A method for excavating space-time sequence characteristics of width learning base station cooperation beam alignment

Abstract

The application provides a width learning base station cooperative beam alignment method for excavating space-time sequence characteristic information. Each base station utilizes a local width learning network, based on a distributed learning training architecture, to cooperatively learn the mapping relationship of the perception beam response of multiple base stations to corresponding narrow beams, so as to complete the prediction of the optimal downlink transmission narrow beam. In order to excavate the perception beam information of the past several sampling time points and effectively predict the narrow beam at the current time point, the application uses the time sequence of the perception beam response to obtain the characteristic nodes of the width learning network, and designs a cyclic width learning method. Further, a gating width learning method is proposed, and a forgetting gate is designed to control the learned sequence information. The application solves the adaptation problem of the model to the fast time-varying scene, realizes the lightweight of the model on the base station side, can reduce the central processing unit computing pressure, and achieves a reasonable compromise between the cooperation overhead and the beam selection prediction performance, and realizes the beam alignment with high performance under the condition of a small sample.

Description

Technical Field

[0001] This invention relates to a method for cooperative beam alignment of millimeter-wave MIMO base stations based on width learning and adaptive sensing beams, belonging to the fields of wireless communication network optimization and intelligent communication. Background Technology

[0002] The digital transformation is an inevitable trend, and future communications need to meet requirements such as ultra-high throughput, ultra-high bandwidth, ultra-low latency, and high reliability. Due to the large bandwidth advantage of the millimeter-wave band, and the fact that massive MIMO (Multiple Input Multiple Output) beamforming can effectively compensate for millimeter-wave propagation fading using array gain, millimeter-wave massive MIMO technology is considered one of the key technologies for future communications. Considering that millimeter-wave massive MIMO requires narrow beams for data transmission, it is highly sensitive to channel shielding effects. Cooperative multi-point transmission has become the main mode for achieving performance gains in millimeter-wave massive MIMO systems. However, the potentially large array gain requires high alignment of the beam and transmission path; otherwise, communication quality will be lost. Moreover, the frequent changes in millimeter-wave channels require frequent beam training at the transceiver end.

[0003] Existing research has demonstrated that uplink pilot signals transmitted by users can help base stations perceive environmental characteristics between themselves and users. By utilizing the wide-beam responses received by multiple base stations, an implicit representation of the user's location can be obtained. Since beam selection is determined by the beam-domain equivalent channel state information (CSI), and CSI information is directly related to the user's location, a mapping relationship can be established between the user-perceived beam responses received by multiple base stations and beam selection. Current narrow-beam prediction based on perceived beams mainly utilizes the perceived beam response at the current moment. However, in real-world communication environments, the location of mobile users generally exhibits spatial continuity over a period of time, and the interval for perceived beam training may be longer than the interval for transmission beam updates. Therefore, designing a suitable scheme to mine perceived beam information from several past sampling moments to effectively predict the narrow beam at the current moment is the focus of this invention. Summary of the Invention

[0004] Objective: To explore the changing patterns of user positions in real-world dynamic scenarios involving continuous user movement and rapid time-varying conditions, thereby optimizing beam alignment methods. This invention proposes a low-overhead cooperative beam alignment scheme. To ensure beam alignment is always applicable to the current rapid time-varying scenario, the network structure of the width learning method is improved. The feature nodes of the width learning network are obtained using time series of perceived beam responses, and a Recurrent Broad Learning (RBL) method is designed. To more effectively mine spatiotemporal sequence feature information, a Gated Broad Learning (GBL) method is designed, drawing inspiration from Long Short-Term Memory (LSTM) prediction methods. Under small sample conditions, this invention addresses the problems of traditional beam alignment methods, such as the inability to predict network performance, slow algorithm convergence, and inability to quickly provide beam configuration decisions. The proposed algorithm obtains the gain of spatiotemporal sequence feature information on beam selection performance, aiming to solve the model's adaptation problem to rapid time-varying scenarios. It achieves lightweight AI models on the base station side, reducing the computational burden on the central processing unit while achieving a reasonable trade-off between cooperative overhead and beam selection prediction performance, thus achieving high-performance beam alignment under small sample conditions.

[0005] Technical solution: A width-learning base station cooperative beam alignment method for mining spatiotemporal sequence features, comprising the following steps:

[0006] S1. Construct a communication scenario for continuous user movement, build a millimeter-wave MIMO system model for multi-base station cooperative transmission, and construct a beam alignment problem model with the goal of maximizing system efficiency and rate.

[0007] S2 proposes a low-overhead width learning cooperative beam alignment scheme. Each base station utilizes its local width learning network and, based on a distributed learning training architecture, collaboratively learns the mapping relationship between the perceived beam response of multiple base stations and the corresponding narrow beam to complete the prediction of the optimal downlink transmission narrow beam.

