CN109714086A

CN109714086A - Optimization MIMO detection method based on deep learning

Info

Publication number: CN109714086A
Application number: CN201910063733.2A
Authority: CN
Inventors: 陈倩; 张舜卿; 徐树公; 曹姗
Original assignee: University of Shanghai for Science and Technology
Current assignee: University of Shanghai for Science and Technology
Priority date: 2019-01-23
Filing date: 2019-01-23
Publication date: 2019-05-03
Anticipated expiration: 2039-01-23
Also published as: CN109714086B

Abstract

A kind of optimization MIMO detection method based on deep learning, by constructing the end-to-end mode of MIMO, the imperfect channel state information of the signal y (t) and estimation that are received according to the receiving end MIMOInput of the multiple time-domain vector of model as deep neural network (DNN) is obtained, obtains the estimated value of transmitting terminal bit stream using DNNWith the prior art according toDo hard decision obtain send bit stream estimated value compare, the present invention can be under faulty channel information, accuracy and detection rates are improved, guarantee the detection performance for realizing low bit error rate under low complexity algorithm, while there is good robustness in the case where containing intrinsic channel errors.

Description

Optimization MIMO detection method based on deep learning

Technical field

The present invention relates to a kind of technology of MIMO detection field, specifically a kind of deep learning that is based on is in imperfect letter Optimization MIMO detection method under road information condition.

Background technique

The information detecting method of the receiving end of existing mimo system includes linear and nonlinear method.For example, linear MIMO Detector includes force zero (ZF) or least mean-square error (MMSE) balanced device, and non-linear MIMO detectors are often relied on and are based on The maximum likelihood (ML) of minimum range detects.Since neural network can be fitted arbitrarily complicated function, neural network is carried out Extremely complex Nonlinear Mapping.Due to the strong correlation between being output and input in mimo system, as can by deep learning with MIMO detection technique combines, and computation complexity can be greatly lowered while reducing bit error rate.

Summary of the invention

The present invention In view of the above shortcomings of the prior art, proposes a kind of detection side optimization MIMO based on deep learning Method learns the obtained signal in receiving end and faulty channel information H simultaneously using deep neural network, obtains channel mistake Mapping relations between the distribution and incoming bit stream and the symbol received of difference, rather than according toHard decision is done to obtain Send the estimated value of bit stream.The present invention can improve accuracy and detection rates, guarantee under faulty channel information The detection performance of low bit error rate is realized under low complexity algorithm, while being had very well in the case where containing intrinsic channel errors Robustness.

The present invention is achieved by the following technical solutions:

The optimization MIMO detection method based on deep learning that the present invention relates to a kind of, by constructing the end-to-end transmission mould of MIMO Type, the imperfect channel state information of the signal y (t) received according to the receiving end MIMO and estimationObtain the multiple time domain of model to The input as deep neural network (DNN) is measured, obtains the estimated value of transmitting terminal bit stream using DNN

The end-to-end mode of the MIMO includes: N_tA transmitting antenna and N_rA receiving antenna, by using vertical point Layer space -time code technology is t, modulation size 2 with time slot^MCarry out message transmission, in which: transmitting terminal bit information flowSymbol is obtained through ovennodulation f (, M)That is x (t)=f (b (t), M), receiving end receive symbol Number it isThat is y (t)=Hx (t)+n (t), in which:Indicate flat between transmitting and receiving antenna pair Smooth Rayleigh fading coefficient,It is the additive white Gaussian noise with zero-mean and unit variance, i.e.,Receiving end is according to imperfectAnd receive the decoding process of symbol further progress are as follows:Wherein:It is to select to obtain from all possible brewed symbols, g () indicates the mapping relations of symbol detection, then obtains the detection bit of corresponding t time slot, i.e. estimated value by demodulation are as follows:

The detection bit that the demodulation obtains corresponding t time slot refers to: minimizing the bit error rate by general Optimization Framework (BER), the modeling of BER minimization problem is obtained:Wherein:

Indicate that the flat Rayleigh fading coefficient between transmitting and receiving antenna pair, T are that channel condition H keeps static Period, and H is independently varied between different fading periods, and h () indicates the combined process of symbol detection and demodulation, Subscript n indicates n-th of fading period；And the modeling of above-mentioned BER minimization problem is used for the end-to-end mode of MIMO Detection solves

The imperfect channel state information of the estimationBy will answer time-domain vector switch to the real number time domain to Amount

Wherein:R () and I () respectively indicates complex vector located Real and imaginary parts,The as imperfect channel state information (CSI) estimated of receiving end,Δ H is indicated The evaluated error of channel, if carrying out channel estimation using maximum likelihood method, Δ H passes through with zero-mean and same association side The multiple Gauss of the independent same distribution formula (i.i.d.) of poor matrix is distributed to model, i.e. and Δ H~WhereinN_pIndicate the quantity of frequency pilot sign,

E_pIndicate the power of frequency pilot sign.

