A method for blind recognition of low-density parity-check code rate based on deep learning

Abstract

A method for blind rate recognition of low-density parity-check codes based on deep learning has two stages: training and verification. In the training stage, the transmitter selects a code rate from the code rate set, generates a code word sequence C through low-density parity-check coding of the original information sequence S, and obtains a transmission symbol sequence X through binary shift keying modulation of the code word sequence C. The code word sequence X reaches the receiving end through an additive white Gaussian noise channel, and the receiving end obtains a received symbol sequence Y as training input; the training label is a vector P. A rate recognition neural network model is constructed, Y is used as the training data input model, and the probability vector P corresponding to its rate category is used as the model output. The model is trained using the stochastic gradient descent algorithm, the classification cross entropy loss is calculated, and the model weights and biases are updated to obtain an optimized rate recognition neural network model; in the verification stage, a verification sequence Y is generated. v , input the optimized model to perform blind bit rate recognition. Compared with the traditional bit rate recognition method, the performance of the present invention is significantly improved in the low signal-noise range.

Description

Technical Field

[0001] The present invention discloses a method for blindly identifying a code rate of a low-density parity-check code by applying deep learning in a wireless communication system, and belongs to the field of combining deep learning with wireless communication. Background Art

[0002] In the past few years, with the improvement of neural network architecture and optimization of algorithms, deep learning has been rapidly developed and applied as an important technology in the field of artificial intelligence. Its achievements in computer vision, natural language processing and other tasks have attracted much attention from scientific researchers. In the field of wireless communications, researchers are also committed to integrating deep learning technology with wireless communication technology. Intelligent communication technology based on deep learning has been considered to be one of the potential key technologies for the next generation of mobile communications (see "Deep Learning Techniques for Advancing 6G Communications in the Physical Layer", IEEE Wireless Communications, Vol. 28, No. 5, Oct. 2021).

[0003] By applying adaptive coding and modulation technology in complex communication environments, the channel coding parameters can be adjusted in real time to adapt to the time-varying channel state, thereby achieving efficient and reliable transmission. In traditional schemes, the transmitter needs to use a control channel to send channel coding and modulation parameter information to the receiver, which reduces the transmission efficiency of the system. In order to solve this problem, the technology of the receiver directly using the received signal itself to blindly identify the modulation type and channel coding parameters has received widespread attention in recent years. In addition, this type of technology also has broad application prospects in the field of radio reconnaissance, which is used to identify the coding and modulation parameters of intercepted signals and help decode them. The article (see "Over-the-Air Deep Learning Based Radio Signal Classification", IEEE Journal of Selected Topics in Signal Processing, Vol. 12, No. 1, Feb. 2018) studies the use of deep learning technology for blind modulation type recognition. The results show that compared with traditional technologies, the modulation type blind recognition technology based on convolutional neural network (CNN) has better recognition performance. Low-density parity check (LDPC) code is a channel coding technology with excellent performance. It has been widely used in wireless transmission protocols such as WIFI and 5G. The article (see "Novel blind identification of LDPC codes using average LLR of syndrome a posterioriprobability", IEEE Transactions on Signal Processing, Vol. 62, No. 3, Sep. 2014) proposes an effective LDPC code rate blind identification technology, which has good recognition performance under high signal-to-noise ratio, but the performance decreases significantly under low signal-to-noise ratio. How to improve the rate recognition accuracy under low signal-to-noise ratio needs to be solved urgently. Summary of the invention

