A multi-user power allocation method based on one-dimensional convolutional neural network

Abstract

The present invention discloses a multi-user power allocation method based on a one-dimensional convolutional neural network, comprising: 1. Initializing a single-cell multi-user system model; 2. Using an improved multi-proportional fractional programming algorithm based on quadratic transformation to collect a data set for neural network training; 3. Designing a specific one-dimensional convolutional neural network structure; 4. Training the one-dimensional convolutional neural network using a mean square error loss function; 5. Saving the one-dimensional convolutional neural network structure. The present invention improves the multi-proportional fractional programming method based on quadratic transformation, and uses a one-dimensional convolutional neural network for learning. The present invention uses the channel state information of each user in the cell as input, and uses the power vector of each user under the multi-proportional fractional programming algorithm as output, to perform supervised learning training on the one-dimensional convolutional neural network. The present invention can reduce the computational complexity while achieving a near fractional programming algorithm.

Description

Technical Field

[0001] The invention belongs to the field of resource allocation in a mobile communication system, and in particular relates to a multi-user power allocation method based on a one-dimensional convolutional neural network. Background Art

[0002] With the commercial development of the fifth generation of mobile communication technology, the number of base stations and mobile users continues to increase, and the contradiction between limited communication resources and unlimited user demand gradually emerges. The unreasonable allocation of resources will cause losses to system performance. Therefore, resource allocation is one of the research focuses in the field of wireless communications.

[0003] Common optimization objectives of power allocation algorithms include maximizing system sum rate, maximizing spectrum efficiency, minimizing system power consumption, and ensuring fairness among users. These objectives and corresponding constraints together form a non-convex optimization model. The common solution method is to alternately fix certain parameters and continuously iterate until the algorithm converges. Existing traditional power allocation algorithms often involve more complex calculations such as matrix inversion or matrix singular value decomposition in the process of model solving, which will further increase the scale of non-convex optimization problems and the complexity of iterative calculations.

[0004] Deep learning technology, which has developed rapidly in recent years, has the advantages of low computational complexity and good adaptability. It has a huge advantage in low computational complexity in learning the nonlinear mapping relationship from input to strategy, which is of great significance for real-time decision-making in wireless communication systems. Summary of the invention

[0005] The present invention aims to provide a multi-user power allocation method based on a one-dimensional convolutional neural network to solve the technical problem of achieving performance close to that of a fractional programming algorithm while reducing computational complexity.

[0006] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0007] A multi-user power allocation method based on a one-dimensional convolutional neural network comprises the following steps:

[0008] S1: Initialize the downlink single-cell multi-user system model. The base station is equipped with multiple antennas, and the user side is a single antenna. The base station can obtain the channel state information of all users. The normalized achievable data rate of user i is:

[0009]

[0010] where h i is the channel vector of the ith user, is the transpose of the i-th user channel vector, H is the transpose operator, and w iis the precoding vector of the ith user, is the noise variance of the ith user side, p i is the transmission power allocated to the i-th user, satisfying 0≤p i ≤P max , p j is the transmission power allocated to the jth user, satisfying j≠i and 0≤p j ≤P max , P max is the maximum transmit power of the base station;

[0011] S2: Collect data sets, including the channel state information matrix H of each user and the power allocation result p* under the multi-proportional fractional programming FP algorithm based on quadratic transformation;

[0012] S3: Design the number of layers and convolution kernels of the one-dimensional convolutional neural network, and initialize the weights and biases of the one-dimensional convolutional neural network;

[0013] S4: training a one-dimensional convolutional neural network, taking the mean square error as the loss function form, to minimize the error between the predicted power vector p and the power allocation vector p* under the multi-proportional fractional programming algorithm based on quadratic transformation in step S2;

[0014] S5: When the number of iterative training is greater than or equal to 2000 or the loss function error is less than 10 -3 , save the weights and biases of the current one-dimensional convolutional neural network.

[0015] Furthermore, S2 specifically includes the following steps:

[0016] S2-1: Initialize auxiliary variables γ and y;

[0017] S2-2: Initialize iteration counter counter=0;

[0018] S2-3: Update the i-th parameter y in the auxiliary variable y i :

[0019]

[0020] where γ i is the i-th parameter in the auxiliary variable γ, K is the number of users in the base station;

[0021] S2-4: Update the i-th parameter γ in the auxiliary variable γ i :

[0022]

[0023] S2-5: Update the i-th power result p in the power vector p i :

[0024]

[0025] S2-6: counter counter=counter+1;

[0026] S2-7: Repeat S2-3 to S2-6 until counter ≥ 1000;

[0027] S2-8: Output the power allocation result p* under the multi-proportional fractional programming FP algorithm based on quadratic transformation.

