Multi-feature beam forming wireless communication model training method based on Fourier spectrum
By adopting the Fourier spectrum multi-eigen model and focus loss constraint method in beamforming technology, the calculation complexity and spectrum efficiency problems in multi-user scenarios are solved, and more efficient beamforming effects and system transmission rate are achieved.
Patent Information
- Application Number
- CN202410974037.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing beamforming technology has high computational complexity and severe structural interference in multi-user scenarios. The traditional algorithm is costly and the spectrum efficiency is single. The deep learning model is not effective enough, making it difficult to achieve high performance and low cost at the same time.
Using the multi-eigen beam model training method based on Fourier spectrum, a millimeter wave communication system is built through the MATLAB simulation platform, Fourier spectrum features are extracted as deep neural network input, and two-stage design is designed in combination with CNN network and focal loss function. Methods to improve model robustness and generalization and reduce power loss.
It significantly improves the beamforming effect, improves the system transmission rate and spectrum efficiency, reduces power loss, and enhances the stability and generalization capabilities of the model.
Smart Images

Figure CN119010964A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence, and relates to a multi-feature beamforming wireless communication model training method based on Fourier spectrum, and relates to data transmission, range expansion, and millimeter wave systems in the field of wireless communications. Background Art
[0002] With the explosive growth of modern wireless communication devices, the electromagnetic environment is becoming more and more complex. In order to better improve the user experience, the function of the antenna system is crucial. Not only does it require the communication system to have a fast transmission speed, but the power emitted from the base station must be transmitted to the receiving end with the least loss. This requirement is also increasing. Therefore, the development of beamforming technology is urgent. At present, the existing and practical beamforming-related algorithms are all based on traditional algorithms, such as maximum ratio synthesis, minimum mean square error, and zero forcing algorithms. Among them, the MRC method is a simple and effective traditional algorithm that uses the signals of multiple receiving antennas and linearly combines them through appropriate weights to maximize the signal-to-noise ratio of the received signal. The MMSE algorithm attempts to minimize the mean square error between the received signal and the desired signal, and adjusts the relative contribution of the signal by selecting appropriate weights. The ZF algorithm aims to eliminate the mutual interference in the received signals of multiple antennas, so that the receiving end is only affected by the signal radiated by the transmitting antenna. These traditional methods can improve the effect of beamforming to a certain extent, and each has its own strengths. MMSE beamforming can achieve a higher rate than MRC beamforming, but the cost is too high (Reference [1]: Liu L, Yu W. Massive connectivity with massive MIMO—Part II: Achievable rate characterization [J]. IEEE Transactions on Signal Processing, 2018, 66 (11): 2947-2959.). In view of the high computational complexity, strong noise and delay of the existing traditional algorithms, some people proposed to use an end-to-end model based on sparse autoencoder in the beamforming system, but only consider a single receiver (Reference [2]: Tao J, Wang Q, Luo S, et al. Constrained deep neural network based hybrid beamforming formillimeter wave massive MIMO systems [C] / / ICC 2019-2019 IEEE International Conference on Communications (ICC). IEEE, 2019: 1-6.).After that, a hybrid beamforming model was designed based on unsupervised deep learning to replace the traditional algorithm, achieving good beamforming effect, reducing computational complexity, improving rate, reducing cost and effectively improving beamforming effect, but only considering the case of one user (Reference [3]: Lin T, Zhu Y. Beamforming design for large-scale antenna arrays using deep learning [J]. IEEE Wireless Communications Letters, 2019, 9 (1): 103-107.). In order to solve the problems of computational complexity and structural interference of multiple users, Zhang P et al. proposed a two-stage design method to continuously estimate new CSI information as the input of the model to reduce the impact of data delay (Reference [4]: Zhang P, Pan L, Laohapensaeng T, et al. Hybrid beamforming based on an unsupervised deep learning network for downlink channels with imperfect CSI [J]. IEEE Wireless Communications Letters, 2022, 11 (7): 1543-1547.).