[0004] The present invention proposes a method for blindly identifying the code rate of low-density parity-check codes based on deep learning. The receiving end does not rely on the code rate parameters transmitted by the control channel, but only extracts the characteristics of the received signal itself to identify the channel coding code rate. Compared with traditional algorithms, the present invention can effectively improve the recognition accuracy in low signal-to-noise ratio areas.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A method for blindly identifying the code rate of low-density parity check codes based on deep learning uses a neural network model to calculate the weighted checksum of the received sequence and perform feature processing and classification, and finally outputs its code rate category. This method includes two stages: training and verification. The specific implementation steps are as follows: The transmitter and the receiver agree on an LDPC code rate set θ = {r1, r2, ..., r i ,...,r t}, the number of corresponding codeword check nodes are {m1,m2,...,m i ,...,m t In the training phase, the transmitter first transmits a random information sequence S of length k through a code rate of r. i The LDPC code encoding method generates a codeword sequence C of length n. The codeword sequence C generates an energy-normalized modulation symbol sequence X through the binary phase shift keying (BPSK) modulation method. The modulation symbol sequence X is transmitted to the receiver through the additive white Gaussian noise (AWGN) wireless channel. The noise distribution has a mean of 0 and a variance of σ 2 The symbol sequence Y received by the receiver is used as the training input; the training label is a vector P of length t. When the encoding rate of Y is r i When the value P of the vector at position i is i =1, and the rest of the positions are 0. Within the continuous signal-to-noise ratio range, the encoding method corresponding to each bit rate in θ is used to generate equal amounts of data to form a training data set. The training data set is then input into the bit rate recognition neural network model, and the Stochastic Gradient Descent (SGD) algorithm is used to train the model weights and biases, and finally an optimized bit rate recognition model is obtained; in the verification stage, the test data set is generated using the same method as the training data set, and the optimized bit rate recognition model is used for verification.

[0007] The code rate recognition neural network model consists of a check feature calculation module, a rearrangement and splicing module, a convolutional neural network module and a fully connected neural network module; the check feature calculation module consists of t parallel calculation channels, each of which contains a variable node layer, a check node layer and a feedforward layer; the rearrangement and splicing module consists of t parallel rearrangement layers and a splicing layer; the convolutional neural network module consists of a two-dimensional convolution layer, a pooling layer and a batch normalization layer; the feedforward classification module consists of a feedforward layer and a random drop layer. The calculation process of the model is as follows: first, the check feature calculation module calculates the weighted checksum:

[0008] Calculate the channel log likelihood ratio (LLR) for each symbol in Y:

[0009]

[0010] In the ith channel of this module, use r i The corresponding check equation calculates the check node LLR:

[0011]

[0012] In the formula, tanh(.) is the hyperbolic tangent function, Π is the multiplication symbol, and tanh -1 (.) is the inverse hyperbolic tangent function, N i (c) is the set of variable nodes connected to the check node c. It is input into the feedforward layer. The number of input nodes of the feedforward layer of each channel is {m1,m2,...,m i ,...,m t}, the number of output nodes is n node ; In the rearrangement and concatenation module, t one-dimensional weighted checksum features are combined into two-dimensional features and concatenated into three-dimensional features along the first dimension; then the features are input into the convolutional neural network model for processing, and finally the features are classified in the feedforward classification module and the normalized probability vector P' for rate recognition is output. The method for updating the weights and biases of the rate recognition model is as follows: after the received symbol sequence Y is input into the rate recognition model, the model outputs a normalized probability vector P' of length t, and the output is combined with the label to calculate the classification cross entropy loss Loss:

[0013]

[0014] Where P i is the i-th element of the label vector P, P i ' is the i-th element of the model output vector P', and exp(.) is an exponential function. The SGD algorithm is used to update the weights and biases of the rate recognition model according to the value of Loss. When the maximum number of training iterations is reached or the value of the classification cross entropy loss Loss no longer changes, the training is stopped, and the optimal model weights and biases are saved to obtain the optimized rate recognition model.

[0015] The testing phase will verify that the sequence Y v Use the optimized bit rate recognition model to perform bit rate classification tasks. The specific implementation process is as follows:

[0016] 1) Enter the verification sequence Y for code rate identification v ;

[0017] 2) The bit rate recognition model calculates and outputs the normalized probability vector P v ;

[0018] 3) P vThe bit rate category corresponding to the maximum probability is output as the discrimination result.

