Model training method and multi-base station cooperative beam tracking and prediction method and device

Abstract

The application provides a model training method, a multi-base-station cooperative beam tracking and prediction method and device. The model training method comprises the following steps: obtaining a training sample data set, each training sample data in the training sample data set comprising an actual optimal base station serial number corresponding to a time instant and communication with a vehicle, actual optimal beam pair attribute information corresponding to the time instant, actual radar echo intensity received by a plurality of base stations corresponding to the time instant, and actual beam pair attribute information corresponding to the time instant; based on the training sample data set, performing supervised training on an initial deep neural network model to determine a multi-base-station cooperative beam tracking and prediction model; and the supervised training is used for learning nonlinear correlation between adjacent time instants. The application does not require additional signaling interaction, reduces beam switching overhead, improves beam tracking accuracy, and improves beam tracking and switching efficiency.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a model training method, a multi-base station cooperative beam tracking prediction method and apparatus. Background Technology

[0002] With the rapid development of wireless communication technology, users' demands for faster data transmission rates and larger data storage capacities are constantly increasing. While millimeter-wave bands offer abundant spectrum resources and greater bandwidth, their short wavelengths result in limited diffraction and refraction capabilities, and their high-frequency characteristics lead to higher path loss. Therefore, millimeter-wave communication typically requires beamforming technology to generate directional gain to compensate for the high path loss and achieve long-distance signal transmission. Millimeter-wave communication is highly advantageous in high-speed mobile scenarios, enabling low-latency communication tasks such as real-time video and streaming video transmission, as well as high-resolution beamforming for position tracking and high-precision positioning.

[0003] Therefore, when millimeter-wave communication is applied to high-dynamic massive multiple-input multiple-output (mMIMO) scenarios, the number of transceiver antennas increases, the number of beams also rises, and the communication beam becomes narrower and more directional. This makes it more prone to misalignment in high-dynamic scenarios, thus requiring multiple beam alignments to ensure communication performance. However, frequent beam switching significantly increases the signaling overhead of beam tracking and channel cost estimation, severely straining normal communication resources and reducing system spectral efficiency. Therefore, reducing the pilot and signaling overhead caused by beam tracking and switching in high-dynamic mMIMO scenarios is particularly important.

[0004] In related technologies, a beam tracking method based on radar echo is designed for integrated communication and sensing scenarios. Specifically, the radar is responsible for sensing the channel environment. The radar echo contains a large amount of key information related to the channel state and the location of the transceiver. Based on the channel environment information and key information sensed by the radar, beam tracking is assisted, thereby reducing pilot and signaling resource overhead.

[0005] However, existing integrated communication and sensing scenarios typically consider a single base station, and the sensing accuracy of a single base station is limited and easily affected by noise or interference, thus reducing beam tracking accuracy. Furthermore, obtaining radar echoes requires the existence of a Line of Sight (LOS) between the base station and the terminal. However, when the LOS is blocked by an obstacle, the base station's ability to sense the user's location drops sharply, thereby reducing the efficiency of beam tracking and handover. Summary of the Invention

[0006] This invention provides a model training method, a multi-base station cooperative beam tracking prediction method, and an apparatus to address the shortcomings of existing integrated communication and sensing scenarios where beam tracking accuracy is reduced due to the reliance on radar sensing information from a single base station. It achieves improved accuracy in user location perception by using a multi-base station cooperative beam tracking prediction model, and mitigates the problem of reduced beam tracking accuracy caused by radar echo obstruction. Without requiring additional signaling interaction, it significantly reduces beam switching overhead, greatly improves beam tracking accuracy, and significantly enhances the efficiency of beam tracking and switching.

[0007] This invention provides a model training method applied to a millimeter-wave V2I network scenario, wherein the millimeter-wave V2I network scenario is a high-dynamic scenario of multiple base stations serving a vehicle with integrated sensing and communication, wherein the channel between one of the base stations and the vehicle has a LOS path, and the channels between the other base stations and the vehicle have NLOS paths; the model training method includes:

[0008] A training sample dataset is obtained, wherein each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the corresponding time, the actual optimal beam pair attribute information for the corresponding time, the actual radar echo intensity received by the multiple base stations at the corresponding time, and the actual beam pair attribute information for the corresponding time; wherein, the actual beam pair attribute information for the corresponding time includes the optimal beam pair attribute information between each base station and the vehicle at the corresponding time, and the actual optimal beam pair attribute information for the corresponding time is the optimal beam pair attribute information selected from each of the optimal beam pair attribute information at the corresponding time;

[0009] Based on the training sample dataset, the initial deep neural network model is subjected to supervised training to determine the multi-base station cooperative beam tracking prediction model; wherein, the supervised training is used to learn the nonlinear correlation between channels at adjacent time points.

[0010] According to a model training method provided by the present invention, the process of obtaining the training sample data for each training sample in the training sample dataset includes:

[0011] In the case that at the corresponding time the vehicle sends SRS signals to all beam directions of each base station, each base station receives each SRS signal, and each base station determines the optimal beam pair attribute information and RSRP value between the vehicle and the corresponding time based on the received SRS signals, the actual optimal beam pair attribute information and the actual optimal base station number at the corresponding time are obtained.

[0012] When the multiple base stations respectively send DFRC signals to the vehicle in a directional manner, the multiple base stations are instructed to report the radar echo intensity and the optimal beam pair attribute information at the corresponding time.

[0013] The training sample data is determined based on the actual optimal beam pair attribute information at the corresponding time, the actual optimal base station number at the corresponding time, the actual radar echo intensity reported by the multiple base stations at the corresponding time, and the optimal beam pair attribute information between each base station and the vehicle at the corresponding time.

[0014] According to a model training method provided by the present invention, the step of performing supervised training on an initial deep neural network model based on the training sample dataset to determine a multi-base station cooperative beam tracking prediction model includes:

[0015] The training sample dataset is preprocessed to determine the preprocessed training sample dataset;

[0016] The preprocessed training sample dataset is used to perform supervised training on the initial deep neural network model to obtain the optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations; the optimal prediction result includes the optimal predicted base station number at the next time step and the optimal predicted beam pair attribute information at the next time step.

[0017] Determine the error value between the predicted optimal result and the actual optimal result, wherein the actual optimal result includes the actual optimal base station number at the next time step and the actual optimal beam pair attribute information at the next time step;

[0018] If the error value meets the preset error requirement, the intermediate deep neural network model corresponding to the error value is determined as the multi-base station cooperative beam tracking prediction model.

[0019] If the error value does not meet the preset error requirement, the intermediate deep neural network model corresponding to the error value is subjected to supervised training using the preprocessed training sample dataset to determine the multi-base station cooperative beam tracking prediction model.

[0020] According to a model training method provided by the present invention, the step of preprocessing the training sample dataset to determine the preprocessed training sample dataset includes:

[0021] When the actual optimal beam pair attribute information is the actual optimal beam angle or the actual optimal transceiver beam number, the training sample dataset is divided into a first training sample data subset containing the actual optimal beam angle and a second training sample data subset containing the actual optimal transceiver beam number.

[0022] The first training sample data subset is subjected to mean-variance standardization, and the second training sample data subset is subjected to One-Hot encoding to determine the preprocessed training sample dataset.

[0023] According to a model training method provided by the present invention, the step of performing supervised training on the initial deep neural network model using the preprocessed training sample dataset to obtain the optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations includes:

[0024] The first training sample data subset after mean-variance standardization is used to perform supervised regression and classification training on the initial deep neural network model to obtain the first optimal prediction result output by the intermediate deep neural network model after the preset number of training times. The first optimal prediction result includes the optimal predicted base station number for communication between the vehicle and the next time step and the optimal predicted beam angle for the next time step.

[0025] Using the second training sample data subset encoded with One-Hot, supervised classification training is performed on the initial deep neural network model to obtain the second optimal prediction result output by the intermediate deep neural network model after the preset number of training iterations. The second optimal prediction result includes the predicted optimal base station number for communication between the vehicle and the next time step and the predicted optimal transceiver beam number for the next time step.

