Method and system for training a beamforming deep neural network based on continuous learning

Abstract

The application discloses a beamforming deep neural network training method and system based on continuous learning, and the system comprises a memory set module, a mixing module, a beamforming matrix recovery module and a neural network module. In a new time slot, the memory set is mixed with a current time slot sample set for model training in the training process of each time slot, and the memory set and the current time slot sample set are sent into the trained model to generate a predicted output after the training period of the current time slot ends. The model calculates a performance function value for the predicted values of all samples, and according to the performance of each sample, samples with poor performance are selected from the memory set and the current time slot sample set in different proportions to update the memory set. When entering a new time slot again, the above steps are iteratively executed to obtain optimal model weights. The beamforming deep neural network training method based on continuous learning is suitable for a dynamically changing actual communication environment and can maintain good communication performance in the case of channel distribution change.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of machine learning and mobile communication technology, in particular to a beamforming deep neural network training method and system based on continuous learning. BACKGROUND

[0002] Multiple-input multiple-output (MIMO) systems can effectively improve the capacity of communication networks, and beamforming technology, as a key technology of MIMO systems, has received extensive attention. With the development of artificial intelligence technology, data-driven deep learning and model-driven deep learning are used to solve the contradiction between the performance and the calculation delay of beamforming technology. Purely data-driven deep learning methods will cause high training complexity and low learning accuracy. Beamforming deep neural network is a model-driven deep learning framework that balances performance and algorithm complexity, reduces latency while ensuring performance, and provides real-time beamforming services for MIMO systems. However, existing work is based on the assumption that test samples and training samples follow the same distribution, and if the environment distribution mutates, it will suffer great performance loss. SUMMARY

[0003] The purpose of the present application is to provide a beamforming deep neural network training method and system based on continuous learning.

[0004] To achieve the above purpose, the present application provides the following technical scheme: a beamforming deep neural network training method based on continuous learning, comprising the following steps:

[0005] S1, obtaining channel data of a current time slot t, and constructing a training mapping pair D of the current time slot t based on the channel data corresponding to the uplink power vector label value t ;

[0006] S2, based on known channel sample data M t , stacking the training mapping pair D t of the time slot t and the channel sample data M t in front and back and randomly shuffling to form a training mapping pair set M t ∪D t , using the channel data of each sample in the training mapping pair set M t ∪D t as input and the uplink power vector label value corresponding to each channel data in the training mapping pair set as output to construct and iteratively train a neural network model; at the same time, a least mean square error function is used as a loss function for supervised learning during the iterative training process until all training mapping pair sets M t ∪D t are traversed, obtaining a neural network model and the optimal weight Θ of the neural network model;

[0007] S3. Using the trained neural network model, map the training mappings to the set M. t ∪D t The channel information of each sample is input into the neural network model to obtain the predicted uplink power allocation vector;

[0008] S4. Input the predicted uplink power allocation vector into the beamforming moment recovery module and use the closed-form solution to obtain the predicted beamforming matrix.

[0009] S5, based on the predicted beamforming matrix, and the training map pair set M t ∪D t The predicted beamforming matrix and rate are calculated using a performance function and compared with the training mapping pair set M. t ∪D t The optimal ratio of the sum and the negative ratio, based on this negative ratio, is used to target M according to a preset method A. t D t Select mapping pairs that meet the preset conditions, stack and shuffle these mapping pairs to obtain channel sample data M. t+1 and channel sample data M t+1 Updated to channel sample data M t Then return to step S1.

[0010] Furthermore, in step S1 described above, the training mapping pair D t The channel matrix H of the time slot t data is obtained as follows: k (k = 1, 2, ..., K) Take the modulo value (k = 1, 2, ..., K) and the uplink power vector label value P generated by the WMMSE weighted least mean square error algorithm. k The training mapping pairs that make up this time slot Where k is the user index, P k These are the sample label values, used by the model to calculate the difference between the loss function and the predicted values.

[0011] Furthermore, in step S2 above, when training the neural network model, the maximum number of iterations is set to M, the optimizer is RMSprop, the learning rate is l, and the training mapping pair set M is... t ∪D t The model is divided into several batches, and each batch is iterated. During the iteration process, gradient descent is used to update the model weights and optimize the global loss function. This process is applied to the training mapping pair set M. t ∪D t After the traversal is complete, the training process of the neural network model ends, the model converges, the global loss function reaches its minimum value, and the model weights Θ reach their optimal values.