[0028] Furthermore, in S3, since the one-dimensional convolutional neural network can only process real numbers, and the user channel state information is in complex form, the input of the one-dimensional convolutional neural network is defined as:

[0029] v=abs(vec(H H H)) 2

[0030] The vec function implements H H The vectorization of H, the abs function realizes the modulus of each complex number, H is the channel state information matrix of the user, H H The transpose of the user's channel state information matrix, the input size of the one-dimensional convolutional neural network is K 2 ×1, where K is the number of users in the base station; therefore, a one-dimensional convolutional neural network is used to learn the mapping relationship from the processed input v to the output power vector, which includes an input layer, three convolution modules, two fully connected layers, and an output layer.

[0031] Further, S4 specifically includes the following steps:

[0032] S4-1: Based on the characteristics of deep learning, the channel state information between the user and the base station is collected, and the corresponding power vector p* is obtained using the multi-proportional fractional programming FP algorithm based on quadratic transformation, and the training data obtained in S4-1 is repeated;

[0033] S4-2: The training data samples obtained in step S4-1 are randomly divided into a training set, a validation set, and a test set in a ratio of 4:1:1;

[0034] S4-3: Train the one-dimensional convolutional neural network with the initialization structure and construct the loss function L FP :

[0035]

[0036] Among them, P out is the output result of the one-dimensional convolutional neural network, P FP It is the optimal power allocation result obtained by FP algorithm.

[0037] Furthermore, S5 specifically means that the number of training iterations is greater than or equal to 2000 or the loss function is less than the threshold 10 -3 Stop training and save the weights and biases of the current model.

[0038] The multi-user power allocation method based on a one-dimensional convolutional neural network of the present invention has the following advantages:

[0039] 1. Compared with the multi-proportional fractional programming algorithm based on quadratic transformation in step S2, step S4 of the present invention can reduce the computational complexity while achieving similar performance, and shorten the computational time when there are a large number of users.

[0040] 2. The present invention designs a one-dimensional convolutional neural network for power allocation based on user status information, improves the characteristics of traditional convolutional neural networks for extracting two-dimensional information, and enhances the power allocation result prediction ability of the system model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the one-dimensional convolutional neural network structure of the present invention;

[0042] FIG2( a ) is a schematic diagram showing a comparison of the system and rate of a one-dimensional convolutional neural network algorithm, a fractional programming algorithm and an equal power allocation algorithm under minimum mean square error precoding according to the present invention;

[0043] FIG2( b ) is a schematic diagram showing a comparison of the system and rate of the one-dimensional convolutional neural network algorithm, the fractional programming algorithm and the equal power allocation algorithm under zero-forcing precoding according to the present invention.

[0044] Figure 3 Schematic diagram of the variation of computing time with the number of users based on one-dimensional convolutional neural network algorithm, fractional programming algorithm and equal power allocation algorithm. DETAILED DESCRIPTION

[0045] In order to better understand the purpose, structure and function of the present invention, a multi-user power allocation method based on a one-dimensional convolutional neural network of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0046] The present invention is further described below in conjunction with embodiments:

[0047] Taking the downlink single-cell multi-user scenario as an example, assuming that the cell base station is equipped with 8 antennas, the user side has a single antenna, and the maximum transmission power of the base station is 30dbm. The multi-user power allocation method based on a one-dimensional convolutional neural network described in the present invention includes the following steps:

[0048] S1: Initialize the downlink single-cell multi-user system model. The base station is equipped with multiple antennas, and the user side is a single antenna. The base station can obtain the channel state information of all users. The normalized achievable data rate of user i is:

[0049]

[0050] where h i is the channel vector of the ith user, is the transpose of the i-th user channel vector, H is the transpose operator, and w i is the precoding vector of the ith user, is the noise variance of the ith user side, p i is the transmission power allocated to the i-th user, satisfying 0≤p i ≤P max , p j is the transmission power allocated to the jth user, satisfying j≠i and 0≤p j ≤P max , P max is the maximum transmit power of the base station;

[0051] S2: Collect data sets, including the channel state information matrix H of each user and the power allocation result p* under the multi-proportional fractional programming algorithm based on quadratic transformation;

[0052] S2-1: Initialize auxiliary variables γ and y;

[0053] S2-2: Initialize iteration counter counter=0;

[0054] S2-3: Update the i-th parameter y in the auxiliary variable y i :

[0055]

[0056] where γ i is to initialize the i-th parameter in the auxiliary variable γ, and K is the number of users in the base station.

