Channel Adaptive High-Frequency Dual-Function Waveform Design Method Assisted by Multimodal Information

Abstract

The present invention discloses a channel adaptive high-frequency dual-functional waveform design method assisted by multi-modal information, belonging to the field of wireless communication technology. It includes constructing a complex-valued beam squint adaptive precoding network model CSP-Net, training CSP-Net by inputting complex-valued communication and sensing channel multi-modal data, and jointly outputting the parameters of a time delay device, a phase shifter, and power allocation to achieve the dual-functional waveform optimization design of communication and sensing integration in a high-frequency system. The present invention utilizes the multi-modal information of communication and sensing channels, can solve the problem that it is difficult to balance communication and sensing performance in the high-frequency band, and reduces the computational complexity in the process of hybrid precoding signal processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication, and relates to the integrated sensing and communications (ISAC) dual-functional waveform optimization design technology. In particular, it relates to a channel adaptive high-frequency dual-functional waveform design method assisted by multi-modal information. By designing a task-specific deep neural network model, it fully utilizes the multi-modal information of radio frequency communication and sensing channels. Through the adaptive active adjustment of communication and sensing channels, it completes the low-complexity high-frequency hybrid precoding signal processing to achieve the dual-functional waveform design with high performance gain, and is applied to the integrated sensing and communications dual-functional waveform optimization in the high-frequency millimeter wave / sub-terahertz band. Background Art

[0002] With the development of technology, intelligent transportation such as vehicle-to-everything (V2X) and low-altitude economy will become an inevitable trend. High-rate communication and high-precision sensing are the basis for ensuring the safety and efficiency of intelligent transportation. The integrated sensing and communications (ISAC) dual-functional waveform design technology deeply integrates time-frequency space domain resources, and realizes both communication and sensing functions by transmitting one signal, reduces hardware and power costs, and improves the efficiency of communication and sensing. It will become the technical cornerstone to support the high-security and high-efficiency intelligent transportation network.

[0003] High-frequency band (millimeter wave, sub-terahertz) transmission systems can not only provide a larger bandwidth for communication to improve the channel capacity, but also provide a larger resolution in terms of distance, angle, etc. for sensing tasks to improve the sensing accuracy. Therefore, it is expected to support the massive data interaction requirements and extremely high sensing accuracy requirements in intelligent transportation application scenarios. However, in high-frequency systems, non-ideal factors such as beam squint effect are more obvious. The beam squint effect causes difficulties in beam alignment on each subcarrier, resulting in the loss of antenna array gain, thus reducing the signal-to-noise ratio at the receiving end and leading to the deterioration of the air interface waveform communication performance. To overcome the influence of the beam squint effect, for example, the invention patent with the patent number CN202010388522.9 proposes new signal processing architectures such as time-delay-phase hybrid precoding to mitigate the beam squint effect. However, the waveforms generated by these schemes cannot take into account the function of sensing and positioning.

[0004] For sensing tasks, the beam squint effect in high-frequency systems can provide a larger sensing range and achieve frequency-dependent angle estimation due to the expanded beam coverage. Therefore, in existing waveform design studies, through the parameter design of hybrid precoding, the pattern of beam squint can be controlled to cover the range to be sensed, enabling the waveform to accurately calculate parameters such as angle and distance. For example, the invention with the patent number CN202310202012.1 provides a method for actively controlling beam squint to achieve low-overhead angle sensing of the target by the waveform. However, these waveform schemes cannot guarantee reliable services for communication users.

