An intelligent beam alignment method based on wide and narrow beam mapping

Abstract

The application belongs to the technical field of deep learning and wireless communication, and particularly relates to an intelligent beam alignment method based on wide-narrow beam mapping. The application utilizes the correlation between wide beams and narrow beams and the powerful nonlinear representation capability of a convolutional neural network (CNN) to establish a mapping model from wide beam detection results to optimal narrow beams based on the CNN. In this way, only the wide beams need to be detected and the results input into the mapping model, and the optimal narrow beams can be inferred. In addition, by utilizing the correlation between the wide beams, the application further reduces the number of wide beams to be detected to infer the optimal narrow beams. Simulation results show that the beam detection algorithm proposed by the application is superior to the benchmark algorithm in terms of beam alignment accuracy, beamforming gain and throughput.

Description

Technical Field

[0001] This invention patent belongs to the fields of wireless communication and deep learning technology, specifically relating to an intelligent beam alignment method based on wide and narrow beam mapping. Background Technology

[0002] Compared to traditional sub-6G microwave communication, millimeter-wave communication has become a key technology for 5G and next-generation wireless communication due to its ability to provide richer bandwidth resources and faster data transmission rates. To find the transmit and receive beam pairs that match the strongest channel path of millimeter waves for data transmission, existing beam management techniques typically use beam training methods to achieve strict beam alignment.

[0003] Traditional beam management techniques typically employ exhaustive beam search (EBS) or hierarchical beam search (HBS) to obtain optimal transmit and receive beam pairs. However, when the number of beams is large, both traditional methods often incur significant beam training overhead. Therefore, there is a need to find efficient beam management techniques with low beam training overhead for millimeter-wave communication.

[0004] Intelligent beam training is a potential technology for reducing training overhead. Machine learning (ML)-based intelligent beam training can automatically extract and apply environmental or historical training information to limit the search area for beam training. Furthermore, because ML-based intelligent beam training outperforms traditional beam training in prediction accuracy, beam tracking, and beam fault recovery, it is being widely researched for its ability to identify and learn complex movement patterns and track environmental dynamics. Existing advanced technologies mainly include:

[0005] (1) Original DNN-based Beam Training (ODBT) algorithm. This algorithm is a deep learning algorithm for training a portion of narrow beams in a uniform space for narrow beam space. It uses a fully connected deep neural network (DNN) to predict the optimal narrow beam. However, due to the large number of narrow beams, the angle space covered by a small number of narrow beams during beam training is small, which leads to poor prediction accuracy. Additional narrow beam training is often required to improve accuracy. At the same time, because each neuron in a fully connected network is fully connected, stacking multiple layers will lead to excessively high computational cost and is prone to overfitting.

[0006] (2) Calibrated Beam Training (CBT) based on Convolutional Neural Network (CNN). To reduce overhead, this algorithm uses a significantly reduced number of wide beams (compared to all narrow beams) for beam training to predict the optimal narrow beam. Furthermore, to improve prediction accuracy, considering the continuity of user movement over time, the authors propose a beam training algorithm based on LSTM and CBT, LCBT. CBT utilizes all wide beams, leaving room for further reduction in the number of wide beams to be trained. LCBT involves additional narrow beam training and, due to the need for continuous historical information from the UE, is unsuitable for scenarios such as initial user beam connection, random user movement, and multi-user scheduling, where base stations often cannot obtain continuous historical user information. Simultaneously, while dual-channel CNNs perform well in intelligent signal recognition due to their focus on I / Q component features for distinguishing different signals, beam training often prioritizes signal strength for optimal beam differentiation. The dual-channel CNNs used in CBT and LCBT actually reduce training efficiency and accuracy. Summary of the Invention

[0007] Based on the above background, this invention proposes an intelligent beam alignment algorithm based on Beam Mutual Information and Wide-to-Narrow Beam Mapping (BMI-WNBM). It maps the detection results of wide beams to the optimal narrow beam. Furthermore, by utilizing the correlation between wide beams, the number of detected wide beams is further reduced to infer the optimal narrow beam. Specifically, the algorithm first uniformly divides the entire wide beam space into multiple regions, each containing the same number of wide beams. Then, the training data is preprocessed to calculate the mutual information between each wide beam and the optimal wide beam, called Beam Mutual Information (BMI). Finally, the algorithm selects the wide beam with the largest mutual information value in each region to form a wide beam training set, which is then input into a CNN for predicting the optimal narrow beam.