[0003] The current beamforming technology is based on the contradiction between performance and cost of traditional algorithms, and cannot reach a relatively high level at the same time. The beamforming model based on deep learning mainly considers spectrum efficiency, and the network input feature content is single, and then the precoding matrix is obtained. The current effect is not significant enough. Therefore, in order to improve the above shortcomings, the present invention proposes a multi-feature beamforming model training method based on Fourier spectrum, which adds features according to the input data, and uses focal loss to constrain the loss function of the model, so that the effect of the beamforming model is significantly improved. Summary of the invention
[0004] In order to overcome the problems of single data features and poor beamforming model effect in existing methods, the present invention proposes a multi-feature beamforming wireless communication model training method based on Fourier spectrum, which uses a new data processing method, focal loss and CNN network to reduce power loss and improve the effect of beamforming model.
[0005] The technical solution adopted by the present invention to solve the technical problem is:
[0006] A method for training a wireless communication model with multi-feature beamforming based on Fourier spectrum, comprising the following steps:
[0007] S1: Build a millimeter wave communication system based on the MATLAB simulation platform, define the number of receiving antennas, the number of transmitting antennas and the number of RF chains, and generate data;
[0008] S2: In the case of generated data, the required training set and test set are obtained through channel estimation, which contains the channel matrix h H and the signal-to-noise ratio γ;
[0009] S3: The overall process is designed in two stages: online deployment stage and offline training stage. In the training stage, the Fourier spectrum features in the data set are extracted to form the input of the deep neural network C, realizing multi-feature input;
[0010] S4: According to the structure of the deep neural network C, the data processing steps are reasonably increased to improve the robustness of the unsupervised model. The network outputs the precoding matrix V that can change the antenna state. RF , design the loss function, update the training parameters according to the optimization goal, and get the new V RF Matrix, to achieve better beamforming effect;
[0011] S5: Use the test set data to test the trained beamforming model and judge the beamforming effect by the spectrum efficiency.
[0012] The beneficial effects of the present invention are as follows: frequency domain features are obtained by Fourier transform, and the deep neural network C has more sample features as input, which improves the model effect; using the CNN network, normalization is performed after each convolution, which is conducive to better fitting and improving the model effect; designing the loss function, adding focal loss for constraint, solving the balance problem, and improving the stability of the numerical value; using a two-stage design method, an offline training stage, and an online deployment stage, this design takes into account the real-time nature of the data and increases the generalization of the model. The present invention proposes a multi-feature beamforming model training method based on Fourier spectrum, which can reduce power loss and increase the system transmission rate. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 A flow chart of the method provided by the present invention.
[0014] Figure 2 This is the MISO system structure diagram.
[0015] Figure 3 Diagram for a two-stage design approach.
[0016] Figure 4 The flowchart of the training method of the wireless communication model with multi-feature beamforming based on Fourier spectrum. DETAILED DESCRIPTION
[0017] In combination with the above-mentioned drawings, the method of the present invention is systematically and clearly described below by taking the MISO system as an example.
[0018] Reference Figure 1 to Figure 4 ,In order to minimize the power loss, a multi-feature beamforming wireless communication model training method based on Fourier spectrum is proposed, which includes the following steps:
[0019] S1: Build a MISO system based on the MATLAB simulation platform, such as Figure 2 As shown in 1, define the number of receiving antennas, the number of transmitting antennas and the number of RF chains to generate data:
[0020] A simulated MISO system includes a base station, a radio frequency chain, a phase shifter, and an antenna structure, wherein the number of base stations is 1, the number of receiving antennas is 1, the number of radio frequency chains, phase shifters, the number of transmitting antennas, and the number of paths are all selectable, and the number of radio frequency chains is generally set to be much smaller than the number of antennas; in this embodiment, the data stream sent by the base station is represented by the symbol s, and the dimension of the s vector is N s ×1, where N s Represents the number of data streams, and the number of radio frequency chains is N RF Indicates that the number of transmitting antennas is N t Indicates that the number of receiving antennas is N r Indicates that, based on the MISO system, N is defined r =1. Since hybrid beamforming includes digital beamforming and analog beamforming, the transmitted signal will be changed by two sets of precoding, where the analog precoding is composed of V RF Indicates that digital precoding is composed of V D indicates (because there is only one RF chain, V D is actually a scalar), because of the maximum transmit power constraint, V D The optimal value of It is concluded that the channel model for transmitting signals in the system adopts the millimeter wave Saleh-Valenzuela channel model. The channel path has one LOS and L-1 NLoS paths, where L represents the number of paths.
