CN112311486B

CN112311486B - Method and device for accelerating wireless network interference prediction convergence

Info

Publication number: CN112311486B
Application number: CN201910688702.6A
Authority: CN
Inventors: 彭涛; 董卫国
Original assignee: Beijing University of Posts and Telecommunications
Current assignee: Beijing University of Posts and Telecommunications
Priority date: 2019-07-29
Filing date: 2019-07-29
Publication date: 2021-08-03
Anticipated expiration: 2039-07-29
Also published as: CN112311486A

Abstract

The invention discloses a method and a device for accelerating wireless network interference prediction convergence. The method comprises the following steps: executing wireless resource routine scheduling, acquiring wireless resource allocation information of each user and wireless measurement data of each user, and forming a training set of each user based on the wireless resource allocation information of each user and the wireless measurement data of each user; respectively training the training set of each user for multiple times by using a machine learning algorithm to obtain a plurality of training models of each user; predicting the interference condition of each user by using each training model of each user to obtain multiple prediction results, and determining an inaccurate user set for interference prediction of each user based on the fluctuation of the multiple prediction results; and aiming at each user, constructing a user tuple containing the user and each user with inaccurate interference prediction in the user set, and executing radio resource enhanced scheduling to increase the user multiplexing times in the user tuple.

Description

Method and device for accelerating wireless network interference prediction convergence

Technical Field

The invention relates to the technical field of interference prediction, in particular to a method and a device for accelerating convergence of wireless network interference prediction.

Background

Interference prediction has many practical and potential applications. In a wireless network, if the interference situation between users can be predicted, effective interference avoidance and resource allocation can be performed, and data transmission strategies such as Modulation and Coding Scheme (MCS) allocation can be guided. In the current network environment which is becoming more intensive, effective interference prediction is particularly important, the interference prediction plays an important guiding role in wireless resource management, and system capacity and user experience can be effectively improved.

The interference prediction is based on a mathematical relationship between user signal strength and interference in the wireless communication. In wireless communication, wireless resources are divided according to two dimensions of Time and frequency, the minimum scheduling unit of the Time dimension is a subframe, one subframe is called a Transmission Time Interval (TTI), the minimum scheduling unit of the frequency dimension is a Resource Block (RB), and users using the same RB in the same TTI may generate interference with each other, which affects the quality of a wireless channel of the users. Therefore, interference prediction between users can be realized according to whether each user uses the same radio resource and the channel quality of the user at that time.

The interference prediction algorithm uses a machine learning algorithm. Machine learning algorithms are classified into supervised learning and unsupervised learning algorithms. For the supervised machine learning algorithm, data of a training set are divided into features and labels, the label value of any feature can be predicted by learning the data, and common algorithms include algorithms such as neural network and linear regression. The machine learning algorithm can train the training set data to obtain a training model, and then the model is used for predicting the channel quality. Wireless network interference modeling is the basis for wireless network resource allocation. The more comprehensive and accurate the interference information is, the more effective the network can be guided to carry out reasonable resource allocation, and the more favorable the network performance is. The existing methods for acquiring the interference strength mainly include the following three methods: the first method is to establish an interference matrix based on sweep frequency data, and the method is based on the sweep frequency data analysis of a specific sampling position of a cell; secondly, an interference matrix is established based on a measurement report message of the mobile phone, and interference in the wireless network is analyzed through the signal intensity of each cell in the measurement report of the mobile phone; and thirdly, predicting the interference intensity based on a machine learning algorithm to establish an interference matrix.

In the prior art, frequency domain information in frequency sweep data is complete, but a generated interference matrix can only reflect the interference condition of a sampling point, if the interference condition changes, a measuring point needs to be redeployed, and the measuring cost is too high. The mobile phone measurement report contains user interference source information, but only contains interference information of a plurality of adjacent cells with strong interference signals, and when the network is dense, the interference information missing in the established interference matrix is more. The machine learning algorithm is used for predicting the interference intensity, a comprehensive and accurate interference matrix can be obtained for a large-data-volume sample, but the algorithm has excessive requirements on the sample data volume, long feedback time is needed, and the accuracy cannot be guaranteed under the condition that the sample data volume is small.

And predicting the interference intensity by using a machine learning algorithm, wherein the interference prediction result gradually tends to converge as the sample data amount gradually increases along with the increase of time, and the convergence speed is low. For a large-data-volume sample, a comprehensive and accurate interference matrix can be obtained by utilizing machine learning interference prediction, the requirement on the large-data-volume sample means that longer information collection time is needed, and the interference prediction accuracy cannot be guaranteed under the condition that the sample data volume is smaller.

Disclosure of Invention

The embodiment of the invention provides a method and a device for accelerating wireless network interference prediction convergence.

The technical scheme of the embodiment of the invention is as follows:

a method of accelerating convergence of wireless network interference predictions, comprising:

executing wireless resource routine scheduling, acquiring wireless resource allocation information of each user and wireless measurement data of each user, and forming a training set of each user based on the wireless resource allocation information of each user and the wireless measurement data of each user;

respectively training the training set of each user for multiple times by using a machine learning algorithm to obtain a plurality of training models of each user;

predicting the interference situation of each user by using each training model of each user to obtain multiple prediction results, and determining an inaccurate user set for interference prediction of each user based on the fluctuation of the multiple prediction results;

and aiming at each user, constructing a user tuple containing the user and each user with inaccurate interference prediction in an inaccurate interference prediction user set of the user, and executing radio resource enhanced scheduling to increase the multiplexing times of the users aiming at the same transmission time interval and the same resource block in the user tuple.

In one embodiment, the wireless measurement data packet signal to interference and noise ratio.

In one embodiment, the forming a training set for each user based on the radio resource allocation information of each user and the radio measurement data of each user includes:

generating a first table, the first table comprising: in any transmission time interval, a user scheduling set on each resource block and the signal-to-interference-and-noise ratio of each user in the user scheduling set;

generating a training set for each user based on the first table, wherein the training set for each user comprises: and in any transmission time interval, scheduling a user set of the same resource block and the signal-to-interference-and-noise ratio of the user at the same time as the user.