[0019] The present invention makes full use of the coding structure information of the codeword, and extracts the features of the codeword corresponding to each candidate coding method by calculating the LLR of the check node. By taking the weighted sum of the check log-likelihood ratio as the feature, adding trainable feature weights, and enriching the number of features, the performance is significantly improved in the low signal-to-noise ratio area compared with the traditional code rate recognition algorithm, and it has broad application prospects in both military and civilian dual-use fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the blind recognition model of LDPC channel coding rate of the present invention;

[0021] Figure 2 The simulation results of the code rate blind recognition and comparative recognition method for LDPC codes with n=480 in the code rate set {1 / 2, 2 / 3, 3 / 4, 5 / 6} in the IEEE802.22 standard;

[0022] Figure 3 The simulation results of the blind recognition and comparative recognition method of the LDPC code with n=960 in the code rate set {1 / 2, 2 / 3, 3 / 4, 5 / 6} in the IEEE802.16e standard are shown. DETAILED DESCRIPTION

[0023] The present invention is further described below in conjunction with the accompanying drawings and actual embodiments.

[0024] The channel coding involved in the present invention is any code length n, and the code rate set is θ={r1,r2,...,r i ,...,r t} LDPC code, and construct a blind recognition method for LDPC code rate through deep learning technology to solve the problem that the traditional method has poor performance when the signal-to-noise ratio is low in Gaussian white noise channels. The following is a detailed description.

[0025] A method for identifying LDPC code rate based on deep learning, the specific implementation steps are as follows:

[0026] Step 1: The transmitter and the receiver agree on a set of coding rate parameters θ = {r1, r2, ..., r i ,...,r t}, the number of corresponding codeword check nodes are {m1,m2,...,m i ,...,m t The transmitter first processes a random information sequence S of length k using a code rate of r. iThe LDPC channel coding generates a codeword sequence C of length n. The codeword sequence C generates an energy-normalized modulation symbol sequence S = -2×C+1 through BPSK modulation, where 0 is encoded as symbol +1 and 1 is encoded as symbol -1. After passing through the AWGN channel to the receiving end, the received sequence is Y = S + N as the training input, where the mean of N is 0 and the variance is σ 2 , the signal-to-noise ratio is defined as E b / σ 2 , E b represents the average bit energy; the training label is a vector P of length t, when the encoding rate of Y is r i When the value P of the vector i is i = 1, and the rest of the positions are 0. According to the above process, within a continuous signal-to-noise ratio range, the coding method corresponding to each bit rate within θ is applied to generate equal amounts of data to form a training data set.

[0027] Step 2: Use the check feature calculation module, rearrangement and concatenation module, convolutional neural network module and feedforward classification module to jointly construct a LDPC channel coding rate blind recognition model based on deep learning, and give the activation function, weight and bias of the model. The overall model is as follows: Figure 1 shown.

[0028] The verification feature calculation module consists of t parallel calculation channels, each of which contains a variable node layer, a verification node layer and a feedforward layer connected in sequence; the rearrangement and splicing module is composed of t parallel rearrangement layers and a splicing layer after the verification feature calculation module; the convolutional neural network module consists of a two-dimensional convolution layer, a pooling layer and a batch normalization layer; the feedforward classification module consists of a feedforward layer and a random dropout layer. The activation function of the model uses the ReLU (see "Deep SparseRectifier Neural Networks", Journal of Machine Learning Research, 2011) function: y = max(x, 0). The initialization weights of the model have a mean of 0 and a variance of Normal distribution, the initial value of the bias is b v The model first calculates t weighted checksums by the check feature calculation module: Calculate the LLR of each symbol in Y:

[0029]

[0030] In the ith channel of this module, use r i The corresponding check equation calculates the check node LLR:

[0031]

[0032] In the formula, tanh(.) is the hyperbolic tangent function, Π is the multiplication symbol, and tanh -1 (.) is the inverse hyperbolic tangent function, N i (c) is the set of variable nodes connected to the check node c. It is input into the feedforward layer. The number of input nodes of the feedforward layer of each channel is {m1,m2,...,m i ,...,m t}, the number of output nodes is n node ; In the rearrangement and concatenation module, t one-dimensional weighted checksum features are combined into two-dimensional features and concatenated into three-dimensional features along the first dimension; then the features are input into the convolutional neural network model for extraction and processing, and finally the features are classified in the feedforward classification module and the recognition probability P' is output.

