User scheduling and simulation beam selection optimization method based on machine learning
阅读说明：本技术 基于机器学习的用户调度和模拟波束选择优化方法 (User scheduling and simulation beam selection optimization method based on machine learning ) 是由 赵赛 邹章晨 唐冬 黄高飞 于 2021-07-28 设计创作，主要内容包括：本发明公开了一种基于机器学习的用户调度和模拟波束选择优化方法,该方法包括下述步骤：获取每个用户信道特征向量；模拟波束匹配：依次输入每个用户信道特征向量至波束预测模型确定模拟波束,利用多个波束分类器将基站和所选用户之间的下行信道划分为多个不同的波束类,利用超平面预测信道的类别并为每个用户选择最佳的模拟波束；经过匹配模拟波束后,判断用户集中所有用户是否完成匹配,当所有用户匹配完成时,根据模拟波束调度集进行信道的调度,其中模拟波束调度集为通过输出用户集对应的模拟波束集合。本发明减少所适用的大型系统所需的计算能力,具有较高的兼容性,减少通信系统搭建的成本,减少在多用户情况下用户匹配信道的时延。(The invention discloses a user scheduling and simulated beam selection optimization method based on machine learning, which comprises the following steps: acquiring a characteristic vector of each user channel; and (3) analog beam matching: sequentially inputting each user channel characteristic vector to a beam prediction model to determine a simulated beam, dividing a downlink channel between a base station and a selected user into a plurality of different beam classes by using a plurality of beam classifiers, predicting the type of the channel by using a hyperplane, and selecting the optimal simulated beam for each user; and after the analog wave beam is matched, judging whether all the users in the user set finish matching, and when all the users finish matching, scheduling the channel according to an analog wave beam scheduling set, wherein the analog wave beam scheduling set is an analog wave beam set corresponding to the output user set. The invention reduces the computing power required by the applicable large-scale system, has higher compatibility, reduces the construction cost of the communication system, and reduces the time delay of the user matching channel under the condition of multiple users.)
1. A user scheduling and simulation beam selection optimization method based on machine learning is characterized by comprising the following steps:
acquiring a characteristic vector of each user channel;
and (3) analog beam matching: sequentially inputting each user channel characteristic vector to a beam prediction model to determine a simulated beam, dividing a downlink channel between a base station and a selected user into a plurality of different beam classes by using a plurality of beam classifiers, predicting the type of the channel by using a hyperplane, and selecting the optimal simulated beam for each user;
the beam prediction model is trained through machine learning, and training data comprise a user channel characteristic vector and a label for identifying a simulation beam index corresponding to the user channel characteristic vector;
and after the analog wave beam is matched, judging whether all the users in the user set finish matching, and when all the users finish matching, scheduling the channel according to an analog wave beam scheduling set, wherein the analog wave beam scheduling set is an analog wave beam set corresponding to the output user set.
2. The method of claim 1, wherein the user channel eigenvector comprises path loss, 2L real-valued characteristics of L complex path gain, and L transmission angles of the base station, where L is the number of channel channels between the base station and the user.
3. The method of claim 1, wherein the beam prediction model is trained by machine learning, and comprises the following steps:
user clustering: dividing K users into user set N by using K-means algorithm according to user channel correlations；
After user clustering is carried out according to the channel correlation threshold, the user with the largest channel energy in one cluster sends a signal of the user, and different analog beams are distributed to the user selected for transmission;
generating channel training samples based on a millimeter wave channel model, normalizing the channel training samples, converting the channel training samples into feature vectors, evaluating effective channel gain from analog beam code words to users according to KPI (Key performance indicator), generating corresponding tags of analog beam indexes for each feature vector, and dividing the feature vectors and the corresponding tags into a training set and a verification set;
let the labels of the M training samples be represented asEach training sample tmCorresponding label r [ m ]]Training a beam multi-classifier by using a support vector machine algorithm;
and verifying the trained beam classifier by using the verification set, and obtaining a beam prediction model after the training is finished when the verification accuracy reaches a preset threshold value.
4. The method according to claim 3, wherein the method for optimizing user scheduling and analog beam selection based on machine learning is characterized in that the method for evaluating the effective channel gain from the analog beam codeword to the user according to the KPI indicator specifically uses a KPI function of the signal-to-noise ratio of user k:
wherein the content of the first and second substances,representing the effective channel gain of the analog beam codeword b to user k, P representsTransmission power at base station, hkRepresenting the channel between the base station and user k,the code word b of the analog wave beam selected by the base station to the user k is represented, and the code word b of the analog wave beam is a codebookMiddle (b) analog beam, σ2The variance is indicated.
5. The method of claim 1, wherein the beam classifier is obtained using a support vector machine algorithm, specifically expressed as:
s.t.yi(WTφ(ti)+b)≥1-ξi,i＝1,2,…,M；
wherein w represents the parameter vector of the hyperplane, C represents the penalty coefficient, and phi represents the input characteristic t of the sampleiKernel function mapped to high dimensional space, B represents threshold of parameter vector of w hyperplane, ξiIndicating an edge error, y, that allows the ith sample to be misclassifiediIndicating a sample label.
