Self-adaptive LDPC (Low Density Parity Check) decoding method based on model driving
By constructing a model-driven adaptive LDPC decoder, combining multiple enhancement factors and residual connections, the decoding method of LDPC code is optimized, solving the problems of decoding complexity and performance decay, and achieving efficient bit error rate performance.
Patent Information
- Application Number
- CN202411077358.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2024-11-29
AI Technical Summary
The existing LDPC coding and decoding method based on the minimum sum (MS) algorithm has performance degradation in improving the decoding complexity, and the training complexity of neural networks is high, resulting in slow convergence of iterative decoding when the code length increases. During the training process Gradient disappearance problems are prone to occur.
Adaptive LDPC decoding method based on model driver is adopted, and the model-driven decoder with a connection structure consistent with the Tanner diagram is constructed, and the prototype LDPC code is expanded using multiple enhancement factors to train the network to obtain the final model-driven decoding. The machine, considering the normalization factor and the offset correction factor, introduces residual connections to optimize the iterative process.
Reduces the decoding complexity, improves the decoding accuracy, reduces the training complexity of neural networks, enhances the generalization ability of different code lengths and code rates, avoids the problem of gradient disappearance, and optimizes the errors of communication systems. Bitrate performance.
Smart Images

Figure BDA0004982712790000031 
Figure BDA0004982712790000032 
Figure BDA0004982712790000034
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of channel decoding, and in particular relates to an adaptive LDPC decoding method based on model driving. Background Art
[0002] The purpose of a complete communication system is to send the information people need from the source to the destination through the channel. How to combat the influence of noise in the channel and improve the transmission capacity and reliability of the communication system during the whole process has been a long-term research direction in the field of communication. Obviously, increasing the signal transmission power at the transmitter and adding the relay power of the signal in the channel can improve the signal to noise ratio (SNR), but the application of this method requires changing the existing communication network architecture, which is costly, and high-power signals will also introduce new noise to a certain extent. Forward Error Correction (FEC), as an error control code (ECC), is a coding reorganization of the original information before entering the channel under certain rules. Compared with the state before coding, it changes the relatively independent and weakly correlated state between codewords, so that the code elements have a specific correlation. Through this property, the deviation of the signal generated during transmission can be checked and corrected at the receiving end, so it can effectively reduce the impact of error information on the receiving end, thereby improving the transmission efficiency of the system. FEC adds redundancy during information transmission and uses the remaining bits to recover the data affected by noise at the receiving end. It can effectively resist channel noise and avoid changes to the system architecture. Among many channel coding technologies, low-density parity check (LDPC) code has become a mainstream technology due to its efficient coding method and excellent decoding performance close to the Shannon limit. It is widely used in a variety of modern communication systems, such as Wi-Fi, optical communication systems and 5G networks. In modern LDPC codes, the check matrix generally adopts a quasi-cyclic (QC) structure. Codes with this structure are called QC-LDPC codes. At present, most of the QC-LDPC codes with excellent performance are constructed by extending the prototype graph. Therefore, this type of QC-LDPC code constructed by extending the prototype graph is also called Protograph Based LDPC (PB-LDPC) code. Protograph LDPC expands a smaller matrix or base graph (BG) into a full matrix or graph. The prototype graph LDPC code has a good structural design, can achieve better performance, and is more suitable for efficient decoding implementation. Among the decoding methods of LDPC codes, the SP (Sum-Product) algorithm can achieve the best performance, but its high decoding complexity hinders its practical application. In order to solve this problem, simplified algorithms based on SP have been widely studied.Among them, from the perspective of the balance between performance and complexity, the most prominent algorithms are two improved versions of the simplified Min-Sum (MS) algorithm, the Normalized MS (NMS) algorithm and the Offset MS (OMS) algorithm. The basic idea of the MS algorithm is to replace the complex multiplication operation in the SP algorithm with a simple comparison