[0026] According to a model training method provided by the present invention, when the vehicle is moving in a straight line and the ULA antenna array of each base station is parallel to the direction of movement of the vehicle, the actual radar echo intensity received by the plurality of base stations at the corresponding time includes the radar echo intensity reflected by the vehicle at the corresponding time and the radar echo intensity reflected by the obstacle at the corresponding time.

[0027] This invention also provides a multi-base station cooperative beam tracking prediction method, applied to a millimeter-wave V2I network scenario. The millimeter-wave V2I network scenario is a high-dynamic scenario where multiple base stations serve a vehicle, and the channel between one of the base stations and the vehicle has a LOS path, while the channels between the remaining base stations and the vehicle have NLOS paths. The multi-base station cooperative beam tracking prediction method includes:

[0028] The current radar echo intensity received by each base station at the current moment, the current transmitted beam attribute information received by each base station at the current moment, and the multi-base station cooperative beam tracking prediction model trained according to the aforementioned model training method are obtained.

[0029] The current radar echo intensity and the current transmitted beam attribute information are input into the multi-base station cooperative beam tracking prediction model for tracking and prediction, to obtain the target optimal base station number for communication between the target time and the vehicle, and the target optimal beam pair attribute information at the target time.

[0030] The target optimal base station number and the target optimal beam pair attribute information are sent to the target optimal base station corresponding to the target optimal base station number and the vehicle, respectively, to prompt the target optimal base station and the vehicle to communicate.

[0031] The present invention also provides a model training apparatus, applied to the millimeter-wave V2I network scenario described in the aforementioned model training method, comprising:

[0032] The first acquisition unit is used to acquire a training sample dataset. Each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the corresponding time, the actual optimal beam pair attribute information for the corresponding time, the actual radar echo intensity received by the plurality of base stations at the corresponding time, and the actual beam pair attribute information for the corresponding time. The actual beam pair attribute information for the corresponding time includes the optimal beam pair attribute information between each base station and the vehicle at the corresponding time, and the actual optimal beam pair attribute information for the corresponding time is the optimal beam pair attribute information selected from each of the optimal beam pair attribute information at the corresponding time.

[0033] The model training unit is used to perform supervised training on the initial deep neural network model based on the training sample dataset to determine the multi-base station cooperative beam tracking prediction model; wherein, the supervised training is used to learn the nonlinear correlation between channels at adjacent time points.

[0034] The present invention also provides a multi-base station cooperative beam tracking device, comprising:

[0035] The second acquisition unit is used to acquire the current radar echo intensity received by each base station at the current time, the current transmitted beam attribute information received by each base station at the current time, and the multi-base station cooperative beam tracking prediction model trained according to the aforementioned model training method.

[0036] The prediction and tracking unit is used to input the current radar echo intensity and the current transmitted beam attribute information into the multi-base station cooperative beam tracking prediction model for tracking and prediction, and to obtain the target optimal base station number for communication between the target and the vehicle at the target time, as well as the target optimal beam pair attribute information at the target time.

[0037] The information sending unit is used to send the target optimal base station number and the target optimal beam pair attribute information to the target optimal base station corresponding to the target optimal base station number and the vehicle, respectively, so as to prompt the target optimal base station and the vehicle to communicate.

[0038] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the model training methods or multi-base station cooperative beam tracking prediction methods described above.

[0039] The present invention provides a model training method, a multi-base station cooperative beam tracking prediction method, and an apparatus. In the model training method, an edge server acquires a training sample dataset by obtaining the actual optimal base station number communicating with the vehicle at each time step, the actual optimal beam pair attribute information at each time step, the actual radar echo intensity reported by multiple base stations at each time step, and the actual beam pair attribute information reported at each time step. Based on this training sample dataset, an initial deep neural network model is subjected to supervised training to determine the multi-base station cooperative beam tracking prediction model. Since the actual beam pair attribute information at a given time step includes the optimal beam pair attribute information between each base station and the vehicle at that time step, and the actual optimal beam pair attribute information at that time step is the optimal beam pair attribute information selected from all the optimal beam pair attribute information at that time step, the multi-base station cooperative beam tracking prediction model determined by training the training sample dataset can improve the base station's perception accuracy of the user's location and mitigate the problem of reduced beam tracking accuracy caused by radar echo obstruction by obstacles. Without requiring additional signaling interaction, it not only significantly reduces beam switching overhead but also greatly improves beam tracking accuracy and efficiency. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 This is one of the flowcharts illustrating the model training method provided by the present invention;

[0042] Figure 2 This is a schematic diagram of a millimeter-wave V2I network scenario provided by the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of the DNN model provided by the present invention;

[0044] Figure 4 This is the second flowchart of the model training method provided by the present invention;

[0045] Figure 5 This is the third flowchart illustrating the model training method provided by this invention;

[0046] Figure 6 This is the fourth flowchart illustrating the model training method provided by this invention;

[0047] Figure 7 This is a flowchart illustrating the multi-base station cooperative beam tracking prediction method provided by the present invention;

[0048] Figure 8 These are simulation results comparing the method of this invention with other methods.

[0049] Figure 9 This is a simulation prediction result diagram based on adaptive codebook provided by the present invention;

[0050] Figure 10 This is a simulation prediction result diagram based on a fixed codebook provided by the present invention;

[0051] Figure 11 This is a schematic diagram of the structure of the model training device provided by the present invention;

[0052] Figure 12 This is a schematic diagram of the structure of the multi-base station cooperative beam tracking device provided by the present invention;

[0053] Figure 13 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0055] The following is combined Figures 1-13 This invention describes a model training method, a multi-base station cooperative beam tracking prediction method, and an apparatus. The execution entity of the model training method can be an edge server (ES). This edge server is connected to multiple base stations and also establishes a connection with a high-speed moving vehicle. Furthermore, this edge server possesses at least model training, information transmission, information reception, and information processing capabilities. The following description uses an edge server as the execution entity as an example.

[0056] The model training method provided by this invention is applied to millimeter-wave vehicle-to-infrastructure (V2I) network scenarios. Millimeter-wave V2I network scenarios are highly dynamic scenarios where multiple base stations serve vehicles, integrating sensing and communication. One base station has a Line-of-Sight (LOS) path with the vehicle, while the other base stations have non-line-of-sight (NLOS) paths with the vehicle. (Refer to...) Figure 1 This is a flowchart illustrating the model training method provided by the present invention, as shown below. Figure 1 As shown, the model training method includes the following steps:

[0057] Step 110: Obtain the training sample dataset. Each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the corresponding time, the actual optimal beam pair attribute information for the corresponding time, the actual radar echo intensity received by multiple base stations for the corresponding time, and the actual beam pair attribute information for the corresponding time.

[0058] The actual beam pair attribute information at a given time includes the optimal beam pair attribute information between each base station and the vehicle at that time. The actual optimal beam pair attribute information at a given time is the optimal beam pair attribute information selected from all the optimal beam pair attribute information at that time, which can be considered as the optimal beam pair attribute information between the actual optimal base station and the vehicle at that time. Furthermore, each training sample in the training sample dataset corresponds to one time point. This means that a training sample can be acquired at each time point according to a continuous time interval data acquisition rule, or according to a non-continuous time interval data acquisition rule. No specific limitation is made here. Moreover, in the millimeter-wave V2I network scenario, there are multiple vehicles, and each vehicle is moving at high speed. Additionally, the actual optimal beam pair attribute information can be the actual optimal beam angle or the actual optimal transceiver beam sequence number.

[0059] Specifically, in the case where multiple base stations are actually multiple Road Side Units (RSUs) and each RSU serves K vehicles moving at high speed, the following can be referred to: Figure 2 The diagram shown illustrates a millimeter-wave V2I network scenario. Figure 2 In this system, each RSU is equipped with a downlink dual-function radar communication (DFRC) system. Each DFRC system employs massive MIMO technology and has two ULA antenna arrays, one for transmitting and the other for receiving DFRC signals. The number of transmitting and receiving antennas are N respectively. t and Nr N t and N r Both K and K are positive integers greater than 0.

