Communication beam selection method based on sub6GHz channel and partial millimeter wave pilot

Abstract

The present invention discloses a communication beam selection method based on sub6GHz channels and partial millimeter-wave pilots, which predicts the optimal beam of the user through the uplink sub6GHz channel and the downlink millimeter-wave pilot sent by a small number of activated antennas. Specifically, the present invention designs a new dual-input neural network architecture that can fuse sub-6GHz band channel information and partial information of millimeter-wave band channels, and output the probability that each beam is the optimal beam. At the same time, in order to enable the model to have the ability to quickly generalize to different base station scenarios, the meta-learning MAML algorithm is used to update the model parameters, so that the pre-trained model has a more powerful learning ability and can quickly converge to the ideal performance through a small number of samples in the new base station scenario. Compared with traditional beam scanning, the algorithm proposed in the present invention reduces the time overhead, and also avoids the huge training overhead caused by obtaining the CSI of the complete millimeter-wave channel.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications and deep learning, and specifically relates to a communication beam selection method based on sub6GHz channels and partial millimeter-wave pilot frequencies. Background Art

[0002] Millimeter wave and massive MIMO are one of the key technologies of 5G mobile communications. Millimeter wave communications in the 30-300GHz frequency band can provide a huge transmission bandwidth and are the key compensation for the current sub-6GHz frequency band. In order to combat the severe path loss in the millimeter wave frequency band, a large number of antennas are usually deployed at the transmitting end to form a highly directional beam to the communication user.

[0003] In actual communications, codebooks are usually used to implement beamforming, and each codeword in the codebook corresponds to a beam. At present, the codebook-based beamforming method mainly uses an exhaustive search algorithm. The base station and the user end traverse the codewords in the codebook respectively to find the beam with the largest received power as the optimal beam. However, as the number of antennas on the base station side increases, the codebook also increases, and the use of exhaustive search will bring a lot of time overhead. Subsequent researchers adopted a hierarchical search algorithm based on the exhaustive search, dividing the search into two stages: sector-level scanning and beam-level scanning. At the sector level, the base station will use a wide beam to scan the space at all angles. After determining the best sector, it will use a narrow beam to determine the user's optimal beam. The hierarchical search algorithm has been greatly improved compared to the exhaustive search method, but it still has a considerable time overhead and there is still room for optimization.

[0004] With the development of deep learning technology, researchers have begun to try to use deep learning algorithms for beam training to reduce time overhead costs. Since the complete acquisition of millimeter wave channel CSI requires a huge training overhead, people try to use out-of-band information (OOB) for beam training, that is, using the channel CSI obtained in another frequency band to assist the transmission of the current frequency band. Some researchers have simulated that the proportion of paths with common angles in the propagation paths with frequencies from 900MHz to 90GHz exceeds 90%, which shows that there is a correlation between the channel information through the frequency band below 6GHz and the millimeter wave channel information. The existing algorithms for beam training through sub6 channels have achieved certain results, but the performance is relatively general in complex scenarios, and the model does not have good generalization, resulting in the model often needing to be retrained when it is in a new communication scenario.

[0005] The algorithm proposed in the present invention proposes to use part of the millimeter-wave pilot channel to correct the prediction error of the sub6 channel, and the activated millimeter-wave antenna is adaptively selected by the input sub6 channel information. In different scenarios, the base station will choose to activate different antennas to assist in prediction. At the same time, in order to make the model generalize and adapt to more communication scenarios, the MAML algorithm based on meta-learning is used to update the model parameters during the model parameter training phase. The pre-trained model based on meta-learning has better learning ability, so that when the model is deployed in other communication scenarios, it can achieve ideal prediction performance only through rapid iteration of small samples, thereby improving the deployment feasibility of the model. Summary of the invention

[0006] In order to reduce the downlink beam training overhead of the millimeter wave communication system, the present invention proposes a communication beam selection method based on sub6GHz channels and partial millimeter wave pilots. This algorithm uses partial millimeter wave pilot channels to correct the prediction error of the sub6 channel, and the activated millimeter wave antenna is adaptively selected by the input sub6 channel information. At the same time, the meta-learning algorithm MAML is used to improve the generalization performance of the model.

[0007] To achieve the purpose of the present invention, the technical solution adopted by the present invention is: a communication beam selection method based on sub6GHz channels and partial millimeter wave pilots, and its training process includes the following steps:

[0008] Step (1): Construct n task subtasks, each of which corresponds to the communication environment of a base station. Each Task includes a support set and a query set. Build a beam selection prediction model, including sub6 module, antenna selection module, millimeter wave module, classification module, and initialize model parameters Θ;

[0009] Step (2): For subtask T i , input the sub6 channel CSI in the support set into the sub6 module to obtain the sub6 feature vector. Input the sub6 feature vector into the antenna selection module to obtain the activated millimeter wave antenna index. Obtain the pilot of the activated antenna in the support set, input it into the millimeter wave module, and obtain the millimeter wave feature vector; fuse the sub6 feature vector with the millimeter wave feature vector, input it into the classification module, and obtain the probability that each beam is the optimal beam;

[0010] Step (3): Calculate the loss function of the support set and update the model parameters in reverse. After L times of gradient update, the updated model parameters are obtained.

[0011] Step (4): Based on model parameters Make predictions on the query set and calculate the loss function. Perform reverse gradient descent based on the loss function of all tasks in the batch and update the model parameters Θ (t+1) ;

[0012] Read the subtask of the next batch and repeat steps (2) to (4) until the model converges.

[0013] Step (5): Based on the probability vector output by the model, obtain the Top-1 optimal beam and the Top-3 optimal beam.

[0014] Preferably, the specific steps of step (1) include:

[0015] (1.1) Subtask division: The basic training unit of meta-learning is Task, i.e., subtask. Each subtask corresponds to a communication scenario in which the base station is located. Each subtask is divided into a support set and a query set.

[0016] (1.2) The support set is Composition, of which represents the sub6GHz channel matrix between the user and the base station, N s Indicates the number of base station antennas in the sub6 frequency band, N c Indicates the number of carriers. represents the millimeter wave channel information between the user and the base station, N m Indicates the number of base station antennas in the millimeter wave frequency band. spt represents the codeword index corresponding to the best beam between the user and the base station. The query set is composed of The meaning of each part is the same as that of the support set.

[0017] The support set is used for innerloop training of the model, and the query set is used to evaluate the reasoning ability of the model after the innerloop and to perform outerloop training on the model.

[0018] Preferably, the specific steps of step (2) include:

[0019] (2.1) sub6 module: X sub6 Converted into a 1-dimensional vector, input into the sub6 module, and output the feature vector y s . The module consists of L s The layer fully connected network is composed of

[0020]

[0021] where Θ s = {W s , b s} represents the weight and bias of the fully connected layer, The nonlinear function representing the model can be written as

[0022]

[0023] in Represents the Relu activation function.

[0024] (2.2) Antenna selection module: The antenna selection module is used to select and activate a small number of millimeter wave antennas to assist sub6 channel information in beam selection. The module inputs the sub6 feature vector y s , output antenna selection vector v. Vector n and millimeter wave channel matrix X mm Multiply them to get the partial millimeter wave channel information X′ corresponding to the selected antenna mm .

[0025] The antenna selection module consists of L v Layer fully connected network, antenna selection vector Where 0 represents an unselected antenna and 1 represents a selected antenna. v can be expressed as

[0026]

[0027] Where C(·) represents a coding function that can convert the previous The largest element is encoded as 1, and the rest are encoded as 0. Indicates the number of selected antennas, Since C(·) is not differentiable, it hinders the back propagation of the neural network, so a differentiable vector is set Approximate expression v, where

[0028]

[0029]

[0030] When the model performs forward reasoning, v is obtained through the encoder C(·), and when the model reverses the gradient, the encoder is skipped and used The gradient of replaces v. For We expect it to have indivual 0. In order to The shape is the same as we expected, so we need to Imposing certain punishment

[0031]

[0032] The millimeter wave channel information X′ corresponding to the selected antenna mm It can be expressed as follows

[0033] X′ mm =v T ·X mm

[0034] (2.3) Millimeter wave module: The millimeter wave channel matrix X′ of the activated antenna mm Convert to a one-dimensional vector, input to the millimeter wave module, and output the feature vector y m . The module consists of L m Layer fully connected network.

[0035] (2.4) Classification module: Sub6 feature vector y s and the millimeter wave feature vector y m The module is composed of L p The optimal beam probability vector is output Expressed as

[0036]

[0037] Preferably, the specific steps of step (3) include:

[0038] (3.1) Calculate the model output probability vector With label y spt The loss function between

[0039]

[0040] Add the antenna selection module loss function calculated in step (2.2) As support set loss function

[0041]

[0042] (3.2) Update model parameters using reverse gradient descent

[0043]

[0044] The training process is repeated L times to obtain the model parameters based on the support set training. Where T i represents the i-th subtask, Represents the support set data of the i-th subtask.

[0045] Preferably, the specific steps of step (4) include:

[0046] (4.1) Based on model parameters Verify the model's performance on the query set. Input into the model to get the predicted probability

[0047]

[0048] Calculate predicted probabilities with query set label y qry The loss function between

[0049]

[0050]

[0051]

[0052] (4.2) Traverse all subtasks in the Batch and update the model parameters Θ with reverse gradient

[0053]

[0054] Will After substituting, we can derive a Batch parameter update formula as follows:

[0055]

[0056] Preferably, the specific steps of step (5) include:

[0057] (5.1) Top-1 optimal beam: select the probability vector output by the model The index of the maximum value is the index of the optimal beam in the codebook.

[0058] (5.2) Top-3 optimal beam: select the probability vector output by the model The top three values ​​in the index are the optimal beam candidates. When the communication environment SNR is poor, the prediction accuracy of the Top-1 optimal beam will decrease to a certain extent. Beam scanning can be performed from the Top-3 optimal beam candidate pool to determine the optimal beam.

[0059] Beneficial effects of the present invention:

[0060] 1) Using machine learning algorithms, the optimal beam selection can be directly predicted by the algorithm, avoiding the time overhead caused by searching for beams. Using out-of-band information for beam training avoids the training overhead caused by obtaining millimeter wave channel information;

[0061] 2) The model is based on sub6 channel information combined with some millimeter wave pilot information. Compared with the case based only on sub6 channel, the performance of the algorithm of the present invention is better;

[0062] 3) The algorithm can adaptively select activated millimeter-wave antennas and selectively compensate for different communication scenarios, further improving model performance;

[0063] 4) The model parameters are trained based on the meta-learning MAML algorithm, so that the model has good generalization ability and can converge quickly with only a small sample training set, which improves the deployability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The flowchart of the beam selection algorithm based on sub-6Ghz channels and some millimeter wave channels;

[0065] Figure 2 This is a schematic diagram of the algorithm's online reasoning process;

[0066] Figure 3 Schematic diagram of algorithm prediction performance under different SNRs;

[0067] Figure 4 Schematic diagram of the generalization ability of the model based on MAML optimization. DETAILED DESCRIPTION

[0068] The present technical solution is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0069] like Figure 1-4 As shown, the present invention is based on the communication beam selection method of sub6GHz channels and partial millimeter-wave pilots, and models the beam training problem as a multi-classification problem. The model consists of a sub6 module, an antenna selection module, a millimeter-wave module and a classification module.

[0070] 1. Simulation environment construction

[0071] The simulation communication environment is built based on the open source scene - DeepMIMO dataset "01" scene. The "01" ray tracing scene is an outdoor scene with two streets and an intersection. There are 18 base stations in this scene, all of which are 6 meters high and evenly distributed on both sides of the street. There are 1,184,923 communication users in the scene, and each user moves randomly in a random direction in the scene at a speed of 0 to 20 km / h. There are buildings on both sides of the street, and the height of the buildings is 60 meters.

[0072] All base stations can work in the 3.5 GHz band and the 28 GHz band at the same time, and the number of antennas in the 3.5 GHz band is N. s , the number of antennas in the 28GHz band is N m , adopting a planar antenna array distribution.

[0073] The communication channel between the base station and the user can be expressed as

[0074]

[0075] where α l is the attenuation coefficient of the lth path, is the departure azimuth of the lth path, The departure elevation angle of the lth path. l is the phase of path l, τ l is the propagation delay of path l, B is the signal bandwidth, K is the number of carriers, and a(·) is the channel response vector.

[0076] set up represents the downlink channel vector of the kth subcarrier in the sub-6GHz band, s s [k] represents the pilot signal sent by the kth subcarrier. Then the received signal y s [k] can be expressed as

[0077]

[0078] Where n s [k] is Gaussian distributed The least square method is used to estimate the channel for each subcarrier, which is expressed as follows:

[0079]

[0080] Similarly, for the millimeter wave frequency band, represents the downlink channel vector of the kth subcarrier in the sub-6GHz band, s m [k] represents the pilot signal sent by the kth subcarrier. Then the received signal y m [k] can be expressed as

[0081]

[0082] Where f is the codeword in the codebook F. Similarly, the least square method is used to estimate the millimeter wave channel matrix.

[0083] The AR for downlink transmission can be defined as

[0084]

[0085] Where P s represents the transmit power, represents the noise power, and k represents the kth carrier. Let R(h m , f) The f with the largest value is called the optimal beam.

[0086] Construct sub6 channel information X sub6 , millimeter wave channel information X mmAnd the optimal beam y is used as the training set.

[0087] The simulation environment setting parameters are as follows:

[0088] 1) The base station antennas are evenly distributed in a planar array, with N antennas in the sub6 frequency band. s =2×2, the number of millimeter wave band antennas N m =16×16; the user side uses a single full-phase antenna

[0089] 2) Base station operating frequencies are 3.5GHz and 28GHz

[0090] 3) Millimeter wave uses DFT codebook

[0091] 4) Number of carriers K = 32

[0092] 2. Training process of millimeter wave communication beam selection algorithm based on sub6Ghz channel and partial pilot

[0093] The basic training unit of meta-learning is Task, or subtask. Each subtask corresponds to a communication scenario in which a base station is located. Each subtask is divided into a support set and a query set.

[0094] Support set by Composition, of which represents the sub6GHz channel matrix between the user and the base station, N s Indicates the number of base station antennas in the sub6 frequency band, N c Indicates the number of carriers. represents the millimeter wave channel information between the user and the base station, N m Indicates the number of base station antennas in the millimeter wave frequency band. spt represents the codeword index corresponding to the best beam between the user and the base station. The query set is composed of The meaning of each part is the same as that of the support set.

[0095] The support set is used for innerloop training of the model, and the query set is used to evaluate the reasoning ability of the model after the innerloop and to perform outerloop training on the model.

[0096] X sub6 Converted into a 1-dimensional vector, input into the sub6 module, and output the feature vector y s . The module consists of L s The layer fully connected network is composed of

[0097]

[0098] where Θ s = {W s, b s} represents the weight and bias of the fully connected layer, The nonlinear function representing the model can be written as

[0099]

[0100] in Represents the Relu activation function.

[0101] The antenna selection module is used to select and activate a small number of millimeter wave antennas to assist sub6 channel information in beam selection. The module inputs the sub6 feature vector y s , output antenna selection vector v. Vector n and millimeter wave channel matrix X mmm Multiply them to get the partial millimeter wave channel information X′ corresponding to the selected antenna mm .

[0102] The antenna selection module consists of L v Layer fully connected network, antenna selection vector Where 0 represents an unselected antenna and 1 represents a selected antenna. v can be expressed as

[0103]

[0104] Where C(·) represents a coding function that can convert the previous The largest element is encoded as 1, and the rest are encoded as 0. Indicates the number of selected antennas, Since C(·) is not differentiable, it hinders the back propagation of the neural network, so a differentiable vector is set Approximate expression v, where

[0105]

[0106]

[0107] When the model performs forward reasoning, v is obtained through the encoder C(·), and when the model reverses the gradient, the encoder is skipped and used The gradient of replaces v. For We expect it to have indivual 0. In order to The shape is the same as we expected, so we need to Imposing certain punishment

[0108]

[0109] The millimeter wave channel information X′ corresponding to the selected antenna mmIt can be expressed as follows

[0110] X′ mm =v T .X mm

[0111] The mmWave channel matrix X′ of the activated antenna mm Convert to a one-dimensional vector, input to the millimeter wave module, and output the feature vector y m . The module consists of L m Layer fully connected network.