[0021] The method for generating the randomly generated transmission signal described above is:
[0022] x=V RF V D s (1)
[0023] Where s is the data stream sent by the base station, and the dimension is N s ×1; V RF For analog precoding, the dimension is N t ×N RF, the signal is processed by adjusting the analog beamforming realized by the phase shifter; V D is digital precoding, dimension is N RF ×N s , perform digital beamforming on s; finally, x is the initial signal without passing through the channel, with a dimension of N t ×1;
[0024] S2: In the case of generated data, the required training set and test set are obtained through channel estimation, which contains the channel matrix h H And the signal-to-noise ratio γ:
[0025] The input of the beamforming model is CSI information. The present invention uses the channel matrix h H and the signal-to-noise ratio γ, where the channel matrix h H In the millimeter wave Saleh-Valenzuela channel model, h is obtained by channel estimation method. H The dimension is defined as N r ×N t , the formula is:
[0026]
[0027] Where N t is the number of transmitting antennas; L is the number of paths, α l represents the complex gain of the lth path, represents the antenna array response vector at the base station, represents the azimuth deviation angle associated with the lth path, and when l = 1, it represents the component on the LOS path; Determines the response characteristics of the antenna array in different directions;
[0028] Signal-to-noise ratio γ calculation formula:
[0029]
[0030] Where P is the power of signal s, σ 2 is the covariance value between the transmitted signal and the received signal. The input of the designed hybrid beamforming model can be obtained by the above method;
[0031] S3: The overall process is designed in two stages: online deployment stage and offline training stage. In the training stage, the Fourier spectrum features in the data set are extracted to form the input of the deep neural network C, realizing multi-feature input:
[0032] The overall process is as follows Figure 3As shown in the figure, the offline training phase is the training part, and the online deployment phase is the testing part. In the offline training phase, the signal is first simulated, and then the required data set is obtained through the estimator. The method for extracting the Fourier spectrum features of the data is:
[0033] Reference Figure 4 , the data input to the model is discrete Fourier transformed, where h H Get the corresponding frequency domain signal, the DFT transformation formula is:
[0034]
[0035] Among them, h H (n) is the nth channel information matrix in complex form, N is the discrete sampling length, and the mth frequency domain component H(m) is obtained. The frequency domain signal is in complex form and contains amplitude and phase related information. The Fourier amplitude spectrum and Fourier phase spectrum of the signal are further calculated:
[0036] Fourier Amplitude Spectrum:
[0037]
[0038] Fourier phase spectrum:
[0039]
[0040] Where Re(·) means extracting the real part in the complex form, and Im(·) means extracting the imaginary part in the complex form;
[0041] In order to form a model with better performance, features are extremely important in the process of training the neural network model. The richer the information, the more levels, scales, and dimensions it contains, the more the neural network learns during the training process, and the better the model performance. Therefore, the obtained amplitude and phase spectra are changed to the format of the same dimension as the original data set, and then combined with the original data, feature fusion, and multi-feature data are obtained as the input of the neural network. The fusion expression is as follows:
[0042] A=[h H ,H,λ] (7)
[0043] Among them, A is the new data set after fusion;
[0044] S4: According to the structure of the deep neural network C, the data processing steps are reasonably increased to improve the robustness of the unsupervised model. The network outputs the precoding matrix V that can change the antenna state. RF , design the loss function, update the training parameters according to the optimization goal, and get the new V RF Matrix to achieve better beamforming effect:
[0045] The label definition of beamforming is difficult to determine, so the overall model adopts an unsupervised form. The neural network C is set as a CNN network. The data set goes through 2-3 rounds of BN layer, flattening layer, convolution layer, activation layer, and pooling layer. The activation layer uses the ReLU activation function. The output of the network is the precoding matrix V RF , the designed loss function contains the parameter V RF , add focal loss to constrain, solve the balance problem, and improve the numerical stability;
[0046] Precoding matrix V RF The output representation of complex values is:
[0047] V RF =e j·θ =cosθ+j·sinθ (8)
[0048] Where θ represents the phase of the beamforming coefficient;
[0049] The best analog beamforming V is obtained through neural network training RF The value is then used to calculate the subsequent spectrum efficiency and construct the convolutional neural network training. The convolution output expression of the kth layer in the forward propagation is as follows:
[0050] y k =σ(y k-1 *W k +b k ) (9)
[0051] where y k represents the output of the convolutional layer at layer k, y k-1 represents the output of the k-1th layer, W k represents the trainable weight matrix, b k Represents bias, σ represents the activation function. The activation function used here is ReLU, and the expression is as follows:
[0052]
[0053] The pooling layer pools the results obtained by convolution. The pooling layer can reduce the number of parameters. The maximum pooling is used. The maximum pooling can extract the most obvious features. Its expression is as follows:
[0054] y k =pool(y k-1 ) (11)
[0055] Where pool() refers to the process of reducing the input tensor according to the pooling area size and pooling standard;
[0056] The preliminary design of the loss function loss is specifically expressed as:
[0057]
[0058] Where N represents the total number of training samples, γ n , V RF,n represents the signal-to-noise ratio, channel matrix and simulated beamforming precoding matrix between the Nth sample, N t represents the number of transmitting antennas; then the focal loss is added, expressed as:
[0059] loss = λ·(1-p t ) β ·loss1 (13)
[0060] In which p is set t =loss1 / 20, β is the focusing parameter, which adjusts the weight reduction rate; λ is the weighting factor, which solves the balance problem; loss1 is obtained by formula (12);
[0061] After the loss function designed by formula (13), the latest matrix V is obtained RF , achieving better beamforming effect;
[0062] S5: Use the test set data to test the trained beamforming model and judge the beamforming effect by spectrum efficiency:
[0063] Reference Figure 2 , deploy the trained beamforming model to the test phase, use the test set to evaluate the model effect, and finally measure the model effect by the spectrum efficiency value SE;
[0064] The calculation formula of spectrum efficiency SE is:
[0065]
[0066] Where γ represents the signal-to-noise ratio, N t represents the number of transmitting antennas, h H represents the channel matrix, V RF 、V D They are analog and digital precoding respectively;
[0067] The above beamforming model is formed by combining steps S3 and S4 for training, and the effect of the model is measured according to the result of spectrum efficiency.
[0068] As described above, this is a multi-feature beamforming model training method based on Fourier spectrum of the present invention. Most existing beamforming algorithms are implemented by traditional algorithms. Compared with existing traditional algorithms and a small number of beamforming models based on deep learning, the present invention uses Fourier transform to obtain more features, merges with the original data, and finally becomes the input of the network, adds focal loss constraint loss function, balances data, and improves the beamforming effect on the original basis.
[0069] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.