6. The method of claim 1, wherein different penalty constant values are applied to the majority class and the minority class for optimization, and the beam classifier hyperplane optimization is represented as:
s.t.yi(WTφ(ti)+b)≥1-ξi,i＝1,2,…,M；
where w represents the parameter vector of the hyperplane and phi represents the input feature t of the sampleiKernel function mapped to high dimensional space, B represents threshold of parameter vector of w hyperplane, ξiIndicating an edge error, y, that allows the ith sample to be misclassifiediRepresents a sample label, C+、C-Penalty constants, I, for positive and negative classes, respectively+、I-Respectively representing a majority and a minority sample set.
7. The method of claim 6, wherein the dual processing is performed on the optimized representation of the beam classifier hyperplane, specifically:
0≤αi≤C+,i∈I+
0≤αj≤C-,j∈I-
wherein alpha isiAnd alphajIs a pair of Lagrangian relaxation variables, i.e. dual variables, K (t)i,tj) Representing a gaussian kernel function.
8. The machine learning-based user scheduling and simulated beam selection optimization method of claim 7, wherein lagrangian relaxation variables are iteratively updated using a sequence minimization algorithm, a pair of lagrangian relaxation variables are updated in each iteration, and the other lagrangian relaxation variables are unchanged.
9. A system for optimizing user scheduling and simulated beam selection based on machine learning, comprising: the device comprises a channel characteristic vector acquisition module, an analog beam matching module and a channel scheduling module;
the channel characteristic vector acquisition module is used for acquiring a channel characteristic vector of each user;
the analog beam matching module is used for simulating beam matching, sequentially inputting each user channel characteristic vector to the beam prediction model to determine analog beams, dividing downlink channels between the base station and the selected users into a plurality of different beam classes by using a plurality of beam classifiers, and selecting the optimal analog beam for each user by using the classes of the hyperplane prediction channels;
the beam prediction model is trained through machine learning, and training data comprise a user channel characteristic vector and a label for identifying a simulation beam index corresponding to the user channel characteristic vector;
the channel scheduling module is used for scheduling channels according to an analog beam scheduling set, wherein the analog beam scheduling set is an analog beam set corresponding to a user set through output;
a base station and a transmitting precoder are also arranged;
the base station is provided with NBSAn antenna and NRFThe base station adopts a full-array mixed structure, each radio frequency chain is connected with a base station antenna through an analog phase shift network, and a Saleh-Vallenzuela channel model is adopted to describe the channel response of the millimeter wave system;
the transmitting precoder comprises an analog precoder and a digital precoder, the analog precoder is realized on a radio frequency chain through a phase shift network, a predefined codebook is adopted, and the digital precoder is applied to baseband digital data;
the effective channel gain for the base station to select the analog beam codeword b to user k is expressed as:
where, P denotes the transmission power at the base station,is the analog beam code word b selected by the base station for the user k, and the analog beam code word b is a codebookMiddle (b) analog beam, σ2The variance is indicated.
10. The system of claim 9, wherein the achievable sum rate for the downlink multi-user MIMO-mmWave based system is represented as:
whereinRepresenting the user k signal-to-interference-plus-noise ratio (SINR),the method specifically comprises the following steps:
whereinIs the interference between the users and is,the method specifically comprises the following steps:
wherein wiRepresents the analog beam selected at the base station side;
maximization and rate are achieved by jointly optimizing user scheduling and analog beam selection, and the method specifically comprises the following steps:
the first constraint condition is:
the second constraint condition is as follows:
the third constraint condition is as follows:
the fourth constraint condition is as follows:
wherein the beam allocation matrix For identifying whether user k is allocated as beam b in the beam allocation matrix.
Technical Field
The invention relates to the technical field of wireless communication, in particular to a user scheduling and analog beam selection optimization method based on machine learning.
Background
Millimeter wave (mmWave) communication technology and large-scale Multiple Input Multiple Output (MIMO) systems are key technologies to cope with explosive growth of data traffic in fifth generation (5G) mobile communication systems. For massive MIMO-mmWave systems, the traditional all-digital beamforming method is hardly applicable in practical applications, because in all-digital beamforming, each antenna is equipped with a Radio Frequency (RF) chain, and each RF chain occupies a dedicated baseband processor, so that all-digital beamforming makes the complexity and power consumption of the system difficult to bear in the case of a large number of antennas. Hybrid beamforming, which divides beamforming into a low-dimensional digital part and a radio frequency analog part, is a low-cost massive MIMO technique. In the RF analog part, each RF is connected to all antennas (a subset of all antennas) through one interface, i.e. a fully connected (partially connected) array structure. The design of analog beamforming has important significance for improving the system performance. The design of analog beamforming is to select the appropriate analog beam for the radio frequency chain based on a predefined codebook.
