Deep learning based isac hybrid beamforming method

Abstract

The application discloses an ISAC hybrid beamforming method based on deep learning, comprising: performing hybrid beamforming on communication and sensing functions respectively, all users in a cell reporting measured information to a base station to obtain downlink channel information, the base station deploying a sensing receiving antenna and measuring a sensing channel when receiving a sensing signal; data splicing and processing are performed on the downlink channel and the sensing channel to obtain a real number form channel matrix, user data is randomly generated and channel CSI is collected as a data set; an ISAC hybrid beamforming neural network model is trained at the base station end based on the data set; after the training at the base station end is completed, a digital precoding matrix and an analog precoding matrix are generated by using the trained neural network model on real-time channel data, and the method can jointly design the digital and analog beamforming matrices of the ISAC scene, is superior to an existing iterative optimization algorithm in terms of communication rate or sensing estimation rate, and has lower time complexity in an inference stage.

Description

Technical Field

[0001] This invention belongs to the field of integrated sensing technology, specifically relating to an ISAC hybrid beamforming method based on deep learning. Background Technology

[0002] Integrated sensing and communication (ISAC) aims to achieve traditional sensing functions on the basis of wireless communication with minimal or no additional dedicated sensing modules, such as through signal co-design or hardware sharing. In future ISAC applications, extremely high sensing accuracy and communication speed are required. Compared to low- and mid-frequency bands, the 30-300 GHz millimeter-wave band offers a continuous large bandwidth, significantly improving sensing and communication performance. Millimeter-wave signals suffer from severe path loss, necessitating compensation using multiple-input-output (MIMO) technology. Therefore, beamforming technology for MIMO antennas has become a key research focus in ISAC.

[0003] For hardware-sharing-based ISAC systems, existing research has focused on joint beamforming of communication and sensing to optimize their performance as much as possible. However, several challenging issues remain: First, optimizing the global performance of communication and sensing is a difficult problem to solve because it is NP-hard. Most existing studies have chosen suboptimal solutions, such as null-space projection, which projects the sensing signal onto the null space of communication channel interference. This method excessively guarantees communication performance but sacrifices the performance of the sensing system. Other methods try to make the sensing signal as close as possible to the expected beam pattern, using a lower bound on the communication rate as a constraint, but this often results in a loss of communication performance. None of these methods can achieve a balance in global performance. Second, traditional optimization problems require highly complex convex optimization or iterative algorithms, which introduce significant computational delays, especially in large-scale MIMO scenarios with a large number of transmit antennas. Third, in practical communication systems, hybrid beamforming architectures are often used to reduce hardware complexity, and designing analog and digital beamforming matrices is a more complex but also more realistic problem for ISAC scenarios.

[0004] Based on the above analysis, existing ISAC beamforming algorithms still have significant limitations in solving beamforming problems in ISAC systems, such as difficulty in achieving global performance optimization of communication and sensing in integrated sensing scenarios, high computational complexity, and incomplete beamforming matrix design.

[0005] Based on the aforementioned technical problems in existing technologies, this invention proposes a deep learning-based ISAC hybrid beamforming method. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a deep learning-based ISAC hybrid beamforming method.

[0007] The present invention adopts the following technical solution:

[0008] A deep learning-based ISAC hybrid beamforming method includes:

[0009] Step 1: In a communication scenario where a single integrated sensing and communication (ISAC) base station serves multiple downlink users and senses multiple point targets, hybrid beamforming is applied to both communication and sensing functions. All users in the cell report the measured information to the base station to obtain downlink channel information. The base station then obtains the channel measurements reported by all users. ,in, For the number of users, The number of transmitting antennas is used to simultaneously measure the sensing channel. ,in, To determine the number of point targets to be sensed, the base station deploys sensing receiving antennas and measures the sensing channel when receiving sensing signals.

[0010] Step 2: The downlink channel and sensing channel are spliced ​​and processed to obtain a channel matrix in real number form. User data is randomly generated and channel CSI is collected as a dataset.

[0011] Step 3: The base station trains the ISAC hybrid beamforming neural network model based on this dataset: the loss function is defined as:

[0012] ……(1),

[0013] In equation (1), For the loss of communication rate, For the loss of perception rate, As a penalty item, It is the L2 regularization function, used to alleviate overfitting during model training. , and These are the communication rate weighting coefficient, the sensing rate weighting coefficient, and the penalty term weighting coefficient, respectively, where:

[0014] ...(2),

[0015] In equation (2), B represents the communication rate, and B represents the bandwidth.