In one embodiment, the performing radio resource scheduling to increase the number of times of user multiplexing for the same transmission time interval and the same resource block in the user tuple comprises:

a1, resource allocation is carried out resource allocation resource by resource block, if a user tuple exists to apply for wireless resources, the wireless resources are allocated by taking the user tuple as a unit, the user tuple is allocated for each resource block in sequence, wherein, the user tuple with a plurality of times of multiplexing needs of users is preferentially selected, if the resource allocation is finished, the users still need to apply for the resources and the resource blocks still remain, the step A2 is executed, and if the resource blocks do not remain, the step A3 is executed;

a2, for the rest resource blocks, resource allocation is carried out in sequence by resource block, and users are randomly allocated to each resource block in sequence until each resource block has at most two users;

and A3, carrying out resource allocation by user, randomly selecting users to carry out resource allocation, wherein for the selected users, calculating the average value of the multiplexing requirement times of the users on each resource block, and selecting the resource block with the maximum multiplexing requirement times to carry out resource multiplexing.

In one embodiment, the method further comprises:

uniformly determining the multiplexing requirement times of each user tuple as a preset parameter;

and determining the multiplexing demand times of the user tuples according to the fluctuation of the multiple prediction results, wherein more multiplexing demand times are determined for tuples with higher fluctuation.

An apparatus for accelerating convergence of wireless network interference predictions, comprising:

a training set obtaining module, configured to perform radio resource regular scheduling, obtain radio resource allocation information of each user and radio measurement data of each user, and form a training set of each user based on the radio resource allocation information of each user and the radio measurement data of each user;

the training module is used for respectively training the training set of each user for multiple times by utilizing a machine learning algorithm to obtain a plurality of training models of each user;

the set determination module is used for predicting the interference situation of each user by utilizing each training model of each user to obtain multiple prediction results, and determining an inaccurate user set of interference prediction of each user based on the fluctuation of the multiple prediction results;

and the scheduling module is used for constructing a user tuple containing the user and each user with inaccurate interference prediction in the user interference prediction inaccurate user set aiming at each user, and executing radio resource enhanced scheduling to increase the multiplexing times of the users aiming at the same transmission time interval and the same resource block in the user tuple.

In one embodiment, the training set obtaining module is configured to generate a first table, where the first table includes: in any transmission time interval, a user scheduling set on each resource block and the signal-to-interference-and-noise ratio of each user in the user scheduling set; generating a training set for each user based on the first table, wherein the training set for each user comprises: and in any transmission time interval, scheduling a user set of the same resource block and the signal-to-interference-and-noise ratio of the user at the same time as the user.

In one embodiment, the scheduling module is configured to perform the following steps:

In one embodiment, the scheduling module is configured to uniformly determine the multiplexing requirement times of each user tuple as a preset parameter; and determining the multiplexing demand times of all user tuples according to the fluctuation of the multiple prediction results, wherein more multiplexing demand times are determined for the user tuples with higher fluctuation.

As can be seen from the above technical solutions, the embodiments of the present invention include: executing wireless resource routine scheduling, acquiring wireless resource allocation information of each user and wireless measurement data of each user, and forming a training set of each user based on the wireless resource allocation information of each user and the wireless measurement data of each user; respectively training the training set of each user for multiple times by using a machine learning algorithm to obtain a plurality of training models of each user; predicting the interference condition of each user by using each training model of each user to obtain multiple prediction results, and determining an inaccurate user set for interference prediction of each user based on the fluctuation of the multiple prediction results; aiming at each user, constructing a user tuple containing the user and each user with inaccurate interference prediction in an inaccurate interference prediction user set of the user, executing radio resource enhanced scheduling to increase the multiplexing times of the users in the user tuple, increasing the multiplexing of the user and the user with inaccurate interference prediction by targeted scheduling of the inaccurate interference prediction user set, and rapidly improving the diversity of data and realizing the acceleration of interference prediction convergence.

Moreover, wireless resource allocation data and measurement data generated in the network operation process are effectively utilized, the accuracy of an interference prediction result is analyzed through big data analysis and a machine learning algorithm, user scheduling is carried out aiming at the condition that the interference prediction is inaccurate, and quick, accurate and complete interference prediction is realized.

Drawings

Fig. 1 is a flow chart of a method for accelerating convergence of interference prediction in a wireless network according to the present invention.

Fig. 2 is an exemplary flow chart of a method for accelerating convergence of wireless network interference predictions in accordance with the present invention.

FIG. 3 is a graphical illustration of prediction accuracy for various users in accordance with the present invention.

Fig. 4 is a diagram illustrating union recognition rates of respective users according to the present invention.

FIG. 5 is a diagram illustrating intersection recognition rates of respective users according to the present invention.

FIG. 6 is a graph of single-interference predicted Root Mean Square Error (RMSE) versus simulation time in accordance with the present invention.

FIG. 7 is a diagram of multiple interference prediction (RMSE) versus simulation time in accordance with the present invention.

Fig. 8 is a block diagram of an apparatus for accelerating convergence of interference predictions for a wireless network in accordance with the present invention.

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.

For simplicity and clarity of description, the invention will be described below by describing several representative embodiments. Numerous details of the embodiments are set forth to provide an understanding of the principles of the invention. It will be apparent, however, that the invention may be practiced without these specific details. Some embodiments are not described in detail, but rather are merely provided as frameworks, in order to avoid unnecessarily obscuring aspects of the invention. Hereinafter, "including" means "including but not limited to", "according to … …" means "at least according to … …, but not limited to … … only". In view of the language convention of chinese, the following description, when it does not specifically state the number of a component, means that the component may be one or more, or may be understood as at least one.

Aiming at the limitation of the algorithm in the prior art, the embodiment of the invention provides a user scheduling method for accelerating interference prediction convergence. The embodiment of the invention can effectively utilize the wireless resource distribution data and the measurement data generated in the network operation process based on the characteristics of centralized processing and global wireless resource management of a cloud wireless access network (C-RAN) network, analyze the accuracy of an interference prediction result through big data analysis and a machine learning algorithm, perform user scheduling aiming at the condition of inaccurate interference prediction and realize quick, accurate and complete interference prediction. Specifically, aiming at the defects of the current interference prediction algorithm, the invention provides a scheduling method for accelerating interference prediction convergence, which can quickly and accurately realize interference prediction in a wireless network. The method is based on a mathematical model of interference in wireless communication, and utilizes a machine learning algorithm to realize rapid convergence of interference prediction.

