CN109728824B

CN109728824B - LDPC code iterative decoding method based on deep learning

Info

Publication number: CN109728824B
Application number: CN201811488099.9A
Authority: CN
Inventors: 郭锐; 冉凡春
Original assignee: Hangzhou Dianzi University
Current assignee: Hangzhou Dianzi University
Priority date: 2018-12-06
Filing date: 2018-12-06
Publication date: 2023-03-28
Anticipated expiration: 2038-12-06
Also published as: CN109728824A

Abstract

The invention provides an LDPC code iterative decoding method based on deep learning, and relates to the technical field of information transmission. Firstly, a standard BP iterative decoder is adopted to estimate coding bits and channel noise so as to obtain estimated values of the channel noise and information bits, then a neural network DNN is used to remove noise estimation errors of the standard BP decoder so as to obtain more accurate estimation of the channel noise, then the obtained noise estimation is processed at a channel receiving end and input into the decoder again, and iteration is carried out continuously. The invention solves the technical problem that the channel noise influences the decoding performance in the prior art. The invention has the beneficial effects that: more accurate estimation of the channel noise is obtained, and the decoding signal-to-noise ratio is improved, so that the decoding performance is improved, and the algorithm complexity is reduced.

Description

LDPC code iterative decoding method based on deep learning

Technical Field

The invention relates to the technical field of information transmission, in particular to a BP-DNN decoding method based on deep learning.

Background

With the development of technology, deep learning is widely applied in the fields of computer vision, natural language processing, automatic vehicle driving and the like, and has achieved unusual achievements in information transmission. The information has interference problems, especially noise interference, in the transmission process, so that the information has errors at a receiving end. Therefore, a coding scheme with good performance and a suitable decoding algorithm are required. The invention discloses a hard decision decoding method of LDPC codes, which is an invention patent application file named as 'an LDPC hard decision decoding method and a decoder based on deep learning' and is published in China patent application No. CN106571831A, published in 2017, 4 and 19. The method comprises the following steps: 1. taking a set of (X, Y) as a set of tagged data; 2. establishing an LDPC decoding sample set; 3. establishing a deep learning decoding model; 4. pre-training a deep learning decoding model; 5. and decoding and outputting the deep learning decoding model. The decoding method adopts a hard decision method, firstly, the decoding performance is possibly reduced compared with a soft decoding method, the soft decoding method makes a decision through multiple comparisons, and the hard decision method makes a decision through a single comparison. Secondly, in order to achieve the decoding error rate of the neural network to the error rate required in the text, the algorithm may be too time-consuming, and an overfitting situation may occur, so that the performance of the decoding algorithm is reduced. In addition, the influence of channel noise on the decoding performance in the communication process cannot be ignored.

Disclosure of Invention

In order to solve the technical problem that channel noise influences the decoding performance in the communication process in the prior art, the invention provides an LDPC code iterative decoding method based on deep learning, which is a decoding method with low complexity and robustness.

The technical scheme of the invention is as follows: an LDPC code iterative decoding method based on deep learning comprises the following steps: the method comprises the following steps: carrying out LDPC code coding on an information sequence X, wherein the coded information sequence is U = XG, and the information sequence U is divided into a training set and a test set; step two: modulating an information sequence U of a training set, and obtaining a transmission symbol Y at a channel output end by a modulated transmission symbol S through a noise adding channel of channel noise N; step three: the transmission symbol Y obtains an information sequence estimated value through a BP decoder

The evaluation value based on the subtraction of the channel transmission symbol S from the transmission symbol Y->

Obtain a channel noise estimate>

Channel noise estimate>

Inputting the signal into a neural network to be updated to obtain a channel noise estimated value->

The updated transfer symbol pick>

Inputting the data into a BP decoder, and performing iterative decoding to continuously update and decode the data at the input end of the BP decoder and the input end of the neural network until the BP-DNN decoder is iterated; step four: and inputting the data of the test set into the BP decoder, so that the data of the input end of the BP decoder and the data of the input end of the neural network are continuously updated and decoded until the iteration is finished.

Preferably, the information sequence U of the training set is BPSK modulated, and the modulated channel transmission symbol S is transmitted.

Preferably, the transmission symbol Y is soft demodulated by BPSK to obtain an estimate of the channel transmission symbol S

/>

Preferably, the channel noise N and the channel noise estimation value are used

A channel residual noise value R can be obtained,

n is the codeword length, which is a loss function of the neural network.

Compared with the prior art, the invention has the beneficial effects that: and estimating the coded bits by adopting a standard BP decoder, and then removing noise estimation errors of the BP decoder by using a neural network to obtain more accurate estimation of channel noise. Meanwhile, the iteration between the BP iterative decoder and the neural network DNN gradually improves the decoding signal-to-noise ratio, thereby improving the decoding performance and reducing the algorithm complexity.

Drawings

FIG. 1 is a flow chart of the present invention.

Detailed Description

The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings.

Example 1:

as shown in fig. 1, an LDPC code iterative decoding method based on deep learning includes the following methods,

the method comprises the following steps: and performing LDPC code coding on the binary information sequence X. The LDPC code is represented by its check (H) matrix, and is conventionally encoded. And performing elementary transformation on the H matrix, and converting into a generation (G) matrix. The encoded information sequence is U = XG. The information sequence U is divided into a training set and a test set. The ratio of training set to test set data is nine to one. In this embodiment, an LDPC code with a 3/4 code rate is taken as an example, the check matrix H is a 24 × 18 matrix, and the generator matrix G is also a (24, 18) LDPC code with a code rate of 3/4, where 24 is a codeword length and 18 is an information bit length.