[0060] In addition, Figure 2 In this configuration, each vehicle is also equipped with a ULA antenna array consisting of M antennas to receive downlink DFRC signals transmitted by the RSU; as the computing center for multi-base station collaboration, the ES is mainly responsible for training and predicting deep learning network models; M is a positive integer greater than 0; the downlink DFRC signal transmitted by the RSU can be represented as:

[0061]

[0062] In equation (1), This represents the downlink DFRC data stream sent by the RSU to K vehicles at time n. k,n (t) represents the downlink DFRC data stream sent by the RSU to the k-th vehicle at time n; This represents the transmit beamforming matrix at time n, where each column represents the beamforming vector of a vehicle, f k,n Let represent the beamforming vector of the k-th vehicle at time n, where k ∈ [1, K].

[0063] The radar echo reflection signal received by the RSU from the vehicle is:

[0064]

[0065] In equation (2), p represents the antenna array gain. k,n β represents the transmission power of the RSU to the k-th vehicle at time n. k,n μ k,n τ k,n These represent the radar echo reflection coefficient, Doppler frequency shift, and time delay of the k-th vehicle, respectively. This indicates that the mean is zero and the variance is σ. 2 Gaussian noise; a(θ) k,n ) represents the response of the transmitting ULA antenna array, b(θ) k,n The response of the ULA antenna array at the receiving end is represented as:

[0066]

[0067]

[0068] Equations (3) and (4) assume that the antenna spacing of the ULA antenna array is half a wavelength.

[0069] It is worth noting that although the RSU receives radar echoes from multiple vehicles simultaneously, the inter-beam interference can be ignored because the antenna array responses at different angles in a large-scale MIMO system are nearly orthogonal. Therefore, the reflected echoes from different vehicles do not interfere with each other. The radar echo from the k-th vehicle received by the RSU at time n can be expressed as Equation (5):

[0070]

[0071] The downlink DFRC signal received by the RSU from the vehicle at the nth time and the kth vehicle through beam combining can be expressed as Equation (6):

[0072]

[0073] In equation (6), the superscript H indicates the transpose operation. This represents the channel coefficient during the communication process. Represents path loss, 2πd k,n / λ represents the phase of the channel; H k,n This represents the millimeter-wave channel matrix between the RSU and the vehicle; f k,n and w k,n These represent the transmitting beamforming vector and the receiving beam combining vector, respectively; z c (t) represents Gaussian noise with a mean of zero and a variance of .

[0074] Specifically, path loss Referring to the path loss model for rural macrocell (RMa) scenarios proposed in 3GPP TR 38.901, the path loss includes both LOS and NLOS paths; the total path loss is calculated as shown in Equation (7):

[0075] PL = P LOS PL RMa-LOS +(1-P LOS )PL RMa-NLOS (7)

[0076] In equation (7), P LOS PL represents the probability of a LOS path. RMa-LOS and PL RMa-NLOS These represent the path losses for the LOS path and the NLOS path, respectively.

[0077] It should be noted that w k,n Beamforming vectors are generated on the vehicle side based on the angle of arrival (AOA). In scenarios where the vehicle is moving in a straight line, the AOA is the same as the angle of departure (AOD). Therefore, the beamforming vector can be generated using the AOD predicted on the RSU side. k,nIt is a wideband eSV channel containing LOS paths, and a Rice factor is introduced to control the power ratio of LOS and NLOS paths. In this channel, the arrival delay of each path is different, thus ensuring that the channel is wideband. The wideband eSV channel used in this invention is shown in equation (8):

[0078]

[0079] In equation (8), N cl N represents the number of eSV channel clusters. ray This represents the number of paths in each cluster, therefore the total number of NLOS paths is N. cl ×N ray The number of LOS paths is set to 1, so the total number of channel paths is N. cl ×N ray +1; a k,n,j,i ρ represents the complex gain corresponding to the i-th path in the j-th cluster between the RSU and the k-th vehicle at time n; k,n,j,i =v k,n cosθ k,n,j,i f c / c represents the Doppler frequency shift; a(θ) and u(θ) represent the ULA antenna array responses on the RSU side and the vehicle side, respectively; and Let AOD and AOA represent the path i in the j-th cluster between the RSU and the k-th vehicle at time n, respectively; K represents the Rice factor; accordingly, These represent the complex gain coefficient, Doppler frequency shift, AOD, and AOA corresponding to the LOS path, respectively.

[0080] Assume s k,n If the power of (t) is in units of 1, then the received signal-to-noise ratio of the k-th vehicle at the n-th time is defined by equation (9):

[0081]

[0082] Correspondingly, the total speed of the K vehicles is given by equation (10):

[0083]

[0084] As can be seen from equation (10), for a specific channel H, the optimal simulated beamforming vector f k,n and beam combining vector w k,n The design will determine the spectral efficiency of the simulated beamforming system; k,n and w k,nThe design is divided into two categories based on whether the angle domain can be quantized: one is based on a fixed codebook, and the other is based on an adaptive codebook. The fixed codebook quantizes the entire angle space into a finite number of angle directions, each direction corresponding to a codeword, which corresponds to a beamforming or beam combining vector. The adaptive codebook corresponds to continuously changing angle directions, without using a fixed number of codewords for quantization. The beam selection will be more flexible, and theoretically, the beam with the best performance can be found to obtain the maximum beamforming gain. Currently, there are many codebook design schemes, such as the 3c beamcodebook proposed by the Institute of Electrical and Electronics Engineers (IEEE) protocol, the beam-steering beamcodebook obtained by quantizing angles, and the beamcodebook based on Discrete Fourier Transform (DFT). This invention considers both fixed codebook and adaptive codebook beamforming methods. The fixed codebook used is the DFT codebook, as shown in Equation (11).

[0085] The element in the m-th row and k-th column of the DFT codebook matrix is ​​as follows:

[0086]

[0087] Where M is the number of antennas and K is the number of beams contained in the codebook.

[0088] Based on this, the millimeter-wave V2I network scenario provided by this invention considers the impact of obstacle obstruction on radar echoes by introducing NLOS paths. However, since receiving radar echoes from vehicles requires LOS paths as a basis, a high-dynamic scenario integrating sensing and communication is designed where some LOS paths are obstructed. Specifically, only one RSU has an LOS path between it and the vehicle, while the remaining RSUs only have NLOS paths due to obstacle obstruction. Furthermore, since the power of LOS paths is much greater than that of NLOS paths, the communication between the vehicle and the RSU with an LOS path at each time step is determined. Here, the RSU communicating with the vehicle via the LOS path at each time step is identified as the actual optimal base station providing the best service. Based on this, by traversing all beam pair directions between each RSU and the vehicle at each time step, the actual optimal base station number communicating with the vehicle via the LOS path at each time step, the actual optimal beam pair attribute information at each time step, the actual radar echo intensity received by multiple base stations at each time step, and the actual beam pair attribute information at each time step are obtained. This is used to obtain the training sample dataset.

[0089] Step 120: Based on the training sample dataset, supervised training is performed on the initial deep neural network model to determine the multi-base station cooperative beam tracking prediction model; wherein, supervised training is used to learn the nonlinear correlation between channels at adjacent time points.

[0090] Specifically, the network structure diagram of the initial deep neural network (DNN) model can be as follows: Figure 3 As shown, in Figure 3 In DNN, the model consists of different layers, which can be divided into input layer, hidden layer and output layer according to their position and function. Each layer contains a large number of neurons. The neurons in different layers are interconnected by weights. The layers in the DNN model are fully connected, that is, any neuron in the g-th layer is connected to any neuron in the g+1-th layer; g is a positive integer greater than 0.