Generally, in a multi-user system, when the number of users is greater than the number of service resources, user scheduling is required to further improve the spectrum efficiency of the system. At present, the performance of the joint design of user scheduling and beam selection in a lens antenna array multi-user large-scale MIMO system is better; the local optimal solution based on the differential convex function (DC) planning is derived through the combined design of user scheduling and analog beams in the multi-user hybrid mmWave system. However, the local optimal solution is iterative, with the solution highly dependent on the initial iteration values. In addition, the existing beam selection technology has a low-complexity solution based on a greedy method, but when the system scale is large, the calculation complexity is high.
Furthermore, due to the duality of the beam allocation matrix, the base station allocates beams to users, which is a non-convex NP problem, and the global optimization scheme based on exhaustive search has exponential computational complexity, which is unacceptable when the scheduling scale is large. Furthermore, the locally optimal solution based on Successive Convex Approximation (SCA) is iterative, the solution of which is highly dependent on the initial iteration values. Therefore, it is necessary to provide a method for selecting a beam that achieves better performance and lower complexity when the system is large in scale.
Disclosure of Invention
In order to overcome the defects and shortcomings of the prior art, the invention provides a user scheduling and analog beam selection optimization method based on machine learning, which realizes the maximum system reach and speed under the user scheduling constraint, the analog beam selection constraint and the resource capacity constraint by combining the user scheduling and the analog beam selection, reduces the calculation capacity required by a large-scale system, has higher compatibility, reduces the construction cost of a communication system, and reduces the time delay of a user matching channel under the condition of multiple users.
In order to achieve the purpose, the invention adopts the following technical scheme:
the invention provides a user scheduling and simulated beam selection optimization method based on machine learning, which comprises the following steps:
acquiring a characteristic vector of each user channel;
and (3) analog beam matching: sequentially inputting each user channel characteristic vector to a beam prediction model to determine a simulated beam, dividing a downlink channel between a base station and a selected user into a plurality of different beam classes by using a plurality of beam classifiers, predicting the type of the channel by using a hyperplane, and selecting the optimal simulated beam for each user;
the beam prediction model is trained through machine learning, and training data comprise a user channel characteristic vector and a label for identifying a simulation beam index corresponding to the user channel characteristic vector;
and after the analog wave beam is matched, judging whether all the users in the user set finish matching, and when all the users finish matching, scheduling the channel according to an analog wave beam scheduling set, wherein the analog wave beam scheduling set is an analog wave beam set corresponding to the output user set.
As a preferred technical solution, the user channel feature vector includes path loss, 2L real-valued features of L complex path gain, and L transmission angles of the base station, where L is the number of channel channels between the base station and the user.
As a preferred technical solution, the beam prediction model is trained by machine learning, and the specific steps include:
user clustering: dividing K users into user set N by using K-means algorithm according to user channel correlations；
After user clustering is carried out according to the channel correlation threshold, the user with the largest channel energy in one cluster sends a signal of the user, and different analog beams are distributed to the user selected for transmission;
generating channel training samples based on a millimeter wave channel model, normalizing the channel training samples, converting the channel training samples into feature vectors, evaluating effective channel gain from analog beam code words to users according to KPI (Key performance indicator), generating corresponding tags of analog beam indexes for each feature vector, and dividing the feature vectors and the corresponding tags into a training set and a verification set;
let the labels of the M training samples be represented asEach training sample tmCorresponding label r [ m ]]Training a beam multi-classifier by using a support vector machine algorithm;
and verifying the trained beam classifier by using the verification set, and obtaining a beam prediction model after the training is finished when the verification accuracy reaches a preset threshold value.
As a preferred technical solution, the method for evaluating the effective channel gain from the analog beam codeword to the user according to the KPI indicator specifically adopts a KPI function of the signal-to-noise ratio of the user k:
wherein the content of the first and second substances,representing the effective channel gain of the analog beam codeword b to user k, P represents the transmission power at the base station, hkRepresenting the channel between the base station and user k,the code word b of the analog wave beam selected by the base station to the user k is represented, and the code word b of the analog wave beam is a codebookMiddle (b) analog beam, σ2The variance is indicated.
As a preferred technical solution, the beam classifier is obtained by using a support vector machine algorithm, and is specifically represented as:
s.t.yi(WTφ(ti)+b)≥1-ξi,i＝1,2,…,M；
wherein w represents the parameter vector of the hyperplane, C represents the penalty coefficient, and phi represents the input characteristic t of the sampleiKernel function mapped to high dimensional space, B represents threshold of parameter vector of w hyperplane, ξiIndicating an edge error, y, that allows the ith sample to be misclassifiediIndicating a sample label.