The deep neural network include: an input layer, hidden layers be made of four full connection (Dense) layers with One output layer, in which: hidden layer activation primitive is Rectified Linear Unit (ReLU), optimizer Adam；Output Layer activation primitive is sigmoid, and loss function is cross entropy, is exported to detect bitLabel is original information bits b.

Technical effect

Compared with prior art, the present invention need to only train the data under a kind of state of signal-to-noise to obtain training pattern, just general Change to other signal-to-noise ratio and there is excellent detection performance, or even has more than the gain of 5dB.With very strong generalization ability.

Secondly, the bit error rate that the present invention obtains be superior in the case where imperfect CSI and perfection CSI conventional method and DetNet: under 4 × 4MIMO BPSK modulation case, when BER is 2 × 10-3, this method is about 3.5dB better than DetNet, 4dB is above better than ZF and MMSE.It has good robustness.

Third, the present invention brought by handling capacity have compared to traditional MMSE detection method and ML detection method it is huge Raising, while detection efficiency is high.

Detailed description of the invention

Fig. 1 is MIMO downlink transfer model schematic of the present invention；

Fig. 2 is DNN network model schematic diagram of the present invention；

Fig. 3 is BER vs Eb/No signal of the embodiment under perfect channel condition under 4 × 4MIMO BPSK modulation system Figure；

Fig. 4 is that BER vs Eb/No of the embodiment under imperfect channel condition under 2 × 2MIMO QPSK modulation system shows It is intended to；

Fig. 5 is that BER vs Eb/No of the embodiment under imperfect channel condition under 4 × 4MIMO BPSK modulation system shows It is intended to.

Specific embodiment

As shown in Figure 1, for a kind of optimization MIMO detection method based on deep learning that the present embodiment is related to, including it is following Step:

Step 1) initiation parameter: respectively in perfect channel condition (i.e. N_p·E_P→ ∞,) and imperfect CSI feelings (i.e. N under condition_p·E_P=400, Δ H~) obtain training data: 2 × 2MIMO QPSK modulation and 4 are considered respectively The result of × 4MIMO BPSK modulation case, it is assumed that the decline of channel it is spatially uncorrelated to receiver but with transmitting phase It closes, i.e., typical downlink channel in mobile communication system.

The training data is real time-domain vectorWith b (t).

Step 2) neural metwork training: the network parameter for the DNN that the present embodiment uses is as shown in the table:

Table 1DNN network parameter allocation list

	DNN
		Input layer	45/89
First layer	Dense 512-Relu
		The second layer	BN
Third layer	Dense 256-Relu
		4th layer	Dense 128-Relu
Layer 5	Dense 64-Relu
		Output layer	4 road sigmoid
Parameter amounts to	215372

The present embodiment selects 5 layers of DNN network to be trained: for 2 × 2MIMO, input matrix 4*5, for 4 × 4MIMO, input matrix 8*9.

The training, the test data set total 7.2 × 10 of use⁵A symbol, wherein training dataset b (t) is 5.4 ×10⁵A symbol, validation data setIt is 1.8 × 10⁵A symbol,InFor transmitting terminal bit information flow Receiving end reception symbol y (t) that b (t) is generated through the end-to-end mode of MIMO turns real number field through complex field and obtains.

The present embodiment converges on 25 or so epoch in training.

The output of the DNN network isLabel is set as original information bit, and loss function is selected Crossentropy, i.e. cross entropy

Step 3) is by the obtained detection symbols of this method and traditional squeeze theorem method (ZF, Baseline 1), most Small mean square error detection method (MMSE, Baseline 2), DetNet (Baseline 3) and maximum likelihood ratio (ML, Baseline 4) method compares respectively, as a result as shown in Figure 3.

Under perfect channel condition under 4 × 4MIMO BPSK modulation system, as seen from Figure 3, the BER of this method is various MMSE and ZF are below under SNR, performance has even been got well more than 5dB.This also illustrates that this method has very strong generalization ability.

It is 10 for BER^-3, this method channel model in be better than DetNet 4.5dB.

