle θ 0,n The Cramer-Rao lower bound of is as follows:

[0049]

[0050] Where c is the speed of light, Represents the noise variance contained in the perceived radar echo signal.

[0051] Furthermore, the CAttn-GRU network is trained using a multi-objective loss function:

[0052]

[0053] Among them, λ1, λ2 and λ3 represent the weights of the uplink user rate constraint, downlink user rate constraint and power constraint respectively, R UL Indicates the target value of the uplink user rate (artificially set constant), R DL Indicates the target value of the downlink user rate (artificially set constant).

[0054] Beneficial effects: In order to obtain higher spectrum efficiency, the base station is allowed to act as both a radar transceiver and a communication transceiver. The BS receives and transmits signals from multiple uplink users and downlink users, reusing the same time and frequency resources. The downlink transmission signal is an ISAC signal, which is used both to transmit information to downlink users and to perform sensing tasks for point target detection. The BS also simultaneously receives uplink communication signals and processes radar echo signals. However, in this setting, not only is there interference between the sensing and communication functions, but there is also coupling between uplink and downlink transmissions, which greatly increases the complexity of the communication and perception ISAC design. By adopting SIC technology for the ISAC system, it is assumed that the SI at the BS, or more precisely, the direct signal coupling link between transceivers is suppressed to an acceptable level.

[0055] In addition, in the traditional paradigm, effective predictive beamforming schemes require high computational complexity due to the need to continuously track the motion parameters of dynamic targets. To address this challenge, a deep learning (DL)-based framework is introduced. Full-duplex radar communication systems based on deep learning show great potential in signal processing and beamforming. Deep learning algorithms can automatically extract deep features of signals and improve the accuracy of target recognition and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 It is a schematic diagram of uplink and downlink signals;

[0058] Figure 3 This is the CAttn-GRU network structure diagram. DETAILED DESCRIPTION

[0059] A joint transmit beamforming method for a full-duplex radar communication system based on deep learning comprises the following steps:

[0060] Step 1: Build an ISAC system

[0061] Consider an ISAC system with two dual-function full-duplex base stations FD BS equipped with two uniform linear arrays (ULAs) to receive communication signals from K single-antenna uplink users and send downlink ISAC signals through the same time-frequency resources. t The downlink ISAC signal transmitted by the unit uniform linear array ULA communicates with L single antenna downlink users simultaneously to sense and detect dynamic targets. The radar echo signal and the uplink communication signal are transmitted through a r The receiving uniform linear array ULA of each unit is received at the BS end.

[0062] Step 2, full-duplex signal model, such as Figure 2 As shown, including uplink communication and downlink communication

[0063] Consider the uplink signal, such as Figure 1 As shown, the base station receiving signal includes the user uplink communication signal passing through the channel, the base station downlink communication signal echo, the base station sensing signal echo, the undesired interference and the noise. The calculation formula of the base station receiving signal is:

[0064]

[0065] in, represents the antenna gain, γ k,n represents the uplink communication channel coefficient, γ l,n represents the downlink communication channel coefficient, γ l＝0,nrepresents the downlink communication channel coefficient of the sensing target, τ 0,n and ν 0,n Respectively represent the time delay and Doppler frequency shift of the perceived target, τ l,n and ν l,n They represent the time delay and Doppler shift of downlink communication users respectively. represents the beamforming vector associated with downlink user l, l∈{1,…,L}, Represents the unit power data symbol of user l, that is, here, Represents the covariance matrix The dedicated radar signal is used to extend the DoF of the transmitted signal x to achieve enhanced perception performance. Assuming that the unit power signal of user l and dedicated radar signals Independent of each other, Indicates the unit power data symbol of uplink communication user k, a t (θ l,n ) represents the steering vector of the transmitting array in the direction of the downlink communication user l, a r (θ l,n ) represents the steering vector of the receiving array in the direction of downlink user l, a r (θ 0,n ) represents the steering vector of the receiving array to the perceived target, z represents the undesired interference, and n represents the noise.

[0066] In the above formula, the downlink beamforming is designed by Once F0 is determined, a dedicated radar signal can be generated. where a t (θ k,n ) represents the guiding vector of the transmitting array in the θ direction, which can be written as:

[0067]

[0068] a r (θ k,n ) represents the steering vector of the receiving array in the θ direction, which can be written as:

[0069]

[0070] Among them, N t Indicates the number of base station transmitting antennas, N r Indicates the number of base station receiving antennas.