As a preferred technical scheme, different penalty constant values are respectively adopted for the majority class and the minority class for optimization processing, and the optimization expression of the hyperplane of the beam classifier is as follows:
s.t.yi(WTφ(ti)+b)≥1-ξi,i＝1,2,…,M；
where w represents the parameter vector of the hyperplane and phi represents the input feature t of the sampleiKernel function mapped to high dimensional space, B represents threshold of parameter vector of w hyperplane, ξiIndicating an edge error, y, that allows the ith sample to be misclassifiediRepresents a sample label, C+、C_Penalty constants, I, for positive and negative classes, respectively+、I_Respectively representing a majority and a minority sample set.
As a preferred technical solution, the dual processing is performed on the optimized representation of the hyperplane of the beam classifier, specifically:
0≤αi≤C+,I∈I+
0≤αj≤C_,J∈I_
wherein alpha isiAnd alphajIs a pair of Lagrangian relaxation variables, i.e. dual variables, K (t)i,tj) Representing a gaussian kernel function.
As an optimal technical scheme, a sequence minimization algorithm is used for updating Lagrange relaxation variables in an iteration mode, a pair of Lagrange relaxation variables is updated in each iteration mode, and other Lagrange relaxation variables are unchanged.
The invention provides a user scheduling and simulated beam selection optimizing system based on machine learning, which comprises: the device comprises a channel characteristic vector acquisition module, an analog beam matching module and a channel scheduling module;
the channel characteristic vector acquisition module is used for acquiring a channel characteristic vector of each user;
the analog beam matching module is used for simulating beam matching, sequentially inputting each user channel characteristic vector to the beam prediction model to determine analog beams, dividing downlink channels between the base station and the selected users into a plurality of different beam classes by using a plurality of beam classifiers, and selecting the optimal analog beam for each user by using the classes of the hyperplane prediction channels;
the beam prediction model is trained through machine learning, and training data comprise a user channel characteristic vector and a label for identifying a simulation beam index corresponding to the user channel characteristic vector;
the channel scheduling module is used for scheduling channels according to an analog beam scheduling set, wherein the analog beam scheduling set is an analog beam set corresponding to a user set through output;
a base station and a transmitting precoder are also arranged;
the base station is provided with NBSAn antenna and NRFThe base station adopts a full-array mixed structure, each radio frequency chain is connected with a base station antenna through an analog phase shift network, and a Saleh-Vallenzuela channel model is adopted to describe the channel response of the millimeter wave system;
the transmitting precoder comprises an analog precoder and a digital precoder, the analog precoder is realized on a radio frequency chain through a phase shift network, a predefined codebook is adopted, and the digital precoder is applied to baseband digital data;
the effective channel gain for the base station to select the analog beam codeword b to user k is expressed as:
where, P denotes the transmission power at the base station,is the analog beam code word b selected by the base station for the user k, and the analog beam code word b is a codebookSimulation of the middle bBeam, σ2The variance is indicated.
As a preferred technical solution, the achievable sum rate based on the downlink multi-user MIMO-mmWave system is represented as:
whereinRepresenting the user k signal-to-interference-plus-noise ratio (SINR),the method specifically comprises the following steps:
whereinIs the interference between the users and is,the method specifically comprises the following steps:
wherein wiRepresents the analog beam selected at the base station side;
maximization and rate are achieved by jointly optimizing user scheduling and analog beam selection, and the method specifically comprises the following steps:
the first constraint condition is:
the second constraint condition is as follows:
the third constraint condition is as follows:
the fourth constraint condition is as follows:
wherein the beam allocation matrix For identifying whether user k is allocated as beam b in the beam allocation matrix.
Compared with the prior art, the invention has the following advantages and beneficial effects:
(1) the invention realizes the maximum system reach and speed under the user scheduling constraint, the simulated beam selection constraint and the resource capacity constraint by combining the user scheduling and the simulated beam selection, reduces the computing power required by the applicable large-scale system, has higher compatibility, reduces the construction cost of the communication system, and reduces the time delay of the user matched channel under the condition of multiple users.
(2) According to the invention, different punishment constant values are respectively adopted for the majority classes and the minority classes during the training of the beam prediction model, the optimization processing still has higher performance under the condition of unbalanced data set, the dual processing is carried out on the optimized representation of the hyperplane of the beam classifier, the Lagrange relaxation variable is iteratively updated by using the sequence minimization algorithm, the precision of the hyperplane is improved, and the classification accuracy of the beam classifier is further improved.
(3) The invention adopts the K-means algorithm to cluster the users according to the channel correlation, solves the problem that the data set of the users and the beam mapping in the training process of the beam prediction model is overlarge, and achieves higher prediction accuracy while reducing the number of training sets by clustering the users.