In the embodiment of the present invention, the data structure of the training set is characterized by whether the interfering users use the same radio resource, and the label is a Signal to interference plus noise ratio (SINR) corresponding to the situation.

The method provided by the embodiment of the invention comprises four steps of data acquisition, model training, inaccurate interference prediction user set detection and user scheduling, and the aim of accelerating the convergence speed of the interference prediction is realized by sequentially executing and continuously iterating the four steps until no user with inaccurate interference prediction exists in the system. Wherein: the data acquisition refers to acquiring wireless information in a wireless network, including allocation data of wireless resources and wireless measurement information, as a training set of a machine learning algorithm. The model training is to train the training set for multiple times through machine learning algorithms such as neural network and linear regression, wherein the characteristic value input by the algorithm is the allocation condition of wireless resources, the label value corresponds to wireless measurement data, namely the signal-to-interference-and-noise ratio of a user on any resource block at any time, and the obtained model can predict wireless measurement information according to the allocation condition of the wireless resources. And predicting the user signal-to-interference-and-noise ratio measurement information under the mutual interference between every two users by the obtained training models, obtaining prediction results of a plurality of label values for the same characteristic value, measuring the volatility of the prediction results, considering that the interference prediction between every two users is inaccurate if the fluctuation degree is greater than a preset threshold value, analyzing all user combinations and finally realizing the user set detection with inaccurate interference prediction. Finally, the multiplexing of the users and the interference users with inaccurate interference prediction is increased through the targeted scheduling of the user set with inaccurate interference prediction, the diversity of data can be rapidly promoted, and the acceleration of interference prediction convergence is realized.

As shown in fig. 1, the method includes:

step 101: and executing the conventional scheduling of the wireless resources, acquiring the wireless resource allocation information of each user and the wireless measurement data of each user, and forming a training set of each user based on the wireless resource allocation information of each user and the wireless measurement data of each user.

Step 102: and training the training set of each user for multiple times by using a machine learning algorithm to obtain multiple training models of each user.

Step 103: and predicting the interference situation of each user by using each training model of each user to obtain multiple prediction results, and determining an inaccurate user set for interference prediction of each user based on the fluctuation of the multiple prediction results.

Step 104: and for each user, constructing a user tuple containing the user and each interference prediction inaccurate user in the interference prediction inaccurate user set of the user, and executing radio resource enhancement scheduling to increase the users aiming at the same transmission time interval and the same resource block in the user tuple.

Such as: suppose user 1, whose set of users whose interference prediction is inaccurate, includes (user 2, user 3, user 4, user 5); the user tuple for user 1 then includes: user tuple 1 (user 1, user 2); user tuple 2 (user 1, user 3); user tuple 3 (user 1, user 4) and user tuple 4 (user 1, user 5). For another example: suppose user 2, whose set of users whose interference prediction is inaccurate, includes (user 1, user 5, user 6, user 7); the user tuple for user 2 comprises: user tuple 1 (user 2, user 1); user tuple 2 (user 2, user 5); user tuple 3 (user 2, user 6) and user tuple 4 (user 2, user 7).

The conventional scheduling of radio resources is a radio resource scheduling mode in the prior art at present. In one embodiment, performing the regular scheduling of radio resources comprises at least one of: performing wireless resource scheduling by adopting a Proportional Fair algorithm (Prooral Fair); performing radio resource scheduling using a Round Robin (RR) algorithm; performing radio resource scheduling using a Maximum carrier-to-interference ratio algorithm (Maximum C/I) algorithm, and so on. The above exemplary description describes a typical example of the conventional scheduling of radio resources, and those skilled in the art will appreciate that this description is only exemplary and is not intended to limit the scope of the embodiments of the present invention.

In one embodiment, the data packet signal to interference and noise ratio is measured wirelessly. In one embodiment, forming a training set for each user based on the radio resource allocation information of each user and the radio measurement data of each user comprises: generating a first table, the first table comprising: in any transmission time interval, a user scheduling set on each resource block and the signal-to-interference-and-noise ratio of each user in the user scheduling set; generating a training set for each user based on the first table, wherein the training set for each user comprises: and in any transmission time interval, scheduling a user set of the same resource block and the signal-to-interference-and-noise ratio of the user at the same time as the user.

Performing radio resource scheduling to increase the number of user multiplexes for the same transmission time interval and the same resource block in the user tuple comprises:

In one embodiment, the method for determining the multiplexing requirement times for different user tuples is further included. The method comprises the following steps: uniformly determining the multiplexing requirement times of each user tuple as a preset parameter; and determining the multiplexing demand times of the user tuples according to the fluctuation of the multiple prediction results, wherein more multiplexing demand times are determined for tuples with higher fluctuation.

Fig. 2 is an exemplary flow chart of a method for accelerating convergence of wireless network interference predictions in accordance with the present invention. The flow chart of the method provided by the invention is shown in figure 2, and the method mainly comprises four steps of data acquisition, model training, inaccurate interference prediction user set detection and user scheduling, and achieves the acceleration of the convergence speed of the interference prediction through continuous iteration of the four steps until no user with inaccurate interference prediction exists in the system. The invention reduces the data quantity requirement of the sample and accelerates the convergence of the interference prediction by rapidly enriching the diversity of the sample data.

The following is a detailed description of the above four steps.

(1) Data acquisition

The data acquisition refers to acquiring wireless information in a wireless network, including allocation data of wireless resources and wireless measurement information, as a training set of a machine learning algorithm. Due to the centralized management characteristic of the C-RAN network, the wireless resource allocation information in the network is easy to obtain, and meanwhile, the remote wireless radio head can obtain corresponding wireless measurement data. For any TTI, the C-RAN network can count the user scheduling condition on each RB and the SINR value corresponding to the scheduled user. The scheduling conditions of all users on each RB of the TTI used for data transmission and the SINR values of the corresponding users are recorded, and a data set in the form of table 1 can be obtained.

Table 1 is a C-RAN wireless data information table.