Step two: in order to make the encoded information suitable for transmission in the channel, the training set of the information sequence U is modulated, and BPSK modulation is used in this embodiment. And obtaining a transmission symbol S through modulation. The transmission symbol S results in an output symbol Y through an AWGN (white gaussian noise) channel, Y = S + N. N is the channel noise of the AWGN channel.

Step three: the receiving end of the BP decoder carries out BPSK soft demodulation on the received output symbol Y first. Arranging data into matrix Y _i ＝[Y ₀ ,Y ₁ ,…,Y ₂₂ ,Y ₂₃ ]The index i is the number of data sets, determined by the information bit length k, and has a total of 2 ¹⁸ And (4) grouping. The obtained data Y _i And decoding is performed through a BP decoder in sequence. Obtaining estimated value of channel transmission symbol S through BPSK soft demodulation

Evaluation value based on BP decoding and information sequence X>

Evaluation value for transmitting a symbol S over a channel>

The channel noise estimate is available at the output of the BP decoder->

To be->

Is a basic input unit and is input into a neural network (DNN). In a neural network, the channel noise estimate is based on a forward propagation algorithm and a back propagation algorithm>

Classifying under the action of sigmoid function of an output layer to obtain an updated channel noise estimation value->

Upon obtaining the updated channel noise estimate->

After that, the residual noise R can be obtained. Residual noise->

I.e. the difference between the channel real noise and the channel noise estimate obtained by the neural network. In order to improve the decoding performance of the BP decoder, a smaller residual noise R is required. The smaller R, the less channel estimation noise>

The closer to the channel true noise N, the greater the value of>

The closer to the true channel the symbol S is transmitted, the less the effect of channel noise on decoding performance is reduced in the BP decoder. Thus, the selection will +>

As a loss function of the neural network, the neural network steps down the loss function value by a back propagation algorithm such that R → 0, <' > is greater or less than>

Thereby achieving the purpose of reducing the residual noise R. n is the codeword length, and n is 24 in this embodiment. At this point, the first round of decoding ends.

At the input of the BP decoder, the updated noise estimation value is subtracted by the output symbol Y from the second iteration decoding

I.e., the input update of the BP decoder is &>

(subscript i represents the ith update, i is 0. Ltoreq. M, and m is the total number of cycles)The total number of iterations of the BP-DNN decoder is set to 10. Update the input by Y _i And as the input of a new round of decoding to the BP decoder, the data of the input end of the BP decoder and the input end of the neural network are continuously updated and decoded, and meanwhile, the iteration between the BP and the DNN gradually improves the decoding signal-to-noise ratio, so that the decoding performance is improved until the iteration of the BP-DNN decoder is finished.

Step four: and after the third step, completing the training of the neural network, and finally testing the decoding performance of the neural network by using test set data. And inputting the data of the test set into the BP decoder, so that the data of the input end of the BP decoder and the data of the input end of the neural network are continuously updated and decoded until the iteration is finished. Meanwhile, in the testing stage, the total iteration number m can be tried to be changed, so that the BP-DNN decoding performance can be tested. The decoding performance of the testing neural network is consistent with the process of training the neural network, but the data of the training set is changed into the data of the testing set, and the description is omitted.

Claims

1. An LDPC code iterative decoding method based on deep learning is characterized by comprising the following steps:

the method comprises the following steps: carrying out LDPC code encoding on an information sequence X, wherein the encoded information sequence is U = XG, dividing the information sequence U into a training set and a test set, representing the LDPC code by a check matrix thereof, and carrying out elementary transformation on the check matrix to convert the check matrix into a generation matrix G;

step two: modulating an information sequence U of a training set, and obtaining a transmission symbol Y at a channel output end by a modulated transmission symbol S through a noise adding channel of channel noise N;

step three: after BRSK soft demodulation is carried out on the transmission symbol Y, an information sequence estimation value is obtained through a BP decoder

Evaluation value for subtracting a channel transmission symbol S from a transmission symbol Y>

Deriving a channel noise estimateValue->

The transmission symbol Y is subjected to BPSK soft demodulation to obtain an estimation value ^ of the channel transmission symbol S>

Channel noise estimate pick-up>

Input into the neural network to obtain an updated channel noise estimate->

To be->

Is a basic input unit, and is input into a neural network, wherein the channel noise estimation value is subjected to a forward propagation algorithm and a backward propagation algorithm>

The updated transfer symbol pick>

Inputting the data into a BP decoder, and performing iterative decoding to continuously update and decode the data at the input end of the BP decoder and the input end of the neural network until the BP-DNN decoder is iterated;

step four: and inputting the data of the test set into the BP decoder, so that the data of the input end of the BP decoder and the data of the input end of the neural network are continuously updated and decoded until the iteration is finished.

2. The LDPC code iterative decoding method based on deep learning of claim 1, wherein: and carrying out BPSK modulation on the information sequence U of the training set, and transmitting the symbol S through the modulated channel.

3. The deep learning based LDPC code iterative decoding method according to claim 1, wherein: from the channel noise N and the channel noise estimate

The channel residual noise value R,. Sup.,. Can be found>

N is the codeword length, which is a loss function of the neural network. />