[0033] Step 3: Use the training data set generated in step 1 to train the bit rate recognition model using the SGD algorithm, set the maximum number of training iterations, batches, and learning rate, and calculate the classification cross entropy loss Loss through the model output and the preset label input loss function:

[0034]

[0035] Where P i is the i-th element of the label vector P, P i ' is the i-th element of the model output vector P', and exp(.) is the exponential function. The optimizer uses Adam (see "Deep Learning and MindSpore Practice", Tsinghua University Press, 2020). The obtained loss value is used to optimize the model's trainable weights and biases through the back propagation algorithm.

[0036] When the classification cross entropy loss Loss of the channel coding rate recognition model no longer decreases or reaches the maximum number of training iterations, the training is stopped, the optimal weights and biases are saved, and the optimized channel coding rate recognition model is obtained.

[0037] Step 4: The receiver verifies the sequence Y v The optimized channel coding rate recognition model is used for blind rate recognition. The whole process is implemented as follows:

[0038] 1) Enter the verification sequence Y for code rate identification v ;

[0039] 2) The bit rate recognition model calculates and outputs the normalized probability vector P v ;

[0040] 3) P v The bit rate category corresponding to the maximum probability is output as the discrimination result.

[0041] The present invention is further illustrated by the following specific examples.

[0042] Embodiment 1, a method for blindly identifying channel coding rate based on deep learning, the specific implementation steps are as follows:

[0043] Step 1: Channel coding uses the LDPC code with n=480 in the 802.22 standard. The transmitter and the receiver agree on an LDPC coding rate set θ={1 / 2, 2 / 3, 3 / 4, 5 / 6}, which includes 4 coding rates. Each code rate corresponds to its coding matrix G i and check matrix H i : G i The shapes are (480,240), (480,320), (480,360), (480,400), respectively. i The shapes are (480,240), (480,160), (480,120), (480,80) respectively. Four sets of training input data are generated under each signal-to-noise ratio. The signal-to-noise ratio ranges from -1dB to 7dB, the interval is 0.5dB, and the number of codewords in each set is 200. The above encoding method is used according to the corresponding r i ,i∈[1,4] generates the labels of the training data set. The amount of data in each round of training data set is 200×4×17=13600.

[0044] Step 2: The verification feature calculation module consists of 4 parallel calculation channels, each channel inputs a received symbol sequence, the number of input nodes of the feedforward layer of each channel is 240, 160, 120 and 80, and the number of output nodes is 1024; the convolution module consists of four two-dimensional convolution-activation function-batch normalization structures, the size of the convolution kernel is (1, 1), and the convolution step is 1; the feedforward classification module consists of three layers of feedforward layers. The model weights follow a normal distribution with a mean of 0 and a variance of 0.05, and the bias is initialized to 0.

[0045] Step 3: Input the model output and the training data set label into Loss to calculate the loss value, use the SGD algorithm to update the model weights and biases, use Adam as the optimizer, and set the learning rate to 0.001; the maximum number of training rounds is 200, and new inputs and labels are randomly generated after each round of training. When the model's training loss value no longer decreases or reaches the maximum number of training rounds, stop training, save the optimal weights and biases, and obtain an optimized bitrate recognition model.

[0046] Step 4: The receiver verifies the sequence Y v Use the optimized LDPC channel coding rate blind identification model to perform rate blind identification, and input the verification sequence Y that needs to be rate identified v , calculate the output normalized probability vector P v ', and obtain the recognition result r'.