TABLE 1

In Table 1, the second row of data indicates the set of users { UE }_i,...,UE_jAt TTI_tAt a moment using RB_rThe corresponding set of SINR values is { gamma }_i,...,γ_j}。

Table 1 shows a data set for storing global radio information in a C-RAN network, in which data of all users is contained.

(2) Model training

Machine learning algorithms can model data and predict outputs from data inputs. There are two basic types of machine learning: supervised learning and unsupervised learning. Input data for supervised learning requires manual tagging, while unsupervised learning does not. The machine learning algorithm learns the training set data to obtain a machine learning model, so that a prediction function can be realized.

In the embodiment of the invention, model training is to train a training set for multiple times through machine learning algorithms such as neural networks, linear regression and the like, wherein a characteristic value input by the algorithm is the allocation condition of wireless resources, a label value corresponds to wireless measurement data, namely, the signal-to-interference-and-noise ratio of a user on any resource block at any time, and the obtained model can predict wireless measurement information according to the allocation condition of the wireless resources.

In the present invention, the SINR of the user level is predicted by the machine learning model, so the data in table 1 needs to be formatted and preprocessed to screen the training set of each user. With user UE_iFor example, the global data set in the screening table 1 includes UEs_iScheduled related data, reserved only with the UE_iMultiplexing situation of users scheduling same video resource and UE (user equipment)_iThe corresponding SINR. The resulting data set format is shown in table 2 below, where table 2 is the test set data format for interference prediction for a single user.

TABLE 2

x_m,nE.g {0,1}, and represents the user UE in the mth row of data of the training set_nWhether in TTI_tTime of day and user UE_iMultiplexing resource RB_rThe multiplexing is 1, the non-multiplexing is 0, gamma_iFor a user UE_iThe corresponding SINR value. For user UE_iTraining set of (2), rank UE_iAlways 1.

On the other hand, fast fading of the network is unavoidable, and SINR is interfered with by fluctuation of time and frequency, but within a certain time range, the influence of fast fading should fluctuate around an expected value, and the prediction of SINR is actually the prediction result under the influence of the expected value. Therefore, the SINR value can be regarded as independent of frequency domain and time, and time and frequency domain information in the data can be deleted.

Therefore, the time-frequency information in table 2 is deleted, and a final data set table is constructed as shown in table 3 below, where table 3 is a single-user interference prediction test set data format table.

TABLE 3

Taking the data set shown in the form of table 3 as a training set of a machine learning algorithm, the multiplexing condition of users as a characteristic value, and SINR as a label value, and performing multiple times of machine learning training on the data set of each user

(3) Identification of users with inaccurate interference prediction

The machine learning algorithm usually adopts a gradient descent method to adjust the coefficient, perform fitting of the function and measure the fitting effect by the cost function. Since the cost function in the fitting process is a non-convex function, under the condition of multiple dimensions, a large number of saddle points and local minimum values exist, and the initial coefficient is randomly generated in the initialization stage of the gradient descent method, so that different local minimum values or saddle points are finally involved each time, and the finally fitted functions are different and have different prediction capabilities, namely the prediction results of the same input by the model obtained by multiple times of training of the same data set are different. For a data set which is abundant enough, a model obtained by multiple times of training tends to be convergent, fitting accuracy should be similar, namely, the fluctuation of a prediction result is small, and the fluctuation is large for a data set which is poor in diversity.

For a data set with a small sample size, the trained model is often inaccurate. In a data set, multiple user multiplexing resources often exist in any RB of any TTI, any user is interfered by more than one user, and the accuracy of interference prediction of each interference user is difficult to distinguish. In consideration of a single interference user scenario, assuming that only two users are multiplexed in a certain RB of a TTI and any one user is only affected by one interference user, it is easier to identify a user with inaccurate interference prediction in this scenario. Under the scene of a single interference user, the SINR of the user is interfered by an interference user, namely, the label value is only influenced by one characteristic, so that the interference of other characteristics can be eliminated, and the prediction accuracy degree of the characteristic by the data set is detected. Meanwhile, on the premise that the data set is rich enough, the prediction model tends to be convergent, the user SINR prediction result under the influence of any interference user should tend to be convergent, the fluctuation is small, and the fluctuation is large when the data diversity is insufficient. Therefore, the data set can be trained for multiple times, and the prediction accuracy of the user can be judged by utilizing the fluctuation of the prediction effect of the model obtained by training each time.

And training the current data set for multiple times, predicting the SINR of the user under the interference of a single interference user, predicting the fluctuation of the SINR value by using a multiple training model, and if the fluctuation is high, determining that the model prediction is inaccurate for the prediction of the characteristic dimension, so that the multiplexing frequency of the user and the interference user needs to be increased.

In case, consider a user UE_iPresence of interfering user UE_jTraining the current data set for multiple times, and predicting the UE by using the model obtained by each training_jUser Equipment (UE) under interference_iThe SINR of (1). Let gamma_k,jObtained at UE for k-th training_jUser Equipment (UE) under interference_iSINR of (1), t being the total number of training times, σ_jIndicating the volatility of the data. There are many methods for measuring data fluctuation, such as the method of the worst value difference, the variance or standard deviation, the percentage measurement, etc.

The maximum difference value is judged by comparing the maximum value and the minimum value in the data, and the calculation formula is shown as the following formula (1):

the standard deviation is calculated as shown in the following formula (2):

the percentage measurement method analyzes the percentage value of the difference value between the maximum value and the minimum value in the parameter mean value, and the calculation mode is shown as the following formula (3):

if the volatility σ is_jIf the value is larger than a certain threshold value epsilon, the user UE is considered to be_iAt a user UE_jSINR prediction under interference is inaccurate, i.e. UE is in current data set condition_iAnd UE_jInter-interference prediction is inaccurate, UE_jAnd UE_iUsers that are not accurate in predicting mutual interference. Wherein the threshold epsilon is a preset system parameter.

According to the above methods and criteria, all users whose interference predictions are inaccurate can be detected.

An intelligent scheduling algorithm (i.e., radio resource enhanced scheduling) targeted at accelerating interference prediction is described below.