[0091] Furthermore, each layer in a DNN model can be viewed as a feature extractor, learning certain features from the data input to that layer and passing these features to the next layer. The input layer is the first layer of the DNN model, serving as the input to the entire model. It feeds the input data into the DNN for feature extraction and then passes the extracted features to the hidden layers. The hidden layers are the neural network layers between the input and output layers, and are the core of the DNN. They assign corresponding weights to each input value and combine these weights to calculate a value that is passed to the next layer. Their role is to transform complex functions and approximate the final decision function. The number of hidden layers can be multiple; more layers result in a more expressive model, but also higher complexity. The last layer of the DNN model is the output layer, which performs classification or regression prediction on the input data based on the learned features.

[0092] Based on this, the initial deep neural network model is trained in a supervised manner using the training sample dataset, that is, the nonlinear correlation between channels at adjacent time points is learned, thereby determining the multi-base station cooperative beam tracking prediction model.

[0093] The model training method provided by this invention involves an edge server acquiring a training sample dataset by obtaining the actual optimal base station sequence number for communication between the edge server and the vehicle at each time step, the actual optimal beam pair attribute information at each time step, the actual radar echo intensity reported by multiple base stations at each time step, and the actual beam pair attribute information at each time step. Based on this training sample dataset, an initial deep neural network model is trained in a supervised manner to determine a multi-base station cooperative beam tracking prediction model. Since the actual beam pair attribute information at a given time step includes the optimal beam pair attribute information between each base station and the vehicle at that time step, and the actual optimal beam pair attribute information at that time step is the optimal beam pair attribute information selected from all the optimal beam pair attribute information at that time step, the multi-base station cooperative beam tracking prediction model determined by training the training sample dataset can improve the accuracy of base station perception of user location and mitigate the problem of reduced beam tracking accuracy caused by radar echo obstruction by obstacles. Without requiring additional signaling interaction, this method significantly reduces beam switching overhead, greatly improves beam tracking accuracy, and significantly enhances the efficiency of beam tracking and switching.

[0094] It is understandable that, for each training sample in the training sample dataset in step 110, the process of obtaining one training sample includes:

[0095] Under the circumstances that the vehicle sends SRS signals to all beam directions of each base station at the corresponding time, each base station receives each SRS signal respectively, and each base station determines the optimal beam pair attribute information and RSRP value between itself and the vehicle at the corresponding time based on the received SRS signals, the actual optimal beam pair attribute information and the actual optimal base station number at the corresponding time are obtained.

[0096] When multiple base stations send DFRC signals to the vehicle in a directional manner, the multiple base stations are instructed to report the radar echo intensity and the attribute information of the optimal beam pair at the corresponding time.

[0097] Training sample data is determined based on the actual optimal beam pair attribute information at the corresponding time, the actual optimal base station number at the corresponding time, the actual radar echo intensity reported by multiple base stations at the corresponding time, and the optimal beam pair attribute information between each base station and the vehicle at the corresponding time.

[0098] Specifically, at a given moment, the vehicle first traverses all transmit beams of each base station, that is, the vehicle transmits a Sounding Reference Signal (SRS) to all beam directions of each base station. Simultaneously, each base station receives SRS signals across all beam directions at that moment, selecting the attribute information corresponding to the maximum Reference Signal Receiving Power (RSRP) value at that moment as the optimal beam pair attribute information between the current base station and the vehicle. Then, each base station reports the determined optimal beam pair attribute information and RSRP value to the edge server. The edge server compares each optimal beam pair attribute information and RSRP value to determine the actual optimal beam pair attribute information at that moment. Simultaneously, the edge server can instruct each base station to report radar echo intensity and the optimal beam pair attribute information. After receiving the actual radar echo intensity and the reported optimal beam pair attribute information from each base station at that moment, the edge server determines the actual optimal beam pair attribute information and the actual radar echo intensity at that moment. This process obtains the training sample data corresponding to that moment.

[0099] It should be noted that for beamforming based on a fixed codebook, the actual optimal beam pair attribute information is reflected by the actual transmit and receive beam numbers; for beamforming based on an adaptive codebook, the actual optimal beam pair attribute information is reflected by the actual departure angle and the actual arrival angle; and the determined actual optimal beam pair attribute information between the actual optimal base station and the vehicle can provide conditions for subsequent directional transmission of DFRC signals. In addition, the formula for calculating the RSRP value is shown in equation (12).

[0100]

[0101] In equation (12), y represents the SRS signal received by the base station, and ||·||2 represents the 2-norm operation.

[0102] It should be noted that base stations (e.g.) Figure 4 In the RSU1, the DFRC signal is directionally transmitted to the corresponding beam pair attribute information and the corresponding radar reflection echo is received. The base station here can be the actual optimal base station (i.e., not obstructed) or a non-actual optimal base station (i.e., obstructed). If the base station is the actual optimal base station, the transmitted DFRC signal contains actual data signals, and the DFRC signal can achieve both normal communication and vehicle tracking. If the base station is not the actual optimal base station, the transmitted DFRC signal does not contain actual data signals, and the DFRC signal only performs vehicle tracking. For Figure 4For RSU1, there is a LOS path between it and the vehicle, so it can receive radar echo signals reflected from the vehicle. However, RSU2 is blocked by an obstacle and has no LOS path with the vehicle, so RSU2 can only receive echoes from the obstacle. Furthermore, the radar echo intensity received by each base station at different vehicle locations varies significantly. Therefore, the actual radar echo intensity reported at each moment can indirectly reflect the actual optimal location distribution of the base station, vehicle, and obstacles in the millimeter-wave V2I network scenario. This nonlinear correlation is difficult to analyze using deterministic modeling methods; therefore, this invention employs machine learning methods for analysis. Moreover, all participating base stations are connected to the edge server via stable communication links.

[0103] Based on this, it can be understood that, when the vehicle is moving in a straight line and the ULA antenna array of each base station is parallel to the direction of the vehicle's movement, the actual optimal radar echo intensity received by the base station at each moment includes the radar echo intensity reflected by the vehicle at each moment and the radar echo intensity reflected by the obstacle at each moment.

[0104] It is understandable that the specific implementation process of step 120 may include:

[0105] First, the training sample dataset is preprocessed to determine the preprocessed training sample dataset. Then, the preprocessed training sample dataset is used to perform supervised training on the initial deep neural network model to obtain the optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations. The optimal prediction result includes the predicted optimal base station number for the next time step and the attribute information of the predicted optimal beam pair for the next time step. Next, the error value between the predicted optimal result and the actual optimal result is determined. The actual optimal result includes the actual optimal base station number for the next time step and the attribute information of the actual optimal beam pair for the next time step. If the error value meets the preset error requirement, the intermediate deep neural network model corresponding to the error value is determined as the multi-base station cooperative beam tracking prediction model. If the error value does not meet the preset error requirement, the preprocessed training sample dataset is used to perform supervised training on the intermediate deep neural network model corresponding to the error value to determine the multi-base station cooperative beam tracking prediction model.

[0106] Specifically, refer to Figure 4 , Figure 5 and Figure 6The edge server first preprocesses the acquired training sample dataset, and then feeds the preprocessed training sample dataset into the initial deep neural network model in batches for supervised training. Each preprocessed training sample data is fed into the DNN model and calculated layer by layer until the output layer is reached. After the DNN model has been trained a preset number of times, the error between the predicted optimal result and the actual optimal result can be calculated using a loss function for the intermediate deep neural network model after the preset number of training times. Common loss functions include mean square error (MSE) and cross entropy loss. Backpropagation calculates the gradient of the loss function with respect to each layer of the DNN model according to the chain rule. Weight update uses the weight gradient obtained in backpropagation to update the weights of each layer.

[0107] Furthermore, when the prediction metric of the DNN model is a discrete value such as the predicted sequence number of the optimal base station or the predicted transceiver beam number, the loss function can be defined as the cross-entropy value L between the One-Hot encoded vectors of the predicted optimal result and the actual sequence number or actual transceiver beam number in the actual optimal result. cross The calculation formula is shown in equation (13):

[0108]

[0109] In equation (13), n represents the number of training iterations in a training batch, and m represents the number of predicted result categories. This represents the optimal prediction result for the k-th class after the i-th supervised training. Indicates and This represents the actual optimal result of the size comparison.

