CN113518457B - Power distribution strategy based on one-dimensional deep convolutional neural network - Google Patents
Power distribution strategy based on one-dimensional deep convolutional neural network Download PDFInfo
- Publication number
- CN113518457B CN113518457B CN202110437441.8A CN202110437441A CN113518457B CN 113518457 B CN113518457 B CN 113518457B CN 202110437441 A CN202110437441 A CN 202110437441A CN 113518457 B CN113518457 B CN 113518457B
- Authority
- CN
- China
- Prior art keywords
- neural network
- user
- base station
- convolutional neural
- dimensional
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 32
- 238000013528 artificial neural network Methods 0.000 claims abstract description 44
- 238000004891 communication Methods 0.000 claims abstract description 20
- 238000012549 training Methods 0.000 claims description 26
- 230000006870 function Effects 0.000 claims description 23
- 238000000034 method Methods 0.000 claims description 11
- 238000005562 fading Methods 0.000 claims description 10
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 9
- 238000002347 injection Methods 0.000 claims description 8
- 239000007924 injection Substances 0.000 claims description 8
- 238000012360 testing method Methods 0.000 claims description 8
- 230000005540 biological transmission Effects 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 210000004027 cell Anatomy 0.000 claims description 5
- 230000011218 segmentation Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 230000001413 cellular effect Effects 0.000 claims description 4
- 210000002569 neuron Anatomy 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 abstract description 9
- 238000013468 resource allocation Methods 0.000 abstract description 7
- 238000011160 research Methods 0.000 abstract description 4
- 238000011217 control strategy Methods 0.000 abstract description 3
- 230000007547 defect Effects 0.000 abstract description 3
- 230000000694 effects Effects 0.000 abstract description 2
- 238000005516 engineering process Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000000280 densification Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
Abstract
The invention discloses a power distribution strategy based on a one-dimensional deep convolutional neural network, belongs to the field of communication systems, and provides a power control strategy based on the one-dimensional deep convolutional neural network, aiming at the defects of the existing power distribution algorithm based on the deep neural network, which not only can realize online decision, but also can achieve a good performance approximation of the traditional algorithm, and the network prediction capability of the power control strategy is superior to that of the existing deep neural network based on a full-connection structure. The invention researches a resource allocation strategy of replacing the traditional algorithm by the one-dimensional convolutional neural network, learns the power allocation effect obtained based on the traditional algorithm by a supervised learning mode, realizes quick and reliable online decision, overcomes the defect of limited learning capacity and has higher prediction capacity compared with the traditional power allocation algorithm based on deep learning.
Description
Technical Field
The invention belongs to the field of communication systems, and particularly relates to a power distribution strategy based on a one-dimensional deep convolutional neural network.
Background
With popularization and promotion of internet technology and base station densification concepts, the increasing service requirements and requirements on capacity and rate of cellular networks gradually increase, how resource allocation in a communication system becomes a problem to be considered, more specifically how power allocation becomes a problem to be considered, a traditional power allocation algorithm is an iterative water filling algorithm, the principle is that among associated users of a given base station, users with good channel quality allocate more power, users with poor channel quality allocate little power or do not allocate power, so as to realize utility function maximization of a single base station, however, convergence performance of the algorithm is poor, i.e. convergence is low, slow convergence can lead to higher computational complexity, which limits application range of the algorithm, and deep learning technology can help to solve.
At present, deep learning technology is applied to different fields including image classification, natural language processing and voice recognition, but in recent years, due to the convenience and timeliness of channel sample acquisition in a wireless communication network, the application of deep learning in wireless communication is also more and more advantageous, and a neural network in deep learning can reach a satisfactory nonlinear approximation to a traditional algorithm, so that the deep learning technology is widely studied, and the current neural network comprises a Deep Neural Network (DNN), a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN). Many researches for realizing maximum minimization of an objective function through parameter control by using a neural network in a communication system are focused, for example, a deep neural network is used for realizing maximization of a spectrum efficiency or energy efficiency target through an approximation secondary gradient algorithm, or a target for realizing maximization of a system and a rate through an approximation IPM algorithm; in the aspect of convolutional neural network, local features are extracted by using a convolutional filter, and the maximization of spectral efficiency, energy efficiency or sum rate is realized in a supervised learning mode.