[0008] S3. In order to make beam alignment always applicable to the current fast time-varying scenario, the network structure of the width learning network was improved. The feature nodes of the width learning network were obtained by using the time series of the perceived beam response, and the RBL method was designed.

[0009] S4. In order to more effectively mine spatiotemporal sequence feature information, we draw on the LSTM prediction method, add a gating mechanism to the RBL method, and design a forget gate to control the learned sequence information.

[0010] S5, each base station inputs the collected user wide-beam response into the local mapping network and interacts with the central processing unit to predict its own beam selection.

[0011] Furthermore, step S1 specifically includes:

[0012] In a millimeter-wave cell-free MIMO downlink system, the system bandwidth is B. w There are K subcarriers, B base stations, and M antennas at each base station. These base stations cooperate to serve U single-antenna users via Orthogonal Frequency Division Multiple Access (OFDMA). The sets of base stations, users, and subcarriers are represented as follows: and user The set of subcarriers is in This represents the size of the subcarrier set for user u, i.e., the number of subcarriers. For practical considerations, each base station uses U radio frequency chains and hybrid beamforming for downlink transmission. A typical scenario involves a single base station using a single beam to serve one user, and multiple users moving randomly within a certain area. From the base station... To users The antenna-subcarrier downlink channel is Where L represents the number of distinguishable propagation paths, α b,u,l ,τ b,u,l ,θ b,u,l and φ b,u,l It is the complex gain, propagation delay, azimuth angle, and elevation angle of path l, f k Indicates subcarrier The center frequency, It is the array guidance vector.

[0013] Considering that downlink transmission is based on typical hybrid precoding, the analog beams of each base station are selected from the standard Discrete Fourier Transform (DFT) codebook. Base station b selects the first from F List As an analog beam Serving user u. Define a matrix composed of analog beams from multiple base stations arranged in a block diagonal configuration. The collaborative center control node can collect the beam domain equivalent CSI of each base station and, based on the Maximum Ratio Transmission (MRT) criterion, design the beam domain on the subcarriers. The digital precoding vector of the service user u is in

[0014] User u on subcarrier The received signal on can be represented as in For data symbols, For user u, the receiver noise on subcarrier k, σ 2 This is the noise power. It satisfies... Furthermore, the system employs equal power allocation between users and between subcarriers, which has Where P is the total transmit power of the base station. The received SINR of user u on subcarrier k can be expressed as: The time-varying tracking period is defined as T, meaning the system retrains the beam to update the precoding design every T periods. The first T periods... r During this time, the base station performs beam training to search for an optimal beam. For example, a beam is found that can meet the maximum received signal strength, and the remaining time is used to transmit data using the trained beam. Therefore, the actual effective rate of user u... It can be represented as:

[0015]

[0016] The beam selection problem can be formulated as maximizing the system's efficiency and rate through beam selection. Optimization issues:

[0017]

[0018] The following conversion can be performed:

[0019]

[0020] We obtain a low-complexity suboptimal solution to the original problem. For the transformed problem, if we have the complete beam domain CSI amplitude, i.e. This can maximize the narrow beam equivalent rate. Because the required beam training time is relatively large, that is A smaller value may lead to poor overall performance. How can we ensure that c... b,u Minimize T as much as possible without significant loss. r This became the key to solving the problem.

[0021] In communication scenarios, continuous user movement inevitably forms a reasonable trajectory. This invention applies a two-dimensional background field cellular automaton model to divide the communication scenario into a uniform grid, establishing a user movement model. The basic cellular automaton consists of four parts: cells, cell state space, cell neighbors, and cell rules. The neighbor type is selected as Moore type, meaning that a user located in the central cell can move along the eight surrounding directions, such as... Figure 2As shown. Assuming the user's current position is (i,j), the probability of the user's next position at any given time satisfies P. i-1,j-1 +P i-1,j +P i-1,j+1 +P i,j-1 +P i,j +P i,j+1 +P i+1,j-1 +P i+1,j +P 1+1,j+1 =1. In real-world dynamic scenarios, user movement generally follows a certain trajectory, and the speed is limited, making it impossible for there to be a large-scale instantaneous transfer.

