CN110855331A

CN110855331A - Machine learning-based proactive antenna inclination angle adjusting mechanism

Info

Publication number: CN110855331A
Application number: CN201910816598.4A
Authority: CN
Inventors: 高晖; 姜百淳; 许文俊; 曹若菡
Original assignee: Beijing University of Posts and Telecommunications
Current assignee: Beijing University of Posts and Telecommunications
Priority date: 2019-08-30
Filing date: 2019-08-30
Publication date: 2020-02-28

Abstract

Aiming at the problem that the intelligent antenna inclination angle adjustment in a cell network needs to meet the requirement of communication real-time performance, a Machine Learning (ML) -based proactive antenna inclination angle adjustment (PATAA) mechanism is provided. The user position at the next moment is predicted through an Echo State Network (ESN), and then the inclination angle group of the base station antenna at the next moment is obtained through NN fitting. The online training of the NN can save the online solving time; in the mechanism, the base station directly performs PATAA with ideal spectrum efficiency performance before acquiring the real position of the user, so that the inclination angle adjustment time can be further shortened, and the effect of meeting the real-time requirement of a communication system is achieved.

Description

Machine learning-based proactive antenna inclination angle adjusting mechanism

Technical Field

The invention relates to user position prediction in a three-sector scene and antenna inclination angle group prediction on the basis of the user position prediction, in particular to a method for realizing proactive antenna inclination angle adjustment through a machine learning method, and belongs to the technical field of wireless communication.

Background

A large-scale multiple input multiple output system (LS-MIMO) is a key technology in a fifth generation mobile communication system (5G), which can significantly improve Energy Efficiency (EE) and Spectral Efficiency (SE) of the system using spatial multiplexing. As the most important linear signal processing techniques, Zero Forcing (ZF), Maximum Ratio Transmission (MRT), and Minimum Mean Square Error (MMSE) precoding/detection methods are widely used in the uplink/downlink of the MIMO system, which can improve the throughput of users and systems by performing a preprocessing operation using known spatial channel information. On the basis of horizontal two-dimensional adaptive beamforming, simple three-dimensional beamforming (3D-BF) can reduce inter-cell interference by adjusting Antenna Tilt Angle (ATA), thereby further improving system performance.

At present, ATA adjustment mechanisms researched in the leading-edge field are all established on fixed user distribution, the condition that the user distribution also changes (users continuously move in an actual scene) is not considered, in the research, ATA is predicted based on the positions of the users, and the application range is wider.

The prediction of the user position belongs to the problem of Time Series Prediction (TSP), and the Echo State Network (ESN) is widely applied in the fields of TSP, channel estimation, spectrum prediction, nonlinear control and the like due to the characteristics of simple structure and feedback connection. The core of the ESN is a randomly generated and unchangeable reserve pool, and the output weight of the ESN is the only part needing to be adjusted, so that the ESN has few parameters and high training speed, and the network can be trained by simple linear regression.

In addition, the communication system generally has a high requirement for real-time performance, and the calculation of the corresponding parameters (such as ATA, etc.) of the system by using an algorithm alone may not meet the requirement of low latency 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 the Neural Network (NN) to obtain experience, and more accurate system parameters can be obtained when the variables of a new state are input to the NN input end, and the time can be greatly shortened.

Disclosure of Invention

The invention considers a three-base-station three-sector scene and carries out prediction based on the actual position of multiple users, and proposes a proactive antenna inclination angle adjustment (PATAA) based on Machine Learning (ML), namely, the user position at the next moment is predicted through an Echo State Network (ESN), and then an NN fitting is used for obtaining a base station antenna inclination angle group with ideal proven performance at the next moment, thereby realizing PATAA. By means of offline training and prediction of two layers of neural networks, namely ESN and common NN, the time for obtaining the optimal inclination angle at the next moment can be shortened, so that the base station can perform proactive inclination angle adjustment with ideal spectrum efficiency performance, and meanwhile, the real-time requirement of a communication system is met.

Drawings

Fig. 1 is an abstract attached drawing.

FIG. 2 is a diagram of a three-base-station three-sector system model.

FIG. 3 is a flowchart of a method for verifying the concept of PATAA.

FIG. 4 is a Normalized Mean Square Error (NMSE) curve as a function of training set size at different ESN mid-layer neuron numbers.

Fig. 5 shows the mean square error of the prediction set as a function of the number of hidden neurons at different training step sizes when P is 1.

FIG. 6 is a comparison of spectral efficiency corresponding to maximum spectral efficiency of poor search and ML fitting inclination.

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.

Fig. 1 is an abstract drawing, and the detailed description thereof is referred to the abstract part of the specification.