(4) Intelligent scheduling algorithm with accelerated interference prediction as target

First, a scheduling objective is analyzed. The C-RAN network architecture has comprehensive control and cognition on all nodes in the whole network, wireless resources are managed in a centralized mode, and users can conduct centralized scheduling. The embodiment of the invention is based on the characteristic of centralized management of the C-RAN, aims to improve the convergence rate of SINR prediction by enriching the data set, performs intelligent scheduling aiming at the condition of inaccurate interference prediction among users after the accuracy analysis of the interference prediction among the users in the network, and dynamically enriches the diversity of the data set so as to improve the convergence rate of the SINR prediction.

Only if there are data samples of the user multiplexed resources with the interfering user, it is possible to predict the interference generated by the interfering user. For the predicted accurate interference users, increasing the multiplexing frequency cannot bring new information. In the intelligent scheduling algorithm of the embodiment of the invention, in order to enrich the effectiveness of the data set, users with inaccurate scheduling prediction need to be pertinently scheduled, so that the reuse times of each user and the users with inaccurate interference prediction in a wireless system are increased, and the effective sample data volume is improved.

Next, the enhanced scheduling policy of the present invention will be described in detail.

The service requirements of users have certain randomness, the service requirements of each user in each TTI are different, the data quantity requested to be transmitted is different, and the TTI system has the advantages thatSo that the number of RBs to be used is different. Therefore, the multiplexing requirement among users is difficult to be satisfied at one time in one TTI, and resource scheduling of multiple TTIs is often required to satisfy the multiplexing requirement. Therefore, it is considered to set a user multiplexing number matrix, which is used to store the multiplexing number required by each user, and record the change of multiplexing requirement in the resource allocation process of consecutive TTIs. If there are n users in the system, the multiplexing requirement matrix is marked as C_n×nAs shown in the following formula (4):

c_i,jrepresenting a user UE_iAnd User Equipment (UE)_jThe frequency of the required multiplexing in between. Obtaining the interference user prediction condition of each user by using the screening scheme described above, if the user UE_iAnd User Equipment (UE)_jInter-interference prediction accuracy matrix C_n×nMiddle corresponding element c_i,jSet to 0, c if prediction is not accurate_i,jSetting a non-zero value c, wherein the non-zero value c is a preset system parameter and belongs to c corresponding to two users under the same base station_i,jIs always 0.

C_n×nThe number of reuses required by user interference pairs in any system is recorded. Although the user reuse number matrix C_n×nThe same multiplexing times c are set for user tuples with inaccurate predictions at first, but the service requirements of users in each TTI are different, so the numerical values in the matrix gradually change first in the scheduling process, and the multiplexing times of the user requirements gradually decrease. Resource scheduling requiring multiple TTIs, C_n×nCan the elements in (1) be reduced to 0 in total. Element c_i,jWhen the number of the UE decreases to 0, the UE is temporarily considered as the user UE_iAnd UE_jThe interference between the two is predicted accurately.

In the embodiment of the invention, the service requirement of the current TTI of each user is represented by the number of RBs required by each user in the network, and the number of RBs required by each user is determined by the size of data transmitted in the TTI and the historical average transmission rate of each RB. With N_iRepresenting a user UE_iNumber of RBs required, UE_iOne RB per allocation then UE_iThe value is decremented by 1. User Equipment (UE)_iRequired number of RBs N_iWhen the value is 0, it indicates that the user UE is a UE_iIn this TTI, the user UE no longer needs radio resources_iThe service requirements are met.

The enhanced scheduling process of the embodiment of the invention comprises three steps of primary allocation, secondary allocation and supplementary allocation. And allocating users for the first time RB by RB, and allocating resources for user tuples with inaccurate interference prediction for the RB by RB. And distributing users for RB by the secondary distribution, and distributing resources for user tuples with accurate interference prediction for RB by RB. And (4) supplementary allocation and user-by-user allocation of RBs, and resource RBs are allocated to all users in the system one by one until the user requirements of the TTI are met.

The goal of the initial allocation is to make the reuse of user tuples with inaccurate interference prediction. And resource allocation is carried out on the RBs one by one, if user tuples with inaccurate predictions exist in the system to apply for wireless resources, the wireless resources are allocated by taking the interference user tuples as units, and the interference user tuples with inaccurate predictions are sequentially allocated for the RBs. Preferentially selecting the user tuples with a large number of user multiplexing requirements, as shown in the following formula (5):

wherein, c_i,jRepresenting a user UE_iAnd UE_jThe required number of times of multiplexing, i and j, are user IDs, respectively.

Subtracting 1 from the value of the corresponding position in the user multiplexing time matrix after each distribution, c_i,j←c_i,j-1. Simultaneous user UE_iAnd UE_jCorresponding RB requirement number N_i、N_jMinus 1, N_i←N_i-1，N_j←N_j-1. When all RBs in the system are used up or the multiplexing requirement between every two users which do not meet the service requirement is zero, the primary allocation is completed. After the initial allocation is completed, there are and only two users multiplexed on the allocated RB.

After the initial allocation is finished, if the RB still has no user in the systemIf the user is used, secondary allocation is carried out, and the aim is to fully utilize radio resources so that users can be used on each RB. After the initial allocation is finished, the number of multiplexing demands between the remaining users which do not meet the service demands in the system is 0. In the secondary allocation process, users are allocated RB by RB, user tuples are randomly allocated on each RB, and the users UE_iThe RB requirement number N corresponding to one RB used more than one_iMinus 1, N_i←N_i-1, until the RB is exhausted or the full user needs are met. So far, the total user traffic demand is met or there are only two users on all RBs. And finally performing supplementary distribution if the service requirement of the user is not met.

And if the users with the service requirements which are not met exist in the system, the service requirements of the users are met by continuously distributing resources to the users one by one. The allocation under consideration of the present scheduling algorithm aims to generate more multiplex numbers, and thus users tend to use RBs that can generate more multiplex numbers. User Equipment (UE) to be allocated_iIn RB_uAnd summing up the number of multiplexing of all the allocated users and the demand thereof, wherein the formula is shown as the following formula (6):

in the above equation, the user UE_jIndicates that a resource block RB has been allocated_uA user of c_i,jRepresenting a user UE_iAnd resource block RB_uUser of (UE) of (2)_jThe number of required reuses.

