CN113839743A - RLL code decoder for wireless communication receiving end - Google Patents

Abstract

The invention provides an RLL code decoder for a wireless communication receiving end, which aims at the field of visible light communication. A seq2seq model in the RNN network is adopted and named as seq2seq decoder. Conventional RLL code encoding and decoding relies on a code table, which is prone to errors occurring during transmission. The invention converts the mapping relation between the sending information and the receiving signal in the RLL code into the mapping relation between the neural network training label and the characteristic, and can directly use the seq2seq decoder in the decoding of the RLL code in the visible light communication through good training. Compared with the traditional code table lookup decoding method, the seq2seq decoder brings about 2-2.5 dB (E)_b/N₀) Performance gain of (2). Second, seq2seq decoders are capable of achieving low error rates approaching maximum a posteriori probability (MAP) decoding. In addition, the invention can realize multiframe in one time slotAnd parallel decoding is carried out to improve the system throughput. Finally, the design steps of the present invention are applicable to RLL codes of other codeword lengths in VLC, such as 8B10B, which shows great application potential in visible light communications.

Description

RLL code decoder for wireless communication receiving end

Technical Field

The invention belongs to the field of design of channel decoding methods in wireless communication, and mainly relates to design of a channel decoder based on a deep learning technology.

Background

Visible Light Communication (VLC) uses Light Emitting Diode (LED) flashing for Communication, has the advantages of high bandwidth, high speed, high stability, safety and environmental protection, and is an important research direction in the Communication field. Compared with wireless communication, VLC has a visible light spectrum of 380-780 nm, and the spectral width is larger than 4 multiplied by 10⁶GHz is more than 10000 times of the existing wireless communication frequency spectrum, so VLC has great development potential in the communication field.

In the practical application process of VLC, various artificial noises and thermal noises are received, and various light pollutions also cause interference to light signals, which seriously reduces the reliability of VLC. In order to effectively suppress information errors caused by inter-channel interference, channel coding techniques, such as RLL Codes and Error Correction Codes (ECC), have been widely used in VLC. The RLL encoder receives any bit stream of a sending end, maps K source bits into code words with the length of N through a code table, and ensures that the number of 0 in the output code words is the same as the number of 1 after encoding, namely, direct current balance. For decoding of RLL codes, a widely used decoding method is to recover the transmitted bits using a conventional table lookup decoding method after the receiving end generates hard decision bits. If the generated hard decision bits cannot be mapped to any one codeword in the code table, the corresponding transmitted codeword is estimated according to the minimum hamming distance. Because the table look-up decoding method belongs to a hard decision decoding method, the method can not fully utilize the information of the channel, the decoding performance is very limited, and the error correction capability can not be further improved.

The invention provides a design method of an RLL code decoder based on a Sequence to Sequence (Sequence) model, which is used for improving the decoding performance of an RLL code in VLC. In the IEEE 802.15.7 standard, VLC uses three types of RLL codes, which are the manchester code, the 4B6B code, and the 8B10B code, respectively. For convenience of illustration, the present invention is illustrated with 4B6B codes, and fig. 5 shows a table of 4B6B codes for VLC in IEEE 802.15.7 standard.

Disclosure of Invention

Aiming at a decoding method of an RLL code (4B6B code) in visible light communication, the invention aims to convert a 4B6B code decoding problem into a seq2seq deep learning problem by using a seq2seq model based on an RNN (Current Neural networks) technology, and designs a seq2seq decoder to replace a hard decision device and a traditional RLL table look-up decoder in a visible light communication receiving end. The simulation result shows that the decoding performance of the decoder exceeds that of the traditional table lookup decoding method, the decoding performance of the Maximum A Posteriori (MAP) criterion can be achieved, and the throughput of a VLC system can be improved.

The technical scheme of the invention is an RLL code decoder for a wireless communication receiving end, which is adopted to replace a hard decision device and a traditional RLL table look-up method decoder in a visible light communication receiving end; the RLL decoder comprises: an RNN encoder and an RNN decoder;

the RNN encoder is an RNN network using GRU units, the input layer size is N, and the hidden layer has two layers { h }₁,h₂The size of an output layer is 1, that is, an output vector C, where N is the size of a received signal in visible light communication and is also the length of the sequence after the RLL coding;

copying K parts of the output of the RNN encoder, and transmitting the K parts of the output to an RNN decoder;

the RNN decoder is also an RNN network using GRU units, and has an input layer with a size of K and two hidden layers with a size of h₁,h₂The output layer size is K, where K is the RLL code message bit length; the output layer of RNN decoder uses Sigmoid activation function, the rest layers of RLL decoder use ReLu activation function, and the loss function is

Wherein

Representing the decoded output of the ith neuron in the RNN decoder output layer, i.e. the source transmits the ith bit x_iAnd i ∈ {1,2, …, K }, and finally, based on the soft estimate obtained by the RNN decoder output, obtaining a decoding result using the following decision rule:

the decoded bit is judged to be 1, otherwise, the decoded bit is judged to be 0.