[0047] According to the process of Example 1, 20,000-frame verification sequences are generated at each signal-to-noise ratio and each bit rate to verify the model recognition effect. Figure 2 This is the simulation curve for model verification. The horizontal axis is the signal-to-noise ratio, and the vertical axis is the bit rate recognition accuracy, ACC. Figure 2 The solid line represents the average LLR method rate blind recognition accuracy curve used for comparison, the star point represents r = 1 / 2, the circle point represents r = 2 / 3, the cross point represents r = 3 / 4, and the triangle point represents r = 5 / 6; the corresponding dotted line represents the rate blind recognition accuracy curve of the present invention. Figure 2 It can be seen that the LDPC code rate blind identification method based on deep learning proposed in the present invention has significantly improved performance in the low signal-to-noise ratio area compared with the average LLR method proposed in the article (see "Novel blind identification of LDPC codes using average LLR of syndrome a posteriori probability", IEEE Transactions on SignalProcessing, Vol. 62, No. 3, Sep. 2014").

[0048] Embodiment 2, a method for blindly identifying channel coding rate based on deep learning, the specific implementation steps are as follows:

[0049] Step 1: Channel coding uses the LDPC code with n=960 in the 802.16e standard. The transmitter and the receiver agree on an LDPC coding rate set θ={1 / 2, 2 / 3, 3 / 4, 5 / 6}, which includes 4 coding rates. Each code rate corresponds to its coding matrix G i and check matrix H i : G i The shapes are (960,480), (960,640), (960,720), (960,800), H i The shapes are (960,480), (960,320), (960,240), (960,160) respectively. Four sets of training input data are generated under each signal-to-noise ratio. The signal-to-noise ratio ranges from -1dB to 7dB, the interval is 0.5dB, and the number of codewords in each set is 200. The above encoding is used according to the corresponding r i ,i∈[1,4] generates the labels of the training data set. The amount of data in each round of training data set is 200×4×17=13600.

[0050] Step 2: The verification feature calculation module consists of 4 parallel calculation channels, each channel inputs the received symbol sequence Y, the number of input nodes of the feedforward layer of each channel is 480, 320, 240 and 160, and the number of output nodes is 1024; the convolution module consists of four two-dimensional convolution-activation function-batch normalization structures, the size of the convolution kernel is (1, 1), and the convolution step is 1; the feedforward classification module consists of three layers of feedforward layers. The model weights follow a normal distribution with a mean of 0 and a variance of 0.05, and the bias is initialized to 0.

[0051] Step 3: Input the model output and the training data set label into Loss to calculate the loss value, use the SGD algorithm to update the model weights and biases, use Adam as the optimizer, and set the learning rate to 0.001; the maximum number of training rounds is 200, and new inputs and labels are randomly generated after each round of training. When the model's training loss value no longer decreases or reaches the maximum number of training rounds, stop training, save the optimal weights and biases, and obtain an optimized bitrate recognition model.

[0052] Step 4: The receiver verifies the sequence Y v Use the optimized LDPC channel coding rate blind identification model to perform rate blind identification, and input the verification sequence Y that needs to be rate identified v , calculate the output normalized probability vector P v ', and obtain the recognition result r'.

[0053] According to the process of Example 2, 20,000-frame verification sequences are generated at each signal-to-noise ratio and each bit rate to verify the model recognition effect. Figure 3 This is the simulation curve for model verification. The horizontal axis is the signal-to-noise ratio, and the vertical axis is the bit rate recognition accuracy (ACC). Figure 3 The solid line represents the average LLR method rate blind recognition accuracy curve used for comparison, the star point represents r = 1 / 2, the circle point represents r = 2 / 3, the cross point represents r = 3 / 4, and the triangle point represents r = 5 / 6; the corresponding dotted line represents the rate blind recognition accuracy curve of the present invention. Figure 3 It can be seen that the LDPC code rate blind identification method based on deep learning proposed in the present invention has significantly improved performance in the low signal-to-noise ratio area compared with the average LLR method.