[0008] System Model A

[0009] This invention considers a downlink millimeter-wave communication system, consisting of a system with N n It consists of a base station (BS) with a uniform linear array (ULA) of antennas and a user equipment (UE) with an omnidirectional antenna.

[0010] This invention employs the well-known Saleh-Valenzuela (SV) geometric channel model, and the system's channel matrix h can be expressed as:

[0011]

[0012] Where α represents the large-scale path loss, N n L represents the number of antenna elements and L represents the number of paths. This represents the complex gain of the line-of-sight path (LoS). It is the complex gain of the l-th non-line-of-sight (NLoS) path. This represents the angel of departure (AoD) of the Loss path. Indicates the departure angle of the NLoS path, (·) H This indicates the conjugate transpose. It can be represented as

[0013]

[0014] in, Let λ represent the departure angle of the l-th path, λ represent the signal wavelength, and d represent the antenna spacing, typically d = λ / 2.

[0015] Analog beamforming is a low-cost and practical solution in which all antennas share a single radio frequency chain, so this invention considers analog beamforming and first defines the codebook for analog beamforming.

[0016] This invention considers the Discrete Fourier Transform codebook It is composed of M n It consists of a predefined codeword, that is, Each codeword corresponds to a beam direction φ m It can be represented as

[0017] f m =a(N n ,φ m )

[0018] Among them, the beam direction φ m exist Data collection within the specified range, i.e.

[0019]

[0020] Where Φ represents the entire beam space.

[0021] If f m If selected as an analog precoder, the corresponding received signal y m It can be represented as

[0022]

[0023] Where P is the transmission power, s is the transmission signal, and there is... It is a noise vector, σ 2 Indicates noise power.

[0024] In practical millimeter-wave communication scenarios, each data frame (DF) consists of two phases: the beam training phase (BTP) and the data transmission phase (DTP). This is typically achieved by reducing the time T of the beam training phase. BTP To increase T DTP And T BTP With the selected number of detection beams M T Closely related, T BTP It can be represented as

[0025] T BTP =M T ·T BSW

[0026] Among them, T BSW It is the duration of a beam sweep (BSW) process.

[0027] After beam training, the predicted codewords can be obtained. The corresponding effective Shannon capacity (ESC) is then... It can be represented as

[0028]

[0029] Where W is the system bandwidth, T BTP It is the beam training time, T DF The duration of a data frame is represented by P, where P is the transmit power and h is the channel vector. For the predicted codeword vector, σ 2 This is the noise power. By reducing the number of probe beams M... T This can reduce beam training time and further increase the effective Shannon capacity (ESC).

[0030] B. Design of Intelligent Beam Alignment Technology Based on Wide and Narrow Beam Mapping

[0031] 1) Main Ideas and Algorithm Framework

[0032] Since the number of wide beams is usually significantly less than the number of narrow beams, beam training is performed by probing wide beams in order to reduce overhead.

[0033] The signal strength of different wide beams varies, and the strength is represented by the signal magnitude. Wide beams closer to the user have larger magnitudes, while those farther away have smaller magnitudes. The magnitudes of all detected wide beams can form a mode set. Different user locations correspond to different optimal narrow beams, and also to different wide beam mode sets (i.e., there is a certain mapping relationship between the wide beam mode set and the optimal narrow beam). Therefore, the optimal narrow beam can be predicted by detecting all wide beams and inputting the wide beam mode set into a mapping model.

[0034] Furthermore, it was observed that in specific environments, there is typically a strong correlation between each wide beam and the optimal wide beam. This correlation can be represented by Beam Mutual Information (BMI). The BMI between each wide beam and the optimal wide beam is calculated; a larger BMI indicates a stronger correlation between the wide beam and the optimal narrow beam. Therefore, the optimal narrow beam can be predicted by detecting some wide beams with larger BMIs. To achieve this, the wide beam signal needs to be acquired, and a mode set of some wide beams needs to be constructed using the BMI.

[0035] 2) Beam detection model

[0036] Considering that a wide beam contains c narrow beams, this can be obtained from the wide beam codebook. A wide-beam codeword, Indicates rounding down. (By opening...) One antenna, the m-th wide-beam codeword It can be represented as

[0037]

[0038] Similarly, the direction of the m-th wide beam It can be obtained in the following ways

[0039]

[0040] Φ represents the beam space. Given the corresponding wide-beam channel matrix... Corresponding received signal It can be represented as

[0041]

[0042] If all wide beams are detected for each user location, then all received wide beam signals can be obtained and a wide beam signal mode set, i.e., wide beam mode set Y, can be constructed. w ,Right now

[0043]

[0044] in, This represents the received signal corresponding to the m-th wide beam. The modulus.

