CN111277308A

CN111277308A - Wave width control method based on machine learning

Info

Publication number: CN111277308A
Application number: CN202010045690.8A
Authority: CN
Inventors: 高晖; 贾承璐; 许文俊; 陆月明; 冯志勇
Original assignee: Beijing University of Posts and Telecommunications
Current assignee: Beijing University of Posts and Telecommunications
Priority date: 2020-01-15
Filing date: 2020-01-15
Publication date: 2020-06-12

Abstract

Aiming at the problem that user coverage and spectral efficiency are improved through intelligent beam width control in a millimeter wave (mmWave) -non-orthogonal multiple access (NOMA) network, an intelligent beam width control mechanism based on Machine Learning (ML) is provided, namely angle domain information of users at several moments in the future is predicted through Gaussian Process Machine Learning (GPML), user grouping is conducted through an unsupervised learning method according to the predicted angle domain information, accurate channel state information is obtained through methods such as beam tracking and channel estimation, and finally the optimal beam width is obtained through a Deep Neural Network (DNN) according to the channel state information, the angle domain information and the user grouping information. The beam tracking efficiency can be obviously improved through angle domain information prediction; by training DNN offline, the time for solving online can be saved; the channel state information, the angle domain information and the user grouping information of the users are directly mapped to the optimal beam width through the DNN, so that the user coverage and the spectral efficiency of the system can be remarkably improved.

Description

Wave width control method based on machine learning

Technical Field

The invention relates to a method for realizing flexible beam width control according to angle domain information of a user in a multi-user scene, in particular to a method for realizing high-efficiency transmission and flexible coverage of a millimeter wave (mmWave) -non-orthogonal multiple access (NOMA) system by utilizing Angle Domain Information (ADI) and a Machine Learning (ML), belonging to the technical field of wireless communication.

Background

Due to the abundant spectrum resources in the high frequency band, the millimeter wave (mmWave) frequency band has become a key technology of the fifth generation mobile communication system, however, the implementation of millimeter wave communication still faces severe path loss. Therefore, in order to compensate for the path loss of millimeter wave communication, a millimeter wave communication system usually employs a massive multiple input multiple output (m-MIMO) technique at the transmitting and receiving ends to implement directional transmission, and the energy efficiency and spectral efficiency of the system can be significantly improved by using space division multiplexing. Considering the distribution of users in a three-dimensional space, compared with the horizontal two-dimensional beam forming, the three-dimensional beam forming can realize more flexible coverage by adjusting a horizontal angle and a vertical angle, thereby further improving the system performance.

Due to the large-scale antenna array of millimeter wave transceivers, a purely digital transceiver architecture will consume significant energy and hardware resources. To balance hardware, energy cost, and system performance, a hybrid transceiver architecture is a more practical choice. However, due to its limited number of radio frequency links (RF) and reduced energy consumption of base stations, non-orthogonal multiple access (NOMA) technology is integrated into cellular networks when the number of users is greater than the number of RF. Different users can obviously improve the user coverage and the spectrum efficiency of the system by multiplexing time-frequency space resources.

The traditional two-dimensional (2D) wave width control is based on a linear antenna array, the width of a wave beam is controlled by adjusting the number of active antenna elements, the more the number of the active antenna elements is, the narrower the wave beam is, and the more concentrated the energy is. However, in the mmWave-NOMA scenario, the narrow beam will reduce the range of beam coverage, i.e. reduce the user coverage capability of the system, thereby impairing the spectral efficiency of the system. The three-dimensional (3D) wave width control proposed by the invention is to control the wave beam width in the horizontal and vertical directions by adjusting the antenna array sub-number of the planar antenna array in the x-y two dimensions. The service quality of the user and the maximization of the system performance cannot be guaranteed by the empirical method of 3dB bandwidth. While optimal system performance can be achieved by exhaustive search, the complexity of the system increases exponentially as antenna size, number of RF's, and number of users increase, and in m-MIMO and user-dense cellular scenarios, exhaustive search is not practical.

At present, machine learning is emerging and has been widely applied to communication system design and optimization from the fields of pattern recognition, artificial intelligence and the like. Gaussian Process Machine Learning (GPML), unsupervised learning, and deep learning are important branches of machine learning. The GMPL has the advantages of being easy to implement, capable of obtaining hyper-parameters in a self-adaptive mode, flexible non-parametric reasoning, output of probability significance and the like, and is widely applied to the field of time series prediction. Compared with supervised learning, unsupervised learning does not need labels and is mainly used for mining patterns, structures or rules of data. Deep learning has been widely applied to communication system design and optimization due to its powerful data fitting capability, especially to deal with communication problems that are difficult to express with mathematical models. According to a large amount of historical data, deep learning is applied to establish a mapping relation between system characteristics and system parameters, and therefore system performance is improved.

In addition, the communication system generally has a high requirement on real-time performance, and the simple calculation of the corresponding parameters (such as beam width) of the system through an algorithm may not satisfy the requirement of low delay due to the complexity of the calculation process. Therefore, a Machine Learning (ML) method can be introduced again, the existing data obtained by calculation is used for training a Deep Neural Network (DNN) to obtain experience, and when a variable of a new state is input to the input end of the DNN, more accurate system parameters can be obtained, so that the system performance is obviously improved.