[0022] Furthermore, step S2 specifically includes:

[0023] Each user During the training phase, send (M+1) repeated pilot signals. Each base station Simultaneously perform beam training, using sensing wide beam vectors. And switching between M candidate simulated beam vectors in different directions Received in sequence, This indicates the uplink training power. (Each base station) In subcarrier The received signal of the i-th training sequence obtained above is in This is the received noise vector of the b-th base station on the k-th subcarrier. After pilot matching, the base station can obtain the combined signal received using different beam vectors. Estimate:

[0024]

[0025] In practical systems, CSI resolution for every subcarrier may not be feasible. Furthermore, the beam response varies little across consecutive subcarriers. (Definition) and These are respectively the numbering set of the subcarrier groups belonging to user u and the subcarrier groups belonging to... The set of subcarrier numbers in a group. Then, a subcarrier group... The beam response is:

[0026]

[0027] During the offline phase, each base station calculates the equivalent rate index corresponding to each candidate narrow beam vector in N training samples. The expression is as follows:

[0028]

[0029] in Each base station will receive the sensing beam response and the equivalent rate index of each candidate beam vector All data is sent to the Central Processing Unit (CU). A schematic diagram of a millimeter-wave MIMO base station cooperative beam alignment model is shown below. Figure 3 As shown. Define the nth sample. For base stations Received from user The perceived beam response, where ∠ denotes the complex phase operator, n = 1, ..., N. Then, the cooperative base station model obtains the beam response from the user. The sensing beam response, its nth sample is definition For users The candidate beam response samples. To reduce the learning difficulty, for... Perform one-hot encoding to obtain in

[0030]

[0031] After data arrangement, the label of the width learning network model corresponding to the nth sample is: in Taking the input of the first Q wide-beam sampling points as the input at the current time, then the N sample sets are:

[0032]

[0033] Based on the width learning model, such as Figure 4 As shown, each base station Utilizing local sensing beam response Mapping I group of feature nodes With J group enhanced nodes Where F and E represent the number of feature nodes and augmentation nodes in each group, respectively, and Z... b,u =[Z b,u,1 Z b,u,2 ,…,Z b,u,I [H] is the concatenation matrix of I sets of feature nodes. b,u =[H b,u,1 H b,u,2 ,…,H b,u,J [ ] represents the concatenated matrix of J groups of enhancement nodes. The joint feature enhancement nodes of the base station local width learning network are... Therefore, based on the joint feature enhancement node of multiple base stations and tags The model optimization problem can be expressed as:

[0034]

[0035] in It is an affine transformation matrix.

[0036] To reduce the computational burden on the CU and lower the transmission overhead of the joint feature enhancement node, A is utilized. u With the characteristic of vertical partitioning of data features, the original centralized processing problem is transformed into a distributed optimization problem based on feature dimension partitioning, which can be expressed as:

[0037]

[0038] in These are each base station The local affine transformation matrix,

[0039] Using the ADMM algorithm for solving data with vertical partitioning features, the above problem can be transformed into:

[0040]

[0041]

[0042] in For the introduced auxiliary variable matrix, we have As can be seen, unlike the data partitioning of the sample space, in this problem, multiple base stations share the same user sample space, but each base station only collects a portion of the wide-beam response features from the user to its own location. Therefore, building a more accurate model requires jointly utilizing the features of all base stations. The update steps can be derived into the following three iterative equations:

[0043]

[0044] in The parameter ρ is the penalty coefficient, ρ>0; I IF+JE It is an identity matrix with size and Same; O u W is the standardized dual variable; t is the iteration number. All base stations participating in the federated learning training are weighted and initialized. b,u (0) = O, O u (0) = 0. Repeat the above iterative process t. max This approach can solve the distributed optimization problem of data with partitioning characteristics along feature dimensions, enabling implicit sharing of feature space among base stations, such as... Figure 5 As shown.

[0045] Furthermore, step S3 specifically includes:

[0046] To ensure beam alignment is always applicable to current fast time-varying scenarios, the network structure for width learning was improved, and the RBL method was designed, such as... Figure 6 As shown, the time series of the perceptual beam response is used to obtain the feature node Z of the wide learning network. b,u,i :

[0047]

[0048] Z b,u,0 It is a zero matrix; and represents the connection weights and biases of the feature generation network, respectively; these are randomly generated and not trainable; e i This represents the i-th layer of the feature layer; the diagonal matrix weights γ b,u,i It is also randomly generated and cannot be trained. This represents a column vector with all elements equal to 1. φ(·) represents the ReLU linear activation function. These feature nodes are further mapped to J groups of augmentation nodes H. b,u,j ,Right now

[0049]

[0050] in and h represents the connection weights and biases of the randomly generated feature enhancement network, respectively. j Let denote the j-th layer of the enhancement layer, and ξ(·) denote the nonlinear activation function Tansig. Then, the joint feature enhancement node of the base station local width learning network is...