Drawings
Fig. 1 is a schematic flowchart of a user scheduling and simulated beam selection optimization method based on machine learning in embodiment 1 of the present invention;
fig. 2 is a schematic flow chart of analog beam selection in embodiment 1 of the present invention;
FIG. 3 is a schematic flow chart illustrating clustering of user sets based on K-means in embodiment 1 of the present invention;
fig. 4 is a schematic diagram illustrating a comparison of the system reachability in a user scheduling and simulated beam selection optimization method based on machine learning in embodiment 3 of the present invention;
fig. 5 is a schematic diagram illustrating a comparison between a computational complexity and a user number relationship in a user scheduling and simulated beam selection optimization method based on machine learning in embodiment 3 of the present invention;
fig. 6 is a schematic diagram illustrating the comparison of the effectiveness of the user scheduling and analog beam selection optimization method based on machine learning in embodiment 3 of the present invention under the SNR of 5 dB;
fig. 7 is a schematic diagram comparing average sum rate with the number of users under SNR of 15dB by the optimization method for scheduling users and selecting analog beams based on machine learning in embodiment 3 of the present invention.
Detailed Description
In the description of the present disclosure, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance. Also, the use of the terms "a," "an," or "the" and similar referents do not denote a limitation of quantity, but rather denote the presence of at least one. The word "comprising" or "comprises", and the like, means that the element or item appearing before the word covers the element or item listed after the word and its equivalents, but does not exclude other elements or items. The terms "connected" or "coupled" and the like are not restricted to physical or mechanical connections, but may include electrical connections, whether direct or indirect.
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Example 1
As shown in fig. 1, the present embodiment provides a method for optimizing user scheduling and analog beam selection based on machine learning, which includes the following steps:
a user channel characteristic vector obtaining step: sequentially acquiring each user channel characteristic vector from a user set to be processed, wherein the user channel characteristic vector comprises path loss, 2L real-value characteristics of L complex path gain and L emission angles of a base station, and L is the number of channel channels between the base station and the user;
and (3) simulating beam matching: the method comprises the steps of sequentially inputting a channel characteristic vector of each user to a beam prediction model to determine a simulated beam, specifically, dividing a downlink channel between a base station and a selected user into a plurality of different beam classes by using a plurality of beam classifiers, predicting the channel class by using a hyperplane and selecting the optimal simulated beam for each user, wherein the beam prediction model is obtained by using a plurality of groups of data through machine learning training, and each group of data in the plurality of groups of data comprises the channel characteristic vector of the user and a label for identifying the simulated beam index corresponding to the channel characteristic vector of the user. In practical application, each class corresponds to a different beam vector in the simulated beam codebook, each sample is mapped to an n-dimensional space, the coordinate of each sample is a feature value, two different channels are separated by finding a hyperplane, and when the output of the beam classifier on the user channel feature vector is greater than 0, the best simulated beam matched by the beam classifier for the user is recorded, as shown in fig. 2 specifically.
A scheduling step: and after the analog wave beams are matched, judging whether all users in the user set finish matching, and when all users finish matching, scheduling channels according to an analog wave beam scheduling set, wherein the analog wave beam scheduling set is formed by outputting analog wave beams corresponding to the user set.
In this embodiment, the beam prediction model is obtained by machine learning training using multiple sets of data, and the specific steps include:
user clustering step: dividing K users into user set N according to channel correlation of userss. Dividing K users into user set N by using K-means algorithm according to user channel correlations。
When in actual use, orderAndrespectively expressed as cluster set and cluster center set, let the nth cluster center cnAnd nth cluster SnThe channel correlation between users in (1) is:
in this embodiment, the vector h of the channeliAnd hjTo illustrate the principle of channel correlation, vector hiAnd hjIs inversely proportional to the orthogonality between them. For theThe channel correlation between users can be through hiAnd hjIs measured by the correlation ofiAnd hjThe correlation of (a) is specifically a cosine correlation, i.e.:
wherein corr (h)i,hj) The smaller the correlation between the two channel vectors.
The K-means algorithm is used to maximize the sum of the channel correlations for all clusters, i.e.:
the users with higher channel correlation are divided into the same subset, and after clustering the users according to the channel correlation threshold value, the channel energy | h in one clusterk‖2The largest user will have its signal transmitted. Different analog beams are assigned to the selected users for simultaneous downlink transmission. In addition, those skilled in the art may set the channel correlation threshold value according to actual situations for division, which is not limited herein.
Training data preprocessing step: generating channel training samples based on a millimeter wave channel model, normalizing the channel training samples, converting the channel training samples into feature vectors, evaluating effective channel gain from analog beam code words to users according to KPI indexes, generating corresponding tags of analog beam indexes for each feature vector, and dividing the feature vectors and the corresponding tags into a training set and a verification set. In practical application, M channel training samples are generated, and each channel training sample is described by using 3L +1 real features, specifically including path loss, 2L real-valued features of L complex path gain, and L transmission angles of a base station.
Normalizing the real features of each channel training sample:
whereinDenotes the ith feature of the mth sample,represents the maximum of the i-th feature in the m samples,represents the minimum of the i-th feature in the m samples,represents the average of the i-th feature in m samples,the normalized result of the i-th feature of the m-th sample is represented. In the embodiment, the adverse effect on the training process caused by too large or too small training sample values is avoided by normalizing the real features.