In general, neural networks have the advantage of nonlinear approximation to the performance of traditional algorithms, and low complexity, which is helpful to real-time decision making in a communication system, however, the deep neural networks have limited feature learning capability and insufficient prediction capability based on resource allocation strategies of the deep neural networks, so that research on improved neural networks has great significance to deep learning-based communication resource allocation by improving the prediction capability.
Disclosure of Invention
The invention aims to: the invention aims to provide a power distribution strategy based on a one-dimensional deep convolutional neural network, and more real-time and reliable online power distribution is realized in a communication system.
The technical scheme is as follows: in order to achieve the above purpose, the invention adopts the following technical scheme: the power distribution strategy based on the one-dimensional deep convolutional neural network comprises the following steps:
Step 1, communication modeling;
step2, collecting data sets, wherein the data sets comprise channel state information h k between a base station and a user and an optimal power value p * under a water injection algorithm as a group of data sets;
Step 3, constructing a one-dimensional deep convolutional neural network and initializing the weight of the neural network;
Step 4, training a neural network, constructing a mean square error between a predicted power value p and an optimal power value p * of the neural network as a loss function, and determining an optimization algorithm during training;
And 5, finishing training and storing the neural network when the loss function is smaller than a preset value or the iteration number of training is met.
Further, in the step1, the communication modeling specifically includes: the method comprises the steps of establishing a single-cell multi-user communication environment, wherein the positions of a base station and users in the environment are uniformly distributed, and a downlink channel between the base station and the users is a Rayleigh fading channel.
Further, in the step 2, the collection of 10 ten thousand sets of the data sets is repeated.
Further, the step1 specifically includes the following steps: assuming that considering a single-cell cellular network environment, a base station is deployed in the center of a cell, k users are randomly distributed around the center, and the user sets: k= {1,2,3, K }, these subscribers are served by a centrally located base station.
Further, in the step 1, considering a rayleigh fading channel environment with a mean value of 1, wherein the channel state information from the user K to the base station is h k, K e K, and the channel state information is that the path loss from the user to the base station in the rayleigh fading environment is considered, and the transmission power between the user K and the base station is represented by p k; then in the downlink the SINR (signal to interference plus noise ratio) at user k is expressed asWherein/>Is background noise, h k is channel state information between user k and base station, p k is transmission power from user base station to user k, its achievable rate is denoted as R k＝log2(1+SINRk), system and rate performance is denoted as/>
Further, in the step 3, the one-dimensional convolutional neural network is responsible for learning a mapping relationship between the gain of the input signal channel and the optimal user association or power allocation of the output signal.
Further, the step3 specifically includes the following steps:
Step 1) collecting channel state information between a user and a base station in an environment, operating a water injection algorithm to obtain a corresponding optimal power value p *, and repeating 10 ten thousand times to obtain a data set;
step 2) determining the segmentation ratio of training data and test data;
and 3) designing and constructing a framework structure of the one-dimensional convolutional neural network, and initializing a weight parameter w and a bias parameter b of the neural network.
Further, in the step 4, a neural network is constructed, specifically, a mean square error between the predicted power value p and the optimal power value p * is used as a loss functionAn Adam optimizer is determined to optimize the training neural network.
Further, the step 5 specifically is to save the network when the iteration number epochs =300 or the loss function is smaller than the preset value loss less than or equal to 0.1×10 -3.
The invention aims at research emphasis, namely how to control the downlink transmission power of the base station and the user, so that the sum rate of the whole system can be maximized rapidly and reliably. The resource allocation strategy of replacing the traditional algorithm with the one-dimensional convolutional neural network is researched, the power allocation effect obtained based on the traditional algorithm is learned by a supervised learning mode, quick and reliable online decision is realized, and compared with the traditional power allocation algorithm based on deep learning, the method overcomes the defect of limited learning capacity and has higher prediction capacity.
The beneficial effects are that: compared with the prior art, the power distribution strategy based on the one-dimensional deep convolutional neural network is used for improving the network prediction capability; aiming at the current neural network power distribution strategy based on the full-connection form, the power distribution strategy based on the one-dimensional convolutional neural network is provided, compared with the traditional method, the calculation time is short, the complexity is low, and compared with the neural network based on the full-connection, the prediction capability is stronger, and the real-time and reliable online power distribution is realized in a communication system.