[0051] Furthermore, step S4 specifically includes:

[0052] To more effectively mine spatiotemporal sequence feature information, a GBL method is proposed based on RBL and drawing on the LSTM structure, such as... Figure 7 As shown. An additional forget gate is designed to control the learned sequence information, such as... Figure 8 As shown, the step size s of the forget gate is set, and feature nodes Z are generated. b,u,i i = 1, ..., I:

[0053]

[0054] in and γ and γ represent the connection weights and biases of the feature generation network, respectively. They are randomly generated and not trainable. The diagonal matrix weights γ b,u,i It is also randomly generated and cannot be trained. This represents a column vector with all elements equal to 1. φ(·) represents the ReLU linear activation function. Similarly, these feature nodes are further mapped to J groups of augmentation nodes H. b,u,j ,Right now

[0055]

[0056] in and Let represent the connection weights and biases of the randomly generated feature enhancement network, respectively, and ξ(·) represent the nonlinear activation function Tansig. Then, the joint feature enhancement nodes of the base station local width learning network are:

[0057] Furthermore, step S5 specifically includes:

[0058] During offline training, Regression-Based Learning (RBL) is first used to obtain a basic model, and the number of layers for feature nodes and augmentation nodes is tuned. Then, Gaussian Glycol (GBL) is used to tune the forget gate step size to further improve effectiveness. When the system enters the online execution phase, U users first send uplink pilot signals to perform sensing beam training, with a time cost of T. b Base station Using local sensing beam reception, the sensing beam response is obtained, i.e. Using the first Q wide-beam sampling points x b,u Obtain the joint output of the feature nodes and the augmentation nodes, i.e. After affine transformation, the predicted index values ​​of the simulated beam to be selected are obtained. Based on this predicted value, the location of each base station can be determined. Its analog beam number Furthermore, the CU selects the optimal downlink transmission beam for the base station based on the MRT principle, enabling the base station to perform subsequent downlink data transmission. The effective efficiency of user u during the online execution phase is...

[0059]

[0060] Among them, T b The training time of the sensing beam and the transmission time of the narrow beam are T and B, respectively. w This indicates the system bandwidth. In uplink training based on OFDMA mode, multiple users and BSs can perform beam training simultaneously.

[0061] Beneficial Effects: Compared with existing technologies, the technical solution of this invention has the following beneficial effects: This invention proposes a width-learning base station cooperative beam alignment method for mining spatiotemporal sequence features. It proposes a low-overhead cooperative beam alignment scheme. To ensure beam alignment is always applicable to the current fast-changing scenarios, the network structure of the width-learning method is improved. The time series of perceived beam responses is used to obtain the feature nodes of the width-learning network, and the RBL method is designed. To more effectively mine spatiotemporal sequence feature information, the GBL method is designed, drawing inspiration from the LSTM method. Under small sample conditions, it solves the problems of traditional beam alignment methods, such as the inability to predict network performance states, slow algorithm convergence speed, and inability to quickly provide beam configuration decisions. Compared with centralized training beam alignment, it significantly reduces the network's fronthaul link overhead without significant performance loss, while rapidly adapting to non-stationary environmental changes. The proposed algorithm solves the problem of model adaptation to fast time-varying scenarios, obtains the gain of spatiotemporal sequence feature information on beam selection effect, realizes the lightweighting of AI model on base station side, can reduce the computational pressure of central processing unit, achieve a reasonable trade-off between cooperation overhead and beam selection prediction performance, and achieves high-performance beam alignment under small sample conditions. Attached Figure Description

[0062] Figure 1 This is a flowchart of a width-learning base station cooperative beam alignment method for mining spatiotemporal sequence features according to an embodiment of the present invention.

[0063] Figure 2 This is a schematic diagram of the user movement model in the technical solution of the present invention.

[0064] Figure 3 This is a schematic diagram of the millimeter-wave MIMO base station cooperative beam alignment model in the technical solution of this invention.

[0065] Figure 4 This is a flowchart of the millimeter-wave MIMO base station cooperative beam alignment model based on width learning in the technical solution of this invention.

[0066] Figure 5 This is a schematic diagram illustrating the principle of the base station cooperation algorithm in the technical solution of this invention.

[0067] Figure 6 This is a schematic diagram of the RBL model in the technical solution of this invention.

[0068] Figure 7 This is a schematic diagram of the GBL model in the technical solution of this invention.

[0069] Figure 8 This is a structural diagram of the GBL unit in the technical solution of this invention.

[0070] Figure 9This is a schematic diagram of a millimeter-wave massive MIMO communication scenario in an embodiment of the present invention.

[0071] Figure 10 , 11 This is a performance comparison chart of a width learning base station cooperative beam alignment method for mining spatiotemporal sequence features in an embodiment of the present invention and a traditional method. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following will explain and illustrate in detail, through specific embodiments, a width-learning base station cooperative beam alignment method for mining spatiotemporal sequence features provided by this invention.