Claims
1. A multi-feature beamforming wireless communication model training method based on Fourier spectrum, characterized in that: The method comprises the following steps: S1: Build a millimeter wave communication system based on the MATLAB simulation platform, define the number of receiving antennas, the number of transmitting antennas and the number of RF chains, and generate data; S2: In the case of generated data, the required training set and test set are obtained through channel estimation, which contains the channel matrix h H and the signal-to-noise ratio γ; S3: The overall process is designed in two stages: online deployment stage and offline training stage. In the training stage, the Fourier spectrum features in the data set are extracted to form the input of the deep neural network C, realizing multi-feature input; S4: According to the structure of the deep neural network C, the data processing steps are reasonably increased to improve the robustness of the unsupervised model. The network outputs the precoding matrix V that can change the antenna state. RF , design the loss function, update the training parameters according to the optimization goal, and get the new V RF Matrix, to achieve better beamforming effect; S5: Use the test set data to test the trained beamforming model and judge the beamforming effect by the spectrum efficiency.
2. The method for training a wireless communication model based on Fourier spectrum using multi-feature beamforming according to claim 1, characterized in that: The process of step S1 is as follows: A simulated MISO system includes a base station, a radio frequency chain, a phase shifter, and an antenna structure, where the number of base stations is 1, the number of receiving antennas is 1, the number of radio frequency chains, phase shifters, the number of transmitting antennas, and the number of paths are all selectable, and the number of radio frequency chains is set to be much smaller than the number of antennas. The data stream sent by the base station is represented by the symbol s, and the dimension of the s vector is N s ×1, where N s Represents the number of data streams, and the number of radio frequency chains is N RF Indicates that the number of transmitting antennas is N t Indicates that the number of receiving antennas is N r Indicates that, based on the MISO system, N is defined r =1; Since hybrid beamforming includes digital beamforming and analog beamforming, the transmitted signal will be changed by two sets of precoding, where the analog precoding is composed of V RF Indicates that digital precoding is composed of V D It means that due to the constraint of maximum transmit power, V D The optimal value of It is concluded that the channel model used for transmitting signals in the system is the millimeter wave Saleh-Valenzuela channel model. The channel path has one LOS and L-1 NLoS paths, where L represents the number of paths. The method for generating the above randomly generated transmission signal is: x=V RF V D s (1) Where s is the data stream sent by the base station, and the dimension is N s ×1; V RF For analog precoding, the dimension is N t ×N RF , the signal is processed by adjusting the analog beamforming realized by the phase shifter; V D is digital precoding, dimension is N RF ×N s , perform digital beamforming on s; finally, x is the initial signal without passing through the channel, with a dimension of N t ×1.
3. The method for training a wireless communication model based on Fourier spectrum using multi-feature beamforming according to claim 1 or 2, characterized in that: The process of step S2 is as follows: According to the data generated in step 1, the required training set and test set are obtained through channel estimation, which contains the channel matrix h H and the signal-to-noise ratio γ. The input of the beamforming model is the CSI information, using the channel matrix h H and the signal-to-noise ratio γ, where the channel matrix h H In the millimeter wave Saleh-Valenzuela channel model, it is obtained by the channel estimation method, and the formula is: Where N t is the number of transmitting antennas; L is the number of paths, α l represents the complex gain of the lth path, represents the antenna array response vector at the base station, represents the azimuth deviation angle associated with the lth path, and when l = 1, it represents the component on the LOS path; Determines the response characteristics of the antenna array in different directions; Signal-to-noise ratio γ calculation formula: Where P is the power of signal s, σ 2 is the covariance value between the transmitted signal and the received signal, which is used as the input of the designed hybrid beamforming model.