As described above, the present invention is directed to a seq2seq decoder of RLL code, having the following properties:

1. compared with the traditional table look-up decoding method, the invention can obtain about 2-2.5 dB (E)_b/N₀) Performance gain of (2).

2. The invention fully utilizes the channel information, the receiving end directly inputs the receiving data of the receiving end into the neural network for decoding without hard decision decoding, and the decoding performance of the optimal decoding criterion MAP can be achieved no matter several frames are transmitted, as shown in figure 4.

3. By modifying the sizes of the input layer and the output layer, the multi-frame RLL sequence can be decoded simultaneously, so that the throughput of the system is improved.

4. The invention can also be applied to other RLL code patterns, such as 8B10B codes, and therefore has good universality.

Drawings

Fig. 1 is a schematic diagram of a typical visible light communication model.

Fig. 2 is a schematic diagram of the design flow of a seq2seq decoder for RLL codes.

Fig. 3 is a schematic diagram of a seq2seq decoder neural network.

Fig. 4 shows the simulation results of Seq2Seq decoder.

Fig. 5 is a table of codes of 4B 6B.

Fig. 6 is a seq2seq neural network parameter setting table.

Detailed Description

In order to make the objects, technical means and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following embodiments and accompanying drawings.

One isA typical visible light communication system model is shown in fig. 1. The source bits u are encoded into x and v respectively by an ECC encoder and an RLL encoder. And generating a modulation signal r after OOK modulation or BPSK modulation. Then passes through a Gaussian white noise AWGN channel to receive a signal

Where n is obedient mean 0 and variance σ²White noise of the gaussian distribution. Thereafter, a hard decision detector estimates hard decision bits

Then respectively sending to RLL decoder and ECC decoder to finally obtain estimated value of transmitted message

In the invention, only the performance of the RLL encoder and the RLL decoder is considered, and the influence of ECC is ignored, therefore, the transmission bit is x, and the decoding bit is

The purpose of decoding is to estimate

As close to x as possible.

The decoder of the invention adopts a seq2seq model, and the decoding method of the model is as follows:

step 1: the RLL code decoding problem is a sequence-to-sequence mapping problem, takes into account that RNN networks are good at handling the sequence problem with time dependence, and is a sequence problem due to the decoding mapping problem

N → K unequal length sequence mapping problem, so a neural network based on the seq2seq model of RNN is established, and the model is divided into two parts of RNN-Encoder and RNN-Decoder. The network parameters are: the RNN-Encoder is an RNN network using GRU units. Input layer size N, hidden layerTwo layers { h₁,h₂And 5N, the output layer size is 1, i.e. the output vector C. The output vector C is input to the RNN-Decoder in K copies. The RNN-Decoder is also an RNN network using GRU units, the input layer size is K, and the hidden layer has two layers { h }₁,h ₂5N, and the output layer size is K. If multi-frame parallel decoding is needed to improve the system throughput, the N needs to be modified into a multiple of N, the K needs to be modified into a multiple of K, and other network parameters are kept unchanged. Further, the input of the seq2seq neural network is

Wherein

Received signal after passing through AWGN channel by RLL code coding (without considering ECC coder) for message bit x

The output of the network is the transmission bit x ═ { x corresponding to the received signal₁,x₂,…,x_KSoft estimation of }

Based on derived soft estimates

A hard estimate of x can be obtained by the following decision rule

If it is not

Then

If not, then,

fig. 3 shows a schematic diagram of a seq2seq decoder neural network.

Step 2: setting seq2seq decoder parameters and configuration. Taking the 4B6B code as an example, if single frame transmission is considered, i.e. 4B6B coded sequence is decoded frame by frame, then the network parameters are set as: the input layer size N is 6 and the output layer size K is 4. If multi-frame transmission is considered, i.e. 2 frames or 3 frames of the 4B6B encoded sequence are decoded simultaneously and in parallel, the network parameters are set as follows: the input layer size N is 12,18 and the output layer size K is 8, 12. The input layer and the hidden layer of the RNN-Encoder network and the RNN-Decoder are expressed as sigma by using a ReLu activation function_relu(x) Max {0, x }, the RNN-Decoder output layer uses a Sigmoid activation function, denoted as Sigmoid activation function

And 3, generating a training data set and a testing data set of the seq2seq decoder.