[0112] Sub6 feature vector y s and the millimeter wave feature vector y m The module is composed of Lp layers of fully connected networks. The optimal beam probability vector is output. Expressed as

[0113]

[0114] Calculate the model output probability vector With label y spt The loss function between

[0115]

[0116] Add the antenna selection module loss function calculated in step (2.2) As support set loss function

[0117]

[0118] Update model parameters using reverse gradient descent

[0119]

[0120] The training process is repeated L times to obtain the model parameters based on the support set training. Where T i represents the i-th subtask, Represents the support set data of the i-th subtask.

[0121] Based on model parameters Verify the model's performance on the query set. Input into the model to get the predicted probability

[0122]

[0123] Calculate predicted probabilities with query set label y qryThe loss function between

[0124]

[0125]

[0126]

[0127] Traverse all subtasks in the Batch and update the model parameters Θ with reverse gradient

[0128]

[0129] Will After substituting, we can derive a Batch parameter update formula as follows:

[0130]

[0131] Simulation environment parameters

[0132]

[0133]

[0134] Simulation results analysis:

[0135] Figure 3 Shows when to take The accuracy of the algorithm in predicting the optimal beam when SNR = {0, 5, 10, 15, 20, 25, 30}. The triangle symbol represents the probability that the top-3 optimal beams contain the optimal beam, the square symbol represents the probability that the top-1 optimal beam is equal to the optimal beam, and the asterisk represents the accuracy of prediction based only on sub-6GHz channel information.

[0136] From the prediction results, we can see that the performance of beam prediction based solely on sub-6GHz channel information is poor. When the communication environment quality is good (SNR ≥ 15dB), the prediction accuracy of the Top-1 optimal beam can reach more than 90%. When the environmental interference is severe, the prediction accuracy of the Top-1 optimal beam will decrease to a certain extent.

[0137] The probability that the algorithm predicts that the optimal beam is included in the Top-3 optimal beams can always maintain a very good accuracy when the signal-to-noise ratio drops from 30 to 0dB. This shows that although the noise has caused some interference to the prediction of this algorithm, the algorithm can still accurately determine the rough direction of the optimal beam. If the communication environment where the base station is located has long-term large interference and the base station is sensitive to the achievable rate, you can choose to scan the beam from the Top-3 optimal beam candidate pool and select the beam with the best reception strength as the downlink communication beam.

[0138] Figure 4 The comparison of the learning ability of the model trained based on MAML, the traditional pre-trained model and the randomly initialized model in the new communication environment is shown. It can be seen that although the accuracy of the model trained based on MAML is lower than that of the pre-trained model before training, after 10 epochs, the model prediction performance quickly increased to more than 80%, and in subsequent training, it is always better than the traditional pre-trained model, proving that the model trained based on MAML has stronger learning ability and generalization performance.

[0139] It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. All components not specified in this embodiment can be implemented using existing technologies.