[0005] In summary, for high-frequency transmission systems, existing high-frequency dual-functional waveform design schemes are difficult to balance communication and sensing requirements, fail to fully exploit the multi-modal information of the communication and sensing channels, and have limited integrated performance gain; moreover, the hybrid precoding process for generating dual-functional waveforms involves a large number of iterations and searches, with high computational complexity and long time consumption, making it difficult to be deployed in actual roadside base station high-frequency transmission systems. Summary of the Invention

[0006] To overcome the above deficiencies of the prior art, the present invention provides a channel adaptive high-frequency dual-functional waveform design method assisted by multi-modal information, which is a high-frequency communication and sensing integrated dual-functional design method for actively adjusting channel correlation. By utilizing the multi-modal information of communication and sensing channels, it is used to solve the problem that it is difficult to balance communication and sensing performance in the high-frequency band and reduce the computational complexity of the hybrid precoding signal processing process.

[0007] The principle of the present invention is as follows: First, the correlation between communication and sensing channels has an important impact on the integrated performance gain. Enhancing the correlation of communication and sensing channels helps to expand the performance boundary of communication and sensing integration and achieve a better trade-off of communication and sensing tasks. The true delay network in the high-frequency system hybrid precoding architecture can provide frequency-dependent phase shifts to effectively control the beam squint effect, which is equivalent to actively controlling the correlation between communication and sensing channels. Further, through the joint optimization design of phase shifters and baseband power allocation in hybrid precoding, the generation of dual-functional waveforms can be completed to achieve the coordinated improvement of communication and sensing performance. Finally, by using the inference ability of deep neural networks and the data-driven paradigm, the multi-modal information of communication and sensing channels can be processed through training neural networks to efficiently output hybrid precoding parameters and reduce the computational complexity in traditional optimization and search processes.

[0008] The technical solution provided by the present invention is:

[0009] A multi-modal information assisted channel adaptive high-frequency dual-functional waveform design method designs a deep learning model for the dual-functional waveform optimization design task suitable for the integrated communication and sensing of high-frequency systems: the Complex-valued Squint-aware Precoding Network (CSP-Net). By inputting the complex-valued communication and sensing channel multi-modal data, training the neural network, and jointly outputting the parameters of the time delay device, phase shifter, and power allocation, the optimization design of the dual-functional waveform is realized to enhance the correlation of the communication and sensing equivalent channels, optimize the communication performance (rate), and the sensing performance (the Cramer-Rao bound CRB of target angle estimation). The high-frequency integrated communication and sensing dual-functional waveform design scheme includes the following steps:

[0010] S1: Construct the complex-valued squint-aware precoding network CSP-Net model;

[0011] S2: Use the sample data of the training set to perform offline training and learning on the CSP-Net model according to the custom-designed loss function;

[0012] S3: Preprocess the multi-modal data of the communication channel and sensing channel estimated by the actual system to obtain the complex-valued communication channel sample covariance matrix data and the complex-valued sensing target response matrix that conform to the model input form, and perform power normalization;

[0013] S4: Use the preprocessed data as the input of the CSP-Net model after training and learning, and use the CSP-Net model to sequentially predict the optimal integrated hybrid precoding parameter vector under the current communication and sensing channels;

[0014] S5: Post-process the output vector of the CSP-Net to obtain the time delay, phase shifter, and power allocation values of the hybrid precoding, and complete the determination of the integrated hybrid precoding parameters of the high-frequency system;

[0015] S6: Input the communication symbols to be sent to the user into the hybrid precoding signal processing module to output the dual-functional waveform.

[0016] Furthermore, step S1 specifically includes:

[0017] S11: Build a communication channel feature extraction module CFE-Net to process the communication channel data H in

[0018] S12: Build a sensing channel feature extraction module SFE-Net to process the sensing channel data G in

[0019] S13: Concatenate the output features of the CFE-Net in S11 and the SFE-Net in S12, and build a feature fusion layer FL to further process the communication and sensing channel features obtained after concatenation;

[0020] S14: Build a parameter design module PD-Net to process the features processed by the feature fusion layer in S13 and output the parameter prediction of the hybrid precoding;

[0021] Furthermore, step S2 specifically includes:

[0022] S21: Divide the training data into batches, and in each round of network update iteration, input each batch of training data into the CSP-Net in turn for forward calculation to obtain parameter prediction values;