[0045] 3) Implementation steps of intelligent beam alignment technology based on wide and narrow beam mapping

[0046] S1. Construct the corresponding wide-beam training set.

[0047] This section calculates the beam mutual information of each wide beam based on the concept of mutual information. In this way, the correlation between wide beams is utilized to further reduce the number M of wide beams to be trained. T .

[0048] ① Calculation of beam mutual information

[0049] This section first uses the k-nearest neighbor method to calculate the beam mutual information between each wide beam and the optimal wide beam. If the optimal wide beam in the t-th data frame (DF) It is also the j-th wide beam, i.e., m opt If (t) = j, then the BMI value of the j-th wide beam in the t-th data frame out of a total of S DFs can be expressed as:

[0050] I j (t)=ψ(S)-ψ(N) j (t))+ψ(k)-ψ(n j (t))

[0051] Where ψ(x) represents the digamma function, defined as When x > 0, ψ(x) is strictly increasing. The following section explains how to obtain N. j (t) and n j (t). N j (t) represents the number of times the wide beam index equals the current optimal wide beam index j within S data frames, i.e.,

[0052]

[0053] Where, m opt (t) represents the index of the optimal wide beam in the t-th data frame. count (·) represents a counting function. If its input is true, the output is 1; otherwise, the output is 0.

[0054] n j The meaning of (t) will be explained below. To obtain n jHere, we need to define the distance between the wide-beam signal magnitudes when the j-th wide beam is the optimal wide beam in different data frames (t). If we represent the wide-beam signal magnitudes in the current data frame t (other data frames τ) as... The distance can then be expressed as

[0055]

[0056] The non-negative distance values ​​obtained from the above formula will be used to construct a distance set ordered in ascending order, i.e.

[0057]

[0058] Where, d j (t,τ i ) <d j (t,τ i+1 Because the k-nearest neighbor method is used, therefore in Select the k-th distance n is used as a threshold to obtain j (t), which represents the current optimal wide beam j in less than The number of all neighboring nodes within the range. This can be calculated using the following formula:

[0059]

[0060] Considering all data frames in the sample, the average BMI can be obtained as follows:

[0061]

[0062] ② Wide beam selection based on BMI

[0063] The calculations in the previous section can yield the result related to the wide-beam signal mode set Y. w The corresponding BMI set, i.e.

[0064]

[0065] To ensure the selection of wide-beam training modules is spatially global, a BMI-based approach is used to design the wide-beam training modules. First, the wide-beam space is uniformly divided into M... T A region, in other words, by M T The BMI set I, composed of subsets, can be represented as:

[0066]

[0067] Among them, Z i Represents the i-th region. Depend on Composed of BMI values, i.e.

[0068]

[0069] Then we can obtain the index s of the maximum beam mutual information value in each subset of beam mutual information. i ,Right now,

[0070]

[0071] Finally, the wide-beam training mode set can be constructed using the index obtained from the above formula, as follows.

[0072]

[0073] Figure 1 A specific diagram of beam selection is given, in which the wide, dark sector represents the wide beam selected for beam training. At this time, the beam mutual information of the wide beam is the maximum value of the current region (black dashed sector).

[0074] S2. Construct the corresponding convolutional neural network.

[0075] In this invention, a convolutional neural network (CNN) is used to extract the hidden correlation between the wide-beam training mode set and the optimal narrow-beam index for partially detecting wide beams. The wide-beam training mode set is designed as a 1×M T The image is used as input to the CNN, and the probability of each narrow beam index is the output of the CNN. The proposed CNN model includes input units, convolutional units, flattening units, and prediction units, such as... Figure 2 As shown.

[0076] ① Input Unit

[0077] exist The modulus in it has a large dynamic range, when the modulus value is in Before being input into a convolutional unit, the elements are first processed by the pair with the largest modulus. Normalization can be expressed as

[0078]

[0079] ② Convolutional unit

[0080] Then, the normalized wide-beam training module set is... The input is fed into a convolutional unit, where L is deployed. C A convolutional layer, used to extract from... Feature extraction is performed in the process. The data relationship between layer l and layer (l-1) can be represented as follows:

[0081]

[0082] in, It is the data of the i-th feature channel of the l-th layer, K l It is the convolution kernel of the l-th layer, B l This is the bias vector of the l-th layer. '*' indicates a convolution operation, and N... l-1 f is the number of feature channels in layer l-1. δ (·) is an activation function that has non-linear fitting capabilities.