After normalization, each sample is converted into a feature vector tm，m∈{1,…,M}。
And evaluating the effective channel gain from the analog beam code word to the user by using a KPI index, and specifically adopting a KPI function of the signal-to-noise ratio of the user k:
in the formulaRepresenting the effective channel gain of the analog beam codeword b to user k, P represents the transmission power at the base station, hkRepresenting the channel between the base station and user k,the code word b of the analog wave beam selected by the base station to the user k is represented, and the code word b of the analog wave beam is a codebookMiddle (b) analog beam, σ2The variance is indicated.
In this embodiment, the related indexes specifically include objective indexes such as signal-to-noise ratio and channel capacity.In practical application, there are analog beam codebooksA number of different analog beams are generated,to representRadix of, analog beam index wnThe label information is the label information that maximizes the KPI function for the current channel sample.
Training: let the labels of the M training samples be represented asEach training sample tmCorresponding label r [ m ]]A one-to-many classifier is trained using a support vector machine algorithm (SVM). In the present embodiment, forClass, respectively trainingThe classifier takes the channel feature vector as input and the corresponding label as output, the trained one-to-many classifier classifies the feature vector into preset categories or other categories, namely, the | W | one-to-many classifiers respectively correspond to the | W | preset categories, for the k one-to-many classifier, k belongs to {1,2, …, | W | }, when y is equal to yiWhen +1, sample r [ i]When y is equal to kiWhen-1, sample r [ i]≠k。
In practical application, a support vector machine algorithm is used for obtaining the hyperplane representation of the kth beam classifier:
s.t.yi(WTφ(ti)+b)≥1-ξi,i＝1,2,…,M；
wherein w represents the parameter vector of the hyperplane, C represents the penalty coefficient, phi represents the input characteristic t of the sampleiKernel function mapped to high dimensional space, B represents threshold of parameter vector of w hyperplane, ξiIndicating an edge error, y, that allows the ith sample to be misclassifiediIndicating a sample label.
A verification step: and verifying the trained beam classifier by using the verification set, obtaining a beam prediction model after the training is finished when the verification accuracy reaches a preset threshold, and otherwise, repeatedly executing the training step.
As shown in fig. 3, the user clustering step specifically includes:
selecting a clustering center: inputting user channel characteristics and classification category numbers, and randomly selecting a first clustering center;
clustering: performing clustering processing based on a K-means clustering algorithm, calculating a second clustering center, judging whether the clustering center is changed, namely judging whether the first clustering center is different from the second clustering center, if so, recalculating a third clustering center and executing a clustering step;
a user set generating step: and selecting a user with the largest channel energy from each class to form a user set.
In this embodiment, the training step further includes:
and respectively adopting different penalty constant values for the majority class and the minority class to carry out optimization processing. In practical application, the optimization of the hyperplane of the beam classifier is represented as follows:
s.t.yi(WTφ(ti)+b)≥1-ξi,i＝1,2,…,M；
wherein C is+、C_Penalty constants, I, for positive and negative classes, respectively+、I_Respectively representing a majority and a minority sample set. In practice, in the multi-classification problem, the label wnThe number of samples in the method is far less than that of all other tags, and different penalty constant values are adopted for optimization processing, so that the support vector machine of the embodiment can achieve satisfactory performance under the condition of data set imbalance.
Carrying out dual processing on the optimized representation of the hyperplane of the beam classifier, and iteratively updating Lagrange relaxation variables by using a sequence minimization algorithm (SMO);
in actual application, parameter C+、C_And selecting a large penalty constant for the positive sample and a small penalty constant for the negative sample, thereby obtaining the hyperplane with higher precision. Wherein a penalty constant formula is specifically adopted for selection C+、C_：
In this embodiment, the dual processing on the optimized representation of the hyperplane of the beam classifier specifically includes:
0≤αi≤C+,i∈I+
0≤αj≤C-,j∈I-
wherein alpha isiAnd alphajIs a pair of Lagrangian relaxation variables, i.e. dual variables, K (t)i,tj) Representing a Gaussian kernel function, in particularσ2Parameters representing a gaussian kernel function;
in the present embodiment, σ is obtained by a cross-validation method. In practical applications, σ has an important influence on the accuracy of the hyperplane, and σ determines the flexibility of the beam classifier in fitting data.
In this embodiment, a sequence minimization algorithm (SMO) is used to iteratively update lagrangian relaxation variables, which specifically includes: updating a pair of Lagrangian relaxation variables (α) in each iterationi,αj) Other Lagrangian relaxation variables, { α }nN ∈ {1, …, M }, n ≠ { i, j } } is invariant. In practice, according to a sequence minimization algorithm, αjThe update process of (2) is as follows:
wherein the values of L and H are:
in particular when y isiAnd yjSatisfy { yi＝1,yj＝1}，{yi＝-1,yj＝-1}，{yi＝1,yj1 and yi＝-1,yj1} is a value.