[0023] S22: Calculate the loss function for each batch according to the corresponding true values of the communication and sensing channel data and the parameter prediction values output by the CSP-Net in S21. The loss function includes the communication-sensing channel correlation, communication performance, and sensing performance at the same time;

[0024] S23: Update the network parameters of the CSP-Net based on the loss function calculated in S22 using the backpropagation algorithm;

[0025] Furthermore, step S3 specifically includes:

[0026] S31: For the channel data h on the m-th subcarrier estimated in the actual scenario c,m , calculate its sample covariance matrix

[0027] S32: Aggregate the communication channel sample covariance matrix data on each subcarrier obtained in S31 and the sensing channel matrices on each subcarrier

[0028] S33: Normalize the and obtained in S32 so that the modulus values of the elements in all input data are between [0, 1] to obtain H in and G in . Among them, the communication channel can be obtained through channel estimation, and the sensing channel can be determined according to the prior angular distance and other position information of the target. The input information data involves different modalities.

[0029] Furthermore, step S4 specifically includes: Input the data H in and G in into the CSP-Net model and sequentially output the prediction parameters

[0030] Further, step S5 specifically includes:

[0031] S51: Further perform modulo normalization, dimensional transformation, and quantization operations based on the output of the CSP model, and convert it into the time delay parameter value T with a finite range and finite precision in the actual hybrid precoding architecture ; out ;

[0032] S52: Further perform modulo normalization operation based on the output of the CSP model, and convert it into the phase shifter parameter value with a finite range in the actual hybrid precoding architecture

[0033] S53: Further perform normalization operation based on the output of the CSP model to obtain the final baseband subcarrier power value.

[0034] Further, step S6 specifically includes:

[0035] S61: Modulate the information bits to be transmitted onto the communication symbols of each subcarrier to obtain s = [s1,... s M , where M is the number of subcarriers;

[0036] S62: Input the communication symbol s obtained in S61 into the hybrid precoding signal processing module determined in S5 to obtain the dual-functional waveforms on each subcarrier at the output:

[0037] S63: According to S62, the final output dual-functional waveform is x = [x1,... x M .

[0038] Through the above steps, a low-complexity dual-functional waveform design for communication and sensing integration in high-frequency systems is realized.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] The present invention provides an optimized design scheme for communication and sensing integration dual-functional waveforms suitable for high-frequency (millimeter-wave, terahertz) systems. By using a task-specific deep neural network model, it can efficiently process multi-modal data of communication and sensing channels, infer optimal hybrid precoding parameters, significantly reduce the computational complexity of the design method, reduce the time overhead of the design, and be more robust to the noise of channel estimation, adapting to dynamic environments. It fully exploits the hardware design freedom of hybrid precoding itself, actively regulates the correlation between the equivalent communication and sensing channels by using time delay devices, and lays a foundation for improving the integrated performance gain of waveforms. By adopting the technical solution of the present invention, while ensuring the communication rate, it can improve the angle estimation accuracy of the target, with little modification to the original system hardware architecture and low cost.

[0041] The integrated communication and sensing dual-functional waveform design solution provided by the present invention has the following advantages:

[0042] 1) By using a customized deep learning model, it can efficiently process the input data of communication and sensing channels, replace complex iterative optimization with the forward inference of a deep neural network, reduce the design complexity, and enhance the robustness to noise;

[0043] 2) By designing complex-valued network units, feature channel attention mechanisms, customized heterogeneous network outputs, etc., it can effectively extract channel features and the correlation between communication and sensing channels, adapt to the processing of complex-valued channel multi-modal data and hybrid precoding tasks, and improve the performance of network outputs;

[0044] 3) By using the design freedom of the time delay device itself, it can actively regulate the correlation between communication and sensing channels without relying on external devices, with little modification to the original system, relatively convenient and flexible configuration, and controllable cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is the overall process block diagram of the integrated communication and sensing dual-functional waveform design method based on the CSP-Net model of the present invention.