[0012] Furthermore, obtaining the predicted uplink power vector in step S3 specifically involves: setting the training mapping pair set M... t ∪D t Channel information for each sample The input is fed into the trained neural network model to obtain the predicted uplink power allocation vector π(Θ,h). (i) ), where i is the sample index.

[0013] Furthermore, in step S4 above, obtaining the predicted beamforming matrix using the closed-form solution includes the following sub-steps:

[0014] S4.1 Calculate the beamforming vector of the k-th user constructed from the neural network model's predicted output using the following formula:

[0015]

[0016] Where k is the user index, i is the sample index, and I N It is an identity matrix, σ 2 It is noise power, h k It is the channel vector from the base station to the k-th user, π k (Θ,h (i) ) is the predicted uplink power vector for the k-th user. The k-th element of the eigenvector corresponding to the largest eigenvalue;

[0017] S4.2 Combine the beamforming vectors of all users in user order to obtain the predicted beamforming matrix. Furthermore, in step S5 above, the predicted beamforming matrix and rate versus training mapping pair set M are calculated using a performance function. t ∪D t The optimal ratio of the negative rate is as follows:

[0018]

[0019] Where Θ is the system weight, h (i) It is channel sample data, p (i) It is the uplink power vector that needs to be learned by the neural network model, π(Θ,h) (i) W(π(Θ,h)) is the uplink power allocation vector predicted by the neural network model. (i) R(W(π(Θ,h)) is the predicted beamforming matrix. (i) ));h (i) ) is the sum rate of the predicted beamforming matrix. The training mapping pair set M is obtained through the weighted least mean square error algorithm. t ∪D toptimum and rate, where i is the sample index.

[0020] Further, in the aforementioned step S5, according to the negative ratio, the mapping pairs in M t and D t that meet the preset condition are selected, the mapping pairs are stacked and shuffled to obtain channel sample data M t+1 , and the channel sample data M t+1 is updated to channel sample data M t .

[0021] S5.1, the mapping pairs in M t and D t that are calculated by the performance function are sorted in descending order, the first N-f(t) sample mapping pairs in D t are selected, and the first f(t) sample mapping pairs in M t are selected, where f(t) is a function related to time slot t.

[0022] S5.2, the first N-f(t) sample mapping pairs in D t and the first f(t) sample mapping pairs in M t are stacked and shuffled to obtain channel sample data M t+1 .

[0023] Another aspect of the present application provides a beamforming deep neural network training system based on continuous learning, comprising:

[0024] a memory set module for storing channel sample data;

[0025] a mixing module for constructing training mapping pairs D t of the current time slot t by using channel data of the current time slot t and uplink power vector label values corresponding to the channel data, and receiving channel sample data M t output from the memory set module, and then stacking and shuffling the training mapping pairs D t of the time slot t and the channel sample data M t to form a training mapping pair set M t ∪D t .

[0026] a neural network module comprising an input layer and sequentially connected multiple convolution layers, multiple normalization layers, multiple activation function layers, a flattening layer, and a full connection layer, the neural network module being used for processing the training mapping pair set M t ∪D tThe neural network model is trained by taking each channel data as input, and taking the uplink power vector label value corresponding to each channel data in the training mapping pair set as output, to obtain the uplink power vector corresponding to each channel data.

[0027] The beamforming matrix recovery module comprises a scaling layer and a construction layer connected thereto, the scaling layer serving as the input end of the beamforming matrix recovery module and receiving the uplink power vector corresponding to each channel data output by the neural network module to perform scaling to meet the total power constraint; the construction layer serving as the output end of the beamforming matrix recovery module and processing the scaled uplink power vector by using a closed-form solution to obtain a predicted beamforming matrix and output the same.

[0028] Compared with the prior art, the above technical scheme has the following beneficial effects:

[0029] (1) In the training process of the beamforming deep neural network, the memory set is constructed, and the mixed memory set and the current sample set are trained, so that the model has memory for the trained environment and can cope with the environment distribution mutation, thereby maintaining good communication performance.