Updating alpha with an update formulaiThe update formula specifically includes:
deriving alpha from the updated formulaiThe update values of (a) are:
the beam classifier of this embodiment uses a Gaussian kernel function with a dual variable αiThe classification is performed, let the nth classifier be represented as:
wherein t iskRepresenting the newly input feature vector to be classified, viRepresents a support vector, when Jn(tk) When > 0, select wnAs the best analog beam serving the kth user.
Furthermore, a vector set V is supportednBy a plurality of support vectors viAnd (4) forming.
Example 2
The embodiment also provides a system for optimizing user scheduling and analog beam selection based on machine learning, which includes: the device comprises a channel characteristic vector acquisition module, an analog beam matching module and a channel scheduling module;
in this embodiment, the channel feature vector obtaining module is configured to obtain a channel feature vector of each user;
in this embodiment, the analog beam matching module is configured to simulate beam matching, sequentially input a channel feature vector of each user to the beam prediction model to determine a simulated beam, divide a downlink channel between the base station and the selected user into a plurality of different beam classes by using a plurality of beam classifiers, predict a channel class by using a hyperplane, and select an optimal simulated beam for each user;
in this embodiment, the beam prediction model is trained through machine learning, and the training data includes a user channel feature vector and a tag identifying a simulated beam index corresponding to the user channel feature vector;
in this embodiment, the channel scheduling module is configured to perform channel scheduling according to an analog beam scheduling set, where the analog beam scheduling set is an analog beam set corresponding to a user set through output;
the present embodiment takes a downlink multi-user MIMO-mmWave system as an example to illustrate the principle of analog beam scheduling, and the downlink multi-user MIMO-mmWave system is provided with a Base Station (BS) which serves K users in its working range. Base station is equipped with NBSAn antenna and NRFA Radio Frequency (RF) chain, a base station transmits N to a userSA data stream ofIn NS≤NRF。
In this embodiment, the base station adopts a full-array hybrid structure, and each radio frequency chain is connected to the base station antenna through an analog phase shift network. Each user is configured with an antenna, and the number of users is larger than the system capacity, i.e. K > NRF. The present embodiment describes the channel response of the millimeter wave system using the Saleh-Valenzuela channel model. In addition, the structural configuration of the base station and the description of the channel response of the millimeter wave system are not limited herein, and those skilled in the art may make changes according to the actual situation.
Let the channel between the base station and user k be hk，And h iskConsisting of L finite scattering paths, hkThe concrete expression is as follows:
wherein alpha isk,mIs the complex gain coefficient, ρ, of user k to the mth pathkIs the path loss between the base station and user k, phik,mRepresents the launch angle (AoD), a, of the mth path at user kBS(φk,m)HRepresenting the transmit antenna array response vector of a Uniform Line Array (ULA) and H representing the conjugate transpose.
In this embodiment, aBs(φk,m) The method specifically comprises the following steps:
where d denotes the spacing of the antennas and λ denotes the signal wavelength.
In a hybrid millimeter wave system, the transmission precoder includes an analog precoder and a digital precoder, specifically:
W＝WaWd；
wherein the precoder is simulatedDigital pre-encoder
Analog precoder WaDigital precoder W implemented on a radio frequency chain through a phase shift networkdApplied to baseband digital data. The received signal for user k is represented as:
in which the signal is transmittedAnd isnkIs additive complex Gaussian noise, specifically nk～N(0,σ2)。
Analog precoder WaUsing predefined codebooks
For a digital precoder, the present embodiment considers the baseband precoder WdIs an identity matrix, the effective channel gain from the base station selecting the analog beam codeword b to user k is expressed as:
where P denotes the transmission power at the base station,is the analog beam code word b selected by the base station for the user k, and the analog beam code word b is a codebookMiddle (b) analog beam, σ2The variance is indicated.
In this embodiment, the achievable sum rate of the mm-wave user downlink MIMO system is represented as:
whereinRepresenting the user k signal-to-interference-plus-noise ratio (SINR),the method specifically comprises the following steps:
whereinIs the interference between the users and is,the method specifically comprises the following steps:
wherein wiRepresenting the analog beam selected at the base station side.
The optimization of this embodiment aims to maximize and rate by jointly optimizing user scheduling and analog beam selection, and specifically includes:
in addition, four constraints are required to be satisfied:
the first constraint condition is:
the second constraint condition is as follows:
the third constraint condition is as follows:
the fourth constraint condition is as follows:
wherein the beam allocation matrix For identifying whether user k is allocated as beam b in the beam allocation matrix.
In the present embodiment, the first constraint condition representsIs binary, and in particular, if beam b is assigned to user k,otherwiseThe second constraint is to ensure that only one beam at most can be allocated per user. A third constraint is used to ensure that each beam can be allocated to at most one user. The fourth constraint represents the number of users from K toWith a maximum of N analog beamsRFNon-overlapping allocations.
