CN109728824A

CN109728824A - An iterative decoding method of LDPC codes based on deep learning

Info

Publication number: CN109728824A
Application number: CN201811488099.9A
Authority: CN
Inventors: 郭锐; 冉凡春
Original assignee: Hangzhou Electronic Science and Technology University
Current assignee: Hangzhou Electronic Science and Technology University
Priority date: 2018-12-06
Filing date: 2018-12-06
Publication date: 2019-05-07
Anticipated expiration: 2038-12-06
Also published as: CN109728824B

Abstract

The invention provides an iterative decoding method for LDPC codes based on deep learning, and relates to the technical field of information transmission. First, the standard BP iterative decoder is used to estimate the coded bits and channel noise, so as to obtain the estimated values of the channel noise and information bits, and then the neural network DNN is used to remove the noise estimation error of the standard BP decoder to obtain the channel noise. Then, the obtained noise estimate is processed at the channel receiving end, and is re-inputted into the decoder again, and iteratively. The invention solves the technical problem of the influence of channel noise on decoding performance in the prior art. The beneficial effects of the invention are as follows: obtaining a more accurate estimation of the channel noise, improving the decoding signal-to-noise ratio, thereby improving the decoding performance and reducing the complexity of the algorithm.

Description

A kind of LDPC code iterative decoding method based on deep learning

Technical field

The present invention relates to technical field of information transmission, more particularly, to a kind of decoding side BP-DNN based on deep learning Method.

Background technique

With the development of technology, deep learning obtains in fields such as computer vision, natural language processing, automatic driving vehicles To extensive use, wherein also obtaining original achievement in terms of information transmission.There are interference problems in transmission process for information, especially It is noise jamming, so that information goes wrong in receiving end.Therefore, it is necessary to coding modes of good performance and suitable decoding Algorithm.Chinese patent application publication No. CN106571831A, data of publication of application on April 19th, 2017, entitled " one kind is based on deep The application for a patent for invention file of the LDPC Hard decision decoding method and decoder of degree study " discloses a kind of sentencing firmly for LDPC code Certainly interpretation method.Include the following steps: 1, one group (X, Y) is used as one group of tape label data；2, LDPC decoding sample set is established； 3, the foundation of deep learning Decoding model；4, the pre-training of deep learning Decoding model；5, deep learning Decoding model decodes simultaneously Output.This interpretation method using hard-decision method, first decoding performance compared to soft decoding method may under Drop, soft decoding method is by repeatedly relatively making judgement, and hard-decision method relatively makes judgement by single.Secondly in order to make Neural network error rate of translation reaches in text the required bit error rate, and to may result in algorithm too time-consuming, thereby increases and it is possible to can occur The case where degree fitting, so that the performance of decoding algorithm is declined.In addition shadow of the interchannel noise to decoding performance in communication process Sound is also very important.

Summary of the invention

In order to solve the technical issues of interchannel noise influences decoding performance in communication process in the prior art, the present invention is mentioned For a kind of LDPC code iterative decoding method based on deep learning, the interpretation method with low complex degree, robustness utilizes nerve The powerful computing capability of network is removed the noise estimation error of BP decoder, to promote the decoding of BP decoder Energy.

The technical scheme is that a kind of LDPC code iterative decoding method based on deep learning: step 1: to information Sequence X carries out LDPC code coding, and the information sequence after coding is U=XG, is training set and test set information sequence U points；Step Rapid two: the information sequence U of training set being modulated, modulated transmission symbol S adds channel of making an uproar by interchannel noise N, obtains To the transmission symbol Y of channel output；Step 3: transmission symbol Y obtains information sequence estimated value by BP decoderBy passing Defeated symbol Y subtracts the estimated value of transmission symbol SObtain interchannel noise estimated valueInterchannel noise estimated valueIt is input to Neural network obtains interchannel noise estimated value after being updatedBy updated transmission symbolBP decoding is inputted again Device is iterated decoding, so that the data of the input terminal of BP decoder and neural network input terminal are constantly updated, constantly decodes, Until BP-DNN decoder iteration terminates；Step 4: BP decoder is inputted with the data of test set, so that the input of BP decoder The data of end and neural network input terminal are constantly updated, and are constantly decoded, until iteration terminates.

Preferably, carrying out BPSK modulation to the information sequence U of training set, modulated transmission accords with S.

Preferably, transmission symbol Y obtains the estimated value that transmission accords with S by the soft demodulation of BPSK

Preferably, by interchannel noise N and interchannel noise estimated valueAvailable channel residual noise value R,For the loss function of neural network, n is code word size.

Compared with prior art, the beneficial effects of the present invention are: being estimated using standard BP decoder coded-bit, Then it is removed with noise estimation error of the neural network to BP decoder, obtains the more accurate estimation to interchannel noise. Meanwhile decoding signal-to-noise ratio will be gradually increased in the iteration between BP iterative decoder and neural network DNN, to improve decoding performance With reduction algorithm complexity.

Detailed description of the invention

Attached drawing 1 is flow chart of the present invention.

Specific embodiment

Below with reference to the embodiments and with reference to the accompanying drawing the technical solutions of the present invention will be further described.