Referring to fig. 2, we consider a regular hexagon unit formed by three adjacent sectors of three adjacent base stations, and set that initially, single-antenna users are uniformly distributed in a hexagon, each single-antenna user is served by a base station closest to the user, the user position is set to be unchanged in one time slot (0.5s), the user position moves in each time slot, when the user does not reach the hexagon boundary, the moving rule follows an infinite domain simulation model, when the user reaches the boundary, the speed direction of the user is set to bounce at the original speed according to the reflection direction of similar light, that is, at the moment, the user follows a random walk model. Considering the downlink, it is assumed that the base station can obtain perfect Channel State Information (CSI) and use zero-forcing ZF precoding.

Referring to fig. 3, a mental map of the PATAA concept is validated. Substituting the user real position group obtained based on the mobile model into a combined spectral efficiency (SSE), traversing three base station ATA (ATA) with the granularity of 1 degree to obtain an ATA group which enables the SSE to be maximum, and calculating the average SSE which changes along with the transmission power P; meanwhile, the real position group is used as the input of the NN, the corresponding optimal ATA group is used as the output of the NN, and the NN is trained and parameter-adjusted; then, the Echo State Network (ESN) with the simulation determined parameters is used for predicting the position of the user in the next time slot to obtain a predicted position group, the predicted position group is substituted into the trained NN to obtain the predicted next time slot ATA, and then the average SSE corresponding to different P is calculated. Comparing the two sets of SSEs, if they are in close proximity, the PATAA concept can be verified.

Referring to fig. 4, in the parameter adjusting process of the ESN of the first-layer network, a total of 30 users in the hexagon are set, the total time slot number is 10000, the ESN is used to predict the position of the user in the next time slot, the prediction result is compared with the real user position, and a Normalized Mean Square Error (NMSE) is calculated. NMSE as a function of training set size at different ESN intermediate neuron numbers is shown. The overall trend in observed available NMSE is to decrease and then increase with increasing training set size, and when N is 15 and Nt is 4000, more satisfactory NMSE is obtained and both N and Nt are smaller.

Referring to fig. 5, in the tuning process of the NN in the second layer, the corresponding predicted user position is taken (10,4000), and the remaining 6000 time slots are sampled every fourth time slot, so as to obtain the predicted user positions of 2000 time slots with an interval of 1.5 s. An optimal ATA group of 2000 time slots is obtained in a traversal mode and is used as NN output, a user position (horizontal and vertical coordinates are disassembled) is used as input, and NN is trained. We consider only the single hidden layer NN, whose parameters include the training step size (TS) and the Number of Hidden Layer Neurons (NHLN). NHLN is typically 0.01, since early stops occur almost as soon as NN training begins when TS is 0.007 and NHLN is greater than 23, or NHLN is 23 and TS is greater than 0.007, so that training cannot continue, the abscissa is only 23 and the maximum training step size is only 0.007. From the figure, it can be seen that the smallest full data set MSE can be obtained when TS is 0.007 and NHLN is 23.

Referring to fig. 6, with the trained neural network corresponding to (0.007,23), we obtain a predicted ATA group based on the predicted user location and calculate its average spectral efficiency for 2000 timeslots. As a benchmark reference, we simultaneously calculate the average spectrum efficiency corresponding to the best dip angle by the actual user position. The two are compared in the figure, and the spectral efficiency obtained by predicting the position by the ESN and then fitting the estimated position by the NN is 99.7-99.85 percent of the spectral efficiency obtained by poor search based on the actual position of the user, and the two are very close to each other.

In conclusion, the ATA group with better performance of the next time slot can be obtained by the ML method, so that the PATAA concept based on ML-LP is verified.

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. Aiming at the problem that the intelligent antenna inclination angle adjustment in a cell network needs to meet the real-time requirement of a communication system, a Machine Learning (ML) -based proactive antenna inclination angle adjustment (PATAA) mechanism is provided.

(1) Proactive Antenna Tilt Angle Adjustment (PATAA) mechanism

The PATAA is realized by connecting two Neural Network (NN) modules in series, the first NN module predicts the next time slot user position group by using the Echo State Network (ESN) of the neuron in the intermediate reserve pool adopting a one-way ring connection structure without a self-ring, and the second NN module obtains the antenna inclination angle group based on the next time slot user position by using the single hidden layer NN trained by a gradient descent method.

As above, the "proactive" tilt angle adjustment can be implemented, that is, the base station group can predict the antenna tilt angle group of which the next time slot makes the regional combined spectrum effect approach the maximum (the base station can obtain the spectrum effect greater than 99.5% corresponding to the poor search optimal tilt angle under the perfect channel state information), and then the tilt angle can be directly adjusted before the true position of the user is obtained when the next time comes, so as to shorten the tilt angle adjustment time; and the process of 'foreknowing' is carried out by NN trained under the line, so that the online adjustment time is further saved, and the effect of meeting the real-time requirement of the communication system is achieved.