[0016] ...(3),

[0017] In equation (3), Indicates the rate of perception. It is a constant that depends on the time-domain characteristics of the sensed signal. The variance of the system error during the sensing process;

[0018] ……(4),

[0019] In equation (4), This represents the communication channel from the base station to user k. For communication simulation beamforming matrix, For the digital beamforming matrix corresponding to the symbol of user k, Let k be the variance of the additive white Gaussian noise at user k.

[0020] ……(5),

[0021] In equation (5), This represents the sensing channel from the base station to the sensing target p and the reflected signal. To sense the simulated beamforming matrix, To sense digital beamforming matrix, The variance of the additive white Gaussian noise at the target p;

[0022] Step 4: After the base station completes the training, it uses the trained neural network model to generate digital precoding matrices and analog precoding matrices from the real-time channel data.

[0023] Furthermore, step 1 includes: the base station obtaining channel measurements reported by all users. ,in, For the number of users, Let K be the number of transmit antennas, and let the channel measured by each user k be... The dimension is 1* All channel measurements reported by the user are stitched together in the first dimension. * complex matrix Simultaneously, the sensing channel was measured. ,in, Let the number of point targets be denoted as and the sensing channel reflected by point target p be . The dimension is 1* The sensing channels reflected by all point targets are concatenated according to the first dimension to obtain... * complex matrix .

[0024] Further, step 2 includes:

[0025] Step 2.1, before inputting into the neural network, first... and By directly splicing along the first dimension, we obtain ;

[0026] Step 2.2: Since neural networks typically use real number operations, further extract the complex matrix. The modulus, real part, and imaginary part are concatenated to convert it into a real number form. ,in, , , , Let m and n be the indices of the first and second dimensions of the matrix, respectively, for the training data of the neural network. It is a complex value.

[0027] Furthermore, in step 3, the neural network model includes:

[0028] Channel feature extraction module CSI features used to characterize beamforming design The input is the CSI matrix for communication and sensing, and the output is the CSI features;

[0029] Analog and digital beamforming modules It is used to generate digital and analog beamforming matrices for communication and sensing, where communication and sensing tasks can be performed independently and the analog and digital parts can be decoupled.

[0030] Furthermore, in step 3, the neural network model training employs an unsupervised learning method, and the loss function is defined as:

[0031] ……(1),

[0032] In equation (1), For the loss of communication rate, For the loss of perception rate, As a penalty item, It is the L2 regularization function, used to alleviate overfitting during model training. , and These are the communication rate weighting coefficient, the sensing rate weighting coefficient, and the penalty term weighting coefficient, respectively, where:

[0033] ...(2),

[0034] In equation (2), B represents the communication rate, and B represents the bandwidth.

[0035] ...(3),

[0036] In equation (3), Indicates the rate of perception. It is a constant that depends on the time-domain characteristics of the sensed signal. The variance of the system error during the sensing process;

[0037] ……(4),

[0038] In equation (4), This represents the communication channel from the base station to user k. For communication simulation beamforming matrix, For the digital beamforming matrix corresponding to the symbol of user k, Let k be the variance of the additive white Gaussian noise at user k.

[0039] ……(5),

[0040] In equation (5), This represents the sensing channel from the base station to the sensing target p and the reflected signal. To sense the simulated beamforming matrix, To sense digital beamforming matrix, The variance of the additive white Gaussian noise at the target p is used for perception.

[0041] Furthermore, in step 3, the neural network model training employs an auxiliary optimization method based on the Minimum Mean Square Error (MMSE) algorithm to correct... ; Calculate the effective communication channel:

[0042] ... (6),

[0043] calculate :

[0044] ……(7),

[0045] in:

[0046] ……(8),

[0047] Noise for each user; corrected for:

[0048] ... (9).

[0049] The beneficial effects of this invention are:

[0050] The method described in this invention can jointly design digital and analog beamforming matrices for ISAC scenarios, outperforming existing iterative optimization algorithms in both communication rate and sensing estimation rate, and with lower time complexity in the inference phase. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the ISAC hybrid beamforming method based on deep learning in an embodiment of the present invention.

[0052] Figure 2 This is a schematic diagram of the hybrid beamforming structure in the ISAC scenario of this invention.

[0053] Figure 3 This is a schematic diagram of a parallel beamforming matrix generation network based on an attention mechanism in an ISAC scenario according to an embodiment of the present invention. Detailed Implementation

[0054] To better understand the above-mentioned objectives, features and advantages of the present invention, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present application can be combined with each other.

[0055] Example

[0056] In this application embodiment, a scenario is described where one base station serves multiple downlink users and simultaneously senses multiple targets. The communication and sensing systems are deployed separately and independently, sharing the same time-frequency resources. (Refer to...) Figure 2 The base station is equipped with two antenna panels, each using... and It means that, among them, It is used to transmit downlink communication signals, and also to transmit sensing signals to the sensing target in order to support sensing tasks. To receive the echo signal reflected from the sensing target, this embodiment employs a hybrid beamforming strategy to generate a directional beam in an economical and efficient manner. (Transmitting antenna panel) Each is equipped with One and One radio frequency chain is used for downlink communication and target sensing, and the radio frequency signals of downlink communication and target sensing are superimposed at the antenna. and These are digital and analog beamforming matrices, respectively. and Definition and These are sensing digital and analog beamforming matrices, respectively. and Simulated beamforming uses a phase shifter fully connected scheme, where all RF chains are deflected by a certain phase and connected to all antennas. Therefore, the matrix... and Each element in the table is a weight of the phase shifter, represented as follows: and Its modulus is always 1, only its phase can change, and it has In addition, assuming the system applies downlink power constraints to the hybrid beamforming matrix, ensuring equal transmit power for each user, power allocation need not be considered in subsequent beamforming matrix design. ,in, The maximum power allocated to each downlink user by the system; ,in, The maximum power allocated by the system for sensing;

[0057] like Figure 1 As shown, the deep learning-based ISAC hybrid beamforming method includes:

[0058] Step 1: The base station obtains channel measurements reported by all users. Simultaneously, the sensing channel was measured. ;

[0059] Step 2: The downlink channel and sensing channel are concatenated and processed to obtain a channel matrix in real number form;

[0060] Step 2.1, before inputting into the neural network, first... and By directly splicing along the first dimension, we obtain ;

[0061] Step 2.2: Since neural networks typically use real number operations, extract the complex matrix. The modulus, real part, and imaginary part are concatenated to convert it into a real number form. ,in, , , , This refers to the training data for the neural network;

[0062] Step 3, at the base station end, use To train a hybrid beamforming neural network for the ISAC scene, the neural network model structure is as follows: Figure 3 As shown.

[0063] Step 4: After the base station completes the training, it uses the same data preprocessing method and the trained neural network model to generate digital precoding matrices and analog precoding matrices for the real-time collected communication and sensing channel data.

[0064] In step 3 of the above embodiment, the neural network is... , It consists of two parts, of which:

[0065] Channel feature extraction module CSI key features used to characterize complex beamforming designs; its input is the CSI matrix for communication and sensing, and its output is CSI features, channel feature extraction module. Attention mechanisms were used as the basic units for feature extraction, and residual attention networks were used as... In addition to short-circuit connections, a normalization layer is added after each layer of attention operations to speed up training by normalizing the data to a standard normal distribution.

[0066] Analog and digital beamforming modules Digital and analog beamforming matrices are used to generate communication and sensing matrices, where communication and sensing tasks can be performed independently, and the analog and digital components can be decoupled. The output CSI characteristics are transmitted in parallel to the module. , , and In, generate respectively , , and ;

[0067] First Simulated Beamforming Generation Module The task is to generate ,Depend on The output features are processed by RFBlock. The RFBlock first has a planarization layer that flattens the feature matrix into a vector, and then... There are 3 parallel branches, where each branch consists of two fully connected layers, and the output dimensions of the two fully connected layers are respectively... and Furthermore, batch normalization and... were added between the two fully connected layers. Activation functions are used to improve training efficiency. The output value of each branch represents a phase shift vector from an RF chain to all antennas. By connecting the outputs of these branches in series, we obtain , representing the phase of the simulated beamforming matrix, next, After function This transforms the phase into real and imaginary components, thus producing the final... :

[0068] ,

[0069] Second simulated beamforming generation module for sensing tasks Structure and Similar, the difference is that the number of branches becomes Finally generated ;

[0070] Digital beamforming generation module and First Digital Beamforming Generation Module In generating communication digital beamforming matrix hour, The output features first pass through the BB Block. The main structure is a convolutional network. The feature matrix is ​​first passed to two consecutive convolutional layers, and the kernel size of the two convolutional layers is [missing value]. The number of convolution kernels are 64 and 64 respectively. Each layer is followed by batch normalization and... The activation function is used to improve training efficiency. The output of the second convolutional layer is flattened into a vector through a planarization layer, and finally passed through a fully connected layer. For communication tasks, the output size of the fully connected layer is... ,in, This is the dimension of the communication digital beamforming matrix, where 2 represents the real and imaginary parts of the digital beamforming matrix. The BB Block output is reconstructed as... Finally, regarding To normalize the power, we have:

[0071] ,

[0072] For the second digital beamforming generation module used for sensing tasks The output size of the second fully connected layer of the BB Block is adjusted to ,in, To perceive the dimension of the digital beamforming matrix, 2 represents the real and imaginary parts of the digital beamforming matrix;

[0073] During model training, an unsupervised learning method is used, and the loss function is defined as:

[0074] ,

[0075] In the above formula, For the loss of communication rate, For the loss of perception rate, As a penalty item, It is the L2 regularization function, used to alleviate overfitting during model training. , and These are the communication rate weighting coefficient, the sensing rate weighting coefficient, and the penalty term weighting coefficient, respectively, where:

[0076] ,

[0077] ,

[0078] B represents the communication rate, and B represents the bandwidth. Indicates the rate of perception. It is a constant that depends on the time-domain characteristics of the sensed signal. Let be the variance of the system error during the sensing process, where:

[0079] ,

[0080] ,

[0081] To alleviate overfitting during model training, the common L2 regularization function is selected. , These are all the model parameters that participated in the training;

[0082] For communication tasks, to avoid getting stuck in local optima during the early stages of training, an auxiliary optimization method based on the Minimum Mean Square Error (MMSE) algorithm is used to correct this. First, calculate the effective communication channel:

[0083] ,

[0084] Next calculation The formula is as follows:

[0085] ,

[0086] in:

[0087] ,

[0088] Noise for each user, after correction for:

[0089] .

[0090] This invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims.

Claims

1. A deep learning based ISAC hybrid beamforming method, characterized in that, include: Step 1: In a communication scenario where a single integrated sensing and communication (ISAC) base station serves multiple downlink users and senses multiple point targets, hybrid beamforming is applied to both communication and sensing functions. All users in the cell report the measured information to the base station to obtain downlink channel information. The base station then obtains the channel measurements reported by all users. ,in, For the number of users, The number of transmitting antennas is used to simultaneously measure the sensing channel. ,in, To determine the number of point targets to be sensed, the base station deploys sensing receiving antennas and measures the sensing channel when receiving sensing signals. Step 2: The downlink channel and sensing channel are spliced ​​and processed to obtain a channel matrix in real number form. User data is randomly generated and channel CSI is collected as a dataset. Step 3: The base station trains the ISAC hybrid beamforming neural network model based on this dataset: the loss function is defined as: ……（1）， In equation (1), For the loss of communication rate, For the loss of perception rate, As a penalty item, It is the L2 regularization function, used to alleviate overfitting during model training. , and These are the communication rate weighting coefficient, the sensing rate weighting coefficient, and the penalty term weighting coefficient, respectively, where: ……（2）， In formula (2), denotes the communication rate, B is the bandwidth; ……（3）， In formula (3), represents the perceptual rate, is a constant depending on the time-domain characteristics of the perceptual signal, is the variance of the system error in the perceptual process; ……（4）， In formula (4), denotes a communication channel from the base station to user k, is a communication analog beamforming matrix, is a digital beamforming matrix for the symbols of user k, is the variance of the additive white Gaussian noise at user k; ……（5）， In equation (5), This represents the sensing channel from the base station to the sensing target p and the reflected signal. To sense the simulated beamforming matrix, To sense digital beamforming matrix, The variance of the additive white Gaussian noise at the target p; Step 4: After the base station completes the training, it uses the trained neural network model to generate digital precoding matrices and analog precoding matrices from the real-time channel data.

2. The deep learning based ISAC hybrid beamforming method of claim 1, wherein, Step 2 includes: Step 2.1, concatenate along the first dimension directly and to get ; Step 2.2, Extract the complex matrix The modulus, real part, and imaginary part are concatenated to convert it into a real number form. ,in, , , , Let m and n be the indices of the first and second dimensions of the matrix, respectively, for the training data of the neural network. It is a complex value.

3. The deep learning based ISAC hybrid beamforming method of claim 2, wherein, In step 3, the neural network model includes: Channel feature extraction module , for characterizing the beamforming design, The input is the communication and perception CSI matrix, and the output is the CSI feature; Analog and digital beamforming modules for generating communication and sensing digital and analog beamforming matrices, where communication and sensing tasks can be performed separately, analog and digital parts can be decoupled.

4. The deep learning based ISAC hybrid beamforming method of claim 3, wherein, In step 3, the model training employs an auxiliary optimization method based on the Minimum Mean Square Error (MMSE) algorithm to correct... ; Calculate the effective communication channel: ……（6）， Computing : ……（7）， in: ……（8）， represents the noise for each user; the corrected is: ……（9）。

Images