[0073] like Figure 1 As shown, the method includes:

[0074] Step S1: Construct a communication scenario for continuous user movement, build a millimeter-wave MIMO system model for multi-base station cooperative transmission, and construct a beam alignment problem model with the goal of maximizing system efficiency and rate.

[0075] In this step, the embodiment of the invention uses the open-source Deep MIMO dataset to construct the scene and generate specific channel data. The 'O1' scene in the Deep MIMO dataset has a horizontal main street that is 600m long (along the Y-axis) and 40m wide (along the X-axis), with 12 base stations on each side. The user location grid is located along the main street, 550m long and 35m wide. Starting from the upper right, 15m from the street's starting point, and ending at the lower left, 35m from the street's ending point, the grid consists of 2751 rows, from R1 to R2751. Each row contains 181 user locations, with a 20cm spacing between any two adjacent user locations, and the antenna reception height for each user location is constant at 2m.

[0076] like Figure 9 As shown in this embodiment of the invention, the communication scenario is selected within a 36m × 80m area covered by rows R1066 to R1466 in the Deep MIMO main street. A total of 201 × 181 = 36281 sample data points are collected, with a spacing of 20cm between adjacent points. The system operates in the millimeter-wave band with a carrier frequency of 60GHz, a working bandwidth of 500MHz, and 3 base stations configured with an 8 × 4 uniform planar array. Considering a user set of 2 users, each user is equipped with an omnidirectional single antenna.

[0077] In this embodiment of the invention, it should be noted that, in order to better fit the network, the training dataset is subjected to maximum value normalization. All inputs to the local wide learning network model of base station b are divided by a constant scaling factor Δ of the amplitude. norm Defined as:

[0078]

[0079] The output of the local wide-learning network model will be each candidate beam response sample from base station b. Normalization was performed separately, as follows:

[0080]

[0081] In this embodiment of the invention, it should be noted that each sampling point of the local wide-range learning network of each base station has Each input corresponds to a user set. The amplitude and phase of the user uplink sensing beam response on each subcarrier group; each base station's local wide learning network has M outputs, corresponding to the equivalent rate index of each candidate simulated beam locally. In the simulation, the above 36491 location points are divided into a training set and a test set in a ratio of 8:2, with sizes of 29192 and 7298 respectively, for use by the user set.

[0082] Step S2: Each user During the training phase, send (M+1) repeated pilot signals. Each base station Simultaneously perform beam training, using sensing wide beam vectors. And switching between M candidate simulated beam vectors in different directions Received in sequence, This indicates the uplink training power. (Each base station) In subcarrier The received signal of the i-th training sequence obtained above is in This is the received noise vector of the b-th base station on the k-th subcarrier. After pilot matching, the base station can obtain the combined signal received using different beam vectors. Estimate: definition and These are respectively the numbering set of the subcarrier groups belonging to user u and the subcarrier groups belonging to... The set of subcarrier numbers in a group. Then, a subcarrier group... The beam response is:

[0083]

[0084] During the offline phase, each base station calculates the equivalent rate index corresponding to each candidate narrow beam vector in N training samples. The expression is as follows:

[0085]

[0086] in Each base station will receive the sensing beam response and the equivalent rate index of each candidate beam vector All are sent to the Central Processing Unit (CU).

[0087] Define the nth sample For base stations Received from user The perceived beam response, where ∠ denotes the complex phase operator, n = 1, ..., N. Then, the cooperative base station model obtains the beam response from the user. The sensing beam response, its nth sample is definition For users The candidate beam response samples. To reduce the learning difficulty, for... Perform one-hot encoding to obtain in

[0088]

[0089] After data arrangement, the label of the width learning network model corresponding to the nth sample is: in Taking the input of the first Q wide-beam sampling points as the input at the current time, then the N sample sets are:

[0090]

[0091] Based on the width learning model, each base station Utilizing local sensing beam response Mapping I group of feature nodes With J group enhanced nodes Where F and E represent the number of feature nodes and augmentation nodes in each group, respectively. b,u =[Z b,u,1 Z b,u,2 ,…,Z b,u,I [H] is the concatenation matrix of I sets of feature nodes. b,u =[H b,u,1 H b,u,2 ,…,H b,u,J [ ] represents the concatenated matrix of J groups of enhancement nodes. The joint feature enhancement nodes of the base station local width learning network are... Therefore, based on the joint feature enhancement node of multiple base stations and tags The model optimization problem can be expressed as:

[0092]

[0093] in It is an affine transformation matrix.

