CN110719112B - Parameter self-adaptive RS code decoding method based on deep learning - Google Patents

Abstract

The invention relates to a parameter self-adaptive RS code decoding method based on deep learning, which comprises the following steps: RS code encoding is carried out on the information sequence X, and an encoded information sequence u is obtained; performing BPSK modulation on the information sequence u to obtain a transmission sequence s; the transmission sequence passes through a Gaussian white noise channel to obtain a receiving sequence y; constructing a reliability matrix by using the receiving sequence y, and calculating to obtain a reliability sequence R; dividing the reliability sequence R into a training set and a testing set; training a neural network; aiming at the training set of R and the corresponding receiving sequence y, respectively testing by using LCC decoding algorithms with different eta values to find an eta optimal solution corresponding to the training set receiving sequence y one by one; and taking the SNR and the reliability sequence R as input, and taking the minimum eta value obtained by the test as a classification label, wherein the minimum eta value and the classification label together form a data set for training the neural network.

Description

A Decoding Method of Parameter Adaptive RS Codes Based on Deep Learning

technical field

The invention belongs to the field of error control coding in channel coding, and relates to a parameter adaptive decoding algorithm based on deep learning.

Background technique

Error control is a basic method to control and correct transmission errors caused by channel noise by using codec technology in digital communication to ensure the effectiveness and reliability of communication. Error control coding is widely used in the field of digital communication and has important research significance. Reed-Solomon (RS) code is a kind of error control code, which was first constructed by Reed and Solomon in 1960. It has a simple structure, short decoding delay, rich structural features, and can correct random errors and burst errors. Especially in the application of correcting burst errors, RS codes have better performance than other error control codes. At present, RS codes have been widely used in wireless communication, optical fiber communication, magnetic disk and solid-state storage, etc., and have always been a research hotspot in the industry and academia.

According to the form of demodulator output symbols, the decoding algorithms of RS codes can be divided into two categories: hard-decision decoding algorithms and soft-decision decoding algorithms. Among them, the hard-decision decoding algorithm is relatively simple and easy to implement in hardware, but the algorithm does not make full use of channel soft information, and the decoding performance is lost. Soft-decision decoding trades complexity for the improvement of decoding gain, mainly including KV, BGMD and LCC algorithms. Among them, the LCC algorithm is superior to the former two in terms of complexity and calculation cycle. At the same time, the decoding performance and complexity of the algorithm are closely related to the parameter η of the multiplicity allocation process.

In recent years, deep learning has attracted worldwide attention due to its powerful ability to solve complex tasks. Deep learning methods have also been applied to speech recognition, computer vision, machine translation and other fields, bringing great innovations and breakthroughs to these fields. The combination of deep learning and coding has achieved some research results, and in-depth research has been carried out in the fields of polar codes, BCH codes, and LDPC codes. Most of the current research regards decoding as a classification problem, and uses neural network structures (such as DNN, CNN, RNN) to realize the complete decoding process. The obvious disadvantage of this type of scheme is that the neural network has very poor predictive ability for unlearned codewords (codewords not included in the training set), and the exponential explosion of codewords that may be brought about by the increase in codeword length will cause It greatly affects the performance of the decoder. In addition, some studies propose to use deep learning to assist traditional decoding algorithms to eliminate related noise and predict key parameters to improve the performance of traditional algorithms.

Contents of the invention

The purpose of the present invention is to provide a parameter adaptive decoding method based on deep learning that can reduce complexity and improve efficiency. The technical solution is as follows:

A parameter adaptive RS code decoding method based on deep learning, comprising the following steps:

(1) Encode the information sequence X with RS code to obtain the encoded information sequence u. Perform BPSK modulation on the information sequence u to obtain the transmission sequence s. The received sequence y is obtained after the transmitted sequence passes through the Gaussian white noise channel. Then use the received sequence y to construct a reliability matrix, and calculate the reliability sequence R. The reliability sequence R is divided into training set and test set.

(2) Carry out the training of the neural network. For the training set of R and the corresponding receiving sequence y, the LCC decoding algorithms with different η values are used to test respectively, and the optimal solution of η corresponding to the receiving sequence y of the training set is found one-to-one. Taking SNR and reliability sequence R as input, and taking the minimum η value obtained from the test as the classification label, the two together form the data set for neural network training. The neural network structure uses the MLP structure, and the weight w and bias b of the network are continuously optimized through the back propagation algorithm. Under sufficient data sets and training algebraic conditions, a neural network with high classification accuracy is obtained to complete the training process.