[0046] Figure 2 It is the schematic diagram of the time delay-phase hybrid precoding architecture adopted by the present invention.

[0047] Figure 3 It is the schematic diagram of the overall network structure of the CSP-Net model designed by the present invention.

[0048] Figure 4 It is the process block diagram of the feature fusion method based on the convolutional block attention module in the CSP-Net of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0050] The present invention deploys an integrated communication and sensing dual-functional waveform optimization design method at the communication base station end. The main processes of this solution include: constructing a complex-valued beam squint adaptive precoding network CSP-Net model construction, model offline training, channel multi-modal data preprocessing, CSP-Net inference, output post-processing to determine hybrid precoding parameters, and generating a dual-functional waveform according to the hybrid precoding signal processing module, as Figure 1 shown. The hybrid precoding processing architecture used to generate the dual-functional waveform mainly consists of a digital baseband power distribution, a phase shifter network, a true time delay device network, and a uniform planar antenna array, asFigure 2 As shown below. The specific steps are as follows:

[0051] S1: Refer to Figure 3 as shown, and construct a complex-valued beam squint adaptive precoding network CSP-Net model. The CSP-Net model takes the communication channel data H in and the sensing channel data G in as inputs. The CSP-Net model includes 4 modules: a communication channel feature extraction network (CFE-Net); a sensing channel feature extraction network (SFE-Net); a feature fusion layer (FL); and a parameter design network (PD-Net). The specific structures of each module are as follows:

[0052] S11: The communication channel feature extraction module CFE-Net takes the communication channel data H in as input. The CFE-Net consists of two layers of complex-valued convolutional blocks, and each complex-valued convolutional block contains a complex convolutional layer, a complex activation layer, a complex batch normalization layer, and a complex pooling layer. All parameters in the complex-valued convolutional block are complex numbers. During forward calculation, the real and imaginary parts of the input complex number will be cross-calculated with the real and imaginary parts of the network parameters respectively to obtain a complex output. The parameters of the CFE-Net module can be represented as φ CFE , and the output of the CFE-Net is the communication channel feature V c = φ CFE (H in ).

[0053] S12: The sensing channel feature extraction module SFE-Net takes the sensing channel data G in as input. Similar to the CFE-Net, the SFE-Net consists of two layers of complex-valued convolutional blocks. The parameters of the SFE-Net module can be represented as φ SFE , and the output of the SFE-Net is the communication channel feature V s = φ SFE (G in ).

[0054] S13: The communication channel feature V c obtained in S11 and the sensing channel feature V s, concatenated at the channel dimension and jointly input into the feature fusion layer FL. The feature fusion layer FL consists of a complex convolutional block attention module (CBAM) with parameters φ FL , and the specific structure is referred to Figure 4 as shown. The CBAM module specifically includes two sub-modules: a channel attention module (CAM) and a spatial attention module (SAM). They perform attention mechanisms on channels and spaces respectively, weight-fuse the features of different channels and different spaces, and save parameters and computational complexity. Then, the feature tensor output by the convolutional block attention module CBAM is flattened into a feature vector v, which is used as the output of the feature fusion layer FL, that is, v = φ FL (V c , V s ).

[0055] S14: Obtain the communication and sensing channel feature v through S13, and then input it into the parameter design network PD-Net. The PD-Net first includes two layers of complex fully connected layers, which are used to predict the parameters of the time delay device, and the output is Through post-processing, is converted into the real time delay device parameter T in the hybrid precoding out . Then, the feature v is concatenated with and further input into the next layer of complex fully connected layer in the PD-Net to predict the parameters of the phase shifter, and the output is Through post-processing, is converted into the real time delay device parameter Finally, b, and are concatenated and input into the last layer of complex fully connected layer in the PD-Net to predict the power allocation parameter of the baseband part Through post-processing, it is converted into the power allocation p of the actual baseband. The parameters of the PD-Net network module are represented as φ PD , and its output can be expressed as In summary, the overall network parameters of the CSP-Net can be expressed as φ CSP ={φ CFE , φ SFE , φ FL , φ PD}.