Drawings
FIG. 1 is a block diagram of a one-dimensional convolutional neural network;
FIG. 2 is a block diagram of a deep neural network;
fig. 3 is a system flow diagram.
Detailed Description
The invention is further described below in conjunction with the detailed description.
The resource allocation strategy based on the deep convolutional neural network is better than the prediction capability based on the full-connection type deep neural network, and the power allocation strategy based on the one-dimensional deep convolutional neural network is researched by aiming at a one-dimensional array on the basis of the conclusion.
Assuming that considering a single-cell cellular network environment, a base station is deployed in the center of a cell, k users are randomly distributed around the center, and the user sets: k= {1,2,3 …, K }, which are served by a centrally located base station, taking into account the rayleigh fading channel environment with average value 1, wherein the channel state information of user K to base station is h k, K e K, which is the path loss between users to base station in view of rayleigh fading environment, and the transmission power between them is denoted by p k. Then in the downlink the SINR (signal to interference plus noise ratio) at user k is expressed asWherein/>Is background noise, its achievable rate is denoted as R k：RK＝log2(1+Sk), and the system and rate performance can be denoted as/>Firstly, a utility function maximization problem is provided, a power distribution strategy based on deep learning is provided aiming at the utility function maximization problem, and particularly a power distribution strategy based on a one-dimensional convolutional neural network is provided, and the structural design of the power distribution strategy is detailed in the accompanying drawings and detailed description of the detailed description.
As shown in fig. 3, the training process of the present invention is as follows:
Step 1, communication modeling is carried out, and a communication scene of a single cell and multiple users is established;
step2, collecting a data set comprising channel state information and an optimal power distribution label under a water injection algorithm;
Step 3, designing and constructing a one-dimensional convolutional neural network, and initializing a weight coefficient;
step 4, constructing a loss function and determining an optimization algorithm to train the neural network;
And 5, storing the network when the iteration times or the loss function are smaller than a preset value.
The neural network structure used in the invention is shown in figure 1, and is responsible for learning the mapping relation between the gain of an input signal channel and the optimal user association or power distribution of an output signal, and the whole process comprises two parts of data set collection and neural network training. The method comprises the following specific steps:
step one, collecting channel state information between a user and a base station in an environment, operating a water injection algorithm to obtain a corresponding optimal power distribution label p *, and repeating 10 ten thousand times to obtain a data set;
Step two, determining the segmentation ratio of training data and test data;
Step three, constructing a one-dimensional deep convolutional neural network, and initializing the weight of the neural network;
Training a neural network, constructing a mean square error between a network predicted value p and a label value p * as a loss function, and determining an optimization algorithm during training;
and fifthly, when the loss function is smaller than a preset value or the iteration number of training is met, training is completed and the neural network is saved.
The process of the neural network for optimal resource allocation is described below in a specific example. Assuming that a base station is disposed in the center of the environment in a square area of 1km×1km, 10 user terminals are randomly distributed around the base station, the channel state information between the user and the base station is a path loss under consideration of a rayleigh fading channel, and the value is 128.1+37.6log 10 (d), where d is the euclidean distance between the user and the base station. Aiming at a sum rate formula, a system and rate maximization problem is proposed:
the constraint is/> Meaning that the sum of the power allocated to its associated users by each base station cannot be greater than the total power of the base stations, where the total transmit power p total = 20dBm of the base stations is known. After determining the target problem, a power distribution strategy based on a one-dimensional convolutional neural network is provided, firstly, a data set is collected, channel state information between a user and a base station is collected, a water injection power distribution algorithm is operated, a corresponding optimal power distribution label is obtained, the process is repeated for 10 ten thousands of times, the data set is obtained, and a segmentation ratio 9 of training test data is determined: 1, then designing and constructing a one-dimensional convolutional neural network, wherein the neural network has 7 layers in total, and comprises an input layer, 3 one-dimensional convolutional layers, a flat layer, a full-connection layer and an output layer, the size of a convolutional kernel in the convolutional layer is 3 multiplied by 1, the number of kernels in each layer is 8, 16 and 32 respectively, the number of neurons in the full-connection layer is 3000, and the number of neurons in the output layer is kept consistent with the number of users to be 10. The activation function in the hidden layer is ReLU, the activation function of the output layer is sigmoid function, the predicted power distribution scheme P is output by the output layer, and the mean square error between the target value and the predicted value is constructed as a loss function/>Where m=25 is the number of samples for each batch in the small batch gradient descent algorithm, training period epochs =300, and optimization algorithm selects Adam algorithm.