As shown in figure 4, for 2 × 2MIMO QPSK modulation system under imperfect channel condition, at the BER of 10-2, we Method is more than 5dB better than ZF method still better than MMSE method about 4.5dB.For the 4 × 4MIMO detection for using BPSK to modulate, such as It is 2 × 10 in BER shown in Fig. 5^-3When, this method is about 3.5dB better than DetNet, is above 4dB better than ZF and MMSE.This table Bright this method still has good robustness in imperfect CSI.It is furthermore noted that in the case where perfect CSI, The BER of the ML method of imperfect CSI is considerably higher.But the fluctuation of DNN caused by the imperfection of CSI is similar to other methods. It is noted that the testing result of this method is even better than the base with perfect channel information under the premise of imperfect CSI In the MIMO detection method of DetNet.Although the BER of this method is higher than ML method, complexity ratio ML algorithm is much lower, foot To make up the deficiency of BER.

In addition, using 12GB memory Intel core i5-6500CPU@3.20GHz desktop computer in Python 3.5.2 all MIMO detection algorithms are programmed and run in.Operation test is to detect 7.2 × 10⁵A symbol simultaneously records runing time, It is repeated 3 times, then calculates average throughput.In the following table, the detection effect that 4 × 4MIMO BPSK modulates lower different schemes is compared Rate, including the MIMO detection method and conventional method such as ML method, MMSE method and ZF method based on DNN proposed.

It can be seen from the table, it will be seen that the handling capacity highest of ZF, and this method and ZF algorithm have closely similar handling capacity. Followed by MMSE method, ML method detection efficiency is minimum.It draws a conclusion, this method has the throughput performance close to ZF, together The more accurate detection performance of Shi Shixian.

Above-mentioned specific implementation can by those skilled in the art under the premise of without departing substantially from the principle of the invention and objective with difference Mode carry out local directed complete set to it, protection scope of the present invention is subject to claims and not by above-mentioned specific implementation institute Limit, each implementation within its scope is by the constraint of the present invention.

Claims

1. a kind of optimization MIMO detection method based on deep learning, which is characterized in that by constructing the end-to-end transmission mould of MIMO Type, the imperfect channel state information of the signal y (t) received according to the receiving end MIMO and estimationObtain the multiple time domain of model to The input as DNN is measured, obtains the estimated value of transmitting terminal bit stream using DNN

The end-to-end mode of the MIMO includes: N_tA transmitting antenna and N_rA receiving antenna, by using vertical layered space Time-code technology is t, modulation size 2 with time slot^MCarry out message transmission, in which: transmitting terminal bit information flow Symbol is obtained through ovennodulation f (, M)That is x (t)=f (b (t), M), receiving end receive symbol and areThat is y (t)=Hx (t)+n (t), in which:Indicate flat auspicious between transmitting and receiving antenna pair Sharp fading coefficients,It is the additive white Gaussian noise with zero-mean and unit variance, i.e., Receiving end is according to imperfectAnd receive the decoding process of symbol further progress are as follows:Wherein:It is to select to obtain from all possible brewed symbols, g () indicates the mapping relations of symbol detection, then obtains the detection bit of corresponding t time slot, i.e. estimated value by demodulation are as follows:

2. according to the method described in claim 1, it is characterized in that, the detection bit that the demodulation obtains corresponding t time slot refers to: The bit error rate (BER) is minimized by general Optimization Framework, obtains the modeling of BER minimization problem:

Subject to x (t)=f (b (t), M),

Y (t)=H_nX (t)+n (t),

Wherein:Indicate the flat Rayleigh fading system between transmitting and receiving antenna pair Number, T are that channel condition H is kept for the static period, and H is independently varied between different fading periods, and h () indicates symbol Number detection and demodulation combined process, subscript n indicate n-th of fading period；And the modeling of above-mentioned BER minimization problem is used for institute The end-to-end mode detection of the MIMO stated solves

3. according to the method described in claim 1, it is characterized in that, the imperfect channel state information of the estimationPassing through will Multiple time-domain vector switchs to the real number time-domain vectorWherein:

R () and I () respectively indicate complex vector located real and imaginary parts,The as imperfect channel state information (CSI) estimated of receiving end,Δ H indicates that the estimation of channel misses Difference, if carrying out channel estimation using maximum likelihood method, Δ H passes through the independence with zero-mean and same covariance matrix To model, i.e., multiple Gauss with distribution (i.i.d.) is distributedWhereinN_pIndicate pilot tone symbol Number quantity, E_pIndicate the power of frequency pilot sign.

4. according to the method described in claim 1, it is characterized in that, the deep neural network includes: an input layer, by four The hidden layer and an output layer of a full articulamentum composition, in which: hidden layer activation primitive is ReLU, optimizer Adam；It is defeated Layer activation primitive is sigmoid out, and loss function is cross entropy, is exported to detect bitLabel is original information bits b.