[0110] When the prediction metric of the DNN model is the predicted departure angle of the optimal beam pair, and this prediction metric is a continuous value, the loss function can be defined as the MSE value L between the predicted optimal result and the actual optimal result. MSE The calculation formula is shown in equation (1):

[0111]

[0112] In equation (14), Let y represent the optimal prediction result after the i-th supervised training. i Indicates and The actual optimal result for comparison, where n represents the number of training iterations included in a training batch.

[0113] The error between the predicted optimal result and the actual optimal result can be obtained through the loss function of equation (13) or (14). This error value can be used to update parameters such as the weight set bias during backpropagation. The training objective of the neural network is to find the parameter set θ that minimizes the loss function, as shown in equation (15):

[0114]

[0115] There are many optimization algorithms for deep networks, including stochastic gradient descent (SGD), root mean square propagation (RMSProp), and adaptive moment estimation (Adam). The Adam algorithm is based on the SGD algorithm and is combined with the RMSProp algorithm, which has the advantages of supporting sparse gradients and fast convergence speed. Therefore, this invention uses the Adam algorithm to optimize the DNN model, and obtains the adaptive learning rate of each parameter by calculating the first and second moments of the gradient, as shown in equations (16) to (18):

[0116]

[0117]

[0118]

[0119] In equations (16) to (18), α represents the learning rate, also called the compensation factor. The role of α is to control the weight update ratio; m t This represents the exponentially moving mean of the gradients after the t-th supervised training iteration. Represents the exponential moving average m t The calibration value, v t This represents the squared gradient after the t-th supervised training iteration. Represents the squared gradient v t The standard value of θ t Let θ represent the set of parameters for the loss function after the t-th supervised training. t+1 Let represent the parameter set of the loss function after the (t+1)th supervised training; ε is a very small compensation amount to prevent the denominator of the fraction from being zero; This represents the exponential decay of the first-moment estimate after the t-th supervised training. This represents the exponential decay of the second-order moment estimate after the t-th supervised training. Furthermore, the exponentially moving mean m... t Sum of squared gradients v t The update formulas are shown in equations (19) and (20):

[0120]

[0121]

[0122] In equations (19) and (20), m t-1 Let v represent the exponentially moving mean of the gradients after the (t-1)th supervised training iteration. t-1 Let g represent the squared gradient after the (t-1)th supervised training iteration. t Let represent the gradient of the weights after the t-th supervised training.

[0123] It is understood that the edge server preprocesses the training sample dataset to determine the preprocessed training sample dataset, and the implementation process may include:

[0124] When the actual optimal beam pair attribute information is the actual optimal beam angle or the actual optimal transceiver beam number, the training sample dataset is divided into a first training sample data subset containing the actual optimal beam angle and a second training sample data subset containing the actual optimal transceiver beam number.

[0125] The first training sample data subset is subjected to mean and variance standardization, and the second training sample data subset is subjected to One-Hot encoding to determine the preprocessed training sample dataset.

[0126] Specifically, the edge server obtains the training sample dataset for preprocessing. Depending on the data type, the preprocessing method can divide the training sample dataset into a first training sample data subset and a second sample data subset.

[0127] For the adaptive codebook-based scheme, the features used for DNN model training are the actual departure angle and the actual radar echo power. Both of these are continuous data. Therefore, the mean and variance of each first training sample in the first training sample data subset are standardized, i.e., x norm = (x-μ) / σ, where μ represents the sequence mean of the first training sample data, σ represents the sequence variance of the first training sample data, and x is the preprocessed first training sample data. norm It follows a standard normal distribution; standardization can accelerate the convergence speed of DNN models and improve their prediction accuracy.

[0128] For schemes based on a fixed codebook, the input features of the DNN model are the actual transmitted beam number and the actual radar echo power. Since the actual transmitted beam number is discrete data, One-Hot encoding is used to encode the category of the actual transmitted beam number into a binary feature vector, where the element corresponding to the actual transmitted beam number is set to 1, and the remaining elements are set to 0. One-Hot encoding extends the values ​​of discrete data to Euclidean space, which is beneficial for calculating the correlation between data.

[0129] Furthermore, the preprocessing method for the training set labels is consistent with the above: the optimal base station number providing the best service is encoded using One-Hot encoding; the actual optimal beam angle (based on the adaptive codebook label) is normalized using mean-variance; and the actual optimal transceiver beam number (based on the fixed codebook label) is encoded using One-Hot encoding. Moreover, both the first and second training sample data subsets, after preprocessing, can be directly fed into the DNN model for supervised learning.

[0130] It should be noted that, in terms of vehicle trajectory, the straight-line motion of the vehicle is considered, and the ULA antenna array of the base station is parallel to the direction of vehicle motion. The AOD and AOA are the same. Therefore, the beam combining vector can be generated using the AOD predicted by the base station.

[0131] Understandably, the edge server uses the preprocessed training sample dataset to perform supervised training on the initial deep neural network model, obtaining the optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations. The specific implementation process includes:

[0132] The first training sample data subset after mean-variance standardization is used to perform supervised regression and classification training on the initial deep neural network model. The first optimal prediction result of the intermediate deep neural network model after a preset number of training times is obtained. The first optimal prediction result includes the optimal base station number for communication between the vehicle and the next time step and the optimal beam angle for the next time step.

[0133] Using the second training sample data subset after One-Hot encoding, supervised classification training is performed on the initial deep neural network model to obtain the second optimal prediction result output by the intermediate deep neural network model after a preset number of training times. The second optimal prediction result includes the predicted optimal base station number for communication between the vehicle and the next time step and the predicted optimal transceiver beam number for the next time step.

[0134] Specifically, the initial deep neural network model is the DNN model to be learned. The DNN model uses different prediction methods depending on the beamforming method. For beamforming based on an adaptive codebook, the DNN model performs regression, predicting the optimal beamforming angle (AOD) of the best beam pair. For beamforming based on a fixed codebook, the DNN model performs classification, predicting the transmit / receive beam indices of the optimal beam pair. For the optimal base station indices providing the best service, the DNN model is used uniformly for classification prediction. Therefore, when the input data to the DNN model is the first training sample data subset after mean-variance standardization, the initial deep neural network model undergoes supervised regression and classification training. When the input data to the DNN model is the second training sample data subset after One-Hot encoding, the initial deep neural network model undergoes supervised classification training.

[0135] It should be noted that the DNN model is trained offline using supervised learning. If the optimal beam pair attribute information of multiple base stations at the current time and the actual radar echo intensity received by these base stations are input into the DNN model, the predicted optimal base station number and predicted optimal beam pair attribute information for the next time step can be output. Since the actual optimal base station number for communication with the vehicle at each time step and the actual optimal beam pair attribute information at each time step are known, after a preset number of supervised training iterations, the output predicted optimal result can be compared with the corresponding actual optimal result. This allows determination of whether the intermediate deep neural network model obtained after the preset number of training iterations is a well-trained multi-base station cooperative beam tracking prediction model.

[0136] Furthermore, it should be noted that the training features of the first training sample data subset generated based on the adaptive codebook beamforming method include the radar echo intensity received by each base station and the actual optimal beam angle. Both of these quantities are continuous values, so the data preprocessing for the feature part needs to be normalized by mean and variance. The labels in the first training sample data subset include the actual optimal base station number and the actual optimal beam angle. The actual optimal base station number is a discrete value and needs to be preprocessed by One-Hot encoding, while the actual optimal beam angle is a continuous value and needs to be normalized by mean and variance.

[0137] For the training features of the second training sample data subset based on the fixed codebook beamforming method, including the radar echo intensity received by each base station and the actual optimal transmit beam number, these two quantities are of different types. The radar echo intensity is a continuous value, so it needs to be preprocessed using mean-variance standardization; while the actual optimal transmit beam number is a discrete value, so it needs to be preprocessed using One-Hot encoding. The labels of the second training sample data subset include the actual optimal base station number and the actual optimal transceiver beam number. Both of these quantities are discrete values, so they both need to be preprocessed using One-Hot encoding.