In order to verify the performance of the designed one-dimensional convolutional neural network, the performance of the neural network in the test stage needs to be tested and observed, a test data set is generated as the training data set, and the reliability of the proposed scheme is judged by observing the error between the output predicted value and the label value through the gain of the input channel. Meanwhile, as a comparison object, the training deep neural network is designed, and the structure of the training deep neural network is shown in fig. 2, so that the prediction capability of the proposed network structure in terms of power distribution is verified to be superior to that of a deep neural network scheme.
The specific implementation steps of the power control strategy based on the one-dimensional convolutional neural network are as follows:
And step one, communication modeling, namely establishing a single-cell multi-user communication environment, wherein the path loss between the user and the base station under the Rayleigh fading channel is taken into consideration as channel state information in a scene.
And step two, collecting data sets, namely collecting 10 ten thousand groups of data sets, including channel state information between users and base stations and optimal power distribution labels under a corresponding water injection algorithm.
Step three, determining a segmentation ratio 9 of the training data set and the test data set: 1.
And step four, designing a one-dimensional convolutional neural network framework and initializing weights.
And step five, constructing a loss function and determining an optimization algorithm to train the neural network.
And step six, continuously adjusting the super parameters during training, and storing the network when the iteration times are met.
And step seven, inputting the test set into the trained neural network, and verifying the reliability of the proposed scheme.
The foregoing is merely a preferred embodiment of the present invention, and it will be apparent to those skilled in the art that modifications and variations can be made without departing from the technical principles of the present invention, and the modifications and variations should also be regarded as the scope of the invention.
Claims (1)
1. The power distribution strategy based on the one-dimensional deep convolutional neural network is characterized in that: the method comprises the following steps:
Let the Rayleigh fading channel environment with the mean value of 1 be considered, wherein the channel state information from the user K to the base station is h k, K epsilon K, the channel state information is the path loss from the user to the base station under the Rayleigh fading environment, and the transmission power between the channel state information is represented by p k; then in the downlink the SINR at user k is expressed as Wherein/>Is background noise, h k is channel state information between user k and base station, p k is transmission power from user base station to user k, its achievable rate is denoted as R k＝log2(1+SINRk), system and rate performance is denoted as/>
Step 1, communication modeling; the communication modeling specifically comprises the following steps: establishing a single-cell multi-user communication environment, wherein the positions of a base station and a user in the environment are uniformly distributed, and a downlink channel between the base station and the user is a Rayleigh fading channel; the method specifically comprises the following steps: assuming a single cell cellular network environment, a base station is deployed in the center of a cell, k users are randomly distributed around the center, and a user set is formed: k= {1,2,3, K }, these subscribers are served by a base station at a central location;
step 2, collecting data sets, wherein the data sets comprise channel state information h k between a base station and a user and an optimal power value p * under a water injection algorithm as a group of data sets; repeating the collection of 10 ten thousand sets of said data sets a plurality of times;
Step3, constructing a one-dimensional deep convolutional neural network and initializing the weight of the neural network; the one-dimensional convolutional neural network is responsible for learning the mapping relation between the gain of an input signal channel and the optimal user association or power distribution of an output signal; the method specifically comprises the following steps:
Step 1) collecting channel state information between a user and a base station in an environment, operating a water injection algorithm to obtain a corresponding optimal power value p *, and repeating 10 ten thousand times to obtain a data set;
step 2) determining the segmentation ratio of training data and test data;
Step 3) designing and constructing a framework structure of a one-dimensional convolutional neural network, and initializing a weight parameter w and a bias parameter b of the neural network; the neural network has 7 layers, including an input layer, 3 one-dimensional convolution layers, a flat layer, a full-connection layer and an output layer, wherein the convolution cores in the convolution layers are 3 multiplied by 1, the number of each layer of core is 8, 16 and 32, the number of neurons of the full-connection layer is 3000, and the number of neurons of the output layer is kept consistent with the number of users to be 10; the activation function in the hidden layer is a ReLU, the activation function of the output layer is a sigmoid function, and the predicted power distribution scheme P is output through the output layer;