4. The method for training a wireless communication model based on Fourier spectrum using multi-feature beamforming according to claim 1 or 2, characterized in that: The process of step S3 is as follows: The overall process is designed in two stages, the online deployment stage and the offline training stage. In the training stage, the Fourier spectrum features in the data set are extracted to form the input of the deep neural network C to achieve multi-feature input. The offline training stage is the training part, and the online deployment stage is the testing part. In the offline training stage, the signal is first simulated, and then the required data set is obtained through the estimator. The method for extracting the Fourier spectrum features of the data is to perform discrete Fourier transform on the data input to the model, h H Both γ and γ get corresponding frequency domain signals, and the DFT transformation formula is: Among them, h H (n) is the nth channel information matrix in complex form, N is the discrete sampling length, and the mth frequency domain component H(m) is obtained. The frequency domain signal is in complex form and contains amplitude and phase related information. The Fourier amplitude spectrum and Fourier phase spectrum of the signal are further calculated: Fourier Amplitude Spectrum: Fourier phase spectrum: Where Re(·) means extracting the real part in the complex form, and Im(·) means extracting the imaginary part in the complex form; The obtained amplitude and phase spectra are converted into a format with the same dimension as the original data set, and then combined with the original data to fuse the features and obtain multi-feature data as the input of the neural network.
5. The method for training a wireless communication model based on Fourier spectrum using multi-feature beamforming according to claim 1 or 2, characterized in that: The process of step S4 is as follows: According to the structure of the deep neural network C, the data processing steps are reasonably increased to improve the robustness of the unsupervised model. The network output is the precoding matrix V that can change the antenna state. RF , design the loss function, update the training parameters according to the optimization goal, and get the new V RF Matrix to achieve better beamforming effect: The label definition of beamforming is difficult to determine, so the overall model adopts an unsupervised form. The neural network C is set to CNN1D. The data set goes through 2-3 rounds of BN layer, flattening layer, convolution layer, activation layer, and pooling layer. The activation layer uses the ReLu activation function. The output of the network is the precoding matrix V RF , the designed loss function contains the parameter V RF , add focal loss to constrain, solve the balance problem, and improve the numerical stability; Precoding matrix V RF The output representation of complex values is: V RF =e j·θ =cosθ+j·sinθ (7) Where θ represents the phase of the beamforming coefficient; The best analog beamforming V is obtained through neural network training RF The value is then used to calculate the subsequent spectrum efficiency and construct the convolutional neural network training. The convolution output expression of the kth layer in the forward propagation is as follows: y k =σ(y k-1 *W k +b k ) (8) where y k represents the output of the convolutional layer at layer k, y k-1 represents the output of the k-1th layer, W k represents the trainable weight matrix, b k Represents bias, σ represents the activation function. The activation function used here is ReLU, and the expression is as follows: The pooling layer pools the results obtained by convolution. The pooling layer can reduce the number of parameters. The maximum pooling is used. The maximum pooling can extract the most obvious features. Its expression is as follows: and k =pool(and k-1 ) (10) Where pool() refers to the process of reducing the input tensor according to the pooling area size and pooling standard; The initial design of the loss function loss is expressed as: Where N represents the total number of training samples, γ n , V RF,n Represents the signal-to-noise ratio, channel matrix and simulated beamforming precoding matrix between the Nth sample, and then adds the focal loss, which is expressed as: loss=λ·(1-p t ) β ·loss1 (12) In which p is set t =loss / 20, β is the focusing parameter, which adjusts the weight reduction rate; λ is the weighting factor, which solves the balance problem; After the loss function designed by formula (12), the latest matrix V is obtained RF , to achieve good beamforming effect.
6. The method for training a wireless communication model based on Fourier spectrum using multi-feature beamforming according to claim 1 or 2, characterized in that: The process of step S5 is as follows: Use the test set data to test the trained beamforming model, judge the beamforming effect by spectrum efficiency, deploy the trained beamforming model to the test phase, use the test set to evaluate the model effect, and finally measure the model effect by spectrum efficiency value SE; The calculation formula of spectrum efficiency SE is: Where γ represents the signal-to-noise ratio, N t represents the number of transmitting antennas, h H represents the channel matrix, V RF 、V D They are analog and digital precoding respectively; The beamforming model is formed by combining steps S3 and S4 for training, and the effect of the model is measured according to the magnitude of the spectrum efficiency result.