[0056] S2: Construct the training set data, and use the sample data in the training set to design a loss function to perform offline training and learning on the CSP-Net model constructed in S1. The training set data includes communication channel data and sensing channel data. The offline training and learning specifically includes the following steps:

[0057] S21: Generate a large number of communication and sensing channels with the same dimension as the actual system through random simulation as the training data set. The i-th group of training sample data is the simulated communication channel data and the sensing channel data Then, preprocess the communication channel data and the sensing channel data according to the steps in S3 to obtain the data H in and G in input into the CSP-Net. Divide the preprocessed training data into batches, and in each round of network update iteration, input each batch of training data into the CSP-Net in turn for forward calculation to obtain the parameter prediction value of the hybrid precoding where the parameter prediction value corresponding to the i-th sample is

[0058] S22: Calculate the loss function for each batch according to the corresponding communication and sensing channel data and and the parameter prediction value output by the CSP-Net in S21 Specifically, the loss function L (k) (φ CSP ) of the k-th batch is expressed as:

[0059]

[0060] where N b is the number of samples included in each batch, is the correlation of the communication and sensing equivalent channels calculated according to the existing formula under the time delay of ; is the communication rate calculated according to the output and the communication channel in the i-th sample according to the existing rate calculation formula to measure the communication performance; is the theoretical maximum rate under the current communication channel ; CRB (i) is the angle estimation CRB calculated according to the output and the sensing channel in the i-th sample according to the existing CRB calculation formula to measure the sensing performance; is the theoretical optimal CRB under the current sensing channel ; Cor＊ Indicates the current communication channel and the sensing channel The theoretical maximum channel correlation under

[0061] S23: Calculate the loss reflecting channel adaptation based on the loss function in S22, and update the CSP-Net network parameters based on the existing Back Propagation (BP) algorithm, that is:

[0062]

[0063] where α is the learning rate of the CSP-Net model.

[0064] S3: In the actual deployment stage, obtain the communication channel and the sensing channel through the existing channel estimation technology, and preprocess the data of the two modalities of the communication channel and the sensing channel to obtain the complex-valued communication channel sample covariance matrix data and the complex-valued sensing target response matrix that conform to the model input form, and then perform power normalization. Specifically, it includes the following steps:

[0065] S31: Obtain the communication channel and the sensing channel in the actual scenario through the existing channel estimation technology, where the communication channel data is represented as h c =[h c,1 ,…,h c,M , representing the communication channel on M subcarriers from the transmitter to the user receiver; the sensing channel data is represented as G s =[G s,1 ,…,G s,M , representing the sensing channel on M subcarriers from the transmitter to the echo receiver. For the channel data h c,m on the m-th subcarrier, calculate its sample covariance matrix

[0066] S32: Aggregate the communication channel sample covariance matrix data on each subcarrier obtained in S31, denoted as and the sensing channel matrix on each subcarrier

[0067] S33: Normalize the and obtained in S32 so that the modulus values of all elements in the input data are between [0,1], and obtain the preprocessed communication and sensing channel input data H in and G in .

[0068] S4: Input the data preprocessed in S3 into the CSP-Net model trained and learned in S2, and use the CSP-Net model to sequentially predict the optimal integrated hybrid precoding parameter prediction vector under the current communication and sensing channels. Specifically, that is, input the data H in and G in into the CSP-Net model, and sequentially output the model prediction parameters

[0069] S5: Post-process the output prediction parameters of the CSP-Net to obtain the delay T out of the hybrid precoding, the phase shifter and the power allocation value p out , and complete the integrated hybrid precoding design of the high-frequency system. Specifically, it includes the following steps:

[0070] S51: Further perform modulo normalization, dimension transformation, and quantization operations according to the output of the CSP model to convert it into the time delay parameter values with a limited range and limited precision in the actual hybrid precoding architecture, that is where |·| is the complex modulus operation, t max is the maximum time delay value that the system can reach, reshape(·) is the dimension transformation operation, which converts the output vector into a two-dimensional time delay array value connected to the actual planar antenna array, and Q[·] is the operation function for quantizing the time delay value according to the nearest neighbor rule.