[0083] ③ Flattening unit

[0084] Following the convolutional unit is the flattening unit, which includes pooling and flattening layers. The pooling layers are used to downsample features from the output of the convolutional unit.

[0085]

[0086] Among them, P i f represents the i-th channel of the pooling layer. pool The · symbol represents pooling operations, including max pooling and average pooling. After the pooling layer, a flattening layer is used to obtain the flattened feature map.

[0087]

[0088] Among them, F 0 This represents the flattened layer, also known as the initial fully connected layer, f. flatten (·) is a flattening operation that can convert multidimensional data into one-dimensional data.

[0089] ④ Prediction Unit

[0090] By L F A prediction unit consists of a fully-connected (FC) layer and a softmax activation function layer. The l-th fully-connected layer takes the flattened feature map as input and can be represented as:

[0091]

[0092] Among them, F l W represents the l-th fully connected layer. l f represents the weight vector of the l-th fully connected layer. δ (·) represents the activation function. Let represent the bias vector of the l-th fully connected layer. At the end of the last fully connected layer, the softmax activation function is used to calculate the prediction probability, i.e.,

[0093]

[0094] in, This represents the output probability of the i-th narrow beam. This indicates the last fully connected layer. It is the total number of elements in the last fully connected layer, which is equal to the number of narrow beams, M. n ,Right now The predicted optimal narrow beam index can then be obtained using the following formula, i.e.,

[0095]

[0096] Where, m pre This represents the narrow beam index corresponding to the highest probability obtained, and arg max(·) represents the index corresponding to the maximum value.

[0097] S3. Perform offline training on the constructed CNN network.

[0098] After the convolutional neural network (CNN) is constructed, it needs to be trained using the training data set constructed in part S1. The optimal narrow beam index collected by EBS is used as the label of the CNN. This invention uses the cross-entropy loss function f. loss To train a convolutional neural network, its mathematical expression is:

[0099]

[0100] Where, if the i-th narrow beam is the actual optimal narrow beam, then p i =1; otherwise, p i =0. Using the Adam optimizer with a learning rate of 0.0005, the dropout rate of the neural network is 0.3.

[0101] S4, Online Prediction

[0102] After the CNN is trained, the wide-beam training set constructed in part S1 is input into the CNN to predict the optimal narrow beam. Then the base station transmits data according to the predicted optimal narrow beam.

[0103] The beneficial effects of this invention are that it utilizes the correlation between wide and narrow beams and the powerful nonlinear representation capabilities of convolutional neural networks (CNNs) to establish a CNN-based mapping model from wide beam detection results to the optimal narrow beam. In this way, only the wide beam needs to be detected and the result input into the mapping model to infer the optimal narrow beam. Furthermore, by utilizing the correlation between wide beams, this invention further reduces the number of wide beams detected to infer the optimal narrow beam. Simulation results show that the beam detection algorithm proposed in this invention outperforms benchmark algorithms in terms of beam alignment accuracy, beamforming gain, and throughput. Attached Figure Description

[0104] Figure 1 This is a schematic diagram of the wide beam selection of the present invention.

[0105] Figure 2 This is a schematic diagram of the CNN structure used in this invention.

[0106] Figure 3 This diagram illustrates a comparison of the accuracy of this invention with CBT and ODBT technologies under different beam training numbers.

[0107] Figure 4 This is the cumulative distribution function curve of normalized beamforming for the present invention and CBT and ODBT technologies.

[0108] Figure 5 This is the cumulative distribution function curve of the effective Shannon capacity (ESC) of the present invention and CBT and ODBT technologies. Detailed Implementation

[0109] The practicality of this invention will be illustrated next through simulation examples. First, system parameters and simulation parameters are provided, followed by simulation examples to demonstrate the performance of the algorithm. Furthermore, the algorithm will be compared with existing ODBT and CBT algorithms to illustrate its superior performance.