[0094] To reduce the computational burden on the CU and lower the transmission overhead of the joint feature enhancement node, A is utilized. u With the characteristic of vertical partitioning of data features, the original centralized processing problem is transformed into a distributed optimization problem based on feature dimension partitioning, which can be expressed as:

[0095]

[0096] in These are each base station The local affine transformation matrix is ​​calculated using the iterative formula:

[0097]

[0098] in For the introduced auxiliary variable matrix, The parameter ρ is the penalty coefficient, ρ>0; I IF+JE It is an identity matrix with size and Same; O u W is the standardized dual variable; t is the iteration number. All base stations participating in the federated learning training are weighted and initialized. b,u (0) = 0, O u (0) = 0. Repeat the above iterative process t. max This approach can solve the distributed optimization problem of data with partitioning characteristics along feature dimensions, enabling implicit sharing of feature space among base stations.

[0099] Step S3: To ensure beam alignment is always applicable to the current fast time-varying scenarios, the network structure of the width learning method was improved, and the RBL method was designed. The time series of the perceived beam response was used to obtain the feature node Z of the width learning network. b,u,i :

[0100]

[0101] Z b,u,0 It is a zero matrix. and represents the connection weights and biases of the feature generation network, respectively; these are randomly generated and not trainable; e i This represents the i-th layer of the feature layer; the diagonal matrix weights γ b,u,i It is also randomly generated and cannot be trained. This represents a column vector with all elements equal to 1. φ(·) represents the ReLU linear activation function. These feature nodes are further mapped to J groups of augmentation nodes H. b,u,j ,Right now

[0102]

[0103] in and h represents the connection weights and biases of the randomly generated feature enhancement network, respectively. j Let denote the j-th layer of the enhancement layer, and ξ(·) denote the nonlinear activation function Tansig. Then, the joint feature enhancement node of the base station local width learning network is...

[0104] Step S4: To more effectively mine spatiotemporal sequence feature information, based on RBL, the GBL method is proposed by drawing on the LSTM structure. An additional forget gate is designed to control the learned sequence information. The step size s of the forget gate is set, and feature nodes Z are generated. b,u,i i = 1, ..., I:

[0105]

[0106] in and represents the connection weights and biases of the feature generation network, respectively; these are randomly generated and not trainable; e i This represents the i-th layer of the feature layer; the diagonal matrix weights γ b,u,i It is also randomly generated and cannot be trained. This represents a column vector with all elements equal to 1. φ(·) represents a linear activation function. Similarly, these feature nodes are further mapped to J groups of augmentation nodes H. b,u,j ,Right now

[0107]

[0108] in and h represents the connection weights and biases of the randomly generated feature enhancement network, respectively. j Let denote the j-th layer of the enhancement layer, and ξ(·) denote the nonlinear activation function. Then, the joint feature enhancement node of the base station local width learning network is...

[0109] Step S5: In this embodiment of the invention, regarding the network structure, the simulation sets the number of feature node groups in the wide learning network to 10, with 20 feature nodes per group, and the feature layer uses the linear activation function ReLU; the number of nodes in the enhancement layer is set to 1500, and the activation function of the enhancement layer is designed as Tansig, with the specific form as follows:

[0110]

[0111] Where x represents the independent variable of the function Tansig(x).

[0112] In this embodiment of the invention, it should be noted that, since real and effective samples are difficult to obtain in actual communication scenarios, the main consideration is the limited number of samples collected by base stations, fully demonstrating the advantages of base station cooperative networks with small amounts of data. The number of local sample datasets used for training is increased from 500 to 3500. In this embodiment of the invention, T is set to 96ms. d The ms value is set to 0.48ms. The sensing beam employs a single-antenna quasi-omnidirectional beam. The simulation considers the following four base station cooperative beam alignment schemes:

[0113] (1) Collaborative Scheme: For the proposed Collaborative-RBL (C-RBL) beam prediction scheme, I = 10, F = 20, J = 1, and Q = 3 are set. The number of iterations t for base station collaborative learning is... max For the 5th time, ρ is taken as 0.1, and λ is taken as 2. -9 For the proposed gated structure width learning base station collaborative (C-GBL) beam prediction scheme, the forget gate step size s is set to 3, and other parameters are set the same as the C-RBL scheme. The C-BL scheme uses a basic width learning network, and the parameter settings are the same as the C-RBL scheme.

[0114] (2) Fully Centralized Scheme: For the beam prediction scheme of Fully Centralized-RBL (FC-RBL), samples from multiple users and samples from 3 BSs are aggregated for base station-side learning. The relevant BL parameters refer to the base station cooperation scheme. FC-GBL uses a gated structured width learning network, while FC-BL uses a basic width learning network structure.