Simply by

As a UE_iSelecting RB_uThe standard of (2) is easy to make users use some RBs intensively, so that the interference on the RB is too much, and the strength of the interference is difficult to distinguish. Therefore, the sum is divided by the number of allocated users to find an average value as shown in the following equation (7):

wherein l_uIs RB_uThe number of users already on. Allocating resources for user UE on a user-by-user basis_iComputing

Maximum RB_uAs shown in the following equation (8):

if there is at this moment

If greater than 0, RB will be_uAllocation to user UE_iTherefore, the centralized use of a certain RB by the users can be effectively avoided, and the users are distributed on each RB more evenly. If there is at this moment

Then the RB with less users is selected to be allocated to the user UE_iTo avoid users concentrating on a certain RB. Selecting RB_uThe following formula (9):

RB_u＝argmaxl_u (9)

wherein l_uIs RB_uThe number of users already on.

Allocating resources to all users not meeting the demand, user UE_iThe number of RB requirements N corresponding to one more RB_iMinus 1, N_i←N_i1 until the user needs are met. Resource scheduling over multiple TTIs, as matrix C_n×nAnd (4) reducing the middle element to 0, and detecting users with inaccurate interference prediction for each user again. And if the users with inaccurate interference prediction still exist, performing resource scheduling again, and continuously iterating until convergence. Therefore, the algorithm process proposed by the present invention for accelerating convergence of interference prediction includes:

the first step is as follows: performing conventional radio resource scheduling (e.g., proportional fair)Flat, round robin, maximum carrier-to-interference ratio, etc.). The second step is that: and acquiring wireless resource allocation data and corresponding wireless measurement information to form training set data to be learned. The third step: training the training set data t times by using a machine learning algorithm, wherein t is a positive integer of at least 2. The fourth step: and (5) predicting the interference condition by using t models obtained by t times of training to obtain t prediction results, and identifying an inaccurate interference user set of each user prediction based on the volatility of the t prediction results. The fifth step: setting a user multiplexing demand frequency matrix C_n×nElement c corresponding to the user tuple whose prediction is inaccurate_i,jSet to a non-zero value c and the other elements in the matrix set to 0. Wherein c is a preset system parameter. And a sixth step: user scheduling is carried out according to the enhanced scheduling strategy, and each time user multiplexing is added, the matrix C_n×nOf (1) corresponding to non-zero c_i,jSubtract 1 until matrix C_n×nThe values of all the middle elements are 0, so that the training set data is enriched. And then, repeating the second step to the sixth step until the interference user set with inaccurate prediction obtained in the fourth step is an empty set or a new user is added. Resource scheduling over multiple TTIs, as matrix C_n×nAnd (4) reducing the middle element to 0, and detecting users with inaccurate interference prediction for each user again. And if the users with inaccurate interference prediction still exist, performing resource scheduling again, and continuously iterating until convergence.

Wherein, the enhanced scheduling strategy comprises: step 1, resource allocation is carried out on each RB, if user tuples with inaccurate predictions exist in the system to apply for wireless resources, the wireless resources are allocated by taking the interference user tuples as units, and the interference user tuples with inaccurate predictions are sequentially allocated for each RB. Preference of user tuples with a high number of user reuse requirements, e.g. matrix C_n×nMedian maximum element c_i,jThen it is the user UE_iAnd UE_jThe RB is allocated. And if the resource is still required to be applied by the user and the RB is still remained after the allocation is finished, executing the step 2, and if the RB is not remained, executing the step 3. And 2, carrying out resource allocation on the rest RBs in an RB-by-RB sequence, and randomly allocating users to the RBs in the sequence until each RB has at most two RBs. If the current round is dividedAfter the configuration is completed, the user still needs to apply for the resource, and step 3 is executed. And 3, carrying out resource allocation on a user-by-user basis, and randomly selecting users to carry out resource allocation. For the selected user, calculating the average value of the multiplexing times of the existing users of the user on each RB, wherein the average value is the multiplexing times c of each user on the RB and the selected user_i,jAnd dividing the sum by the number of users, and selecting the RB with the largest average value for resource multiplexing. Until the needs of each user are met.

As shown in fig. 8, the apparatus includes: a training set obtaining module, configured to perform conventional radio resource scheduling, obtain radio resource allocation information of each user and radio measurement data of each user, and form a training set of each user based on the radio resource allocation information of each user and the radio measurement data of each user; the training module is used for respectively training the training set of each user for multiple times by utilizing a machine learning algorithm to obtain a plurality of training models of each user; the set determination module is used for predicting the interference situation of each user by utilizing each training model of each user to obtain multiple prediction results, and determining an inaccurate user set of interference prediction of each user based on the fluctuation of the multiple prediction results; and the scheduling module is used for constructing a user tuple containing the user and each user with inaccurate interference prediction in the user interference prediction inaccurate user set aiming at each user, and executing radio resource enhanced scheduling to increase the multiplexing times of the users aiming at the same transmission time interval and the same resource block in the user tuple.

In one embodiment, the data packet signal to interference and noise ratio is measured wirelessly. In one embodiment, the training set obtaining module is configured to generate a first table, where the first table includes: in any transmission time interval, a user scheduling set on each resource block and the signal-to-interference-and-noise ratio of each user in the user scheduling set; generating a training set for each user based on the first table, wherein the training set for each user comprises: and in any transmission time interval, scheduling a user set of the same resource block and the signal-to-interference-and-noise ratio of the user at the same time as the user. In one embodiment, the scheduling module is configured to perform the following steps: a1, resource allocation is carried out resource allocation resource by resource block, if a user tuple exists to apply for wireless resources, the wireless resources are allocated by taking the user tuple as a unit, the user tuple is allocated for each resource block in sequence, wherein, the user tuple with a plurality of times of multiplexing needs of users is preferentially selected, if the resource allocation is finished, the users still need to apply for the resources and the resource blocks still remain, the step A2 is executed, and if the resource blocks do not remain, the step A3 is executed; a2, for the rest resource blocks, resource allocation is carried out in sequence by resource block, and users are randomly allocated to each resource block in sequence until each resource block has at most two users; and A3, carrying out resource allocation by user, randomly selecting users to carry out resource allocation, wherein for the selected users, calculating the average value of the multiplexing requirement times of the users on each resource block, and selecting the resource block with the maximum multiplexing requirement times to carry out resource multiplexing. In one embodiment, the scheduling module is configured to uniformly determine the number of multiplexing requirements as a preset parameter; and determining the multiplexing demand times of each user tuple according to the fluctuation of the multiple prediction results, wherein the tuple with higher fluctuation determines more multiplexing demand times.