[0030] (2) The performance loss function is defined by using the sum rate and the rate-to-power ratio, and the performance loss function is used as the standard for selecting the memory set sample, so that the problem structure is effectively embedded into the model training process. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a kind of beamforming deep neural network structure diagram.

[0032] Figure 2 It is a flow chart of the beamforming deep neural network training method based on continuous learning. DETAILED DESCRIPTION

[0033] In order to better understand the technical content of the present application, specific embodiments are described below with reference to the accompanying drawings.

[0034] Aspects of the present application are described in this disclosure with reference to the accompanying drawings, which illustrate a number of illustrative embodiments. The embodiments of the present application are not limited to the drawings described. It should be understood that the present application is implemented by any one of the above-mentioned concepts and embodiments, and the concepts and embodiments described in detail below, since the concepts and embodiments disclosed in the present application are not limited to any embodiment. In addition, some aspects disclosed in the present application can be used alone, or in any suitable combination with other aspects disclosed in the present application.

[0035] As Figure 1As shown, the beamforming deep neural network system based on continuous learning provided by the present invention includes a memory set module, a hybrid module, a beamforming matrix recovery module, and a neural network module; the memory set module is used to store channel sample data.

[0036] The hybrid module is used to construct a training mapping pair D for the current time slot t from the channel data of the current time slot t and the corresponding uplink power vector label value of the channel data. t And receive channel sample data M from the memory set module. t Then the training mapping of time slot t is applied to D. t and channel sample data M t The stacked pairs are then randomly shuffled to form a training mapping pair set M. t ∪D t .

[0037] The neural network module includes an input layer and multiple convolutional layers, multiple normalization layers, multiple activation function layers, a flattening layer, and a fully connected layer sequentially connected thereto. This neural network module is used to train a set M of mapping pairs. t ∪D t The neural network model, with each channel data as input and the uplink power vector label value corresponding to each channel data in the set as output, is trained to obtain the uplink power vector corresponding to each channel data.

[0038] The beamforming matrix recovery module includes a scaling layer and a construction layer connected to it. The scaling layer serves as the input to the beamforming matrix recovery module, receiving uplink power vectors corresponding to each channel data output from the neural network module and scaling them to meet the total power constraint. The construction layer serves as the output of the beamforming matrix recovery module, processing the scaled uplink power vectors using closed-form processing to obtain the predicted beamforming matrix, and then outputting it.

[0039] like Figure 2 As shown, the training method for beamforming deep neural networks based on continuous learning includes the following steps:

[0040] S1. Obtain the channel data for the current time slot t, and the corresponding uplink power vector label value for that channel data to construct the training mapping pair D for the current time slot t. t Training mapping pairs D t The channel matrix H of the time slot t data is obtained as follows: k (k = 1, 2, ..., K) Take the modulo value (k = 1, 2, ..., K) and the uplink power vector label value P generated by the WMMSE weighted least mean square error algorithm. k The training mapping pairs that make up this time slot Where k is the user index, P k These are the sample label values, used by the model to calculate the difference between the loss function and the predicted values.

[0041] S2, based on known channel sample data M t The training mapping of time slot t to D t and channel sample data M t The stacked pairs are then randomly shuffled to form a training mapping pair set M. t ∪D t To train the set of mapping pairs M t ∪D t The channel data of each sample in the training mapping pair set is used as input, and the uplink power vector label value corresponding to each channel data in the training mapping pair set is used as output to construct and iteratively train the neural network model. When training the neural network model, the maximum number of iterations is set to M, the optimizer is RMSprop, and the learning rate is l. The training mapping pair set M is used as input. t ∪D t The model is divided into several batches, and each batch is iterated. During the iteration process, gradient descent is used to update the model weights and optimize the global loss function. This process is applied to the training mapping pair set M. t ∪D t After the traversal is complete, the training process of the neural network model ends, the model converges, the global loss function reaches its minimum value, and the model weights Θ reach their optimal values.

[0042] S3, Set the training mapping pairs M t ∪D t Channel information for each sample The input is fed into the trained neural network model to obtain the predicted uplink power allocation vector π(Θ,h). (i) ).

[0043] S4. Input the predicted uplink power allocation vector into the beamforming moment recovery module and use the closed-form solution to obtain the predicted beamforming matrix.