Step 4, training a neural network, constructing a mean square error between a predicted power value p and an optimal power value p * of the neural network as a loss function, and determining an optimization algorithm during training; wherein the loss function Determining an Adam optimizer to optimize the training neural network;
Step 5, when the loss function is smaller than a preset value or the iteration number of training is met, training is completed and the neural network is stored; specifically, the network is saved when the iteration number epochs =300 or the loss function is smaller than the preset value loss less than or equal to 0.1×10 -3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110437441.8A CN113518457B (en) | 2021-04-22 | Power distribution strategy based on one-dimensional deep convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110437441.8A CN113518457B (en) | 2021-04-22 | Power distribution strategy based on one-dimensional deep convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113518457A CN113518457A (en) | 2021-10-19 |
| CN113518457B true CN113518457B (en) | 2024-06-07 |
Family
ID=
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112153616A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Power control method in millimeter wave communication system based on deep learning |
| CN112153617A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Terminal equipment transmission power control method based on integrated neural network |
| CN112153615A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Deep learning-based user association method in multi-cell cellular D2D equipment |
| CN112492686A (en) * | 2020-11-13 | 2021-03-12 | 辽宁工程技术大学 | Cellular network power distribution method based on deep double-Q network |
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112153616A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Power control method in millimeter wave communication system based on deep learning |
| CN112153617A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Terminal equipment transmission power control method based on integrated neural network |
| CN112153615A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Deep learning-based user association method in multi-cell cellular D2D equipment |
| CN112492686A (en) * | 2020-11-13 | 2021-03-12 | 辽宁工程技术大学 | Cellular network power distribution method based on deep double-Q network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109474980A (en) | A kind of wireless network resource distribution method based on depth enhancing study | |
| CN112153616B (en) | Power control method in millimeter wave communication system based on deep learning | |
| CN113596785B (en) | D2D-NOMA communication system resource allocation method based on deep Q network | |
| CN111585816A (en) | Task unloading decision method based on adaptive genetic algorithm | |
| CN112153615B (en) | Deep learning-based user association method in multi-cell cellular D2D equipment | |
| CN112637907B (en) | Combined optimization method for user multi-association and downlink power distribution in millimeter wave network | |
| CN110191489B (en) | Resource allocation method and device based on reinforcement learning in ultra-dense network | |
| CN109195207B (en) | Energy-collecting wireless relay network throughput maximization method based on deep reinforcement learning | |
| CN113473580B (en) | User association joint power distribution method based on deep learning in heterogeneous network | |
| CN107708197A (en) | A kind of heterogeneous network user access of high energy efficiency and Poewr control method | |
| Lei et al. | Learning-based resource allocation: Efficient content delivery enabled by convolutional neural network | |
| CN111625998A (en) | Method for optimizing structure of laminated solar cell | |
| CN111010222B (en) | Deep learning-based large-scale MIMO downlink user scheduling method | |
| CN108990141B (en) | Energy-collecting wireless relay network throughput maximization method based on deep multi-network learning | |
| CN109089307B (en) | Energy-collecting wireless relay network throughput maximization method based on asynchronous dominant actor critic algorithm | |
| CN113242066B (en) | Multi-cell large-scale MIMO communication intelligent power distribution method | |
| CN113518457B (en) | Power distribution strategy based on one-dimensional deep convolutional neural network | |
| CN108965034A (en) | Small-cell base station super-intensive deployment under user-association to network method | |
| CN117119486A (en) | Deep unsupervised learning resource allocation method for guaranteeing long-term user rate of multi-cell cellular network | |
| CN112153617B (en) | Terminal equipment transmission power control method based on integrated neural network | |
| CN113301576B (en) | Cellular network resource allocation method based on improved genetic algorithm | |
| CN113518457A (en) | Power distribution strategy based on one-dimensional deep convolutional neural network | |
| CN112910716B (en) | Mobile fog calculation loss joint optimization system and method based on distributed DNN | |
| CN114423013A (en) | 5G heterogeneous network base station deployment method facing power distribution network | |
| CN113590211A (en) | Calculation unloading method based on PSO-DE algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |