A Blind Identification Method for LDPC Code Rate Based on Deep Learning under Correlated Noise

Abstract

A blind LDPC code rate recognition method based on deep learning under correlated noise uses a multi-task learning strategy to achieve blind LDPC code rate recognition using deep learning (DL), and it consists of two stages: training and verification. In the training stage, LDPC codes with one code rate are selected from the code rate set to generate a codeword sequence. The codeword sequence is modulated by binary phase shift keying to generate a modulation sequence, and the modulation sequence is transmitted through a correlated noise channel to obtain a training data set. Then, a blind LDPC code rate recognition model under correlated noise is constructed using DL. Finally, the training data set is input into the model, and the model is trained using the adaptive moment estimation optimization algorithm. Training stops when the loss value converges or reaches the maximum number of training epochs to obtain an optimized model. In the verification stage, the modulation symbol sequence affected by correlated noise is used for blind LDPC code rate recognition using the trained model. The DL-based blind LDPC code rate recognition method of the present invention has better performance at low signal-to-noise ratios.

Description

Technical Field

[0001] The present invention is a method for blind recognition of LDPC code rate under relevant noise by using deep learning technology, belonging to the field of the combination of wireless communication and deep learning. Background Art

[0002] At present, with the rapid development of artificial intelligence technology represented by Deep Learning (DL), intelligent communication formed by the integration of DL technology and communication has become a new research hotspot in the communication field (see "Research Overview of Intelligent Wireless Communication Technology", Journal of Communications, vol. 41, no. 07, pp. 1-17, 2020). In recent years, the integration of DL technology and channel coding recognition technology has avoided the defects of the above traditional recognition algorithms during recognition, providing some new ideas for the development of channel coding recognition technology of LDPC codes.

[0003] Existing blind recognition of LDPC code rate is carried out under an Additive White Gaussion Noise (AWGN) channel. The article (see "SNR estimation for multi-dimensional cognitive receiver under correlated channel / noise", IEEE Transactions on Wireless Communications, vol. 12, no. 12, pp. 6392–6405, December. 2013) points out that due to operations such as filtering and oversampling, the noise faced in an actual wireless communication system may be correlated noise. In this channel, the noise is no longer completely independent but has a certain correlation, and this noise can be modeled (see "An Iterative BP-CNN Architecture for Channel Decoding", IEEE Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 144-159, Feb. 2018).

[0004] The algorithms designed based on the AWGN channel are applicable to the scenario where the noise follows independent and identically distributed. However, in a correlated noise channel, because the noise has a correlation and no longer satisfies independent and identically distributed, the algorithms designed based on the AWGN channel have problems of performance deterioration in the correlated noise channel. How to design a blind recognition algorithm for LDPC code rate in a correlated noise channel is the problem to be solved by the present invention. Summary of the Invention

[0005] The present invention proposes a blind identification method for the code rate of LDPC codes based on deep learning under correlated noise, which uses DL technology to construct a model for blind identification of the code rate of LDPC codes to solve the problem of poor performance of blind identification of the code rate of LDPC codes caused by correlated noise.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A blind identification method for the code rate of LDPC codes based on deep learning under correlated noise uses a neural network model to reduce correlated noise, and uses deep learning technology to perform feature processing and classification on the weighted checksum of the denoised sequence, and finally outputs the code rate category adopted by the transmitter; it includes two stages: training and verification, and the specific implementation steps are as follows: In the training stage, first select LDPC codes with a code rate of r R =(r1, r2, …, r q ) from the code rate set θ l (1 ≤ l ≤ q) as the channel coding scheme, encode the 0, 1 random sequence c = [c1, c2, …, c k of length k to generate a codeword sequence m = [m1, m2, …, m N of length N, then modulate m through BPSK to generate a modulation sequence x = [x1, x2, …, x N , and the modulation sequence x passes through a wireless channel containing correlated Gaussian noise n = [n1, n2, …, n N to become a received sequence y = [y1, y2, …, y N . Generate a large amount of data according to the above method to form a training data set; then use the training data set as the input of the blind identification model for the code rate of LDPC codes under correlated noise, apply the Adam optimization algorithm to train the network model, and stop training when the loss converges or reaches the maximum number of training epochs, and save the optimal network weights of the model; in the verification stage, use the same method as generating the training data set to generate a verification data set for optimizing the verification of the model.

[0008] The blind identification model for the code rate of LDPC codes mainly includes two parts: a denoising sub-network and a blind identification sub-network for the code rate of LDPC codes; this denoising sub-network is used to estimate the correlated noise received by the received sequence y and perform denoising processing on the received sequence y to obtain a denoised sequence y'; then input the denoised sequence y' into a demodulator to calculate the channel log-likelihood ratio Log-likelihoodRatio, the LLR value L ch , and input the LLR value into the blind identification sub-network for the code rate of LDPC codes. This network can obtain a code rate identification probability vector o = [o1,..., o q; among them, the calculation formula for each neuron in the check layer is as follows:

[0009]

[0010] In the formula, sgn() represents the sign function, min() represents the minimum value operation, and || represents the absolute value operation. represents the code rate r l the set of coding bit coordinates involved in the g-th check equation in the corresponding check matrix; L ch (e) represents L ch the LLR value corresponding to the e-th coding bit in, ω g,l is the corresponding training weight. is corresponding to the code rate r l the neuron output of the g-th check equation; the calculation formula of the pooling layer is

[0011]

[0012] In the formula, is the code rate r l the number of check equations in the corresponding check matrix, Γ l is the code rate r l the corresponding mean feature.

[0013] The noise reduction sub-network uses four fully connected layers to extract features of the relevant noise. Then, the estimated noise n' and the received sequence y form a residual structure, and the received sequence y minus the estimated noise n' can obtain the noise-reduced sequence y'.

[0014] Input the noise-reduced sequence y' into the demodulator to obtain L ch , and then input L ch into the LDPC code rate blind recognition sub-network to complete the LDPC code rate blind recognition task; for the two tasks of noise reduction and code rate recognition, a multi-task joint optimization strategy is used to update the network parameters, and its definition is as follows

[0015] Loss = λLoss1+(1 - λ)Loss2, 0 ≤ λ ≤ 1

[0016] In the above formula, λ represents the weight of the noise reduction loss Loss1 in the total loss, and the calculation formula of Loss1 is as follows, |||| F represents the Frobenius norm;

[0017]

[0018] The calculation formula of the code rate recognition loss Loss2 is as follows

[0019]

[0020] In the above formula, is the value corresponding to r in the code rate label vector. l

[0021] The specific implementation process of the verification process is as follows:

[0022] a: Input the verification sequence to be blindly identified

[0023] b: Output the normalized probability vector o through noise reduction and the LDPC code rate blind identification sub-network;

[0024] c: Use to obtain the code rate identification result, where argmax() represents the input value corresponding to the maximum value of o.

[0025] The present invention makes full use of the technical advantages of DL and the soft information characteristics of the check nodes of LDPC codes. Under correlated noise, the code rate identification performance of the present invention is better than that of the existing LDPC code rate blind identification algorithms at low signal-to-noise ratios, and the advantage becomes more obvious as the noise correlation increases. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a schematic diagram of the LDPC code rate blind identification model based on deep learning of the present invention;

[0027] Figure 2 is the LDPC code rate blind identification simulation and comparison results of the LDPC code with a code length of 648 in the IEEE 802.11n standard under the code rate set θ R ={1 / 2, 2 / 3, 3 / 4, 5 / 6}, where Figure 2 (a) and Figure 2 (b) have noise correlation coefficients of 0.5 and 0.9 respectively;

[0028] Figure 3 is the LDPC code rate blind identification simulation and comparison results of the LDPC code with a code length of 576 in the IEEE 802.16e standard under the code rate set θ R ={1 / 2, 2 / 3, 3 / 4, 5 / 6}, where Figure 3 (a) and Figure 3 (b) have noise correlation coefficients of 0.5 and 0.9 respectively. DETAILED DESCRIPTION OF THE INVENTION

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0030] Refer to Figure 1 ​, A blind recognition method for the code rate of LDPC codes based on deep learning under correlated noise, which uses a neural network model to reduce correlated noise and applies deep learning techniques to process and classify the weighted checksum of the denoised sequence, and finally outputs the code rate category adopted by the transmitter; it includes two stages: training and verification. The specific implementation steps are as follows: In the training stage, first select LDPC codes with code rate r R ={r1,r2,K,r q} from the code rate set θ l (1≤l≤q) as the channel coding scheme, encode the 0, 1 random sequence c=[c1,c2,K,c k of length k to generate a codeword sequence m=[m1,m2,K,m N of length N. Then modulate m through BPSK to generate a modulation sequence x=[x1,x2,K,x N . The modulation sequence x passes through a wireless channel with correlated Gaussian noise n=[n1,n2,K,n N to become a received sequence y=[y1,y2,K,y N . Generate a large amount of data according to the above method to form a training data set; then use the training data set as the input of the blind recognition model for the code rate of LDPC codes under correlated noise, and apply the Adam optimization algorithm to train the network model. Stop training when the loss converges or reaches the maximum number of training epochs, and save the optimal network weights of the model; In the verification stage, use the same method as generating the training data set to generate a verification data set for optimizing the verification of the model.

[0031] The blind recognition model for the code rate of LDPC codes mainly includes two parts: a denoising sub-network and a blind recognition sub-network for the code rate of LDPC codes; this denoising sub-network is used to estimate the correlated noise received by the received sequence y and perform denoising processing on the received sequence y to obtain a denoised sequence y′; then input the denoised sequence y′ into the demodulator to calculate the channel log-likelihood ratio Log-likeihoodRatio, the LLR value L ch , and input the LLR value into the blind recognition sub-network for the code rate of LDPC codes. This network can obtain a code rate recognition probability vector o=[o1,...,o q by passing the LLR value through the check layer, pooling layer, and softmax function in sequence; among them, the calculation formula of each neuron in the check layer is as follows:

[0032]

[0033] In the formula, sgn() represents the sign function, min() represents the minimum value operation, || represents the absolute value operation, represents the code rate r lThe set of coded bit coordinates involved in the g-th check equation in the corresponding check matrix; L ch (e) represents L ch The LLR value corresponding to the e-th coded bit in g,l ω is the corresponding training weight, For the neuron output of the g-th check equation corresponding to the code rate r l The calculation formula of the pooling layer is

[0034]

[0035] In the formula, is the code rate r l The number of check equations in the check matrix corresponding to, Γ l is the code rate r l The corresponding mean feature.

[0036] The denoising sub-network uses four fully connected layers to extract the features of the relevant noise. Then, the estimated noise n' and the received sequence y are constructed into a residual structure, and the received sequence y minus the estimated noise n' can obtain the denoised sequence y'.

[0037] Input the denoised sequence y' into the demodulator to obtain L ch , and then input L ch into the LDPC code rate blind recognition sub-network to complete the LDPC code rate blind recognition task; for the two tasks of denoising and rate recognition, a multi-task joint optimization strategy is used to update the network parameters, which is defined as follows

[0038] Loss = λLoss1 + (1 - λ)Loss2, 0 ≤ λ ≤ 1

[0039] In the above formula, λ represents the weight of the denoising loss Loss1 in the total loss. The calculation formula of Loss1 is as follows, |||| F represents the Frobenius norm;

[0040]

[0041] The calculation formula of the code rate recognition loss Loss2 is as follows

[0042]

[0043] In the above formula, is the value corresponding to r in the code rate label vector l .

[0044] The specific implementation process of the verification process is as follows:

[0045] a: Input the verification sequence to be blindly recognized

[0046] b: Verification sequence After noise reduction and LDPC code rate blind recognition sub-network, the normalized probability vector o is output;

[0047] c: Utilize To obtain the code rate recognition result, where argmax() represents the input value corresponding to the maximum value of o.

[0048] The present invention will be described in detail in the following four steps.

[0049] A blind recognition method for LDPC code rate based on deep learning under correlated noise, and the specific implementation steps are as follows:

[0050] Step 1: Select an LDPC code with code rate r from the code rate set θ R ={r1, r2, K, r q} to perform channel coding on a 0, 1 random information sequence of length k to generate a codeword sequence of length N. Then, modulate the codeword sequence through BPSK to generate a modulated symbol sequence x. The sequence x passes through a correlated noise channel to generate a received sequence y = x + n, and y is used as the input of the built model, where the specific modeling method of n is as follows: l n = Λ

[0051] n 1 / 2 ·n w

[0052] In the formula, Λ represents a correlation coefficient matrix of size N×N, and the element in its a-th row and b-th column can be calculated by the following formula, n w is a Gaussian white noise sequence with a mean of 0 and a variance of σ 2 , and σ 2 represents the noise power.

[0053]

[0054] In the formula, η ∈ is the correlation coefficient and satisfies |η| ≤ 1, and is the set of real numbers. From the definition formula of n, it can be seen that n is a correlated Gaussian noise sequence with a mean of 0, a variance of σ 2 and a correlation coefficient of η. The received signal-to-noise ratio defined in the present invention is the ratio of the energy per information bit to the noise power. According to the above process, a large number of y are generated under different correlation coefficients, different SNRs, and different code rates as the training data set of the model.

[0055] Step 2: The blind recognition model for LDPC code rate based on deep learning mainly consists of a noise reduction sub-network and an LDPC code rate blind recognition sub-network. The schematic diagram of the model structure is as Figure 1 shown.

[0056] The noise reduction sub-network consists of four fully connected layers. After the network estimates the noise n' by extracting features through the fully connected layers, subtracting the estimated noise n' from the received sequence y can obtain the noise-reduced sequence y'.

[0057] Then, the noise-reduced sequence y' is input into the demodulator and demodulated using the following formula to obtain the channel log-likelihood ratio of y' (see "Bitwise Log-likelihood Ratios For Quadrature Amplitude Modulations", IEEE Communications Letters, Vol. 19, No. 6, June 2015)

[0058]

[0059] Subsequently, the channel log-likelihood ratio is input into the LDPC code rate blind recognition sub-network. In this network, the channel log-likelihood ratio first passes through the check layer, and the calculation formula for each neuron in the check layer is as follows

[0060]

[0061] where sgn() represents the sign function, min() represents the minimum operation, || represents the absolute value operation, represents the code rate r l the set of coordinates of the coding bits involved in the g-th check equation in the corresponding check matrix; L ch (e) represents the LLR value corresponding to the e-th coding bit in the channel log-likelihood ratio, ω g,l is the corresponding training weight, corresponding to the code rate r l the neuron output of the g-th check equation.

[0062] Subsequently, it passes through the pooling layer, and the calculation formula of the pooling layer is

[0063]

[0064] where is the number of check equations in the check matrix corresponding to the code rate r l Γ l is the mean feature corresponding to the code rate r l .

[0065] Finally, Γ = [Γ1, Γ2,..., Γ q can obtain the code rate recognition probability vector o = [o1,..., o q through the softmax function, and perform The result of code rate recognition can be obtained.

[0066] Step 3: Use the training data set required to generate the model in Step 1, use the model required for the task generated in Step 2, set the hyperparameters for model training, and train the model using the Adam optimization algorithm (see "Deep Learning and MindSpore Practice", Tsinghua University Press, 2020) and perform joint optimization through the multi-task loss function, where the loss function is as follows

[0067] Loss = λLoss1+(1 - λ)Loss2, 0 ≤ λ ≤ 1

[0068] In the above formula, λ represents the weight of the noise reduction loss Loss1 in the total loss, and the calculation formula of Loss1 is as follows. In the formula, |||| F represents the Frobenius norm.

[0069]

[0070] The calculation formula of the code rate recognition loss Loss2 is as follows

[0071]

[0072] In the above formula, is the value corresponding to r in the code rate label vector l value.

[0073] When the multi-task loss function Loss converges or reaches the maximum number of training epochs, stop training to obtain the optimized model.

[0074] Step 4: Use the optimized model to perform blind recognition on the validation sequence. The overall process is as follows:

[0075] 4.1 Use the same method as in Step 1 to generate the test data set required for the validation phase, and use the optimized model to perform validation work on the test data set;

[0076] 4.2 Pass the validation data set through the optimized model to obtain the normalized probability vector o;

[0077] 4.3 Use the category corresponding to the maximum probability in o as the final code rate discrimination output result.

[0078] The present invention will be further described below through specific embodiments.

[0079] Embodiment 1, A method for blind recognition of LDPC code rate based on deep learning under correlated noise, the specific implementation steps are as follows:

[0080] Step 1: The channel coding method uses the LDPC code with a code length of 648 in the 802.11n standard, and the code rate set consists of 4 code rates (1 / 2, 2 / 3, 3 / 4, and 5 / 6). LDPC codes with different code rates are sequentially selected from the 4 code rate sets to perform channel coding on the information sequences of 0 and 1 to obtain codewords with a code length of 648, and the completed coded codeword sequences are modulated using BPSK to obtain a modulated symbol sequence x. The sequence x passes through a correlated noise channel to obtain a received sequence, where the noise correlation coefficients are selected as 0.5 and 0.9 respectively. The value range of the training signal-to-noise ratio is from -1 dB to 3 dB, with an interval of 1 dB. 6×10 4 codewords are generated as the training data set at each code rate and each signal-to-noise ratio.

[0081] Step 2: In the LDPC code rate blind recognition model based on deep learning, the denoising sub-network consists of four fully connected layers respectively, and the number of neurons in the four fully connected layers is 512, 256, 512, and 648 respectively. The denoising sub-network uses the rectified linear unit as the activation function, and the weights of each fully connected layer in this network follow a Gaussian distribution with a mean of 0 and a variance of 0.05, and the bias of each fully connected layer is initialized to all zeros. The training weight ω g,l of the check layer in the LDPC code rate blind recognition sub-network is initialized to all 1.

[0082] Step 3: The training data set generated in Step 1 is input into the model, and the model is trained using the Adam optimizer. Set the maximum number of training epochs to 100, the data batch size to 256, the learning rate to 0.001, and the proportion λ of the denoising loss Loss1 in the loss function Loss to be set to 0.5. When the maximum number of training epochs 100 is reached or the loss difference of the model is less than 0.0001 after 10 training epochs, save the optimal model weights and biases, and then the optimized LDPC code rate blind recognition model based on deep learning can be obtained.

[0083] Step 4: Use the optimized LDPC code rate blind recognition model based on deep learning to perform code rate blind recognition on the verification sequence, calculate the output normalized probability vector o, and obtain the recognition result r'.

[0084] According to the process of Embodiment 1, a verification data set of 20,000 codewords is generated at each signal-to-noise ratio and each code rate. The value range of the signal-to-noise ratio of the verification sequence changes with the change of the correlation coefficient. When the noise correlation coefficient is 0.5, the value range of the signal-to-noise ratio of the verification sequence is from -1 dB to 10 dB; when the noise correlation coefficient is 0.9, the value range of the signal-to-noise ratio of the verification sequence is from -5 dB to 10 dB. Figure 2 It is a simulation result diagram of the LDPC code with a code length of 648, where the abscissa is SNR and the ordinate is the code rate recognition accuracy. Figure 2The solid line is used to represent the simulation result curve of the existing average LLR algorithm (see "Novel blind identification of LDPC codes using average LLR of syndrome a posteriori probability", IEEE Transactions on Signal Processing, Vol. 62, No. 3, Sep. 2014), the short dashed line is used to represent the simulation result curve of the present invention, the circle points are used to represent the LDPC codes with a code rate of 1 / 2, and the triangular points, cross points and square points are used to represent the LDPC codes with code rates of 2 / 3, 3 / 4 and 5 / 6 respectively. Compared with the average LLR algorithm, the present invention effectively improves the performance of identifying the code rate of LDPC codes in a correlated noise channel, especially as the noise correlation increases, the performance advantage becomes more obvious.

[0085] Embodiment 2, a blind identification method for the code rate of LDPC codes based on deep learning under correlated noise, the specific implementation steps are as follows:

[0086] Step 1: The channel coding method uses the LDPC code with a code length of 576 in the 802.16e standard, and the code rate set consists of 4 code rates (1 / 2, 2 / 3, 3 / 4 and 5 / 6). The transmitter sequentially selects LDPC codes with different code rates from the 4 code rate sets to perform channel coding on the information sequences of 0 and 1 to obtain codewords with a code length of 576, and uses BPSK modulation on the completed coded codeword sequence to obtain a modulated symbol sequence x. The sequence x passes through a correlated noise channel to obtain a received sequence, where the noise correlation coefficients are respectively selected as 0.5 and 0.9. The value range of the training signal-to-noise ratio is from -1 dB to 3 dB, with an interval of 1 dB. 6×10 4 codewords are generated as the training data set at each code rate and each signal-to-noise ratio.

[0087] Step 2: In the blind identification model for the code rate of LDPC codes based on deep learning, the noise reduction sub-network is composed of four fully connected layers respectively, and the number of neurons in the four fully connected layers is 512, 256, 512 and 576 respectively. The noise reduction sub-network uses the rectified linear unit as the activation function, and the weights of each fully connected layer in this network follow a Gaussian distribution with a mean of 0 and a variance of 0.05, and the bias of each fully connected layer is initialized to all zeros. The training weight ω g,l of the parity check layer in the blind identification sub-network for the code rate of LDPC codes is initialized to all 1s.

[0088] Step 3: Input the training data set generated in Step 1 into the model, and use the Adam optimizer to train the model. Set the maximum number of training epochs to 100, the data batch size to 256, the learning rate to 0.001, and the proportion λ of the noise reduction loss Loss1 in the loss function Loss to 0.5. Save the optimal model weights and biases when the maximum number of training epochs 100 is reached or the loss difference of the model is less than 0.0001 after 10 training epochs, and then the optimized LDPC code rate blind recognition model based on deep learning can be obtained.

[0089] Step 4: Use the optimized LDPC code rate blind recognition model based on deep learning to perform code rate blind recognition on the verification sequence, calculate the output normalized probability vector o, and obtain the recognition result r'.

[0090] According to the process of Embodiment 2, a verification data set of 20,000 codewords is generated at each signal-to-noise ratio and each code rate. The value range of the signal-to-noise ratio changes with the change of the correlation coefficient. When the noise correlation coefficient is 0.5, the value range of the verification sequence signal-to-noise ratio is from -1 dB to 10 dB; when the noise correlation coefficient is 0.9, the value range of the verification sequence signal-to-noise ratio is from -5 dB to 10 dB. Figure 3 It is a simulation result graph of LDPC codes with a code length of 576, where the abscissa is SNR and the ordinate is the code rate recognition accuracy. Figure 3 In it, the solid line is used to represent the simulation result curve of the existing average LLR algorithm, the short dashed line is used to represent the simulation result curve of the present invention, the circle points are used to represent the LDPC codes with a code rate of 1 / 2, and the triangle points, cross points, and square points are used to represent the LDPC codes with code rates of 2 / 3, 3 / 4, and 5 / 6 respectively. Compared with the average LLR algorithm, the method proposed in the present invention effectively improves the performance of identifying the LDPC code rate in the correlated noise channel, especially as the noise correlation increases, the performance advantage becomes more obvious.