Embodiment 1:

As shown in Figure 1, a kind of LDPC code iterative decoding method based on deep learning, including following methods,

Step 1: LDPC code coding is carried out to binary information sequence X.LDPC code verifies (H) matrix by it and indicates, with conventional volume Code.H-matrix is subjected to elementary transformation, is converted to generation (G) matrix.Information sequence after then encoding is U=XG.Information sequence U points are training set and test set.The ratio between training set and test set data are nine to one.The present embodiment is with the LDPC code of 3/4 code rate For illustrate, check matrix H is the matrix of 24x18, and generator matrix G is also (24, the 18) LDPC code for obtaining code rate and being 3/4,24 It is information bit length for code word size, 18.

Step 2: in order to make the information after coding be suitble to transmit in the channel, the training set of information sequence U is adjusted System, the present embodiment select BPSK modulation.Transmission symbol S is obtained through ovennodulation.It transmits symbol S and passes through AWGN (Gauss additivity white noise Sound) channel obtains output symbol Y, Y=S+N.N is the interchannel noise of awgn channel.

Step 3: the output symbol Y received is first carried out the soft demodulation of BPSK by BP decoder receiving end.By data preparation For matrix Y_i=[Y₀,Y₁,…,Y₂₂,Y₂₃], subscript i is data group number, is determined by information bit length k, shares 2¹⁸Group.It will obtain Data Y_iBP decoder is passed sequentially through to be decoded.The estimated value that transmission accords with S is obtained by the soft demodulation of BPSKPass through BP decodes to obtain the estimated value with information sequence XThe estimated value of S is accorded with by transmission? BP decoder output can obtain interchannel noise estimated valueWithFor basic input unit, it is input to neural network (DNN) in.In neural network, by propagated forward algorithm and back-propagation algorithm, interchannel noise estimated valueIn output layer The lower classification of sigmoid function effect, obtain updated interchannel noise estimated value Obtaining updated interchannel noise estimated valueAfterwards, available residual noise R.Residual noiseThat is channel The difference for the interchannel noise estimated value that real noise and neural network obtain.In order to promote the decoding performance of BP decoder, need compared with Small residual noise R.R is smaller, channel estimation noiseCloser channel real noise N,Closer to real channel transmission symbol Number S, in this way, influence of the interchannel noise to decoding performance is lowered in BP decoder.Therefore, selection willAs nerve The loss function of network, neural network gradually reduces loss function value by back-propagation algorithm, so that R → 0, With this come achieve the purpose that reduce residual noise R.N is code word size, and the present embodiment n is 24.This is arrived, first round decoding terminates.

Since the second wheel iterative decoding, in BP decoder input terminal, updated estimated value of making an uproar is cut with output symbol YThat is the input of BP decoder is updated to(subscript i indicates that i-th updates, and 0≤i≤m, m are circulation total degree), The present embodiment, setting BP-DNN decoder iteration total degree are 10 times.Input is updated into Y_iBP is decoded as new round decoding The input of device constantly decodes so that the data of the input terminal of BP decoder and neural network input terminal are constantly updated, meanwhile, BP Decoding signal-to-noise ratio will be gradually increased in iteration between DNN, so that decoding performance is improved, until BP-DNN decoder iteration knot Beam.

Step 4: the training to neural network after step 3, is being completed, is finally testing mind with test set data Through network decoding performance.BP decoder is inputted with the data of test set, so that the input terminal of BP decoder and neural network input The data at end are constantly updated, and are constantly decoded, until iteration terminates.Simultaneously in test phase, can attempt change must iteration time Number m, carrys out test b P-DNN decoding performance with this.It is consistent with training neural network process to test neural network decoding performance, only not The data for the data of training set being changed to test set are crossed, are repeated no more.

Claims

1. a kind of LDPC code iterative decoding method based on deep learning, is characterized in that, comprises the following method: Step 1: information sequence X is carried out LDPC code encoding, the information sequence after encoding is U=XG, divides information sequence U into. are the training set and the test set; step 2: modulate the information sequence U of the training set, the modulated transmission symbol S passes through the noise-added channel of the channel noise N, and obtains the transmission symbol Y at the channel output end; step 3: transmits the symbol Y Estimated value of information sequence obtained by BP decoder Subtract the estimated value of the channel transmission symbol S from the transmission symbol Y get the channel noise estimate channel noise estimate The channel noise estimate is obtained after the input to the neural network is updated the updated transmission symbol Then input the BP decoder and perform iterative decoding, so that the data at the input end of the BP decoder and the input end of the neural network are continuously updated and decoded until the iteration of the BP-DNN decoder ends; Step 4: Use the data of the test set The data is input to the BP decoder, so that the data at the input end of the BP decoder and the input end of the neural network are continuously updated and decoded until the iteration ends.

2 . The deep learning-based iterative decoding method for LDPC codes according to claim 1 , wherein: BPSK modulation is performed on the information sequence U of the training set, and the modulated channel transmission symbol S is performed. 3 .

3. a kind of LDPC code iterative decoding method based on deep learning according to claim 1, is characterized in that: transmission symbol Y obtains the estimated value of channel transmission symbol S by BPSK soft demodulation

4. a kind of LDPC code iterative decoding method based on deep learning according to claim 1, is characterized in that: by channel noise N and channel noise estimation value The channel residual noise value R can be obtained, is the loss function of the neural network, and n is the length of the codeword.