Claims

1. A method for blind identification of low-density parity-check code rate based on deep learning, Its characteristics are as follows: A neural network model is used to calculate the weighted checksum of the received sequence and perform feature processing and classification, and finally output its code rate category, which includes two stages: training and verification. The specific implementation steps are as follows: The transmitter and the receiver agree on a Low Density Parity Check LowDensity Parity Check, and the LDPC code rate parameter set θ = {r1, r2, ..., r i ,...,r t }, the number of corresponding codeword check nodes are {m1,m2,...,m i ,...,m t }; In the training phase, the transmitter first transmits a random information sequence S of length k through a code rate of r i The LDPC coding method generates a codeword sequence C of length n. The codeword sequence C generates an energy-normalized modulation symbol sequence X through the binary phase shift keying (BPSK) modulation method. The modulation symbol sequence X is transmitted to the receiver through the additive white Gaussian noise (AWGN) wireless channel. The mean of the noise distribution is 0 and the variance is σ. 2 ; The symbol sequence received by the receiver is Y as the training input; the training label is a vector P of length t. When the encoding rate of Y is r i When the value P of the vector at position i is i =1, and the rest of the positions are all 0; within the set signal-to-noise ratio range, generate equal amounts of data according to the above method to form a training data set; then input the training data set into the rate recognition neural network model, apply the Stochastic Gradient Descent (SGD) algorithm to train the model weights and biases, and finally obtain the optimized rate recognition neural network model; In the verification phase, the verification data set is generated using the same method as the training data set, and is verified using the optimized bit rate recognition model.

2. According to a method for blindly identifying low-density parity-check code rate based on deep learning according to claim 1, it is characterized in that: The code rate identification neural network model is composed of a check feature calculation module, a rearrangement and splicing module, a convolutional neural network module and a feedforward classification module; the check feature calculation module is composed of t parallel calculation channels, each of which includes a variable node layer, a check node layer and a feedforward layer; the rearrangement and splicing module is composed of t parallel rearrangement layers and a splicing layer; the convolutional neural network module is composed of a two-dimensional convolution layer, a pooling layer and a batch normalization layer; the feedforward classification module is composed of a feedforward layer and a random drop layer; the calculation process of the model is as follows: first, the check feature calculation module calculates the weighted checksum of the model input Y: calculate the channel log likelihood ratio Log Likelihood Ratio, LLR of each symbol in Y: In the ith channel of this module, use r i The corresponding check equation calculates the check node LLR: In the formula, tanh(.) is the hyperbolic tangent function, Π is the multiplication symbol, and tanh -1 (.) is the inverse hyperbolic tangent function, N i (c) is the set of variable nodes connected to the check node c; Input it into the feedforward layer. The number of input nodes of the feedforward layer of each channel is {m1,m2,...,m i ,...,m t }, the number of output nodes is n node ; In the rearrangement and concatenation module, t one-dimensional weighted checksum features are rearranged into two-dimensional features and concatenated into three-dimensional features along the first dimension; then the features are input into the convolutional neural network module for processing, and finally the features output by the previous module are classified in the feedforward classification module and the bit rate recognition probability vector P' is output.

3. According to a method for blindly identifying low-density parity-check code rate based on deep learning according to claim 1, it is characterized in that: The method for updating the weight and bias of the rate recognition model is as follows: after the received symbol sequence Y is input into the rate recognition model, the model outputs a normalized rate recognition normalized probability vector P' of length t, and the output is combined with the label to calculate the classification cross entropy loss Loss: Where P i is the i-th element of the label vector P, P i ' is the i-th element of the model output vector P', exp(.) is an exponential function; the SGD algorithm is used to update the weights and biases of the rate recognition model according to the value of the classification cross entropy loss Loss. When the maximum number of training iterations is reached or the value of Loss no longer changes, the training is stopped, and the optimal model weights and biases are saved to obtain the optimized rate recognition model.

4. The method for blindly identifying the rate of low-density parity-check codes based on deep learning according to claim 1, characterized in that: The verification phase will verify the sequence Y v Use the optimized bit rate recognition model to perform bit rate classification tasks. The specific implementation process is as follows: 1) Enter the verification sequence Y for code rate identification v ; 2) The bit rate recognition model calculates and outputs the normalized probability vector P v ; 3) P v The bit rate category corresponding to the maximum probability is output as the discrimination result.

Images