[0115] (3) Fully Distributed Scheme: To demonstrate the necessity for each base station to acquire complete environmental information for the beam selection algorithm, the wide beam responses of other base stations were removed from the model described in the C-RBL scheme, forming a fully distributed base station-side scheme (FD-RBL) based on cyclic structure width learning, as a comparison scheme. FD-GBL uses a gated structure width learning network, while FD-BL uses a basic width learning network structure.

[0116] (4) DNN Scheme: A base station-side centralized (DNN) beam selection scheme based on deep learning was used, employing a DNN to train a mapping network between uplink wide beam response and beam selection. The DNN model used in the simulation had four layers, including two hidden layers with 256 and 256 neurons respectively, and the input and output layers with 256 and 128 neurons respectively. Each DNN model had two hidden layers. The first hidden layer had 200 ReLU neurons, and the number of ReLU neurons in the second layer was the same as the augmentation nodes in the wide learning scheme. The dropout rate and learning rate were set to 0.05 and 0.001 respectively, and the batch size was set to 100. The Adam optimizer was used to update the DNN model under the Cross Entropy Loss Function (CLF). TensorFlow and Keras libraries were used in the simulation.

[0117] The embodiments of the present invention utilize Figure 10 The average effective rate curves for users in the designed schemes were compared. The Genie-Aided curve represents an ideal situation where the base station can perform optimal beam selection without any training overhead, representing the upper limit of rate performance. The Exhaustive curve is set as the average effective rate obtained by beam training using the traditional exhaustive beam scanning search method. The average effectiveness rate on each subcarrier for each user is used as the performance metric for the base station cooperative beam alignment scheme.

[0118]

[0119] Depend on Figure 10It is evident that, regardless of improvements to the width learning network structure, both the base station cooperation scheme and the fully centralized scheme outperform the fully distributed scheme. This indicates that the sensing beam measurement of a single base station does not contain sufficient spatial information to predict the optimal beam. Therefore, sharing spatial feature information through base station cooperation is necessary. Drawing inspiration from the LSTM structure's width learning coordination model, the basic width learning network structure was improved, utilizing spatiotemporal sequence features to enhance the performance of the beam alignment algorithm, especially for the base station cooperation scheme. For the base station cooperation scheme, a small increase in local data storage space can effectively compensate for the complete information obtained by the fully centralized scheme, while avoiding the excessive computational complexity and communication overhead of the fully centralized scheme. Furthermore, as... Figure 10 As shown, the growth trend of the DNN scheme curve is not obvious with the increase of the local training dataset, indicating that DNN has not yet been able to learn the appropriate mapping relationship between input and output when the amount of data is small. This is because DNN often requires a large number of data samples to support it, and therefore does not have an advantage in the case of small samples. Therefore, for dynamic scenarios where users are constantly moving, facing the problems of model lightweighting, few-sample learning, and model adaptation to fast time-varying scenarios, the width learning coordination model proposed in this invention, which borrows from the LSTM structure to explore spatiotemporal sequence features, has research value. Compared with RBL, GBL is more effective; compared with GBL, RBL structure is simpler and does not require tuning of step size s. In practice, RBL can be used first to obtain a basic model, and then GBL can be used to further improve the effectiveness.

[0120] The embodiments of the present invention utilize Figure 11 The convergence performance of the designed scheme was compared. Because the design addresses dynamic scenarios involving continuous user movement, and faces challenges such as model lightweighting, few-shot learning, and model adaptation to rapidly changing scenarios, it is necessary to consider algorithms with fast convergence performance. The convergence performance of the improved width-learning network structure scheme is as follows: Figure 11As shown, each base station has 3000 training samples, and successful beam selection means that all base stations can perform accurate beam prediction. The beam selection success rate of the proposed C-RBL and C-GBL schemes increases with the number of communication rounds between the base station and the CU. Considering the computation time consumed in the real environment, under the condition of a small number of samples, the C-RBL and C-GBL schemes can achieve good performance with 5 interaction iterations. This shows that the proposed distributed base station cooperative beam coordination algorithm is a good iterative algorithm with guaranteed convergence speed. Furthermore, when the number of cooperative iterations increases to 10, the proposed algorithm has almost no performance loss compared to the FC-RBL and FC-GBL schemes. The C-RBL scheme has a significant advantage over the FD-RBL method, with a beam selection success rate improvement of approximately 15.87%; while the C-GBL scheme improves by approximately 15.93% compared to the FD-GBL method.

[0121] Any aspects of this invention not described in detail are well-known to those skilled in the art.