The performance of the embodiments of the present invention is evaluated and compared as follows. The simulation scenario of the embodiment of the invention is explained, and the simulation effect of the identification algorithm of the user with inaccurate interference prediction and the scheduling algorithm for accelerating the interference prediction is involved. First, a simulation scenario will be explained.

The 3GPP proposes a low-cost and low-power Femtocell, also called a Femtocell, suitable for indoor environments. The home base station is applied to wireless communication, and the most prominent advantage of the home base station is that a user can arrange indoors and can provide high channel quality indoors, so that high-quality user experience is brought. The simulation of the paper aims at the influence of co-channel interference between the same layers of the home base station. In the deployment of the home base stations, a single-layer 3GPP Dual-Strip model is considered, the width of the corridor is 10 meters, two rows of rooms are respectively arranged on two sides of the corridor, each row has 10 rooms, the area of each room is 10 meters × 10 meters, and one home base station is deployed, and a total of 40 home base stations are randomly accessed to 2 to 4 users. The home base stations and the users are randomly deployed in corresponding rooms, the minimum distance between the home base stations is 8 meters, the minimum distance between the users and the base stations is 2 meters, the system bandwidth of the whole network is 10MHz (including 50 RBs), and the complete system bandwidth resource which can be used by each home base station is 10 MHz. The user location does not change over a period of time, regardless of the user's mobility.

In the simulation, the propagation loss model used is shown in equation (10):

PathLoss＝38.46+20log₁₀d (10)

in the above equation, d is the distance between the user and the femtocell, and is measured in meters.

The fast fading model in the simulation is an ITU Ped A model, and fast fading data are generated when the moving speed is 3 m/s.

All users in the simulation have one service, the service model is a video service, and the data packet sending time and the packet size of the video are assumed to be in accordance with truncated pareto distribution. The truncated pareto distribution probability density function is shown in equation (11) below:

for the packet transmission time, α is 1.2, k is 2.5, and m is 12, that is, the maximum transmission time interval is 12ms, and the minimum transmission time interval is 25 ms.

For the size of the packet, α is 1.2, k is 20, and m is 250, i.e. the maximum value of the packet is 250Bytes, and the minimum value is 20 Bytes.

1. Algorithmic performance analysis for identifying inaccurate users for interference prediction

In the simulation, the real prediction condition can be directly obtained in the simulation, and the recognition capability can be detected by comparing the results obtained by the result screening scheme. Firstly, a method for acquiring a real prediction situation in the system is introduced. There are n users in the system and,user Equipment (UE)_iIs C ═ UE₁,...,UE_j,...UE_nJ, i ≠ j. In the simulation system, the real user SINR value under the single-user interference scene can be directly obtained as a real value, the result obtained by the prediction of the training model is used as a predicted value, and the real SINR value and the predicted SINR value are compared to identify the user UE_iAnd predicting an inaccurate interference user set, and finally obtaining an inaccurate interference user set D. In this particular example, standard deviation is used to characterize the volatility. To user UE_iT times of training are carried out on the sample data to obtain t training models₁,…model_k,…model_tWherein each training model_kSet of users C for which an inaccurate prediction is detected_kThe subscript k denotes the model serial number. Obtaining a plurality of user sets { C₁,...,C_k,...,C_t}，C_kNamely, the set of interference users with inaccurate prediction obtained by the k training of the user. Because the training models obtained each time are different, the interference user sets with inaccurate prediction identified by each training can be distinguished, and the intersection C is calculated for the interference user sets with inaccurate prediction_interAnd union C_union. Intersect C_interRepresenting users predicting inaccurate interfering users no matter how many times the data is trained, represents a necessity because of the sample's own flaw, C_interAll interfering users present in (a) must be predicted inaccurately. C_unionIndicating all users that may not predict accurately, indicating a likelihood, since the sample's defects may not predict accurately.

Comparing the predicted error user set D with the actual error user set C_interAnd user set C_unionAnd verifying the accuracy. Set D and set C_unionCommon users are recorded as D ═ D &' C_union. The prediction accuracy is calculated using the following formula:

presenting interfering users of set D in set C_unionBecomes the prediction accuracy rate, and is used for representing the prediction accuracy degree of the prediction algorithm. Set of handles C_interThe probability of the interfering users in the set D is called the intersection recognition rate, and is used to characterize the recognition capability of the algorithm to the certainty of the occurrence prediction inaccuracy. Set D and set C_unionCommon users are recorded as D ═ D &' C_union. The union recognition rate is described by the following formula:

set of handles C_unionThe probability of the interfering user in the set D is called union recognition rate, and is used to characterize the recognition ability of the algorithm to the inaccuracy and the likelihood of the interference prediction. Set D and set C_interThe intersection is marked as D ═ D # #C_inter. It is described by the following formula:

the simulation duration is fixed at 3000 TTI, training, predicting and identifying are carried out when the data of each user is taken, and the identification accuracy of each user can be clearly seen as the following graph:

the prediction accuracy was 100% for each user, without any false positives, as shown in fig. 3. For each user, the union recognition rate has a certain fluctuation, the average value is 40%, and the recognition rate is high, as shown in fig. 4. For each user, the intersection recognition rate has a certain fluctuation, the average value is 70%, and the recognition rate is very high, as shown in fig. 5. In conclusion, the algorithm has good identification and detection capability for the user with inaccurate prediction

2. Simulation performance analysis for scheduling algorithms

In order to examine the simulation time and the model accuracy variation curve, a neural network algorithm is used for training the training set of each user under different simulation durations, different scheduling algorithms are compared, the variation situation is observed, and the model accuracy is measured by taking an SINR predicted value and a Root Mean Square Error (RMSE) as measurement standards, as shown in the following formula (15):

where N is the number of samples, SINR_realIs the true value of SINR, SINR_predIs a predicted value of SINR.