[0110] This invention considers a millimeter-wave wireless communication system where the user's movement range is a semi-circular area with a radius of 2m-30m, with the base station located at the center. The base station has a transmission power of 30dBm, a carrier frequency of 60GHz, a system bandwidth of 2.16GHz, and a noise power spectral density of -174dBm / Hz. Furthermore, T... DF ,T BSW The time intervals are 10ms and 15us, respectively. The number of wide beams is 64, the number of narrow beams is 256, and the number of beams to be trained is 8, 16, 32, and 64.

[0111] During the offline training phase, this invention employs Exhaustive Beam Search (EBS) to collect training data. Following the method in S1, wide-beam received signals with fixed directions are selected to form a wide-beam training module set. Simultaneously, the optimal narrow beam obtained from EBS is used as the training label. The wide-beam training module set and the training label together constitute a set of training sample data, totaling 5000 sets during the training phase. After the network is trained, the base station switches to the online prediction phase. During the online testing phase, this invention only needs to detect the wide-beam directions determined in S1 to obtain the corresponding received signals to construct the wide-beam training module set. This set is then input into the trained CNN to directly predict the optimal narrow beam for data transmission, with 20000 test samples.

[0112] Next, we introduce the structure and parameters of the convolutional neural network. Following the input unit is the convolutional unit, which contains three convolutional layers. Each layer has a kernel size of (1, 3), and the activation function is the Tanh function. The number of input and output channels for the three convolutional layers are (1, 16), (16, 64), and (64, 256), respectively. After the convolutional units is the flattening unit, which contains one max-pooling layer and one flattening layer. The max-pooling layer has a kernel size of (1, 2), and the flattening layer transforms the previous multi-dimensional data into one dimension. Following this is the prediction unit, which contains two fully connected layers. The first fully connected layer has an input / output channel count of (256, M...). T The activation function is Tanh (-6) / 2,512); the number of input and output channels in the second fully connected layer is (512,256); and the activation function in the last fully connected layer is Softmax.

[0113] Figure 3 The changes in test accuracy of different algorithms under different numbers of beams trained are shown. It can be seen that the test accuracy of different algorithms improves with the increase of the number of beams trained. It can also be seen that, under different numbers of beams trained, the test accuracy of the algorithm proposed in this invention is higher than that of the baseline algorithm. This advantage is more significant when the number of beams trained is small. Furthermore, it is worth noting that when the number of beams trained is 16, the test accuracy of the algorithm proposed in this invention is even higher than that of the baseline algorithm when the number of beams trained is 64.

[0114] exist Figure 4 The diagram shows the normalized beamforming gain G in various schemes. N The cumulative distribution function (CDF) curve. Normalized beamforming gain reflects how close the predicted beam is to the optimal beam. For example... Figure 4 As shown, the normalized beamforming gain of different algorithms increases with the increase of the number of beams trained. Furthermore, the normalized beamforming gain of the algorithm proposed in this invention is higher than that of the baseline algorithm under different numbers of beams trained. It is worth noting that the test accuracy of the CBT algorithm with 64 beams trained is lower than that of the proposed algorithm with 16 beams trained, but its normalized beamforming gain is higher in this case. This is because the dual-channel CNN model used in the CBT algorithm often results in predicted beams that are suboptimal but close to optimal.

[0115] Figure 5The cumulative distribution function (CDF) curves of the Effective Shannon Capacity (ESC) for different algorithms under different beam training numbers are shown. It can be observed that, in terms of ESC, the proposed algorithm performs best with a training number of 16. This is because ESC depends not only on prediction accuracy and beamforming gain, but also on beam training overhead (i.e., the number of beams trained). Therefore, while increasing the number of beams trained can improve test accuracy and beamforming gain, it also increases beam training overhead, potentially leading to a decrease in ESC.

Claims

1. A smart beam alignment method based on wide and narrow beam mapping for millimeter-wave communication systems, defining the system as including a base station and user equipment, wherein the base station uses... A uniform linear array (ULA) of antennas is used. The user equipment has one omnidirectional antenna. When transmitting a narrow beam, the number of antennas that need to be activated is... , This is the number of antennas that need to be turned on when transmitting a wide beam. Considering that a wide beam contains c narrow beams, the expression is: For each user location, all wide beams are detected, all received wide beam signals are obtained, and a wide beam signal mode set is constructed as follows: ， in, As shown in the following formula, ， in, It's the transmission power. It is a wide-beam channel matrix. It is the m-th wide-beam codeword. It is transmitting a signal. It is a noise vector. It is the number of wide codewords in the constructed wide-beam codebook, and its expression is: ,in The number of narrow-beam codewords in the constructed narrow-beam codebook; The method is characterized by comprising the following steps: S1. Construct a wide-beam training set: The k-nearest neighbor method is used to calculate the beam mutual information between each wide beam and the optimal wide beam, and the optimal wide beam in the t-th data frame is defined. It is also the j-th wide beam, that is The beam mutual information value of the j-th wide beam in the t-th data frame out of a total of S data frames is: ， in, This represents the digamma function, when x >