[0122] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A width-learning base station cooperative beam alignment method for mining spatiotemporal sequence features, characterized in that, Includes the following steps: To construct a communication scenario for continuous user mobility, a millimeter-wave MIMO system model with multi-base station cooperative transmission is built, and a beam alignment problem model is constructed with the goal of maximizing system efficiency and rate. In the millimeter-wave MIMO system model of cooperative transmission, each base station uses a local width learning network and a distributed learning training architecture to collaboratively learn the mapping relationship between the sensing beam response of multiple base stations and the corresponding narrow beam, so as to complete the prediction of the optimal downlink transmission narrow beam. Each base station’s local width learning network feature layer adopts a cyclic structure and uses the time series of sensing beam response to obtain the feature nodes of the width learning network. Each base station adds a gating mechanism to the feature layer of the local width learning network, using a forget gate to control the learned sequence information; Each base station inputs the collected user wide-beam responses into the local mapping network and interacts with the central processing unit to predict its own beam selection.

2. The method for width-learning base station cooperative beam alignment for mining spatiotemporal sequence features according to claim 1, characterized in that, Each base station's local width learning network uses a recurrent structure for its feature layer, employing time series of perceived beam response to obtain the feature nodes of the width learning network, specifically including: Using time series of perceptual beam response to obtain feature nodes of the width learning network : ； in It is a zero matrix. and These represent the connection weights and biases of the feature generation network, respectively. They are randomly generated and not trainable. The first feature layer Layer; Diagonal matrix weights It is also randomly generated and cannot be trained; Represents a column vector whose elements are all 1s; Represents a linear activation function; and users respectively The numbering set of subcarrier groups and belonging to The set of subcarrier numbers for a group; Indicates the input of the wide-beam sampling point; Represents the feature nodes of each group; The number of feature node groups; these feature nodes are further mapped to Group Enhancement Node ,Right now ； in and These represent the connection weights and biases of the randomly generated feature enhancement network, respectively. The first layer represents the enhancement layer. layer; Represents a non-linear activation function; Indicates the number of groups of augmented nodes; Indicates the number of augmentation nodes in each group; then the joint feature augmentation nodes of the base station local width learning network are: ,in Indicates the number of training samples. for The concatenated matrix of group feature nodes, for A cascade matrix of group-enhanced nodes.

3. The method for width-learning base station cooperative beam alignment for mining spatiotemporal sequence features according to claim 1, characterized in that, The gating mechanism specifically involves designing an additional forget gate to control the learned sequence information, and setting the step size of the forget gate. Generate feature nodes : ； in and These represent the connection weights and biases of the feature generation network, respectively. They are randomly generated and not trainable. The first feature layer Layer; Diagonal matrix weights It is also randomly generated and cannot be trained; Represents a column vector whose elements are all 1s; Represents a linear activation function; Indicates the number of wide-beam sampling points. For users The number of subcarrier groups, This represents the number of nodes in each group of feature nodes. Indicates the number of training samples. Indicates the number of groups of feature nodes; These feature nodes are mapped to Group Enhancement Node ,Right now ； in and These represent the connection weights and biases of the randomly generated feature enhancement network, respectively. The first layer represents the enhancement layer. layer; Represents a non-linear activation function; This indicates the number of groups of augmented nodes. This represents the number of nodes in each group of augmentation nodes; therefore, the joint feature augmentation nodes of the base station local width learning network are: ,in for The concatenated matrix of group feature nodes, for A cascade matrix of group-enhanced nodes.

4. The incremental cooperative beam selection method for millimeter-wave MIMO base stations based on wide learning as described in claim 1, characterized in that, Each base station inputs the collected user wide-beam responses into the local mapping network and interacts with the central processing unit to predict its own beam selection, specifically including: During offline training, RBL is first used to obtain a basic model, and the number of layers for feature nodes and augmentation nodes is tuned; then GBL is used to tune the forget gate step size; when the system enters the online execution phase, Each user first sends uplink pilot signals to train the sensing beam, the time cost of which is Base station Using local sensing beam reception, the sensing beam response is obtained, i.e. Before use Wide beam sampling points Obtain the joint output of the feature nodes and the augmentation nodes, i.e. After further affine transformation, the predicted index values ​​of the simulated beam to be selected are obtained. Based on this predicted value, the base stations are determined Its analog beam number Then, CU selects the optimal downlink transmission beam of the base station according to the MRT principle, enabling the base station to perform subsequent downlink data transmission; the effective efficiency of user u during the online execution phase is ； in, These are the training time for the sensing beam and the transmission time for the narrow beam. It is the time-varying tracking period of the channel. Indicates system bandwidth. Indicates the total number of subcarriers. Indicates user The set of subcarriers, For users In subcarrier On the transmit power, For noise power, For base stations To users In subcarrier Channel vector on, For the selected analog beam vector, The simulated beamcodebook size for each base station; in uplink training based on OFDMA mode, multiple users and base stations can perform beam training simultaneously.

Images