Claims

1. A blind recognition method for LDPC code rate based on deep learning under correlated noise, characterized in that: A neural network model is used to reduce the correlated noise, and deep learning technology is used to process and classify the weighted checksum of the denoised sequence, and finally the code rate category used by the transmitter is output; it includes two stages: training and verification. The specific implementation steps are as follows: In the training stage, first, a code rate set θ containing q code rates is selected. R ={r1, r2, ..., r q }Select the code rate as r l The LDPC code (1≤l≤q) is used as the channel coding scheme. For a random sequence of 0 and 1 with length k, c=[c1, c2, ..., c k ] is encoded to generate a codeword sequence m=[m1,m2,...,m N Then m is modulated by Binary Phase Shift Keying (BPSK) to generate the modulation sequence x = [x1, x2, ..., x N ], the modulation sequence x is passed through a series of correlated Gaussian noise n = [n1, n2, ..., n N ] becomes the received sequence y = [y1, y2, ..., y N ], a large amount of data is generated according to the above method to form a training data set; then the training data set is used as the input of the LDPC code rate blind recognition model under correlated noise, and the Adam optimization algorithm is used to train the network model. When the loss converges or the maximum training rounds are reached, the training is stopped and the optimal network weights of the model are saved; In the verification stage, a verification dataset is generated using the same method as the generation of the training dataset for optimizing the verification of the model; the LDPC code rate blind recognition model mainly consists of two parts: a noise reduction sub-network and an LDPC code rate blind recognition sub-network; the noise reduction sub-network is used to estimate the relevant noise received by the received sequence y and perform noise reduction processing on the received sequence y to obtain the denoised sequence y'; then the denoised sequence y' is input into the demodulator to calculate the channel log-likelihood ratio Log-likeihoodRatio, the LLR value L ch , and the LLR value is input into the LDPC code rate blind recognition sub-network, which can obtain the code rate recognition probability vector o = [o1,..., o q by passing the LLR value through the check layer, pooling layer, and softmax function in sequence; among them, the calculation formula of each neuron in the check layer is as follows: where sgn(·) represents the sign function, min(·) represents the minimum operation, and |·| represents the absolute value operation. represents the code rate r l the set of coordinates of the coding bits involved in the g-th parity-check equation in the corresponding parity-check matrix; L ch (e) represents L ch the LLR value corresponding to the e-th coding bit in; ω g,l is the corresponding training weight, is the neuron output corresponding to the g-th parity-check equation for the code rate r l The calculation formula of the pooling layer is wherein, is the code rate r l the number of parity check equations in the corresponding parity check matrix, Γ l is the code rate r l corresponding mean feature.

2. A blind recognition method for LDPC code rate based on deep learning under correlated noise according to claim 1, characterized in that: The noise reduction sub-network uses four fully connected layers to extract features of the relevant noise. Then, the estimated noise n′ and the received sequence y are used to form a residual structure, and the received sequence y is subtracted by the estimated noise n′ to obtain the noise-reduced sequence y′.

3. A blind recognition method for LDPC code rate based on deep learning under correlated noise according to claim 2, characterized in that: Input the denoised sequence y′ into the demodulator to obtain L ch , and then input L ch into the LDPC code rate blind recognition sub-network to complete the LDPC code rate blind recognition task; for the two tasks of denoising and code rate recognition, a multi-task joint optimization strategy is used to update the network parameters, which is defined as follows Loss = λLoss1 + (1 - λ)Loss2, 0 ≤ λ ≤ 1 In the above formula, λ represents the weight of the noise reduction loss Loss1 in the total loss. The calculation formula of Loss1 is as follows, ||·|| F represents the Frobenius norm; The calculation formula for the code rate recognition loss Loss2 is as follows In the above formula, is the value corresponding to r in the code rate label vector l .

4. A blind recognition method for LDPC code rate based on deep learning under correlated noise according to claim 1, characterized in that: The specific implementation process of the verification process is as follows: a: The verification sequence to be blindly recognized for input b: Verification sequence The probability vector o is output after noise reduction and by the LDPC code rate blind recognition sub-network; c: By using obtain the code rate recognition result, where argmax(·) represents the input value corresponding to the maximum value of o.

Images