Disclosure of Invention

The invention considers millimeter wave (mmWave) -non-orthogonal multiple access (NOMA) scenes and predicts based on multi-user historical Angle Domain Information (ADI), and provides an intelligent wave width control method based on Machine Learning (ML), namely, the ADI of users at the future moment is predicted by Gaussian Process Machine Learning (GPML), the unsupervised learning method is used for grouping the users according to the predicted ADI, then the ADI and channel gain information are obtained by beam tracking and channel estimation, and finally the optimal beam width which can enable the system spectrum efficiency to be maximum is obtained by a deep neural network according to the user grouping information, the ADI and the channel gain information. By training the DNN offline, the real-time performance of the system can be improved.

Drawings

Fig. 1 is a flow chart of implementation of bandwidth control.

Fig. 2 is a diagram of a wave width control deep neural network.

FIG. 3 is a system model diagram of a single base station multi-user scenario.

FIG. 4 is a system SE contrast curve for an optimal BC solution without bandwidth control, based on 3dB bandwidth BC, based on machine learning BC, and exhaustive search.

Fig. 5 is a user coverage capability comparison curve of the optimal BC scheme without BC, based on BC with 3dB bandwidth, based on BC of machine learning and exhaustive search, with the ordinate being the proportion of users satisfying the service quality.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the accompanying drawings.

Referring to fig. 1, a flowchart of implementation of the bandwidth control is to obtain an optimal beam width by means of offline exhaustive search based on an offline database (including Angle Domain Information (ADI), channel state information, and user grouping information). And training a Deep Neural Network (DNN) by taking the optimal beam width as a label and corresponding off-line data information as characteristics to obtain a DNN model. In the online implementation process, characteristics such as real-time Angle Domain Information (ADI), channel state information, user grouping information and the like are input into the DNN to obtain the near-optimal beam width, so that the system performance is improved.

Referring to fig. 2, a neural network structure diagram for realizing wave width control takes an optimal beam width as a tag, takes ADI, channel gain and grouping information as input features, the input features obtain corresponding beam width output through DNN, loss is calculated through comparison with the tag (the optimal beam width), and neural network parameters are continuously updated through gradient descent and reverse transfer until a difference between the output of DNN and the tag is lower than a certain threshold.

Referring to fig. 3, a system model diagram of a single base station multi-user scenario, a millimeter wave base station configures a hybrid architecture transmitter. And each user predicts future ADI according to historical ADI and feeds back the prediction result to the base station. And the base station carries out NOMA user grouping according to the predicted ADI, and then carries out beam scanning in each NOMA group to obtain accurate ADI, thereby completing link establishment.

Referring to fig. 4, the average spectral efficiency over 1000 slots is plotted as a function of signal-to-noise ratio (SNR). The average spectral efficiency increases with increasing SNR. The wave width control method based on machine learning is close to the optimal solution obtained through exhaustive search, is obviously superior to the wave width control method based on 3dB wave width and the traditional NOMA transmission method without wave width control, and proves that the wave width control method based on machine learning can obviously improve the system spectrum efficiency.

Referring to fig. 5, an average ratio of users satisfying quality of service (QoS) to the total number of users for 1000 slots. As the signal-to-noise ratio (SNR) increases, the number of users who satisfy QoS gradually increases. The wave width control method based on machine learning is close to the optimal solution obtained through exhaustive search, is obviously superior to the wave width control method based on 3dB wave width and the traditional NOMA transmission method without wave width control, and proves that the wave width control method based on machine learning can obviously improve the user coverage capability of the system.

In summary, we can obtain the beam width with better performance of each time slot by the ML method, thereby verifying the concept of the wave width control based on machine learning.

The above description is only exemplary of the present invention and should not be taken as limiting the invention, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent wave width control method based on Machine Learning (ML), wherein a plurality of users simulating millimeter wave architecture exist in a millimeter wave network, a base station mixing millimeter wave architecture, millimeter wave transmission is carried out between the base station and the users, the base station adjusts the wave beam width of each cluster through the machine learning method to improve the spectrum efficiency and the user coverage capability, and the specific method comprises the following steps:

(1) predicting angle domain information based on a Gaussian process machine learning method: the historical angle domain information is used as input, the angle domain information of a plurality of time slots in the future is predicted, and beam tracking and user grouping of millimeter wave communication are further assisted;

(2) unsupervised learning based user grouping: according to the predicted angle domain information, user clustering is carried out by using an unsupervised learning method, and the base station carries out non-orthogonal multiple access transmission on the users of each cluster;

(3) wave width control based on a deep neural network: obtaining channel gain information through channel estimation, taking an angle domain, channel gain and user clustering information as the input of a deep neural network, taking the optimal beam width of each cluster obtained through exhaustive search as a corresponding label, training the deep neural network offline, and establishing a mapping relation between system parameters and the optimal beam width;

(4) the online realization comprises the following steps: and transferring the offline-trained wave width control neural network model to an actual millimeter wave system, and correcting the model by using a transfer learning method when the network topology in the system is changed violently.