Communication and perception integrated prediction beam forming method
By predicting the angle of the sensed target using a pre-trained SCTMNet model and generating a transmit beamforming vector, the problems of signal overhead and delay in existing technologies are solved, achieving efficient beam tracking and accurate sensed target recognition, which is suitable for 6G communication systems.
Patent Information
- Application Number
- CN202610074645.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-20
- Publication Date
- 2026-04-28
AI Technical Summary
Existing integrated communication and sensing technologies rely on periodic pilot signals, which leads to signal overhead and delay. The beam prediction accuracy is limited, making it impossible to achieve accurate real-time tracking. Furthermore, the generalization ability and practical deployment of high-performance models are constrained.
A pre-trained SCTMNet model is used to predict the angle of the sensed target based on historical data, generate the transmit beamforming vector, separate the echo signal through the orthogonality of the guide vector, and use the SCTMNet model for beam prediction to reduce computational complexity and signaling overhead.
It achieves low-latency, high-precision beam tracking, improves downlink signal reception quality, is suitable for deployment in base stations with limited computing resources, and supports practical applications of 6G communication systems.
Smart Images
Figure CN121940769A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of next-generation information technology, and in particular to a predictive beamforming method that integrates communication and sensing. Background Technology
[0002] The sixth generation of mobile communication (6G) represents the next stage of evolution in wireless communication. Unlike previous generations, the core services of 6G will no longer revolve solely around traditional user-centric communication, but will gradually expand to a wider range of scenarios. This transformation requires wireless networks to not only have efficient information transmission capabilities, but also support multi-dimensional functions such as environmental awareness and data acquisition. Against this backdrop, Integrated Sensing and Communication (ISAC) has become one of the key technologies of 6G networks. By integrating wireless communication and radar sensing functions on a single platform, ISAC technology makes full use of shared hardware and software resources to achieve mutual enhancement between communication and sensing. In wireless communication networks, ISAC allows base stations to sense real-time data such as the location and speed of the target while providing high-speed communication. This helps improve the network's ability to understand the environment, thereby supporting more efficient resource scheduling and more reliable intelligent application services. At the same time, the combination of millimeter wave (mmWave) and massive multiple-input multiple-output (mMIMO) technologies provides ISAC systems with huge bandwidth and high spatial resolution, thus enabling accurate target tracking and high-speed data transmission. However, existing technologies suffer from the following problems: relying on periodic pilot signals results in significant signal overhead and latency; at the same time, the beam prediction accuracy is limited due to the difficulty in accurately describing the complex nonlinear motion trajectory of the perceived target, making it impossible to achieve accurate real-time tracking; most high-performance communication and sensing integrated deep learning models require training with large-scale and diverse labeled data, while in practical applications, building a comprehensive dataset covering all possible scenarios for each base station is costly and difficult to achieve, which restricts the generalization ability and practical deployment of the model. Summary of the Invention
[0003] This invention provides a predictive beamforming method integrating communication and sensing to overcome the problems of significant signal overhead and delay caused by relying on periodic pilots, which makes it impossible to achieve accurate real-time tracking. Furthermore, the high cost and difficulty in implementation of this method restrict the generalization ability and practical deployment of the model.
[0004] To achieve the above objectives, the technical solution of the present invention is as follows: A predictive beamforming method integrating communication and sensing, comprising: S1. The base station uses the echo signal of a single sensing target in historical data as the input signal and obtains the predicted angle of the sensing target relative to the base station at the current moment through a pre-trained SCTMNet model; the predicted angle refers to the angle of arrival of the sensing target relative to the base station at the current moment. S2. Based on the predicted angle between the base station and the sensing target at the current moment, generate a transmit beamforming vector; through the transmit beamforming vector, perform weighted fusion of the integrated communication and sensing signals to be transmitted to obtain a weighted integrated communication and sensing signal, i.e., the target beam signal, and transmit it to the area where the target sensing target is located. S3. All sensing targets receive the weighted communication and sensing integrated signal and return the echo signal of all sensing targets. S4. Based on the echo signals of all sensed targets, the base station sets up a processing mechanism based on the orthogonality of the steering vector to separate and obtain the echo signal of a single sensed target at the current moment. This signal is used as a new input signal. Through the pre-trained SCTMNet model, the predicted angle between the base station and the sensed target at the next moment is obtained, and then the target beam signal at the next moment is obtained.
[0005] Furthermore, using the pre-trained SCTMNet model, the expression for the predicted angle of the perceived target at the current moment is obtained as follows:
[0006] In the formula, For prediction angle; The echo signal of the z-th sensing target in the n-th time slot; This is the mapping function, i.e., the SCTMNet model.
[0007] Furthermore, the SCTMNet model includes a preprocessing layer, a signal embedding layer, an attention weighting layer, a feature extraction layer, and an output layer; The preprocessing layer is used to recover the echo signal of a single sensing target to obtain the recovered echo signal; the recovered echo signal is split into real part and imaginary part by two cascaded fully connected layers; real-valued features are obtained by passing the real part and imaginary part of the recovered echo signal through a fully connected layer. The signal embedding layer is used to perform preliminary feature extraction on real-valued features through three cascaded convolutional layers and a batch normalization layer to obtain initial features; The attention weighting layer is a multi-branch channel attention network used to weight the initial features to obtain multiple attention-weighted features; The feature extraction layer is used to extract deep features from each attention-weighted feature to obtain deep features; The output layer is used to add and fuse deep features to obtain the predicted angle.
[0008] Furthermore, the feature extraction layer includes a convolutional Mamba network module, a convolutional downsampling module, a standard convolutional layer, a first convolutional Transformer network, and a second convolutional Transformer network; The convolutional Mamba network module is used to perform multi-path parallel processing on attention-weighted features to obtain local-global-sequence fusion features; The convolutional downsampling module is used to perform downsampling operations on the local-global-sequence fusion features to obtain downsampled local-global-sequence fusion features; The first convolutional Transformer network is used to extract shallow features from local-global-sequence fusion features to obtain shallow features; The standard convolutional layer is used to normalize and smooth shallow features to obtain refined features; The second convolutional Transformer network is used to extract features from the refined features to obtain deep features.
[0009] Furthermore, the convolutional Transformer network includes depthwise separable convolutions, multi-head attention layers, gating layers, layer normalization layers, feedforward neural network layers, and an output layer; The depthwise separable convolution is used to extract local features from the downsampled local-global-sequence fusion features to obtain local spatial features; The multi-head attention layer is used to calculate the relationships between different locations in the local spatial features to obtain attention features; The gating layer is used to multiply local features and attention features element-wise to obtain gated fusion features; The normalization layer is used to normalize the gated fusion features to obtain normalized features. The feedforward neural network layer is used to perform a nonlinear transformation on the normalized features to obtain the output features of the feedforward neural network. The output layer is used to multiply the normalized features and the output features of the feedforward neural network element by element to obtain the output features of the convolutional Transformer network, namely shallow features or deep features.
[0010] Furthermore, the weighted communication sensing integrated signal is:
[0011] In the formula, For weighted communication sensing integrated signal; For integrated communication and sensing signals; Here is the downlink beamforming matrix, where, Z represents the transmitting antenna, and Z represents the total number of targets to be sensed. The beamforming vector is expressed as:
[0012] In the formula, For beamforming functions; For transmitting antenna; This is the phase offset applied to the m-th antenna.
[0013] Furthermore, the echo signal of all the sensed targets is:
[0014] In the formula, The echo signal of all perceived targets; Let U be the gain coefficient of the sensing array, where U is the number of receiving antennas; The reflection coefficient; The received signal is a delayed version of the transmitted signal; and The first is used for perception. Doppler shift and time delay of the target being sensed; Additive white Gaussian noise with zero mean; In the first The first time slot The actual angle of the sensing target relative to the base station; This is the launch guidance vector; To receive the guide vector; The expressions for the transmit steering vector and the receive steering vector are as follows: ; Based on the echo signal and the orthogonality of the steering vector, the expression for the echo signal of a single sensing target at the current moment is obtained as follows: .
[0015] Furthermore, after step S4, by defining the downlink receiving signal module of the sensing target, the receiving signal-to-noise ratio of the sensing target and the reachability and rate of the communication link are derived to evaluate the receiving performance of the sensing target. The downlink receiving signal module for the sensing target is defined as follows:
[0016] In the formula, The gain of the communication link array; For the first Channel fading between the sensing target and the base station; Based on the defined downlink receiving signal module of the sensing target, the received signal-to-noise ratio of the sensing target is obtained as follows:
[0017] In the formula, For beam alignment accuracy, and ; Based on the received signal-to-noise ratio of the sensing target, the reachability rate of the sensing target is obtained as follows:
[0018] In the formula, For reachable and rate.
[0019] Furthermore, the pre-trained SCTMNet model is an SCTMNet model that has been pre-trained and optimized using mean absolute error as the loss function; the loss function is:
[0020] In the formula, and The perceived target in the k-th training sample In the The echo signal of each time slot and its corresponding true angle; N is the total number of consecutive time slots; K is the communication scenario sample.
[0021] Beneficial Effects: This invention provides a predictive beamforming method integrating communication and sensing. By introducing the SCTMNet model, it directly predicts the angle of the sensing target at the current moment based on the echo signal from the previous moment, thereby generating a transmit beam. This injects "predictability" into beamforming, enabling the beam to point in advance to the direction the sensing target is about to arrive, fundamentally solving the beam inaccuracy problem caused by processing delays and rapid movement of the sensing target, and significantly improving the downlink signal reception quality. In addition, the SCTMNet model does not require dedicated pilots and high-precision, low-latency beam tracking with uplink feedback. While significantly improving the communication signal-to-noise ratio and sensing accuracy, it greatly reduces the computational complexity and signaling overhead of the system. It is particularly suitable for deployment on edge devices with limited computing resources, such as base stations, thus providing an efficient and reliable solution for the practical application of 6G communication systems. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart of the predictive beamforming method of the present invention; Figure 2 This is a schematic diagram of the target angle modeling in an embodiment of the present invention; Figure 3 This is a diagram of the SCTMNet structure in an embodiment of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] This embodiment provides a predictive beamforming method integrating communication and sensing, such as... Figure 1 As shown, it includes: S1. The base station uses the echo signal of a single sensed target from historical data as the input signal and obtains the predicted angle of the sensed target relative to the base station at the current moment through a pre-trained Sensing Combined Transformer Mamba Network (SCTMNet) model. The predicted angle refers to the angle of arrival of the sensed target relative to the base station at the current moment. S2. Based on the predicted angle between the base station and the sensing target at the current moment, generate a transmit beamforming vector; through the transmit beamforming vector, perform weighted fusion of the integrated communication and sensing signals to be transmitted to obtain a weighted integrated communication and sensing signal, i.e., the target beam signal, and transmit it to the area where the target sensing target is located. S3. All sensing targets receive the weighted communication and sensing integrated signal and return the echo signal of all sensing targets. S4. Based on the echo signals of all sensed targets, the base station sets up a processing mechanism based on the orthogonality of the steering vector to separate and obtain the echo signal of a single sensed target at the current moment. This signal is used as a new input signal. Through the pre-trained SCTMNet model, the predicted angle between the base station and the sensed target at the next moment is obtained, and then the target beam signal at the next moment is obtained.
[0026] Preferably, the expression for the predicted angle of the perceived target at the current time step, obtained through the pre-trained SCTMNet model, is as follows:
[0027] In the formula, For prediction angle; The echo signal of the z-th sensing target in the n-th time slot; This is the mapping function, i.e., the SCTMNet model.
[0028] Preferred, such as Figure 3 As shown, the SCTMNet model includes a preprocessing layer, a signal embedding layer, an attention weighting layer, a feature extraction layer, and an output layer; The preprocessing layer is used to recover the echo signal of a single sensing target to obtain the recovered echo signal; the recovered echo signal is split into real part and imaginary part by two cascaded fully connected layers; real-valued features are obtained by passing the real part and imaginary part of the recovered echo signal through a fully connected layer. Specifically, the steps to obtain the recovered echo signal are as follows: S111. Perform radar matched filtering on the echo signal of a single sensed target to obtain time delay estimation and Doppler frequency shift estimation; the expression for the radar matched filtering is:
[0029] In the formula, For time delay estimation; For Doppler frequency shift estimation; The duration of each time interval; The echo signal of a single sensing target; It is a conjugate copy of the transmitted signal; To introduce a phase compensation term for Doppler frequency shift; S112. Based on the time delay estimation and Doppler frequency shift estimation, the echo signal of a single sensing target is recovered to obtain the recovered echo signal; the expression for the recovery process is:
[0030] In the formula, To restore processing gain; This is the result obtained after the original noise has undergone the same matched filtering and recovery process as the signal.
[0031] The signal embedding layer is used to perform preliminary feature extraction on real-valued features through three cascaded convolutional layers and a batch normalization layer to obtain initial features; The attention weighting layer is a multi-branch channel attention network used to weight the initial features to obtain multiple attention-weighted features; Specifically, the multi-branch channel attention network has three branches: the first branch is for average pooling calculation, the second branch is for max pooling calculation, and the third branch is for average pooling calculation. Each branch is followed by a convolutional layer. The steps of the multi-branch channel attention network are as follows: multiply the output of the first branch with the output of the second branch element by element, and then multiply the product with the output of the third branch; the multiplication result is passed through a sigmoid activation function to generate channel attention weights in the range of 0 to 1; multiply the channel attention weights with the initial features to obtain attention-weighted features; The feature extraction layer is used to extract deep features from each attention-weighted feature to obtain deep features; The output layer is used to add and fuse deep features to obtain the predicted angle.
[0032] Preferably, the feature extraction layer includes a convolutional Mamba network module, a convolutional downsampling module, a standard convolutional layer, a first convolutional Transformer network, and a second convolutional Transformer network; The convolutional Mamba network module is used to perform multi-path parallel processing on attention-weighted features to obtain local-global-sequence fusion features; The convolutional Mamba network module is a three-way parallel architecture. The first path consists of a standard convolutional layer for obtaining local features; the second path consists of a depthwise separable convolution, two parallel convolutional layers, and a connection layer for obtaining global features; the third path has two sub-paths. The first sub-path captures long sequence dependencies through the Mamba network layer, while the second sub-path extracts features in different frequency domains through wavelet transform, convolutional layers, and inverse wavelet transform. The outputs of the first and second sub-paths are added element-wise to obtain sequence features. The local, global, and sequence features are then integrated to obtain a local-global-sequence fusion feature. The convolutional downsampling module is used to perform downsampling operations on the local-global-sequence fusion features to obtain downsampled local-global-sequence fusion features; The convolutional downsampling module is a dual-branch design. The first branch is a standard convolution to reduce the spatial size of the feature map, and the second branch is a depthwise separable convolution to generate attention weights. The attention weights are added to the feature map with reduced spatial size by residual addition to obtain the downsampled local-global-sequence fusion features. The first convolutional Transformer network is used to extract shallow features from local-global-sequence fusion features to obtain shallow features; The standard convolutional layer is used to normalize and smooth shallow features to obtain refined features; The second convolutional Transformer network is used to extract features from the refined features to obtain deep features.
[0033] Preferably, the convolutional Transformer network includes depthwise separable convolutions, multi-head attention layers, gating layers, layer normalization layers, feedforward neural network layers, and an output layer; The depthwise separable convolution is used to extract local features from the downsampled local-global-sequence fusion features to obtain local spatial features; The multi-head attention layer is used to calculate the relationships between different locations in the local spatial features to obtain attention features; The gating layer is used to multiply local features and attention features element-wise to obtain gated fusion features; The normalization layer is used to normalize the gated fusion features to obtain normalized features. The feedforward neural network layer is used to perform a nonlinear transformation on the normalized features to obtain the output features of the feedforward neural network. The output layer is used to multiply the normalized features and the output features of the feedforward neural network element by element to obtain the output features of the convolutional Transformer network, namely shallow features or deep features.
[0034] Preferably, the weighted communication sensing integrated signal is:
[0035] In the formula, For weighted communication sensing integrated signal; For integrated communication and sensing signals; Here is the downlink beamforming matrix, where, Z represents the transmitting antenna, and Z represents the total number of targets to be sensed. The beamforming vector is expressed as:
[0036] In the formula, For beamforming functions; For transmitting antenna; This is the phase offset applied to the m-th antenna.
[0037] Preferably, the echo signal of all the sensed targets is:
[0038] In the formula, The echo signal of all perceived targets; Let U be the gain coefficient of the sensing array, where U is the number of receiving antennas; The reflection coefficient; The received signal is a delayed version of the transmitted signal; and The first is used for perception. Doppler shift and time delay of the target being sensed; Additive white Gaussian noise with zero mean; In the first The first time slot The actual angle of the sensing target relative to the base station; This is the launch guidance vector; To receive the guide vector; The expressions for the transmit steering vector and the receive steering vector are as follows: ; Based on the echo signal and the orthogonality of the steering vector, the expression for the echo signal of a single sensing target at the current moment is obtained as follows: .
[0039] Preferably, after step S4, by defining the downlink receiving signal module of the sensing target, the receiving signal-to-noise ratio of the sensing target and the reachability and rate of the communication link are derived to evaluate the receiving performance of the sensing target. The downlink receiving signal module for the sensing target is defined as follows:
[0040] In the formula, The gain of the communication link array; For the first Channel fading between the sensing target and the base station; Based on the defined downlink receiving signal module of the sensing target, the received signal-to-noise ratio of the sensing target is obtained as follows:
[0041] In the formula, For beam alignment accuracy, and ; Based on the received signal-to-noise ratio of the sensing target, the reachability rate of the sensing target is obtained as follows:
[0042] In the formula, For reachable and rate.
[0043] Preferably, the pre-trained SCTMNet model is an SCTMNet model that has been pre-trained and optimized using mean absolute error as the loss function; the loss function is:
[0044] In the formula, and The perceived target in the k-th training sample In the The echo signal of each time slot and its corresponding true angle; N is the total number of consecutive time slots; K is the communication scenario sample.
[0045] In this embodiment, the optimal trade-off between perceptual ability and computational complexity is measured using multiplication-accumulation-addition operations, expressed as:
[0046] In the formula, To evaluate the efficiency of deep learning methods; MACC stands for computational complexity. For sensing capability, the root mean square error of the angle error is defined.
[0047] In the second embodiment, an angle prediction model is constructed using physical kinematics to realize a predictive beamforming method integrating communication and sensing. The specific steps are as follows: S71. Based on the sensor data from the previous moment, the base station obtains the predicted angle of the perceived target at the current moment through an angle prediction model. S72. Generate a transmission beamforming vector based on the predicted angle of the target at the current moment; by transmitting the beamforming vector, weight the integrated communication and sensing signal to be transmitted to obtain a weighted integrated communication and sensing signal, and transmit it toward the target. S73. The sensing target receives the weighted communication sensing integrated signal and returns an echo signal; S74. Based on the echo signal, the base station obtains the echo signal of a single sensing target at the current moment through the orthogonality of the steering vector; it performs radar matched filtering on the echo signal of the single sensing target at the current moment to obtain an estimate of the time delay and an estimate of the Doppler frequency shift. S75. Based on the estimation of time delay, the estimation of Doppler frequency shift, and the predicted angle of the perceived target at the current moment, the actual distance, actual speed, and actual angle at the current moment are obtained through the perceived target state prediction model, which are used to predict the perceived target angle at the next moment.
[0048] Specifically, the angle prediction model is as follows:
[0049] In the formula, For time slot interval; The angle from which the target was perceived in the previous moment; The speed of the target sensed in the previous moment; To determine the distance to the target sensed in the previous moment; This is a Gaussian noise term.
[0050] Specifically, the steps to obtain the actual distance, actual speed, and actual angle at the current moment include: S751. Based on the estimation of time delay and Doppler frequency shift, the echo signal of a single sensing target at the current moment is recovered to obtain the recovered echo signal. S752. Perform radar matched filtering on the recovered echo signal to obtain the measurable time delay and Doppler frequency shift, expressed as:
[0051] In the formula, For round-trip propagation delay; This is a two-way Doppler frequency shift; and These are the signal propagation speed and the carrier frequency, respectively. and The measured Gaussian noise; This refers to random errors or noise generated when measuring time delay; This represents the random error or noise generated when measuring the Doppler frequency shift; Represents a normal distribution; S753. Based on the measurable time delay and Doppler shift, the actual distance, actual velocity, and actual angle at the current moment are obtained through a target state prediction model; for example... Figure 2 As shown, the target state prediction model is as follows:
[0052] In the formula, For the first A perceived target in from Time's up The distance traveled within this time period; For the first The angle of the sensing target relative to the base station, from Time's up The amount of change within this time period; It represents the length of the tangential motion component, and is associated with changes in distance, range, and angle.
[0053] The present invention has the following beneficial effects: This invention presents a predictive beamforming method integrating communication and sensing. By introducing the SCTMNet model, it directly predicts the angle of the sensing target at the current moment based on the echo signal from the previous moment, thereby generating a transmit beam. This injects "predictability" into beamforming, enabling the beam to point in advance to the direction the sensing target is about to arrive, fundamentally solving the beam inaccuracy problem caused by processing delays and rapid movement of the sensing target, and significantly improving the downlink signal reception quality. In addition, the SCTMNet model does not require dedicated pilots and high-precision, low-latency beam tracking with uplink feedback. While significantly improving the communication signal-to-noise ratio and sensing accuracy, it greatly reduces the system's computational complexity and signaling overhead, making it particularly suitable for deployment on edge devices with limited computing resources, such as base stations. Thus, it provides an efficient and reliable solution for the practical application of 6G communication systems.
[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A predictive beamforming method integrating communication and sensing, characterized in that, include: S1. The base station uses the echo signal of a single sensed target in historical data as the input signal and obtains the predicted angle of the sensed target relative to the base station at the current moment through a pre-trained SCTMNet model; the predicted angle refers to the angle of arrival of the sensed target relative to the base station at the current moment. S2. Based on the predicted angle between the base station and the sensing target at the current moment, generate a transmit beamforming vector; through the transmit beamforming vector, perform weighted fusion of the integrated communication and sensing signals to be transmitted to obtain a weighted integrated communication and sensing signal, i.e., the target beam signal, and transmit it to the area where the target sensing target is located. S3. All sensing targets receive the weighted communication and sensing integrated signal and return the echo signal of all sensing targets. S4. Based on the echo signals of all sensed targets, the base station sets up a processing mechanism based on the orthogonality of the steering vector to separate and obtain the echo signal of a single sensed target at the current moment. This signal is used as a new input signal. Through the pre-trained SCTMNet model, the predicted angle between the base station and the sensed target at the next moment is obtained, and then the target beam signal at the next moment is obtained.
2. The predictive beamforming method integrating communication and sensing according to claim 1, characterized in that, The expression for the predicted angle of the perceived target at the current moment, obtained through the pre-trained SCTMNet model, is as follows: In the formula, For prediction angle; The echo signal of the z-th sensing target in the n-th time slot; This is the mapping function, i.e., the SCTMNet model.
3. The predictive beamforming method integrating communication and sensing according to claim 2, characterized in that, The SCTMNet model includes a preprocessing layer, a signal embedding layer, an attention weighting layer, a feature extraction layer, and an output layer. The preprocessing layer is used to recover the echo signal of a single sensing target to obtain the recovered echo signal; the recovered echo signal is split into real part and imaginary part by two cascaded fully connected layers; real-valued features are obtained by passing the real part and imaginary part of the recovered echo signal through a fully connected layer. The signal embedding layer is used to perform preliminary feature extraction on real-valued features through three cascaded convolutional layers and a batch normalization layer to obtain initial features; The attention weighting layer is a multi-branch channel attention network used to weight the initial features to obtain multiple attention-weighted features; The feature extraction layer is used to extract deep features from each attention-weighted feature to obtain deep features; The output layer is used to add and fuse deep features to obtain the predicted angle.
4. The predictive beamforming method integrating communication and sensing according to claim 3, characterized in that, The feature extraction layer includes a convolutional Mamba network module, a convolutional downsampling module, a standard convolutional layer, a first convolutional Transformer network, and a second convolutional Transformer network; The convolutional Mamba network module is used to perform multi-path parallel processing on attention-weighted features to obtain local-global-sequence fusion features; The convolutional downsampling module is used to perform downsampling operations on the local-global-sequence fusion features to obtain downsampled local-global-sequence fusion features; The first convolutional Transformer network is used to extract shallow features from local-global-sequence fusion features to obtain shallow features; The standard convolutional layer is used to normalize and smooth shallow features to obtain refined features; The second convolutional Transformer network is used to extract features from the refined features to obtain deep features.
5. The predictive beamforming method integrating communication and sensing according to claim 4, characterized in that, The convolutional Transformer network includes depthwise separable convolutions, multi-head attention layers, gated layers, layer normalization layers, feedforward neural network layers, and an output layer; The depthwise separable convolution is used to extract local features from the downsampled local-global-sequence fusion features to obtain local spatial features; The multi-head attention layer is used to calculate the relationships between different locations in the local spatial features to obtain attention features; The gating layer is used to multiply local features and attention features element-wise to obtain gated fusion features; The normalization layer is used to normalize the gated fusion features to obtain normalized features. The feedforward neural network layer is used to perform a nonlinear transformation on the normalized features to obtain the output features of the feedforward neural network. The output layer is used to multiply the normalized features and the output features of the feedforward neural network element by element to obtain the output features of the convolutional Transformer network, namely shallow features or deep features.
6. The predictive beamforming method integrating communication and sensing according to claim 1, characterized in that, The weighted communication sensing integrated signal is: In the formula, For weighted communication sensing integrated signal; For integrated communication and sensing signals; Here is the downlink beamforming matrix, where, Z represents the transmitting antenna, and Z represents the total number of targets to be sensed. The beamforming vector is expressed as: In the formula, For beamforming functions; For transmitting antenna; This is the phase offset applied to the m-th antenna.
7. The predictive beamforming method integrating communication and sensing according to claim 1, characterized in that, The echo signals of all the sensed targets are: In the formula, The echo signal of all perceived targets; Let U be the gain coefficient of the sensing array, where U is the number of receiving antennas; The reflection coefficient; The received signal is a delayed version of the transmitted signal; and The first is used for perception. Doppler shift and time delay of the target being sensed; Additive white Gaussian noise with zero mean; In the first The first time slot The actual angle of the sensing target relative to the base station; This is the launch guidance vector; To receive the guide vector; The expressions for the transmit steering vector and the receive steering vector are as follows: ； Based on the echo signal and the orthogonality of the steering vector, the expression for the echo signal of a single sensing target at the current moment is obtained as follows: 。 8. The predictive beamforming method integrating communication and sensing according to claim 1, characterized in that, After step S4, by defining the downlink receiving signal module of the sensing target, the receiving signal-to-noise ratio of the sensing target and the reachability and rate of the communication link are derived to evaluate the receiving performance of the sensing target. The downlink receiving signal module for the sensing target is defined as follows: In the formula, The gain of the communication link array; For the first Channel fading between the sensing target and the base station; Based on the defined downlink receiving signal module of the sensing target, the received signal-to-noise ratio of the sensing target is obtained as follows: In the formula, For beam alignment accuracy, and ; Based on the received signal-to-noise ratio of the sensing target, the reachability rate of the sensing target is obtained as follows: In the formula, For reachable and rate.
9. The predictive beamforming method integrating communication and sensing according to claim 1, characterized in that, The pre-trained SCTMNet model is an SCTMNet model that has been pre-trained and optimized using mean absolute error as the loss function; the loss function is: In the formula, and The perceived target in the k-th training sample is respectively In the The echo signal of each time slot and its corresponding true angle; N is the total number of consecutive time slots; K is the communication scenario sample.