0. Strictly incremental, This represents the number of times the wide beam index equals the current optimal wide beam index j within S data frames, i.e.: ， in, Indicates the first Index of the optimal wide beam in each data frame This represents a counting function that outputs 1 if its input is true, and 0 otherwise. This indicates that the current optimal wide beam j is less than The number of all neighboring nodes within the range is calculated as follows: The distance between the wide-beam signal magnitudes is defined as follows: when the j-th wide beam is the optimal wide beam in different data frames: ， in The magnitude of the wide-beam signal in the current data frame t. For other data frames The wide-beam signal magnitude in the data; the obtained non-negative distance values ​​are then used to construct a distance set arranged in ascending order: ， in, ,exist Select the k-th distance This is used as a threshold to obtain: ， Considering all data frames in the sample, the average beam mutual information is obtained as follows: ， Thus, a wide-beam signal mode set is obtained. The corresponding beam mutual information set, namely: ， A wide-beam training mode set is designed using a beam mutual information-based method. Specifically, the wide-beam space is uniformly divided into... Each region, by Beam mutual information set composed of subsets Represented as: ， in, Represents the i-th region. Depend on It consists of beam mutual information values, i.e. ， This yields the index of the maximum beam mutual information value in each beam mutual information subset. ,Right now ， The wide-beam training mode set is constructed using the obtained indexes as follows: ； S2. Construct a convolutional neural network, including input units, convolutional units, flattening units, and prediction units; the wide-beam training module set is the input of the convolutional neural network, and the output of the convolutional neural network is the probability of each narrow-beam index; The input unit is used for... Normalize: ， The input to the convolutional unit is a normalized set of wide-beam training modules. Convolutional units are deployed with A convolutional layer, used to extract from... Feature extraction is performed in the process, and the data relationship between the l-th layer and the (l-1)-th layer is as follows: ， in, It is the data of the i-th feature channel of the l-th layer. It is the convolution kernel of the l-th layer. It is the bias vector of the l-th layer. This represents the convolution operation. It is the number of feature channels in layer l-1. It is an activation function with nonlinear fitting capability; The flattening unit includes a pooling layer and a flattening layer. The pooling layer is used to downsample features from the output of the convolutional unit. ， in, This represents the i-th channel of the pooling layer. This is a pooling operation. After the pooling layer, a flattening layer is used to obtain the flattened feature map. ， in, This indicates the flattened layer, which is also the initial fully connected layer. It is a flattening operation that converts multidimensional data into one-dimensional data; The prediction unit consists of It consists of a fully connected layer and a softmax activation function layer. The l-th fully connected layer receives the flattened feature map as input. ， in, This indicates the l-th fully connected layer. This represents the weight vector of the l-th fully connected layer. This represents the activation function. Let represent the bias vector of the l-th fully connected layer; at the end of the last fully connected layer, the softmax activation function is used to calculate the predicted probability: ， in, This represents the output probability of the i-th narrow beam. This indicates the last fully connected layer. It is the total number of elements in the last fully connected layer, equal to the number of narrow beams. ,Right now The predicted optimal narrow beam index is obtained using the following formula: ， in, This indicates the narrow beam number corresponding to the highest probability obtained. This indicates the index corresponding to the maximum value; S3. Train the constructed convolutional neural network: The constructed convolutional neural network is trained using the training set built in S1. The optimal narrow beam index collected by exhaustive beam search is used as the label of the convolutional neural network, and the cross-entropy loss function is used to train the convolutional neural network. ， Where, if the i-th narrow beam is the actual optimal narrow beam, then ;otherwise, The Adam optimizer with a learning rate of 0.0005 and a dropout rate of 0.3 were used to obtain the trained convolutional neural network. S4. The trained convolutional neural network is used to predict the optimal narrow beam, and then the base station transmits data according to the predicted optimal narrow beam.

Images