[0138] Furthermore, the training tasks for the DNN model are divided into two categories: when the prediction target is the optimal base station number or the optimal transceiver beam number, the DNN model is trained for classification; when the prediction target is the optimal beam angle, the DNN model is trained for regression. Therefore, for the first and second training sample data subsets determined by the two beamforming methods, the first beamforming method has two prediction metrics, requiring the training of two models: one for classification (predicting the base station number) and one for regression (predicting the transmit beam angle). For the second beamforming method, three models need to be trained, all performing classification tasks (one model for each of the base station number, transmit beam number, and receive beam number).

[0139] The multi-base station cooperative beam tracking prediction method provided by this invention is applied to millimeter-wave V2I network scenarios. These scenarios are highly dynamic, integrating sensing and communication, where multiple base stations serve a vehicle. One base station has a LOS (Low-Insulation / Loss) path with the vehicle, while the other base stations have NLOS (Low-Insulation / Loss) paths with the vehicle. Furthermore, the execution entity of the multi-base station cooperative beam tracking prediction method is an edge server. The functions and application scenarios of the edge server can be referenced in the aforementioned model training method embodiments.

[0140] Reference Figure 7 This invention provides a multi-base station cooperative beam tracking prediction method, such as... Figure 7 As shown, the multi-base station cooperative beam tracking prediction method includes the following steps:

[0141] Step 710: Obtain the current radar echo intensity received by each base station at the current moment, the current transmitted beam attribute information received by each base station at the current moment, and the multi-base station cooperative beam tracking prediction model trained according to the aforementioned model training method.

[0142] Step 720: Input the current radar echo intensity and the attribute information of each current transmitted beam into the multi-base station cooperative beam tracking prediction model for tracking prediction, and obtain the target optimal base station number for communication between the target and the vehicle at the target time, as well as the target optimal beam pair attribute information at the target time.

[0143] Step 730: Send the target optimal base station number and the target optimal beam pair attribute information to the target optimal base station and the vehicle corresponding to the target optimal base station number, respectively, to prompt the target optimal base station and the vehicle to communicate.

[0144] The target time is the next time after the current time; the target optimal beam pair attribute information can be the target departure angle of the target optimal beam pair or the target transceiver beam number. In addition, the current transmission beam attribute information includes the current transmission beam angle or the current transmission beam number.

[0145] Specifically, the edge server inputs the current radar echo intensity and the attribute information of each currently transmitted beam into a trained multi-base station cooperative beam tracking prediction model to predict the target time information. This allows it to obtain the target optimal base station number for communication between the target and the vehicle at that time, as well as the target optimal beam pair attribute information. The edge server then sends the target optimal base station number and target optimal beam pair attribute information to the corresponding target optimal base station and the vehicle, respectively, to prompt them to communicate, thereby completing beam tracking and switching.

[0146] Figure 8 The comparison diagram shows different beam tracking schemes, examining the downlink spectral efficiency of the proposed method versus other comparative methods as a function of downlink transmit power; the number of antennas for the three RSUs is set to N. t =N r =64, the number of vehicle antennas is set to M=16; the five beam tracking schemes are described below:

[0147] (1) The method of this invention is based on adaptive codebook beam tracking predicted by a multi-base station cooperative beam tracking prediction model; the method is based on adaptive codebook beamforming, and the DNN model predicts the optimal service RSU number and the optimal AOD for beamforming at the next moment. In linear motion, AOA and AOD are the same, and the transceiver uses AOD / AOA to generate beamforming vector and beam combining vector, thereby completing communication.

[0148] (2) The method of the present invention is based on fixed codebook beam tracking predicted by a multi-base station cooperative beam tracking prediction model: This method is similar to (1), but uses DFT codebook for beamforming. The result of the DNN model prediction is the optimal RSU number and the optimal beam number of each of the transmitting and receiving ends.

[0149] (3) Multi-base station cooperative beam tracking based on radar-assisted and extended Kalman filter (EKF): This method uses EKF and multi-point cooperation for beam tracking. Since EKF tracks and predicts AOD changes, this scheme also adopts beamforming based on adaptive codebook.

[0150] (4) Relay prediction (adaptive codebook): This method is for the adaptive codebook approach, which uses the AOD information predicted by the DNN as the input for the next prediction and continues to predict in succession.

[0151] (5) Random beamforming (adaptive codebook): This method randomly selects AOD for beamforming at each time step without relying on any reference information.

[0152] from Figure 8 As can be seen, the method of this invention is very close to the performance upper bound. At a downlink transmit power of 20dBm, the spectral efficiency of the DNN prediction method based on adaptive codebooks reaches 86% of the corresponding performance upper bound, while the spectral efficiency of the DNN prediction method based on fixed codebooks reaches 91%. Furthermore, since the fixed codebook quantizes spatial angles, the optimal AOD may not be included in all beam directions generated by codebook quantization; that is, there may be no codeword in the codebook whose corresponding beam direction is completely consistent with the optimal AOD. The adaptive codebook scheme, however, has no limitation on quantization accuracy and may obtain a beam direction consistent with the optimal AOD. Therefore, the performance upper bound of the fixed codebook scheme is slightly lower than that of the adaptive codebook scheme. Compared with the other three schemes, the method of this invention shows a significant improvement. The multi-base station cooperative beam tracking prediction method based on radar-assisted and EKF is greatly affected by the NLOS path. When the LOS path between two of the three RSUs and the vehicle is blocked, the perception information provided by these two base stations has a large error, which will seriously affect the result of multi-base station cooperation. Therefore, this method performs poorly in NLOS scenarios. Compared to the proposed scheme, the method of this invention improves spectral efficiency by 5.4 times and 5.1 times under adaptive codebook and fixed codebook approaches, respectively. The relay prediction scheme leads to error propagation and increases the error in each prediction process, resulting in poor performance. The random beamforming scheme, as the lower bound, performs the worst. In summary, the method of this invention can effectively complete beam tracking tasks when the LOS path is blocked.

[0153] Figure 9 The curves showing the prediction results of the method of this invention as a function of sampling points in the adaptive codebook mode are presented. The number of antennas for the three RSUs is set to N. t =N r =64, the number of antennas on the vehicle is set to M=16, and the downlink transmit power is 20dBm. In Figure 9In the figure, (A) shows the trend of AOD predicted by the DNN with the number of sampling points. It can be seen from the figure that the AOD predicted by the model at each sampling point is very close to the actual AOD. Since AOD is a continuous value, the DNN performs regression prediction, and the predicted value can only approximate the actual value infinitely, but cannot be completely consistent. (B) shows the trend of the best service RSU index with the number of sampling points. Figure 9 As can be seen, the best service RSU index predicted by the model is consistent with the actual best service RSU index at most sampling points. In this case, when the DNN performs multi-class prediction, the predicted value and the true value can be the same.

[0154] Figure 10 The curves showing the prediction results of the proposed method as a function of sampling points under a fixed codebook scheme are presented. The number of antennas for the three RSUs is set to N. t =N r =64, the number of antennas on the vehicle is set to M=16, and the downlink transmit power is 20dBm. In Figure 10 In the diagram, (A), (B), and (C) show the trends of the transmit beam number, receive beam number, and best service unit (RSU) number predicted by the DNN as a function of sampling points. These three prediction metrics are all discrete values; therefore, the DNN performs multi-class prediction. Figure 10 As can be seen, most sampling point DNN models can accurately predict the sequence number of the best serving RSU and the best beam pair.

[0155] The model training apparatus provided by the present invention is described below. The model training apparatus described below and the model training method described above can be referred to in correspondence.

[0156] Reference Figure 11 This is a schematic diagram of the structure of the model training device provided by the present invention, as shown below. Figure 11 As shown, the model training device 1100, applied to the millimeter-wave V2I network scenario described in the aforementioned model training method, includes:

[0157] The first acquisition unit 1110 is used to acquire a training sample dataset. Each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the corresponding time, the actual optimal beam pair attribute information for the corresponding time, the actual radar echo intensity received by multiple base stations at the corresponding time, and the actual beam pair attribute information for the corresponding time. The actual beam pair attribute information for the corresponding time includes the best beam pair attribute information between each base station and the vehicle at the corresponding time, and the actual optimal beam pair attribute information for the corresponding time is the optimal beam pair attribute information selected from each optimal beam pair attribute information at the corresponding time.