[0057] S2-4: Update auxiliary variables γ i The i-th parameter γ i :

[0058]

[0059] S2-5: Update the i-th power result p in the power vector p i :

[0060]

[0061] S2-6: counter counter=counter+1;

[0062] S2-7: Repeat S2-3 to S2-6 until the counter is greater than or equal to 1000;

[0063] S2-8: Output the power allocation result p* under the multi-proportional fractional programming algorithm based on quadratic transformation.

[0064] S3: Design the number of layers and convolution kernels of the one-dimensional convolutional neural network, and initialize the weights and biases of the network. Since the neural network can only process real numbers, and the user channel state information is in complex form, the input of the network is defined as:

[0065] v=abs(vec(H H H)) 2 (5)

[0066] The vec function implements H H The vectorization of H, the abs function realizes the modulus of each complex number. H is the channel state information matrix of the user, and the size of the neural network input v is K 2 ×1, where K is the number of users in the base station. Therefore, consider using a one-dimensional convolutional neural network to learn the mapping relationship from input v to output power vector, which includes an input layer, three convolution modules, two fully connected layers, and an output layer. In order to fit the nonlinear function, the relu function is used to act on the feature output of the convolution layer. The fully connected layer is designed as the last layer of the neural network, because the power output needs to satisfy 0≤p i ≤P max This constraint, therefore, requires adding a custom activation function after the fully connected layer:

[0067] y=min(max(x,0),P max ) (6)

[0068] Among them, x and y represent the input and output of the activation function respectively, P max is the maximum transmit power.

[0069] S4: training the neural network, taking the mean square error as the loss function to minimize the error between the predicted power vector p and the power allocation vector p* under the fractional programming algorithm;

[0070] S4-1: Based on the characteristics of deep learning, the channel state information between the user and the base station is collected, and the corresponding power vector p* is obtained using a multi-proportional fractional programming algorithm based on quadratic transformation. The steps in S4-1 are repeated to obtain training data.

[0071] S4-2: The data samples obtained in step S4-1 are randomly divided into a training set, a validation set, and a test set in a ratio of 4:1:1. S4-3: Train the one-dimensional convolutional neural network with the initialization structure and construct the loss function L FP :

[0072]

[0073] Among them, P out is the output result of the one-dimensional convolutional neural network, P FP It is the optimal power allocation result obtained by FP algorithm.

[0074] S5: When the number of iterative training meets the requirements or the loss function converges, save the current neural network parameters.

[0075] Specifically, the number of training iterations is greater than or equal to 2000 or the loss function is less than the threshold of 10 -3 Stop training and save the network parameters of the current model.

[0076] In order to verify the effectiveness of the proposed method, a simulation experiment was carried out in a downlink single-cell multi-user scenario. Figure 1 The structure of a one-dimensional convolutional neural network is shown in Figure 1, which includes an input layer, three convolution modules, two fully connected layers, and an output layer. The convolution module includes a one-dimensional convolution layer and a pooling function. The number of convolution kernels in each layer is 128, 64, and 32 respectively. The fully connected layer has 1024 and 32 neurons respectively. The number of neurons in the output layer is consistent with the number of users.

[0077] During the simulation, the cell base station is equipped with 8 antennas, the user side is a single antenna, and the maximum transmission power of the base station is 30dbm. Figure 2(a) and Figure 2(b) show the comparison of the system and rate based on the one-dimensional convolutional neural network algorithm, the fractional programming algorithm and the equal power allocation algorithm (EPA) under the minimum mean square error (MMSE) precoding and zero forcing (ZF) precoding. It can be seen from the figure that the system and rate obtained under MMSE precoding are generally better than ZF precoding, and the results obtained under the one-dimensional convolutional neural network and FP algorithms are very similar.