[0071] S52: Further perform modulo normalization operation according to the output of the CSP model to convert it into the phase shifter parameter values with a limited range in the actual hybrid precoding architecture, that is

[0072] S53: Further perform normalization operation according to the output of the CSP model to obtain the final baseband subcarrier power value as where ||·||1 is the L1-norm, and P t is the total transmission power of the system baseband.

[0073] S6: Input the communication symbols to be sent to the user into the hybrid precoding signal processing module and output the dual-functional waveform. Specifically, it includes the following steps:

[0074] S61: Modulate the information bits to be sent onto the communication symbols of each subcarrier to obtain s = [s1,... s M , where M is the number of subcarriers;

[0075] S62: Input the communication symbol s obtained in S61 into the hybrid precoding signal processing module determined in S5 to obtain the dual-functional waveforms on each subcarrier of the output:

[0076] S63: According to S62, the finally output dual-functional waveform is x = [x1,... x M .

[0077] The present invention designs a communication-aware integrated dual-functional waveform optimization design algorithm suitable for channel adaptive adjustment in high-frequency transmission systems. By using multi-modal data of communication and sensing channels, it explores the design freedom of the hardware architecture itself. By designing a real-time delay device, it actively regulates the correlation between communication and sensing equivalent channels to achieve an improvement in integrated performance gain. By using the constructed deep learning model, it realizes the design of low-complexity hybrid precoding signal processing parameters. By customizing the design of complex-valued network structures, feature fusion enhancement, heterogeneous network output and other modules in the network model, it improves the training efficiency and the performance of the model. Through simulation experiments, it can be verified that the proposed scheme can balance communication and sensing performance in high-frequency systems. Compared with traditional design methods, it can further improve the performance boundaries of communication and sensing. This method does not modify the existing high-frequency system waveform generation architecture, can be compatible with existing hardware conditions, and can meet the practical requirements of low operation cost and low complexity of the scheme.

[0078] It should be noted that the purpose of publishing the embodiments is to help further understand the present invention. However, those skilled in the art can understand that various substitutions and modifications are possible without departing from the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection claimed by the present invention is subject to the scope defined by the claims.

Claims

1. A multimodal information-assisted channel adaptive high-frequency dual-function waveform design method, characterized in that: A complex-valued beam squint adaptive precoding network model (CSP-Net) is constructed. The CSP-Net is trained by inputting multimodal data of complex-valued synaesthesia channel, and the parameters of delayer, phase shifter and power allocation are jointly output to realize the dual-function waveform optimization design of synaesthesia integration of high-frequency system. The steps are as follows: S1: Construct a complex-valued beam squint adaptive precoding network model CSP-Net; the input of the CSP-Net model is communication channel data and perception channel data; the output is the parameter prediction of hybrid precoding; The CSP-Net model includes a communication channel feature extraction module, a perception channel feature extraction module, a feature fusion layer, and a parameter design module; the communication channel feature extraction module is used to process communication channel data; the perception channel feature extraction module is used to process perception channel data; the feature fusion layer is used to further fuse and enhance the features after the communication and perception channels are cascaded; the parameter design module PD-Net is used to process the communication and perception channel features after fusion and enhancement, and output the parameter prediction of hybrid precoding; S2: Design loss function and perform offline training on CSP-Net model; The loss function includes communication perception channel correlation, communication performance, and perception performance; including: S21: Divide the training data into batches, input them into CSP-Net in each round of network update iteration, perform forward calculation, and obtain the parameter prediction value of hybrid precoding; S22: Calculate the loss function of each batch according to the corresponding communication and perception channel data and the parameter prediction values ​​obtained in S21; S23: Update the CSP-Net network parameters according to the loss function; S3: preprocess the communication channel and perception channel multimodal data estimated by the actual system, obtain the complex-valued communication channel sample covariance matrix data and the complex-valued perception target response matrix, and perform power normalization; S4: using the preprocessed data as the input of the trained CSP-Net model, and using the trained CSP-Net model to predict the optimal integrated hybrid precoding parameter vector under the current communication and perception channels in turn; S5: Post-process the output vector of CSP-Net to obtain the hybrid precoding delay, phase shifter and power allocation values, and complete the integrated hybrid precoding parameter determination of the high-frequency system; S6: inputting the communication symbol to be sent to the user into the hybrid precoding signal processing module and outputting a dual-function waveform; Through the above steps, the channel-adaptive high-frequency dual-function waveform design assisted by multimodal information is realized.