[0071] Assume that there are I signal-correlated uncorrelated interference sources, located at angles and θ i,n ≠θ 0,n, these interferences will also reflect the sensor signal to the BS, generating undesired interference z, the calculation formula of z is as follows:

[0072]

[0073] In addition, considering the total transmit power constraint Where P max represents the maximum available power budget of the BS. When the FD BS sends x, it receives the uplink communication signal and the target reflection at the same time.

[0074] Considering the downlink signal of the full-duplex radar, it consists of four parts: the base station downlink communication signal, the signal interference between users, the user-received dedicated radar sensing signal, and the noise. The downlink signal received by the user is

[0075]

[0076] in, represents the downlink channel between the lth user and the BS, where represents the transmit antenna gain, γ l,n Represents the downlink communication channel coefficient.

[0077] Step 3: Parameter estimation using traditional channel estimation methods:

[0078] By using spatial filtering technology, the echo of the point target radar in the nth time slot can be obtained from the reflected signal echo, which can be expressed as

[0079]

[0080] in represents the receive beamforming vector for spatial filtering, is the angle of arrival (AoA) estimation technique used to estimate θ 0,n To simplify the formula, we assume item It means the mean is zero and the variance is σ 2 Then we use the classic matched filtering method to obtain the estimated time delay τ 0,n and Doppler shift ν 0,n .

[0081]

[0082] Where, ΔT d is the length of time it takes to receive the signal. Based on the estimated value and Assume that the remaining interference can be ideally eliminated. From this, the angle θ of the perceived target can be deduced 0,nThe measurement model is

[0083]

[0084] G m is the matched filter gain, Show obedience The noise of the distribution has a variance of

[0085] and Respectively represent τ 0,n and ν 0,n These errors are related to the signal-to-noise ratio (SNR), specifically:

[0086]

[0087] where ρ τ and ρ υ is a constant determined by the specific system, and the beamforming vector f 0,n Noise Variance and Therefore, we use a deep learning-based method to predict state evolution and optimize the beamforming matrix f 0,n , thereby improving the estimation accuracy.

[0088] Step 4: System performance evaluation

[0089] The uplink user rate, downlink user rate and Cramé-Rao lower bound (CRB) are used to evaluate the system performance. The signal to interference plus noise ratio (SINR) corresponding to uplink user k is expressed as:

[0090]

[0091] in, represents the uplink channel between the kth user and the BS.

[0092] At this time, the uplink user rate is

[0093] The signal to interference and noise ratio SINR of downlink user l is expressed as:

[0094]

[0095] At this time, the downlink user rate is

[0096] The Cramer-Rao lower bound (CRB) is widely used to characterize the estimation error. The smaller its value, the higher the estimation accuracy. The distance d to the perceived target is calculated based on the Fish information matrix. 0,n and angle θ0,n The Cramer-Rao lower bound of is as follows:

[0097]

[0098] Where c represents the speed of light.

[0099] The uplink user rate is

[0100] Downlink user rate is

[0101] Step 5: Objective function construction and optimization problem solving

[0102] The loss function is to transform the constrained optimization problem into an equivalent unconstrained problem using a penalty method, so that we can use DL technology to solve the resulting unconstrained problem, which is expressed as

[0103]

[0104] Among them, CRB(d 0,n ,f l,n ) and CRB(θ 0,n ,f l,n ) is responsible for ensuring that the output of the neural network can provide a sufficiently low perception performance with an error lower bound, while providing uplink and downlink communication rates that exceed the threshold.

[0105] The constructed CAttn-GRU network includes a computational attention module, a gated recurrent unit module, and a fully connected network; Figure 3 As shown, each channel historical state information data set is input into the computational attention module in the designed network (CAttn-GRU net) to obtain the attention modeling channel state information; each attention modeling channel state information is input into the gated recurrent unit module in the designed network to obtain the global historical channel state information; the global historical channel state information is sent to the fully connected network to predict the radar signal covariance matrix and the communication signal beamforming matrix; the radar signal covariance matrix and the communication signal beamforming matrix will be used to assist the perception signal and beamforming design at the current moment;

[0106] In the training process of CAttn-GRU network, a multi-objective optimization loss function is used to ensure that the uplink and downlink rates and transmission power meet the requirements while ensuring the perception accuracy. Through the deep learning-based method, the beamforming matrix is ​​optimized to improve the system estimation accuracy.

[0107] Step 6: Channel state estimation and signal processing

[0108] The neural network is as follows: Based on the full-duplex signal model, the historical information of the uplink and downlink channel states of each communication user and the sensing target in each time slot is estimated to obtain the uplink channel information estimation result And the downlink channel information estimation result Where λ is the length of historical information, which determines how much historical channel estimation information the designed neural network needs to obtain.