By combining the optimization target and the four constraint conditions of the embodiment, it can be seen that the duality of the beam allocation matrix and the problem that the base station is not convex NP-hard to allocate beams to the user, and this embodiment adopts the optimization technical scheme of embodiment 1 on the basis of the above scheme to reduce the computational complexity allocated to the user beams during the scheduling of the base station and reduce the waiting time delay when the user matches the beams.
Example 3
In this embodiment, a comparison experiment on computational complexity is performed by using the optimization scheme of example 1, and the control scheme specifically includes a global optimal scheme (ES) based on exhaustive search, a local optimal scheme based on differential convex (D.c.) planning, and a latest low-complexity scheme based on a greedy method.
For the machine learning based user scheduling and simulated beam selection optimization method in embodiment 1:
the complexity of the user classification is P (T)ucKNRFNBS) Wherein T isucIs the maximum number of iterations, TucK is smaller than K. Since the training phase of the SVM algorithm is offline, only the complexity of online classification needs to be considered. The complexity of the beam selection of the support vector machine algorithm isWhereinTo representThe base number of (c) is,a set of support vectors representing the svm classifier.
The complexity estimate for this scheme is therefore:
global optimality for exhaustive search:
calculating a feasible set of all users and beam pairs, wherein the calculation complexity is as follows:
whereinIs of the general form:
for the local optimal solution of differential convex programming:
according to the local optimal algorithm based on differential convex programming, the convex optimization problem of each iteration isOptimizing variables andconvex linear constraints, so the computational complexity per iteration isLet the maximum number of iterations be TdThen the computational complexity of the scheme is:
for the latest lowest complexity scheme based on greedy approach:
let the search space dimension of the greedy scheme beThe computational complexity of the scheme is as follows:
combining the calculation complexity of each scheme, and comparing the calculation complexity and the calculation complexity to exhaustively search the complexity CLT of the methodeComplexity CLT of highest user scheduling and simulation beam selection optimization method based on machine learningmThe lowest, namely:
CLTe＞CLTd＞CLTg＞CLTm；
when N is presentBSK andwhen the scheduling time is larger, namely when the scheduling condition with larger capacity is faced, the user scheduling and analog beam selection optimization method based on machine learning has the better effect of reducing the computational complexity and has more advantages on reducing the scheduling time delay.
In this embodiment, the path loss of the kth user isWhere beta is the path loss exponent, DkRepresenting the distance between the base station and user k. In the present embodiment, β is set to 3.76, DkAdopting random variable values which are uniformly distributed between 10-15, and additionally setting NRF＝4，NBS16, K10, for the millimeter wave channel,λ＝5mm，φk,mevenly distributed between 0 and 2 pi. In addition, the number of analog beam codebook at base station end is setThe beam classifier in the beam prediction model adopts SVM classifier and uses 5000 training samples in the experimentIn the generation, the simulation result is realized by adopting montearlo simulation and 1000 channels.
In the present embodiment, an exhaustive search scheme (denoted as "ES"), a d.c. based scheme (denoted as "d.c."), a Greedy approach scheme (denoted as "Greedy"), a scheme based on machine learning user scheduling and analog beam selection optimization approach (denoted as "ML"), and a scheme in which the user randomly selects an analog beam (denoted as "naive").
As shown in fig. 4, for the variation of the average sum rate at different signal-to-noise ratios, the results show that the performance of all six schemes improves as the signal-to-noise ratio increases. The "ES" scheme has the best performance, the "Naive" scheme has the worst performance, and the "Greedy" scheme has better performance than the "ML" scheme. Meanwhile, with the increase of the signal-to-noise ratio, the influence of interference among users is weakened, the characteristics of different channels are more and more obvious, and the performance of the SVM classifier is better and better, so that the gap between the Greedy scheme and the ML scheme in the embodiment 1 is gradually reduced.
As shown in fig. 5, the comparison in complexity with the number of users K for the different schemes is where the SNR is set to 5 dB. The results indicate that the "ML" scheme is less complex than the "Greedy" scheme. Furthermore, the complexity of the "ML" scheme grows slowly with increasing K.
As shown in fig. 6, for the comparison of effectiveness at SNR of 5dB for the different schemes, the result shows that comparing the Cumulative Distribution Function (CDF), the CDF curve of the "ML" scheme is very close to the CDF curve of the "Greedy" scheme, i.e., the "ML" scheme has an effectiveness close to the "Greedy" scheme at SNR of 5 dB.
As shown in fig. 7, the average sum rate comparison with the number of users K for the different schemes, where SNP is set 15 dB. The results show that the average sum rate of the "ML" and "Greedy" schemes increases as K grows, and thus the "ML" scheme has greater potential in dealing with large-scale systems.
Example 4
This embodiment provides a massive MIMO system, which matches beams to users in a working range by using the user scheduling and analog beam selection optimization method based on machine learning in embodiment 1.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.