[0078] Figure 3The figure shows how the calculation time of the one-dimensional convolutional neural network algorithm, the fractional programming algorithm, and the equal power allocation algorithm changes with the number of users. Taking MMSE precoding as an example, the average of 100 implementation results is taken for calculation. Among them, the calculation method based on the neural network only considers the time of online calculation, but does not consider the time of offline training. It can be seen that as the number of users in the cell increases, the calculation time of the fractional programming algorithm increases rapidly, while the method based on the neural network only needs to perform vector addition and multiplication operations after obtaining the trained weights and biases, which greatly reduces the calculation complexity and achieves a shorter calculation time than FP and EPA.

[0079] From the above-mentioned dimensions, it can be seen that compared with the traditional multi-proportional fractional programming algorithm, the multi-user power allocation algorithm based on one-dimensional convolutional neural network proposed in the present invention can effectively learn the nonlinear mapping relationship from user channel state information to power allocation results, while achieving similar system and rate performance, effectively reducing the calculation time.

[0080] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A multi-user power allocation method based on one-dimensional convolutional neural network, characterized in that: The following steps are involved: S1: Initialize the downlink single-cell multi-user system model. The base station is equipped with multiple antennas, and the user side is a single antenna. The base station can obtain the channel state information of all users. The normalized achievable data rate of user i is: where h i is the channel vector of the ith user, is the transpose of the i-th user channel vector, H is the transpose operator, and w i is the precoding vector of the ith user, is the noise variance of the ith user side, p i is the transmission power allocated to the i-th user, satisfying 0≤p i ≤P max , p j is the transmission power allocated to the jth user, satisfying j≠i and 0≤p j ≤P max , P max is the maximum transmit power of the base station; S2: Collect data sets, including the channel state information matrix H of each user and the power allocation result p* under the multi-proportional fractional programming FP algorithm based on quadratic transformation; S3: Design the number of layers and convolution kernels of the one-dimensional convolutional neural network, and initialize the weights and biases of the one-dimensional convolutional neural network; S4: training a one-dimensional convolutional neural network, taking the mean square error as the loss function form, to minimize the error between the predicted power vector p and the power allocation vector p* under the multi-proportional fractional programming algorithm based on quadratic transformation in step S2; S5: When the number of iterative training is greater than or equal to 2000 or the loss function error is less than 10 -3 , save the weights and biases of the current one-dimensional convolutional neural network; The S2 specifically includes the following steps: S2-1: Initialize auxiliary variables γ and y; S2-2: Initialize iteration counter counter=0; S2-3: Update the i-th parameter y in the auxiliary variable y i : where γ i is the i-th parameter in the auxiliary variable γ, K is the number of users in the base station; S2-4: Update the i-th parameter γ in the auxiliary variable γ i : S2-5: Update the i-th power result p in the power vector p i : S2-6: counter counter=counter+1; S2-7: Repeat S2-3 to S2-6 until counter ≥ 1000; S2-8: Output the power allocation result p* under the multi-proportional fractional programming FP algorithm based on quadratic transformation; In S3, since the one-dimensional convolutional neural network can only process real numbers, and the user channel state information is in complex form, the input of the one-dimensional convolutional neural network is defined as: v=abs(vec(H H H)) 2 The vec function implements H H The vectorization of H, the abs function realizes the modulus of each complex number, H is the channel state information matrix of the user, H H The transpose of the user's channel state information matrix, the input size of the one-dimensional convolutional neural network is K 2 ×1, where K is the number of users in the base station; therefore, a one-dimensional convolutional neural network is used to learn the mapping relationship between the processed input v and the output power vector, which includes an input layer, three convolution modules, two fully connected layers, and an output layer; The S4 specifically comprises the following steps: S4-1: Based on the characteristics of deep learning, the channel state information between the user and the base station is collected, and the corresponding power vector p* is obtained using the multi-proportional fractional programming FP algorithm based on quadratic transformation, and the training data obtained in S4-1 is repeated; S4-2: The training data samples obtained in step S4-1 are randomly divided into a training set, a validation set, and a test set in a ratio of 4:1:1; S4-3: Train the one-dimensional convolutional neural network with the initialization structure and construct the loss function L FP : Among them, P out is the output result of the one-dimensional convolutional neural network, P FP It is the optimal power allocation result obtained by FP algorithm.

2. The multi-user power allocation method based on one-dimensional convolutional neural network according to claim 1 is characterized in that: S5 specifically means that the number of training iterations is greater than or equal to 2000 or the loss function is less than a threshold of 10 -3 Stop training and save the weights and biases of the current model.

Images