[0109] Taking the estimated uplink and downlink channel state history information as input, the designed neural network attention module A i Get attention modeling channel state information Η i , where i∈[1,λ]. The formula is as follows:

[0110]

[0111] Among them, d n Representation vector The size of , softmax(·) represents the flexible maximum transfer function.

[0112] After the previous step, the attention channel state information Η in each historical time slot can be obtained i , which will be sent to the network composed of λ gated recurrent units, and then the global historical channel state information C is obtained. For the tth gated recurrent unit, its input signal, input hidden state, reset gate and input hidden state are Η t h t-1 r t z t h t :The formula is as follows:

[0113]

[0114] Among them, W z W r W h is the model parameter, [·,·] represents matrix operation, σ(·) represents the S-type growth function, and tanh(·) is the hyperbolic tangent function. i After passing through the gated cyclic unit group, the global historical channel state information C is obtained, and then the required radar signal covariance F0 and communication signal beamforming matrix f are output through a layer of fully connected network.

[0115] The present invention designs a cross-attention-gated recurrent unit network (CAttn-GRU net), takes the historical state information data set of each channel as input, and sends the global historical channel state information to the fully connected network to predict the radar signal covariance matrix and the communication signal beamforming matrix, which are used to assist the current moment perception signal and beamforming design respectively; during the training process, a multi-objective optimization loss function is used to ensure that the uplink and downlink rates and transmission power meet the requirements while ensuring the perception accuracy. In terms of detection performance, higher transmission power improves the echo reception SNR, and the CRB value decreases with the increase in the number of antennas, which can achieve good perception performance.

Claims

1. A joint transmit beamforming method for a full-duplex radar communication system based on deep learning, characterized in that: The following steps are involved: Step 1: The full-duplex base station receives the echo signal and estimates the angle of the user in each time slot using a traditional channel estimation method; Step 2, constructing a sending and receiving signal model of the FD-ISAC system, and calculating the uplink historical channel state information and the downlink historical channel state information according to the angle of the user in each time slot; Step 3, construct the CAttn-GRU network, including the computational attention module, the gated recurrent unit module, and the fully connected network; The estimated uplink historical channel state information and downlink historical channel state information are input into the computational attention module in the CAttn-GRU network to obtain the attention modeling channel state information; the attention modeling channel state information is input into the gated recurrent unit module to obtain the global historical channel state information; the global historical channel state information is sent to the fully connected network to predict the radar signal covariance matrix and the communication signal beamforming matrix; The radar signal covariance matrix and the communication signal beamforming matrix are used to assist in the current moment perception signal and beamforming design.

2. According to claim 1, a deep learning-based full-duplex radar communication system joint transmit beamforming method is characterized in that: In step 2, the receiving signal model of the FD-ISAC system is constructed, and the calculation formula of the base station receiving signal is expressed as follows: in, represents the antenna gain, γ k,n represents the uplink communication channel coefficient, γ l,n represents the downlink communication channel coefficient, γ l＝0,n represents the downlink communication channel coefficient of the sensing target, τ 0,n and ν 0,n Respectively represent the time delay and Doppler frequency shift of the perceived target, τ l,n and ν l,n They represent the time delay and Doppler shift of the downlink communication user respectively; represents the beamforming vector associated with downlink user l; l∈{1,…,L}, Represents the unit power data symbol of user l, that is, Represents the covariance matrix dedicated radar signal; Assuming the unit power signal of user l and dedicated radar signals Independent of each other, Indicates the unit power data symbol of uplink communication user k, a t (θ l,n ) represents the steering vector of the transmitting array in the direction of the downlink communication user l, a r (θ l,n ) represents the steering vector of the receiving array in the direction of downlink user l, a r (θ 0,n ) represents the steering vector of the receiving array to the perceived target, z represents undesired interference, and n represents noise; In the above formula, downlink beamforming is designed by and F0; once F0 is determined, a dedicated radar signal is generated where a t (θ k,n ) represents the guiding vector of the transmitting array in the θ direction, which can be written as: a r (θ k,n ) represents the steering vector of the receiving array in the θ direction, which can be written as: Among them, N t Indicates the number of base station transmitting antennas, N r Indicates the number of base station receiving antennas; Assume that there are I signal-independent interference sources, located at angles and θ i,n ≠θ 0,n , these interferences will also reflect the sensor signal to the BS, generating undesired interference z, the calculation formula of z is as follows: In addition, considering the total transmit power constraint Where P max represents the maximum available power budget of the BS; definition represents the uplink channel state information between the kth user and the BS; Applying the arrival angle AOA estimation technique at the base station, we can get k,n The estimated results Substitute it into the uplink channel state information expression to obtain the uplink channel estimation result Thereby, the uplink historical channel state information described in step 2 is obtained.