The specific method comprises the following steps: the data can be generated by simulation of the visible light communication model in fig. 1. First a series of transmit bits is generated

The transmit bits are encapsulated into frames at 4 bits per frame, denoted as

Wherein x_iAnd (4) the ith frame is represented, and if the frame is transmitted in a single frame, each frame is 4 bits. M is expressed as a frame number. Transmitting bits

According to the 4B6B code table shown in FIG. 5, the code is coded into code words in units of frames

Wherein v is_iAnd (4) the coded code word of the ith frame is represented, and if the coded code word is transmitted in a single frame, each frame is 6 bits. The coded code words are subjected to OOK/BPSK modulation to obtain sending signals, the sending signals reach a receiving end through a Gaussian channel, and the receiving end receives receiving signals. Will receive informationNumber as training data

Wherein

If the received signal indicates a single frame transmission, the size of each frame is 6. Receiving signal of each frame

Corresponding transmission bit x_iAs a label. Since the channel condition cannot be known in the decoding stage, it is necessary to generate a specific snr data as a training set, and then test the data on test sets with different snrs to simulate the real visible light communication. Through multiple experiments, the training signal-to-noise ratio E is found_b/N₀When set to 1, and the seq2seq decoder requires M to be 6 × 10⁴And the data is frame data, so that a good training effect can be obtained.

And 4, step 4: according to E_b/N₀Generating M frames of training set data, and training a seq2seq model. And (4) inputting the training set data obtained in the step (3) into a neural network, and training a seq2seq decoder. The neural network optimizer uses Adam optimizer, using a default learning rate of 0.01. The network uses the MSE loss function, expressed as

Wherein K is the number of output layer neurons of the seq2seq model,

is the output of the seq2seq model, i.e. is the transmitted bit x_iA soft estimate of (a). And obtaining the trained seq2seq decoder by optimizing the loss function. Experiments show that the training period epoch is 40, and excellent performance can be obtained. The detailed seq2seq decoder parameter set is shown in fig. 6.

And 5: use withoutSame signal-to-noise ratio (E)_b/N₀{ -4, -2,0,2,4,6,8} dB) to evaluate the trained network and obtain soft estimates of the decoded bits

Step 6, outputting a soft estimation value according to the Seq2Seq model

The following rule is used to obtain an estimated decoded value for transmit bit x

The decoding rule is as follows:

wherein the content of the first and second substances,

for the output of the ith neuron of the output layer of the seq2seq decoder,

decoding result for the ith transmitted bit of each frame. The present invention turns the above decoding process into a seq2seq decoder. The performance simulation diagram of the present invention is shown in FIG. 4, and the decoding method of the present invention can meet the requirement of the optimal decoding criterion.

The flow chart of the algorithm from the step 1 to the step 6 is shown in figure 2.

The invention has the advantages that:

1. compared with the traditional table look-up decoding method, the invention can obtain about 2-2.5 dB (E)_b/N₀) Performance gain of (2).

2. The invention fully utilizes the channel information, the receiving end directly inputs the receiving data of the receiving end into the neural network for decoding without hard decision decoding, and the decoding performance of the optimal decoding criterion MAP (maximum a Posteriori probability) can be achieved, as shown in figure 4.

4. By modifying the sizes of the input layer and the output layer, the multi-frame RLL sequence can be decoded simultaneously, so that the throughput of the system is improved.

3. The invention can also be applied to other RLL code patterns, such as 8B10B codes, and therefore has good universality.

Claims (1)

1. An RLL code decoder for a wireless communication receiving end is adopted to replace a hard decision device and a traditional RLL table look-up method decoder in a visible light communication receiving end; the RLL decoder comprises: an RNN encoder and an RNN decoder;

the RNN encoder is an RNN network using GRU units, the input layer size is N, and the hidden layer has two layers { h }₁,h₂The size of an output layer is 1, that is, an output vector C, where N is the size of a received signal in visible light communication and is also the length of the sequence after the RLL coding;

copying K parts of the output of the RNN encoder, and transmitting the K parts of the output to an RNN decoder;

the RNN decoder is also an RNN network using GRU units, and has an input layer with a size of K and two hidden layers with a size of h₁,h₂The output layer size is K, where K is the RLL code message bit length; the output layer of RNN decoder uses Sigmoid activation function, the rest layers of RLL decoder use ReLu activation function, and the loss function is

Wherein

Representing the decoded output of the ith neuron in the RNN decoder output layer, i.e. the source transmits the ith bit x_iAnd i ∈ {1,2, …, K }, and finally, based on the soft estimate obtained by the RNN decoder output, obtaining a decoding result using the following decision rule：

The decoded bit is judged to be 1, otherwise, the decoded bit is judged to be 0.

Images