Claims

1. A communication beam selection method based on sub6GHz channels and partial millimeter wave pilots, characterized in that: The steps include: Step (1): Construct n task subtasks, each of which corresponds to the communication environment of a base station. Each Task includes a support set and a query set; build a beam selection prediction model, including sub6 module, antenna selection module, millimeter wave module, classification module, and initialize model parameters); Step (2): For subtask T i , input the sub6 channel CSI in the support set into the sub6 module to obtain the sub6 feature vector; input the sub6 feature vector into the antenna selection module to obtain the activated millimeter wave antenna index; Obtain the pilot signal that supports the centralized activation of the antenna, input it into the millimeter wave module, and obtain the millimeter wave feature vector; fuse the sub6 feature vector with the millimeter wave feature vector, input it into the classification module, and obtain the probability that each beam is the optimal beam; Step (3): Calculate the loss function of the support set and update the model parameters in reverse order; after L times of gradient update, the updated model parameters are obtained. Step (4): Based on model parameters Make predictions on the query set and calculate the loss function; perform reverse gradient descent based on the loss function of all tasks in the Batch and update the model parameters Θ (t+1) ; Read the subtask of the next batch and repeat steps (2) to (4) until the model converges; Step (5): According to the probability vector output by the model, obtain the Top-1 optimal beam and the Top-3 optimal beam; The specific steps of step (1) include: (1.1) Subtask division: The basic training unit of meta-learning is Task, i.e., subtask. Each subtask corresponds to a communication scenario in which a base station is located. Each subtask is divided into a support set and a query set. (1.2) The support set is Composition, of which represents the sub6GHz channel matrix between the user and the base station, N s Indicates the number of base station antennas in the sub6 frequency band, N c Indicates the number of carriers; represents the millimeter wave channel information between the user and the base station, N m Indicates the number of base station antennas in the millimeter wave frequency band; y spt Represents the codeword index corresponding to the best beam between the user and the base station; the query set is composed of Composition, the meaning of each part is the same as the support set; The support set is used for innerloop training of the model, and the query set is used to evaluate the reasoning ability of the model after the innerloop and to perform outerloop training on the model; The specific steps of step (2) include: (2.1) sub6 module: X sub6 Converted into a 1-dimensional vector, input into the sub6 module, and output the feature vector y s ; The module consists of L s The layer fully connected network is composed of where Θ s = {W s , b s } represents the weight and bias of the fully connected layer, Represents the nonlinear function of the model, written as in Represents the Relu activation function; (2.2) Antenna selection module: The antenna selection module is used to select and activate a small number of millimeter wave antennas to assist sub6 channel information in beam selection; the module inputs the sub6 feature vector y s , output antenna selection vector v; vector n and millimeter wave channel matrix X mm Multiply them to get the partial millimeter wave channel information X′ corresponding to the selected antenna mm ; The antenna selection module consists of L v Layer fully connected network, antenna selection vector Where 0 represents an unselected antenna and 1 represents a selected antenna; v is represented by Where C(·) represents a coding function. The largest element is encoded as 1, and the rest are encoded as 0; Indicates the number of selected antennas, Since C(·) is not differentiable, it hinders the back propagation of the neural network, so a differentiable vector is set Approximate expression v, where When the model performs forward inference, v is obtained through the encoder C(·), and when the model reverses the gradient, the encoder is skipped and used The gradient of replaces v; for Expect it to have 1, 0; in order to The shape is the same as expected, and needs to be Imposing certain punishment The millimeter wave channel information X′ corresponding to the selected antenna mm It is expressed as follows X′ mm =v⊙X mm (2.3) Millimeter wave module: The millimeter wave channel matrix X′ of the activated antenna mm Convert to a one-dimensional vector, input to the millimeter wave module, and output the feature vector y m ; The module consists of L m Layer fully connected network; (2.4) Classification module: Sub6 feature vector y s and the millimeter wave feature vector y m splicing and inputting into the classification module; the module consists of L p Layer fully connected network; output optimal beam probability vector Expressed as The specific steps of step (3) include: (3.1) Calculate the model output probability vector With label y spt The loss function between Add the antenna selection module loss function calculated in step (2.2) As support set loss function (3.2) Update model parameters using reverse gradient descent The training process is repeated L times to obtain the model parameters based on the support set training. Where T i represents the i-th subtask, Represents the support set data of the i-th subtask; The specific steps of step (4) include: (4.1) Based on model parameters Verify the performance of the model on the query set; Input into the model to get the predicted probability Calculate predicted probabilities with query set label y qry The loss function between (4.2) Traverse all subtasks in the Batch and update the model parameters Θ with reverse gradient Will After substituting, we can derive a Batch parameter update formula as follows: The specific steps of step (5) include: (5.1) Top-1 optimal beam: select the probability vector output by the model The index of the maximum value is the index of the optimal beam in the codebook; (5.2) Top-3 optimal beam: select the probability vector output by the model The top three values ​​in the index are the optimal beam candidates. When the communication environment SNR is poor, the prediction accuracy of the Top-1 optimal beam will decrease to a certain extent. Beam scanning is performed from the Top-3 optimal beam candidate pool to determine the optimal beam.

Images