(3) Input the test set of SNR and R, and calculate the optimal solution of η through the trained neural network. Next, LCC decoding is performed using the parameter-adapted η value to obtain the decoded sequence X', and the decoding is finally completed.

Description of drawings

Figure 1 System block diagram of parameter adaptive decoding algorithm based on deep learning

Figure 2 The training process of the neural network

Detailed ways

The parameter adaptive decoding algorithm based on deep learning proposed by the present invention is an algorithm that uses deep learning to assist traditional RS code LCC decoding. According to the analysis, the key parameter η in the LCC algorithm is closely related to the channel environment and reliability information, and the value of η determines the decoding performance and complexity of the LCC algorithm. The problem of solving the optimal solution of η can be regarded as a simple classification problem. The powerful pattern classification ability of the neural network can select the optimal solution of η simply and efficiently. Through the deep learning method, the key parameter η in the LCC algorithm can be adaptively changed with the channel environment and the reliability information of the received sequence, which can improve the decoding efficiency without affecting the decoding performance.

The parameter adaptive decoding algorithm based on deep learning specifically includes two stages, the training stage and the decoding stage.

(1) The training stage realizes the process of constructing and training the neural network. Based on supervised learning, the present invention constructs a multilayer perceptron (MLP) neural network structure, and the number of nodes and the number of layers are determined by the code length. In terms of the training data set, the signal-to-noise ratio and the likelihood ratio are connected as input, and the η optimal solution obtained by the corresponding traditional decoding algorithm test is used as the classification label. After training and debugging, a neural network with excellent classification accuracy was obtained.

(2) In the decoding stage, the signal-to-noise ratio and likelihood ratio of specific codewords are input into the trained neural network to obtain the optimal solution of η. Then perform traditional LCC decoding according to the optimal solution.

The technical solution of the present invention will be further specifically described below in conjunction with the accompanying drawings.

As shown in Figure 1, the parameter adaptive decoding algorithm based on deep learning specifically includes the following steps:

(4) Encode the information sequence X with RS code to obtain the encoded information sequence u. In order to transform the coded information into a mode suitable for transmission in the channel, BPSK modulation is performed on the information sequence u to obtain the transmission sequence s. The received sequence y (y=s+n, n being AWGN channel noise) is obtained after the transmission sequence passes through the white Gaussian noise channel (AWGN). Then use the received sequence y to construct a reliability matrix, and calculate the reliability sequence R. Divide the reliability sequence R into a training set and a test set with a ratio of 9:1.

(5) As shown in Figure 2, the training of the neural network is carried out. For the training set of R and the corresponding receiving sequence y, use the LCC decoding algorithm with different η values to test separately, and find the η optimal solution corresponding to the training set receiving sequence y one-to-one (the minimum value in the case of correct decoding) . Taking SNR and reliability sequence R as input, and taking the minimum η value obtained from the test as the classification label, the two together form the data set for neural network training. The neural network structure uses a multi-layer perceptron (MLP) structure, and continuously optimizes the weight w and bias b of the network through the back propagation algorithm. Under sufficient data sets and training algebraic conditions, a neural network with high classification accuracy is obtained. Complete the training process.

(6) With the test set input of SNR and R, the optimal solution of η is calculated by the neural network after training. Next, LCC decoding is performed using the parameter-adapted η value to obtain the decoded sequence X', and the decoding is finally completed.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (1)

1. A parameter self-adaptive RS code decoding method based on deep learning comprises the following steps:

(1) RS code encoding is carried out on the information sequence X, and an encoded information sequence u is obtained; performing BPSK modulation on the information sequence u to obtain a transmission sequence s; the transmission sequence s passes through a Gaussian white noise channel to obtain a receiving sequence y; constructing a reliability matrix by using the receiving sequence y, and calculating to obtain a reliability sequence R; dividing the reliability sequence R into a training set and a testing set;

(2) Training of the neural network is performed: η is a key parameter in the LCC decoding algorithm and is closely related to channel environment and reliability information, the size of the η value determines the decoding performance and complexity of the LCC algorithm, the LCC decoding algorithms with different η values are used for respectively testing a training set and a corresponding receiving sequence y of R, and an η optimal solution corresponding to the training set receiving sequence y one by one is found, wherein the η optimal solution is the minimum value under the condition of correct decoding; taking SNR and reliability sequence R as input, and taking the minimum eta value obtained by test as a classification label, and forming a data set of neural network training together; the neural network structure uses an MLP structure, the weight w and the bias b of the network are continuously optimized through a back propagation algorithm, and the neural network with high classification accuracy is obtained under the conditions of enough data sets and training algebra, so that the training process is completed.