3. According to claim 2, a deep learning-based full-duplex radar communication system joint transmit beamforming method is characterized in that: In step 2, the signal transmission model of the FD-ISAC system is constructed, and the downlink signal received by the user is expressed as: in, represents the downlink channel state information between the lth user and the BS, where represents the transmit antenna gain, γ l,n represents the downlink communication channel coefficient, represents other downlink user signals other than l, f l′ represents the downlink beamforming vector related to other users other than l; Apply the arrival angle AOA estimation technology at the base station to obtain the l,n The estimated results Substitute it into the downlink channel state information expression to obtain the downlink channel estimation result Thereby, the downlink historical channel state information described in step 2 is obtained.

4. According to claim 3, a deep learning-based full-duplex radar communication system joint transmit beamforming method is characterized in that: Filter the non-perceiving radar signal echo and estimate the angle of the perceived target Get the echo of the perceived radar signal in the nth time slot from the received signal Expressed as in represents the receive beamforming vector for spatial filtering, The existing angle of arrival (AoA) estimation technology is used to estimate θ 0,n Estimate of; Since the receiving beamforming vector of spatial filtering is orthogonal, means the mean is zero and the variance is Then we use the classic matched filtering method to obtain the estimated downlink communication user delay and Doppler shift Where, ΔT d is the length of time to receive the signal; according to the estimated value and Assume that the remaining interference is ideally eliminated; thus derive the perceived radar signal echo for Among them G m is the matched filter gain, Show obedience The noise of the distribution has a variance of f 0,n represents the beamforming vector associated with the sensing target; definition and Respectively represent τ 0,n and ν 0,n The estimation errors are related to the signal-to-noise ratio (SNR), specifically: where ρ τ and ρ υ is a constant determined by a specific system, represents the variance of the noise in the perceived radar signal echo, μ 0,n Represents the Doppler parameter in the perceived radar signal echo.

5. According to claim 4, a deep learning-based full-duplex radar communication system joint transmit beamforming method is characterized in that: The uplink user rate, downlink user rate and Cramer-Rao lower bound (CRB) are used to evaluate the CAttn-GRU network performance. The SINR corresponding to the uplink user k is expressed as: Q represents the covariance matrix of the downlink ISAC signal, CQC H Indicates the interference caused to uplink communication due to the presence of downlink signals; Indicates size N r ×N r The identity matrix of Represents the variance of the noise in the received uplink signal; At this time, the uplink user rate is The downlink user l signal to interference and noise ratio SINR is expressed as: F0 represents the radar signal covariance, Represents the variance of the noise in the received downlink signal; At this time, the downlink user rate is Calculate the distance d of the perceived target based on the Fish information matrix 0,n and angle θ 0,n The Cramer-Rao lower bound of is as follows: Where c is the speed of light, Represents the noise variance contained in the perceived radar echo signal.

6. The method for joint transmit beamforming of a full-duplex radar communication system based on deep learning according to claim 5, characterized in that: The CAttn-GRU network is trained using a multi-objective loss function: Among them, λ1, λ2 and λ3 represent the weights of the uplink user rate constraint, downlink user rate constraint and power constraint respectively, R UL Indicates the uplink user rate target value, R DL Indicates the target value of the downlink user rate.

7. The method for joint transmit beamforming of a full-duplex radar communication system based on deep learning according to claim 2 or 3, characterized in that: Based on the full-duplex signal model, the historical information of the uplink and downlink channel states of each communication user and the sensing target in each time slot is estimated to obtain the uplink channel information estimation result And the downlink channel information estimation result Where λ is the length of historical information, which determines how much historical channel estimation information the designed neural network needs to obtain; Taking the estimated uplink and downlink channel state history information as input, the CAttn-GRU neural network cross-attention module A i Get attention modeling channel state information Η i , where i∈[1,λ]; the formula is as follows: Among them, d n Representation vector The size of , softmax(·) represents the flexible maximum transfer function; After the previous step, the attention channel state information Η in each historical time slot is obtained i , which will be sent to the network composed of λ gated recurrent units, and then the global historical channel state information C is obtained; for the tth gated recurrent unit H t , its input signal h t-1 , input hidden state r t , reset gate z t and the input hidden state h t They are as follows: Among them, W z W r W h are model parameters, [·,·] represents matrix operation, σ(·) represents S-type growth function, tanh(·) is hyperbolic tangent function; H i After passing through the gated cyclic unit group, the global historical channel state information C is obtained, and then the required radar signal covariance F0 and communication signal beamforming matrix f are output through a layer of fully connected network.

Images