[0158] Training unit 1120 is used to perform supervised training on an initial deep neural network model based on a training sample dataset to determine a multi-base station cooperative beam tracking prediction model; wherein, supervised training is used to learn the nonlinear correlation between channels at adjacent time points.

[0159] Understandably, the model training unit 1120 can be specifically used to obtain the actual optimal beam pair attribute information and the actual optimal base station number at the corresponding time, when the vehicle sends SRS signals to all beam directions of each base station at the corresponding time, each base station receives each SRS signal, and each base station determines the optimal beam pair attribute information and RSRP value between itself and the vehicle at the corresponding time based on the received SRS signals; when multiple base stations send DFRC signals to the vehicle in a directional manner, it instructs multiple base stations to report radar echo intensity and optimal beam pair attribute information at the corresponding time; and to determine training sample data based on the actual optimal beam pair attribute information, the actual optimal base station number at the corresponding time, the actual radar echo intensity reported by multiple base stations at the corresponding time, and the optimal beam pair attribute information between itself and the vehicle reported by each base station at the corresponding time.

[0160] Understandably, the model training unit 1120 can also be used to preprocess the training sample dataset to determine the preprocessed training sample dataset; to perform supervised training on the initial deep neural network model using the preprocessed training sample dataset, and to obtain the optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations; the optimal prediction result includes the optimal predicted base station number and the optimal predicted beam pair attribute information for the next time step; to determine the error value between the optimal prediction result and the actual optimal result, where the actual optimal result includes the actual optimal base station number and the actual optimal beam pair attribute information for the next time step; if the error value meets the preset error requirement, the intermediate deep neural network model corresponding to the error value is determined to be a multi-base station cooperative beam tracking prediction model; if the error value does not meet the preset error requirement, the intermediate deep neural network model corresponding to the error value is subjected to supervised training using the preprocessed training sample dataset to determine the multi-base station cooperative beam tracking prediction model.

[0161] Understandably, the model training unit 1120 can also be used to divide the training sample dataset into a first training sample dataset containing the actual optimal beam angle and a second training sample dataset containing the actual optimal transceiver beam number, when the actual optimal beam pair attribute information is the actual optimal beam angle or the actual optimal transceiver beam number; perform mean and variance standardization on the first training sample dataset and One-Hot encoding on the second training sample dataset to determine the preprocessed training sample dataset.

[0162] It is understood that the model training unit 1120 can also be used to perform supervised regression and classification training on the initial deep neural network model using the first training sample data subset after mean-variance standardization, and obtain the first optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations. The first optimal prediction result includes the predicted optimal base station number for communication between the vehicle and the next time step and the predicted optimal beam angle for the next time step. Using the second training sample data subset after One-Hot encoding, the initial deep neural network model is used to perform supervised classification training, and obtain the second optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations. The second optimal prediction result includes the predicted optimal base station number for communication between the vehicle and the next time step and the predicted optimal transceiver beam number for the next time step.

[0163] It is understandable that in the first acquisition unit 1110, when the vehicle is moving in a straight line and the ULA antenna array of each base station is parallel to the direction of the vehicle's movement, the actual radar echo intensity received by multiple base stations at the corresponding time includes the radar echo intensity reflected by the vehicle at the corresponding time and the radar echo intensity reflected by the obstacle at the corresponding time.

[0164] The multi-base station cooperative beam tracking device provided by the present invention is described below. The multi-base station cooperative beam tracking device described below and the multi-base station cooperative beam tracking prediction method described above can be referred to in correspondence.

[0165] Reference Figure 12 This is a schematic diagram of the structure of the multi-base station cooperative beam tracking device provided by the present invention, as shown below. Figure 12 As shown, the multi-base station cooperative beam tracking device 1200 includes:

[0166] The second acquisition unit 1210 is used to acquire the current radar echo intensity received by each base station at the current time, the current transmitted beam attribute information received by each base station at the current time, and the multi-base station cooperative beam tracking prediction model trained according to the aforementioned model training method.

[0167] The prediction and tracking unit 1220 is used to input the current radar echo intensity and the attribute information of each current transmitted beam into the multi-base station cooperative beam tracking prediction model for tracking and prediction, and to obtain the target optimal base station number for communication between the target and the vehicle at the target time, as well as the target optimal beam pair attribute information at the target time.

[0168] The information sending unit 1230 is used to send the target optimal base station number and the target optimal beam pair attribute information to the target optimal base station and the vehicle corresponding to the target optimal base station number, respectively, so as to prompt the target optimal base station and the vehicle to communicate.

[0169] Figure 13 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 13 As shown, the electronic device 1300 may include: a processor 1310, a communication interface 1320, a memory 1330, and a communication bus 1340, wherein the processor 1310, the communication interface 1320, and the memory 1330 communicate with each other through the communication bus 1340. The processor 1310 can call logical instructions in the memory 1330 to execute a model training method, which includes:

[0170] Obtain the training sample dataset. Each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the base station at the corresponding time, the actual optimal beam pair attribute information at the corresponding time, the actual radar echo intensity received by multiple base stations at the corresponding time, and the actual beam pair attribute information at the corresponding time. Among them, the actual beam pair attribute information at the corresponding time includes the best beam pair attribute information between each base station and the vehicle at the corresponding time, and the actual optimal beam pair attribute information at the corresponding time is the optimal beam pair attribute information selected from each of the best beam pair attribute information at the corresponding time.

[0171] Based on the training sample dataset, an initial deep neural network model is subjected to supervised training to determine a multi-base station cooperative beam tracking prediction model; wherein, supervised training is used to learn the nonlinear correlation between channels at adjacent time points. Alternatively, a multi-base station cooperative beam tracking prediction method can be implemented, comprising:

[0172] Obtain the current radar echo intensity received by each base station at the current moment, the current transmitted beam attribute information received by each base station at the current moment, and the multi-base station cooperative beam tracking prediction model trained according to the aforementioned model training method;

[0173] The current radar echo intensity and current transmitted beam attribute information are input into the multi-base station cooperative beam tracking prediction model for tracking and prediction, to obtain the target optimal base station number for communication between the target and the vehicle at the target time, as well as the target optimal beam pair attribute information at the target time.

[0174] The target optimal base station number and the target optimal beam pair attribute information are sent to the target optimal base station and the vehicle corresponding to the target optimal base station number, respectively, to prompt the target optimal base station and the vehicle to communicate.

[0175] Furthermore, the logical instructions in the aforementioned memory 1330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0176] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the execution model training method provided by the above methods, the method comprising:

[0177] Obtain the training sample dataset. Each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the base station at the corresponding time, the actual optimal beam pair attribute information at the corresponding time, the actual radar echo intensity received by multiple base stations at the corresponding time, and the actual beam pair attribute information at the corresponding time. Among them, the actual beam pair attribute information at the corresponding time includes the best beam pair attribute information between each base station and the vehicle at the corresponding time, and the actual optimal beam pair attribute information at the corresponding time is the optimal beam pair attribute information selected from each of the best beam pair attribute information at the corresponding time.

[0178] Based on the training sample dataset, an initial deep neural network model is subjected to supervised training to determine a multi-base station cooperative beam tracking prediction model; wherein, supervised training is used to learn the nonlinear correlation between channels at adjacent time points. Alternatively, a multi-base station cooperative beam tracking prediction method can be implemented, comprising:

[0179] Obtain the current radar echo intensity received by each base station at the current moment, the current transmitted beam attribute information received by each base station at the current moment, and the multi-base station cooperative beam tracking prediction model trained according to the aforementioned model training method;

[0180] The current radar echo intensity and current transmitted beam attribute information are input into the multi-base station cooperative beam tracking prediction model for tracking and prediction, to obtain the target optimal base station number for communication between the target and the vehicle at the target time, as well as the target optimal beam pair attribute information at the target time.

