The content of the invention
In view of the above-mentioned deficiencies in the prior art, it is an object of the present invention to provide a kind of mimo system joint based on deep learning
Precoding and antenna selecting method, the inventive method consider NP-hard problems and traversal in mimo system day line options and searched
The problem of rope computation complexity is too high, it is ensured that realize quick day line selection in the case where obtaining better systems signal to noise ratio
Select.
The purpose of the present invention is achieved through the following technical solutions:A kind of mimo system joint based on deep learning
Precoding and antenna selecting method, comprise the following steps:
(1):The training dataset required for deep neural network, the training are obtained using existing antenna selecting method
Data set includes two parts:Input (input) data set for channel matrix (H) to gather, output (output) data set is antenna
Selection set;Obtain maximum transmission power P;
(2):Initialize deep neural network parameter:Weight w, biasing b, learning rate L, maximum frequency of training I, processing data
Size c, each layer neuron number;
(3):Deep learning model is set up, deep learning model is trained using the training dataset obtained in step 1 and protects
Deposit;
(4):Channel estimation is carried out using pilot frequency system and obtains channel matrix H, utilizes the deep learning mould preserved in step 3
Type, obtains antenna selection set I;
(5):Antenna selection set I is obtained according to step 4, forming corresponding MIMO subsystems progress precoding processing is
HI=H (:, I), it is rightEigenvalues Decomposition is carried out, it is v to make the corresponding characteristic vector of eigenvalue of maximum, is takenAs prelisting
Code vector.
Compared with prior art, beneficial effect of the present invention:The present invention is produced deep by existing antenna selecting method first
The training dataset spent required for study;Then, deep learning model is set up, deep learning model is trained simultaneously using training data
Preserve;Then day line options are completed using the deep learning model preserved;Finally selected MIMO subsystems are carried out optimal
Precoding Design.The present invention designs the precoding of mimo system joint and day line options using depth learning technology, can obtain
Relatively low computation complexity is realized in the case of better systems signal to noise ratio.
Embodiment
In order that the purpose of the present invention and effect are clearer, below in conjunction with the accompanying drawings to the specific embodiment party of the inventive method
Formula is described in detail.
As shown in Figure 1, it is considered to multiple-input and multiple-output (MIMO) system, the system signal includes N number of transmitting antenna from one
Base station (BS) be transferred to a mobile terminal for having M reception antenna, it is assumed that any letter between transmitting antenna and reception antenna
Road is flat fading, then M × N baseband channels complex matrix is denoted as
G=[g1 g2 ... gM]H (1)
WhereinIt is a plural row vector for including channel coefficients between N number of transmitting antenna and m-th of reception antenna, m=
1,2 ..., M.Receiver has channel matrix G information.
Feedback based on receiver, emitter selection K (from N number of) individual transmitting antenna is simultaneously carried out using total power constraint
Precoding, and transmit signalWe equivalently it can use a precoding vector v ∈ VN, submit to restraint | | v
||0=K, wherein
VN={ v ∈ CN:||v||≤1} (2)
The vector of M × 1 that the pulse matching filtering of so down coversion is received is:
WhereinIt is the coloured multiple noise vector of a zero-mean additive, its autocorrelation matrix is denoted as R.Due to y generations
Unknown signaling vector of one, the table under coloured noise interference, it is exactly Minimum Mean Square Error (MMSE) filter to maximize signal to noise ratio wave filter
Ripple device
It is output as
So output signal-to-noise ratio is exactly
WhereinBe M × N conversion after channel matrix.Equation (6) shows prior weight and precoding
The relation of vector.
Our target is K antenna for base station of selection and optimizes precoding vector v to maximize prior weight (6).
It is exactly that we find the solution v of following problem
Due to existing in problem (7) | | v | |0=K bound terms, it is a NP-hard problem to cause problem (7).Solving should
Problem needs to rely on traversal search method, otherwise can not obtain globally optimal solution.And traversal search method can cause the calculating time
Complexity is difficult to calculate optimal antenna selecting plan within coherence time with the increase of antenna amount index in MIMO.
Therefore, the present invention solves antenna selection problem using deep learning (DL).
Fig. 2 gives carries out mimo system day line options flow chart using deep learning.Specifically, it can be described as follows:
A kind of precoding of mimo system joint and antenna selecting method based on deep learning, comprise the following steps:
(1):The training dataset required for deep neural network, the training are obtained using existing antenna selecting method
Data set includes two parts:Input (input) data set for channel matrix (H) to gather, output (output) data set is antenna
Selection set;Obtain maximum transmission power P;
(2):Initialize deep neural network parameter:Weight w, biasing b, learning rate L, maximum frequency of training I, processing data
Size c, each layer neuron number;
(3):Deep learning model is set up, deep learning model is trained using the training dataset obtained in step 1 and protects
Deposit;
(4):Channel estimation is carried out using pilot frequency system and obtains channel matrix H, utilizes the deep learning mould preserved in step 3
Type, obtains antenna selection set I;
(5):Antenna selection set I is obtained according to step 4, forming corresponding MIMO subsystems progress precoding processing is
HI=H (:, I), it is rightEigenvalues Decomposition is carried out, it is v to make the corresponding characteristic vector of eigenvalue of maximum, is takenAs prelisting
Code vector.
Technical scheme is further elaborated below by instantiation.Existing antenna selecting method has
Auxiliary angle algorithm, force search and sparse Principal Component Analysis Algorithm etc., in an experiment, needed for we are obtained using auxiliary angle algorithm
Training dataset.In addition, in an experiment, we are using 8-2 (8 transmitting antennas, 2 reception antennas and selection transmitting day
Line number amount is equal to reception antenna quantity), 16-2,32-2,64-2 carries out method validation.Specifically, using following experiment parameter:
1. maximum frequency of training I spans are 50000-200000, and learning rate L spans are 0.001-
0.000001, it is 50-2000 that data batch size c spans are read every time.
2. each layer biasing b is initialized as 0.1, and each layer weight w is initialized as cutting gearbox, and wherein standard deviation is(NinFor this layer of input number of nodes).
3. deep learning uses five layer networks:Input layer, three hidden layers, output layers.Batch regularization is carried out between each layer
(batch normalization), node in hidden layer is respectively 4N, 2N, 2N (N is transmitting antenna number).
4. linear unit (relu) function is corrected in activation primitive selection, and output layer selection relu6 functions, loss function is used
Mean square deviation function, majorized function selection stochastic gradient descent function.
Fig. 3 sets forth deep learning (DL) and maximum signal to noise ratio and antenna scale relation under auxiliary angle (AA) algorithm
Figure.It can be seen that DL and AA maximum signal to noise ratio performance substantially close to.DL ensure that good system signal noise ratio.
Fig. 4 sets forth deep learning (DL) and AA Riming time of algorithm and antenna scale graph of a relation.Can from figure
To find out, AA exponentially increases with the antenna scale increase calculating time, and the DL calculating times are Millisecond.Calculated needed for DL
Time reduces several times than AA, and DL can quickly realize mimo system day line options.
The present invention is not only limited to above-mentioned embodiment, and persons skilled in the art are according to disclosed by the invention interior
Hold, the present invention can be implemented using other a variety of specific embodiments.Therefore, every design structure and think of using the present invention
Road, does some simple designs for changing or changing, both falls within the scope of the present invention.