[0044] The first step is to calculate the beamforming vector for the k-th user constructed from the neural network model's predicted output using the following formula:

[0045]

[0046] Where k is the user index, i is the sample index, and I N It is an identity matrix, σ 2 It is noise power, h k It is the channel vector from the base station to the k-th user, π k (Θ,h (i) ) is the predicted uplink power vector for the k-th user. yes the kth element of the eigenvector corresponding to the largest eigenvalue of

[0047] Second, combine the beamforming vectors of all users in user order to obtain a predicted beamforming matrix

[0048] S5, first, calculate the predicted beamforming matrix and the rate and training mapping pair set M using the performance function t ∪D t The optimal sum rate and the negative ratio of the rate are as follows:

[0049]

[0050] Wherein, Θ is the system weight, h (i) is the channel sample data, p (i) is the uplink power vector that needs to be learned by the neural network model, π(Θ, h (i) ) is the uplink power allocation vector predicted by the neural network model, W(π(Θ, h (i) )) is the predicted beamforming matrix, R(W(π(Θ, h (i) )) is the sum rate of the predicted beamforming matrix, (i) is the training mapping pair set M obtained by the weighted least mean square error algorithm t ∪D t The optimal sum rate, wherein i is the sample index.

[0051] Second, sort the mapping pairs in M t and D t in descending order, select the first N-f(t) sample mapping pairs from D t , and select the first f(t) sample mapping pairs from M t , wherein f(t) is a function related to time slot t;

[0052] Third, stack and shuffle the first N-f(t) sample mapping pairs selected from D t and the first f(t) sample mapping pairs selected from M t to obtain channel sample data M t+1 , and update the channel sample data M t+1 to channel sample data M t , and then return to step S1.

[0053] Although the present application has been described as above with reference to the preferred embodiments, it is not intended to limit the present application. Those skilled in the art can make various modifications and improvements without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be subject to the claims.​

Claims

1. A beamforming deep neural network training method based on continuous learning, characterized in that, Includes the following steps: S1. Obtain the channel data for the current time slot t, and the corresponding uplink power vector label value for that channel data to construct the training mapping pair D for the current time slot t. t ; S2, based on known channel sample data M t The training mapping of time slot t to D t and channel sample data M t The stacked pairs are then randomly shuffled to form a training mapping pair set M. t ∪D t To train the set of mapping pairs M t ∪D t The neural network model is constructed and iteratively trained using the channel data of each sample as input and the uplink power vector label value corresponding to each channel data in the training mapping pair set as output. Simultaneously, during the iterative training process, the minimum mean square error function is used as the loss function for supervised learning until all training mapping pairs in the set M are traversed. t ∪D t To obtain the neural network model and the optimal weights Θ of the neural network model; S3. Using the trained neural network model, map the training mappings to the set M. t ∪D t The channel information of each sample is input into the neural network model to obtain the predicted uplink power allocation vector; S4. Input the predicted uplink power allocation vector into the beamforming moment recovery module and use the closed-form solution to obtain the predicted beamforming matrix. S5, based on the predicted beamforming matrix, and the training map pair set M t ∪D t The predicted beamforming matrix and rate are calculated using a performance function and compared with the training mapping pair set M. t ∪D t The optimal ratio of the sum and the negative ratio, based on this negative ratio, is used to target M according to a preset method A. t D t Select mapping pairs that meet the preset conditions, stack and shuffle these mapping pairs to obtain channel sample data M. t+1 and channel sample data M t+1 Updated to channel sample data M t Then return to step S1.

2. The beamforming deep neural network training method based on continuous learning according to claim 1, characterized in that, In step S1, the training mapping pair D t Obtain it as follows: The channel matrix H of time slot t data k (k = 1, 2, ..., K) Take the modulo value (k = 1, 2, ..., K) and the uplink power vector label value P generated by the WMMSE weighted least mean square error algorithm. k The training mapping pairs that make up this time slot Where k is the user index, P k These are the sample label values, used by the model to calculate the difference between the loss function and the predicted values.