[0181] The target optimal base station number and the target optimal beam pair attribute information are sent to the target optimal base station and the vehicle corresponding to the target optimal base station number, respectively, to prompt the target optimal base station and the vehicle to communicate.

[0182] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0183] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A model training method, characterized in that, This is applied to millimeter-wave V2I network scenarios, where multiple base stations serve a vehicle in a highly dynamic, integrated communication scenario. In these scenarios, one of the base stations has a LOS path between itself and the vehicle, while the other base stations have an NLOS path between themselves and the vehicle. The model training method includes: A training sample dataset is obtained, wherein each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the corresponding time, the actual optimal beam pair attribute information for the corresponding time, the actual radar echo intensity received by the multiple base stations at the corresponding time, and the actual beam pair attribute information for the corresponding time; wherein, the actual beam pair attribute information for the corresponding time includes the optimal beam pair attribute information between each base station and the vehicle at the corresponding time, and the actual optimal beam pair attribute information for the corresponding time is the optimal beam pair attribute information selected from each of the optimal beam pair attribute information at the corresponding time; Based on the training sample dataset, the initial deep neural network model is subjected to supervised training to determine the multi-base station cooperative beam tracking prediction model; wherein, the supervised training is used to learn the nonlinear correlation between channels at adjacent time points.

2. The model training method according to claim 1, characterized in that, The process of obtaining the training sample data for each training sample in the training sample dataset includes: In the case that at the corresponding time the vehicle sends SRS signals to all beam directions of each base station, each base station receives each SRS signal, and each base station determines the optimal beam pair attribute information and RSRP value between the vehicle and the corresponding time based on the received SRS signals, the actual optimal beam pair attribute information and the actual optimal base station number at the corresponding time are obtained. When the multiple base stations respectively send DFRC signals to the vehicle in a directional manner, the multiple base stations are instructed to report the radar echo intensity and the optimal beam pair attribute information at the corresponding time. The training sample data is determined based on the actual optimal beam pair attribute information at the corresponding time, the actual optimal base station number at the corresponding time, the actual radar echo intensity reported by the multiple base stations at the corresponding time, and the optimal beam pair attribute information between each base station and the vehicle at the corresponding time.

3. The model training method according to claim 1, characterized in that, The step of performing supervised training on the initial deep neural network model based on the training sample dataset to determine the multi-base station cooperative beam tracking prediction model includes: The training sample dataset is preprocessed to determine the preprocessed training sample dataset; The preprocessed training sample dataset is used to perform supervised training on the initial deep neural network model to obtain the optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations; the optimal prediction result includes the optimal predicted base station number at the next time step and the optimal predicted beam pair attribute information at the next time step. Determine the error value between the predicted optimal result and the actual optimal result, wherein the actual optimal result includes the actual optimal base station number at the next time step and the actual optimal beam pair attribute information at the next time step; If the error value meets the preset error requirement, the intermediate deep neural network model corresponding to the error value is determined as the multi-base station cooperative beam tracking prediction model. If the error value does not meet the preset error requirement, the intermediate deep neural network model corresponding to the error value is subjected to supervised training using the preprocessed training sample dataset to determine the multi-base station cooperative beam tracking prediction model.

4. The model training method according to claim 3, characterized in that, The step of preprocessing the training sample dataset to determine the preprocessed training sample dataset includes: When the actual optimal beam pair attribute information is the actual optimal beam angle or the actual optimal transceiver beam number, the training sample dataset is divided into a first training sample data subset containing the actual optimal beam angle and a second training sample data subset containing the actual optimal transceiver beam number. The first training sample data subset is subjected to mean-variance standardization, and the second training sample data subset is subjected to One-Hot encoding to determine the preprocessed training sample dataset.

5. The model training method according to claim 4, characterized in that, The step of using the preprocessed training sample dataset to perform supervised training on the initial deep neural network model, and obtaining the optimal prediction result output by the intermediate deep neural network model after a preset number of training iterations, includes: The first training sample data subset after mean-variance standardization is used to perform supervised regression and classification training on the initial deep neural network model to obtain the first optimal prediction result output by the intermediate deep neural network model after the preset number of training times. The first optimal prediction result includes the optimal predicted base station number for communication between the vehicle and the next time step and the optimal predicted beam angle for the next time step. Using the second training sample data subset encoded with One-Hot, supervised classification training is performed on the initial deep neural network model to obtain the second optimal prediction result output by the intermediate deep neural network model after the preset number of training iterations. The second optimal prediction result includes the predicted optimal base station number for communication between the vehicle and the next time step and the predicted optimal transceiver beam number for the next time step.

6. The model training method according to any one of claims 1 to 5, characterized in that, When the vehicle is moving in a straight line and the ULA antenna array of each base station is parallel to the direction of movement of the vehicle, the actual radar echo intensity received by the plurality of base stations at the corresponding moment includes the radar echo intensity reflected by the vehicle at the corresponding moment and the radar echo intensity reflected by the obstacle at the corresponding moment.

7. A multi-base station cooperative beam tracking prediction method, characterized in that, This is applied to millimeter-wave V2I network scenarios, where multiple base stations serve vehicles in a highly dynamic, integrated communication and sensing scenario. In these scenarios, one of the base stations has a LOS path for communication with the vehicle, while the other base stations have an NLOS path for communication with the vehicle. The multi-base station cooperative beam tracking prediction method includes: The current radar echo intensity received by each base station at the current moment, the current transmitted beam attribute information received by each base station at the current moment, and the multi-base station cooperative beam tracking prediction model trained according to any one of the model training methods of claims 1 to 6 are obtained. The current radar echo intensity and the current transmitted beam attribute information are input into the multi-base station cooperative beam tracking prediction model for tracking and prediction, to obtain the target optimal base station number for communication between the target time and the vehicle, and the target optimal beam pair attribute information at the target time. The target optimal base station number and the target optimal beam pair attribute information are sent to the target optimal base station corresponding to the target optimal base station number and the vehicle, respectively, to prompt the target optimal base station and the vehicle to communicate.

8. A model training device, characterized in that, Applied to the millimeter-wave V2I network scenario described in any one of claims 1 to 6, including: The first acquisition unit is used to acquire a training sample dataset. Each training sample in the training sample dataset includes the actual optimal base station number for communication between the vehicle and the corresponding time, the actual optimal beam pair attribute information for the corresponding time, the actual radar echo intensity received by the plurality of base stations at the corresponding time, and the actual beam pair attribute information for the corresponding time. The actual beam pair attribute information for the corresponding time includes the optimal beam pair attribute information between each base station and the vehicle at the corresponding time, and the actual optimal beam pair attribute information for the corresponding time is the optimal beam pair attribute information selected from each of the optimal beam pair attribute information at the corresponding time. The model training unit is used to perform supervised training on the initial deep neural network model based on the training sample dataset to determine the multi-base station cooperative beam tracking prediction model; wherein, the supervised training is used to learn the nonlinear correlation between channels at adjacent time points.

9. A multi-base station cooperative beam tracking device, characterized in that, include: The second acquisition unit is used to acquire the current radar echo intensity received by each base station at the current time, the current transmitted beam attribute information received by each base station at the current time, and the multi-base station cooperative beam tracking prediction model trained according to any one of the model training methods of claims 1 to 6. The prediction and tracking unit is used to input the current radar echo intensity and the current transmitted beam attribute information into the multi-base station cooperative beam tracking prediction model for tracking and prediction, and to obtain the target optimal base station number for communication between the target and the vehicle at the target time, as well as the target optimal beam pair attribute information at the target time. The information sending unit is used to send the target optimal base station number and the target optimal beam pair attribute information to the target optimal base station corresponding to the target optimal base station number and the vehicle, respectively, so as to prompt the target optimal base station and the vehicle to communicate.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the model training method as described in any one of claims 1 to 6, or the multi-base station cooperative beam tracking prediction method as described in claim 7.

Images