Analyzing the user SINR prediction under single-user interference, as shown in fig. 6:

it can be seen that, in the scenario that users interfere with each other pairwise, the RMSE of the prediction result is faster than the convergence rate of the conventional scheduling algorithm. The algorithm provided by the embodiment of the invention converges at about 4000 TTI, the conventional scheduling algorithm converges at about 8000 TTI, the convergence rate of the embodiment of the invention is higher than that of the common scheduling algorithm in the single-user interference scene, and finally, the time consumption is reduced by about half to achieve convergence.

Also analyzing the user SINR prediction under multi-user interference, as shown in fig. 7: it can be seen that, in the scenario where users interfere with each other in the real interference scenario, the RMSE of the prediction result in the algorithm proposed in the embodiment of the present invention has a faster convergence rate than that of the conventional scheduling algorithm, and in the real interference scenario, any user may be interfered by signals of users in other interfering base stations. The algorithm provided by the embodiment of the invention converges at about 4000 TTI, and the conventional scheduling algorithm converges at about 6000 TTI, so that the convergence speed of the algorithm provided by the embodiment of the invention is higher than that of the common scheduling algorithm in a multi-user interference scene, and finally, the time consumption is reduced by about 33% to achieve convergence. In summary, the scheduling algorithm provided by the embodiment of the present invention achieves the goal of quickly enriching the data set, so that the SINR prediction result of the user reaches convergence in a short time with a small amount of sample data.

The embodiment of the invention provides a scheduling algorithm for accelerating interference prediction convergence, which has the following characteristics and advantages: according to the user scheduling method for accelerating interference prediction convergence, on the premise that physical equipment deployment and a user complete measurement report are not needed, existing data information in a system is effectively utilized, users with inaccurate predictions are effectively identified and detected through big data analysis and a machine learning algorithm, user scheduling is performed in a targeted manner, the time for interference matrix prediction convergence is shortened, and rapid, accurate and complete interference prediction is achieved.

The embodiment of the invention provides an identification scheme for an interference user with inaccurate prediction. The scheme is based on predictability of SINR by a machine learning algorithm and convergence and randomness of the machine learning algorithm.

Moreover, on the basis of identifying the user tuple with inaccurate interference prediction, the embodiment of the invention provides a user scheduling strategy, so that the prediction convergence of SINR is accelerated, a training model with higher prediction accuracy is obtained quickly, and the interference prediction convergence is accelerated.

Examples of the storage medium for supplying the program code include floppy disks, hard disks, magneto-optical disks, optical disks (e.g., CD-ROMs, CD-R, CD-RWs, DVD-ROMs, DVD-RAMs, DVD-RWs, DVD + RWs), magnetic tapes, nonvolatile memory cards, and ROMs. Alternatively, the program code may be downloaded from a server computer or the cloud by a communication network. "exemplary" means "serving as an example, instance, or illustration" herein, and any illustration, embodiment, or steps described as "exemplary" herein should not be construed as a preferred or advantageous alternative. For the sake of simplicity, the drawings are only schematic representations of the parts relevant to the invention, and do not represent the actual structure of the product. In addition, in order to make the drawings concise and understandable, components having the same structure or function in some of the drawings are only schematically illustrated or only labeled. In this document, "a" does not mean that the number of the relevant portions of the present invention is limited to "only one", and "a" does not mean that the number of the relevant portions of the present invention "more than one" is excluded. In this document, "upper", "lower", "front", "rear", "left", "right", "inner", "outer", and the like are used only to indicate relative positional relationships between relevant portions, and do not limit absolute positions of the relevant portions.

The above-listed detailed description is only a specific description of a possible embodiment of the present invention and is not intended to limit the scope of the present invention, and equivalent embodiments or modifications such as combinations, divisions or repetitions of the features without departing from the technical spirit of the present invention are included in the scope of the present invention.

Claims

1. A method for accelerating convergence of interference predictions for a wireless network, comprising:

for each user, constructing a user tuple containing the user and each user with inaccurate interference prediction in an inaccurate interference prediction user set of the user, and executing radio resource enhanced scheduling to increase the multiplexing times of the users aiming at the same transmission time interval and the same resource block in the user tuple;

wherein the performing radio resource enhanced scheduling to increase the number of user multiplexes for the same transmission time interval and the same resource block in the user tuple comprises:

2. The method of claim 1, wherein the wireless measurement data packet signal-to-interference-and-noise ratio is measured.

3. The method of claim 2, wherein the forming a training set for each user based on the radio resource allocation information of each user and the radio measurement data of each user comprises:

4. The method of accelerating convergence of wireless network interference predictions as set forth in claim 1, further comprising:

and determining the multiplexing demand times of all user tuples according to the fluctuation of the multiple prediction results, wherein more multiplexing demand times are determined for the user tuples with higher fluctuation.

5. An apparatus for accelerating convergence of interference predictions for a wireless network, comprising:

a training set obtaining module, configured to perform conventional radio resource scheduling, obtain radio resource allocation information of each user and radio measurement data of each user, and form a training set of each user based on the radio resource allocation information of each user and the radio measurement data of each user;

the scheduling module is used for constructing a user tuple comprising the user and each user with inaccurate interference prediction in an inaccurate interference prediction user set of the user aiming at each user, and executing radio resource enhanced scheduling to increase the multiplexing times of the users aiming at the same transmission time interval and the same resource block in the user tuple;

6. The apparatus of claim 5, wherein the wireless measurement data packet signal to interference and noise ratio is measured.

7. The apparatus for accelerating convergence of interference predictions for a wireless network of claim 6,

a training set acquisition module configured to generate a first table, where the first table includes: in any transmission time interval, a user scheduling set on each resource block and the signal-to-interference-and-noise ratio of each user in the user scheduling set; generating a training set for each user based on the first table, wherein the training set for each user comprises: and in any transmission time interval, scheduling a user set of the same resource block and the signal-to-interference-and-noise ratio of the user at the same time as the user.

8. The apparatus for accelerating convergence of interference predictions for wireless networks according to claim 5,

the scheduling module is used for uniformly determining the multiplexing demand times of all user tuples as a preset parameter; and determining the multiplexing demand times of all user tuples according to the fluctuation of the multiple prediction results, wherein more multiplexing demand times are determined for the user tuples with higher fluctuation.