2. The multimodal information-assisted channel adaptive high-frequency dual-function waveform design method as claimed in claim 1, characterized in that: The complex-valued beam squint adaptive precoding network model CSP-Net constructed in step S1 includes: S11: The communication channel feature extraction module CFE-Net uses the communication channel data H in is the input; CFE-Net consists of two layers of complex-valued convolution blocks, in which all parameters are complex numbers. In the forward calculation, the real and imaginary parts of the input complex numbers and the real and imaginary parts of the network parameters are cross-calculated to obtain the complex output; the output of CFE-Net is the communication channel feature V c =φ CFE (H in ), where φ CFE are the parameters of the CFE-Net module; S12: Perceptual channel feature extraction module SFE-Net is based on the perceptual channel data G in is the input; it consists of two layers of complex-valued convolution blocks; the output of SFE-Net is the communication channel feature V s =φ SFE (G in ), where φ SFE are the parameters of the SFE-Net module; S13: The obtained communication channel feature V c And the obtained perceptual channel feature V s , cascaded in the channel dimension, and input into the feature fusion layer FL together; the feature fusion layer FL includes a complex convolutional block attention module (CBAM) with parameter φ FL ; The feature tensor output by the complex convolution block attention module is used as the output feature vector v of the feature fusion layer FL, expressed as v = φ FL (V c ,V s ); S14: The obtained communication and perception channel features v are input into the parameter design network module PD-Net; PD-Net contains two layers of complex fully connected layers to predict the parameters of the delay device, and the output is Through post-processing Transformed into the real delay parameter T in hybrid precoding out ; Combine feature v with Cascade, further input into the next layer of complex fully connected layer in PD-Net to predict the parameters of the phase shifter, the output is Through post-processing Transformed into real delay parameters in hybrid precoding v, as well as Cascade and input into the last complex fully connected layer in PD-Net to predict the power allocation parameters of the baseband part Transformed into actual baseband power allocation p through post-processing; The output of the PD-Net network module is expressed as φ PD It is the parameter of PD-Net network module; Therefore, the overall network parameters of CSP-Net are expressed as φ CSP ={φ CFE ,φ SFE ,φ FL ,φ PD }.

3. The multimodal information-assisted channel adaptive high-frequency dual-function waveform design method as claimed in claim 2, characterized in that: In S11, each complex-valued convolution block contains a complex convolution layer, a complex activation layer, a complex batch normalization layer, and a complex pooling layer; In S13, the feature fusion layer FL includes a complex convolutional block attention module CBAM, which specifically includes a channel attention (CAM) submodule and a spatial attention (SAM) submodule, which respectively perform channel and spatial attention mechanisms to perform weighted fusion of features from different channels and different spaces.