3. The beamforming deep neural network training method based on continuous learning according to claim 2, characterized in that, In step S2, when training the neural network model, the maximum number of iterations is set to M, the optimizer is RMSprop, and the learning rate is l. The training mapping pair set M is... t ∪D t The model is divided into several batches, and each batch is iterated. During the iteration process, gradient descent is used to update the model weights and optimize the global loss function. This process is applied to the training mapping pair set M. t ∪D t After the traversal is complete, the training process of the neural network model ends, the model converges, the global loss function reaches its minimum value, and the model weights Θ reach their optimal values.

4. The beamforming deep neural network training method based on continuous learning according to claim 3, characterized in that, The predicted uplink power vector obtained in step S3 is as follows: Set the training mapping pairs M t ∪D t Channel information for each sample The input is fed into the trained neural network model to obtain the predicted uplink power allocation vector π(Θ,h). (i) ), where i is the sample index.

5. The beamforming deep neural network training method based on continuous learning according to claim 4, characterized in that, In step S4, obtaining the predicted beamforming matrix using the closed-form solution includes the following sub-steps: S4.1 Calculate the beamforming vector of the k-th user constructed from the neural network model's predicted output using the following formula: Where k is the user index, i is the sample index, and I N It is an identity matrix, σ 2 It is noise power, h k It is the channel vector from the base station to the k-th user, π k (Θ,h (i) ) is the predicted uplink power vector for the k-th user. yes The k-th element of the eigenvector corresponding to the largest eigenvalue; S4.2 Combine the beamforming vectors of all users in user order to obtain the predicted beamforming matrix.

6. The beamforming deep neural network training method based on continuous learning according to claim 5, characterized in that, In step S5, the predicted beamforming matrix and the set M of rate and training mapping pairs are calculated using a performance function. t ∪D t The optimal ratio of the negative rate is as follows: Where Θ is the system weight, h (i) It is channel sample data, p (i) It is the uplink power vector that needs to be learned by the neural network model, π(Θ,h) (i) W(π(Θ,h)) is the uplink power allocation vector predicted by the neural network model. (i) R(W(π(Θ,h)) is the predicted beamforming matrix. (i) ));h (i) ) is the sum rate of the predicted beamforming matrix. The training mapping pair set M is obtained through the weighted least mean square error algorithm. t ∪D t The optimal sum rate, where i is the sample index.

7. The beamforming deep neural network training method based on continuous learning according to claim 6, characterized in that, In step S5, according to the negative ratio, M is targeted separately according to the preset method A. t D t Select mapping pairs that meet the preset conditions, stack and shuffle these mapping pairs to obtain channel sample data M. t+1 and channel sample data M t+1 Updated to channel sample data M t It includes the following steps: S5.1, For M in the sample data calculated by the performance function t and D t Sort the mapping pairs in descending order, starting from D. t Select the first Nf(t) sample mapping pairs from M. t Select the first f(t) sample mapping pairs, where f(t) is a function of time slot t; S5.2, D t Select the first Nf(t) sample mapping pairs from M and from M t The first f(t) samples are selected and their mapping pairs are stacked and shuffled.

8. A beamforming deep neural network training system based on continuous learning, characterized in that, include: The memory set module is used to store channel sample data; The hybrid module is used to construct a training mapping pair D for the current time slot t from the channel data of the current time slot t and the corresponding uplink power vector label value of the channel data. t And receive channel sample data M from the memory set module. t Then the training mapping of time slot t is applied to D. t and channel sample data M t The stacked pairs are then randomly shuffled to form a training mapping pair set M. t ∪D t ; The neural network module includes an input layer and multiple convolutional layers, multiple normalization layers, multiple activation function layers, a flattening layer, and a fully connected layer sequentially connected thereto. This neural network module is used to train a set M of mapping pairs. t ∪D t The neural network model, with each channel data as input and the uplink power vector label value corresponding to each channel data in the set as output, is trained to obtain the uplink power vector corresponding to each channel data. The beamforming matrix recovery module includes a scaling layer and a construction layer connected thereto. The scaling layer serves as the input of the beamforming matrix recovery module, receiving uplink power vectors corresponding to each channel data output from the neural network module and scaling them to meet the total power constraint. The construction layer serves as the output of the beamforming matrix recovery module, using closed-form processing to process the scaled uplink power vectors, obtaining the predicted beamforming matrix, and outputting it.

Images