4. The multimodal information-assisted channel adaptive high-frequency dual-function waveform design method as claimed in claim 2, characterized in that: Step S21 specifically includes: generating a large number of communication and perception channels with the same dimensions as the actual system through random simulation as training data sets; wherein the i-th group of training sample data is the simulated communication channel data and sensing channel data Then the communication channel data and sensing channel data Preprocessing is performed to obtain the data H input into the CSP-Net network in and G in ; Divide the preprocessed training data into batches, and input each batch of training data into CSP-Net in each round of network update iteration, perform forward calculation, and obtain the parameter prediction value of hybrid precoding {T out , p out }, where the parameter prediction value corresponding to the i-th sample is Step S22 specifically includes: according to the corresponding communication and perception channel data and And the parameter prediction value output by CSP-Net in step S21 Calculate the loss function for each batch; The loss function L for the kth batch (k) (φ CSP ) is expressed as: Among them, N b is the number of samples contained in each batch, The delay device is The communication and perception equivalent channel correlation calculated below; is the output of the ith sample and communication channel The communication rate calculated according to the rate calculation formula; The current communication channel Theoretical maximum rate under CRB (i) is the output of the ith sample and sensing channel The calculated angle estimate CRB; is the current sensing channel The theoretical optimal CRB under the condition of and sensing channel Theoretical maximum channel correlation under ; Step S23 specifically calculates the loss reflecting channel adaptation according to the loss function, and updates the CSP-Net network parameters based on the back propagation algorithm, which is expressed as: Among them, α is the learning rate of the CSP-Net model.

5. The multimodal information-assisted channel adaptive high-frequency dual-function waveform design method as claimed in claim 4, characterized in that: In step S3, the data of the communication channel and the sensing channel are preprocessed, which includes the following steps: S31: obtaining the communication channel and the perception channel in the actual scenario through the channel estimation technology, and calculating the sample covariance matrix of the channel data on the m-th subcarrier; S32: Summarize the communication channel sample covariance matrix data on each subcarrier and the perception channel matrix on each subcarrier; S33: Normalize the aggregated communication channel sample covariance matrix data and the perception channel matrix on each subcarrier so that the modulus values ​​of the elements in all input data are between [0, 1], and obtain the preprocessed communication and perception channel input data.

6. The multimodal information-assisted channel adaptive high-frequency dual-function waveform design method as claimed in claim 5, characterized in that: In step S5, the output prediction parameters of CSP-Net are post-processed to obtain the hybrid precoding delay T out , Phase Shifter And the power allocation value p out ; Specifically includes the following steps: S51: Delay parameters for CSP model output Modulo normalization, dimension transformation and quantization operations are performed to convert the delay parameter value with limited range and limited precision in the actual hybrid precoding architecture, which is expressed as: where |·| is the complex modulus operation, t max is the maximum delay value that the system can reach, reshape(·) is a dimension transformation operation that converts the output vector into a two-dimensional delay array value connected to the actual planar antenna array, and Q[·] is an operation function that quantizes the delay value according to the nearest neighbor rule; S52: Phase shifter parameters for CSP model output Perform a modulo normalization operation to convert it into a phase shifter parameter value with a limited range in the actual hybrid precoding architecture, that is, S53: Power allocation parameters for CSP model output Perform normalization operation to obtain the final baseband subcarrier power value where ||·||1 is the L1-norm, P t is the total transmission power of the system baseband.

7. The multimodal information-assisted channel adaptive high-frequency dual-function waveform design method as claimed in claim 6, characterized in that: According to the channel adaptive high frequency dual-function waveform design method, a channel adaptive high frequency dual-function waveform system is implemented, including: Communication channel feature extraction module CFE-Net, used to process communication channel data; The sensory channel feature extraction module SFE-Net is used to process the sensory channel data; The feature fusion layer FL is used to cascade the output features of CFE-Net and SFE-Net, and further process the communication and perception channel features obtained by the cascade; The parameter design module PD-Net is used to process the features after being processed by the feature fusion layer FL and output the parameter prediction of hybrid precoding.

Images