operation to reduce the complexity of the algorithm. The MS algorithm can greatly reduce the implementation complexity of the decoding algorithm, but it also brings the problem of performance degradation. To address this problem, by introducing normalization factors and offset correction factors, the NMS algorithm and the OMS algorithm have achieved performance improvement on the basis of the MS algorithm. Due to their low implementation complexity and excellent decoding performance, the NMS algorithm and the OMS algorithm have been applied to many practical systems. In recent years, machine learning has developed rapidly in the field of communications, and has achieved impressive performance improvements in channel estimation, decoding, and other aspects. Similarly, many researchers have conducted research experiments on NMS and OMS decoders based on machine learning and achieved good performance. However, all of these existing studies have studied the normalization factor α of NMS and the offset correction factor β of OMS separately, which makes the approximate accuracy of the improved MS algorithm not good enough, and there are many limitations in terms of LDCP code rate compatibility and neural network training complexity. Specifically, as the code length increases, the dimension of the parity check matrix also increases, thereby increasing the parameter array of the neural decoder, resulting in extremely high training complexity. In addition, the iterative decoding converges more slowly for longer code lengths, which requires more iterations, such as 50 iterations or even more. This means that a fairly deep neural network is required, which may lead to the problem of gradient vanishing during training. Summary of the invention
[0003] In view of the above shortcomings, the present invention provides an adaptive LDPC decoding method based on model driving.
[0004] A model-driven adaptive LDPC decoding method of the present invention comprises the following steps:
[0005] Step 1: Based on the base graph of the prototype LDPC code and the improved decoding algorithm, construct a connection structure and Tanner Figure 1 A consistent model driven decoder.
[0006] Step 2: Use multiple lifting factors to expand the same PB-LDPC base graph into multiple check matrices to obtain multiple bit sequences with different code rates and code lengths. The bit sequences are modulated by BPSK and then pass through an additive white Gaussian noise (AWGN) channel. The final log-likelihood ratio (LLR) is used as the data set for training the network.
[0007] Step 3: Preprocess the LLR data set obtained in step 2 and input it into the model-driven decoder constructed in step 1, and train the two correction factors simultaneously; after the training, the final model-driven decoder is obtained.
[0008] The model-driven decoder is constructed based on the base matrix of the prototype graph LDPC code. The input layer of the decoding network receives the log-likelihood ratio information passing through the channel. The number of hidden layers depends on the number of iterations of the iterative decoding algorithm. Each iteration corresponds to a hidden layer. Each hidden layer has two sub-layers, which are used for checking nodes to pass messages to variable nodes and for variable nodes to pass messages to check nodes. The neuron connections between the sub-layers are determined by the connection relationship of the Tanner graph of the check matrix expanded from the base matrix; the output layer judges the log-likelihood ratio as bits.
[0009] Furthermore, step 2 is specifically as follows:
[0010] Step 2.1: Assume that the expansion multiple is Z. During the expansion process, the variable nodes, check nodes, and edges will be copied Z times. In order to ensure that the check graph has the structure of the prototype graph and has a quasi-cyclic form, the copied edges will be cyclically replaced between the variable nodes and check nodes of the same type. The nodes of the same type originate from the nodes of the prototype graph before expansion.
[0011] Step 2.2: Directly encode the check matrix H through the lower triangular check structure. For a check matrix H without any check structure, add and permute its rows to make it have a lower triangular coding structure. The resulting matrix H is as follows:
[0012]
[0013] Among them, h m,k+m-1 is the element in the mth row and k+m-1th column of the H matrix.
[0014] Step 2.3: Assume that the first k bits of the encoded bit sequence w are information bits, which are directly assigned by u; the last m bits of the bit sequence w are redundant bits, which are gradually obtained by the check matrix H and the existing bits in the sequence w. The formula is as follows:
[0015]
[0016] Among them, the coded bits w1,w2,…,w k is known, and is composed of the information bit sequence u1,u2,…,u k Assignment is obtained; for the first redundant check bit w k+1 , which is obtained by multiplying the first k elements of the first row of the check matrix with the first k known coded bits and then adding them together; the first redundant bit w is obtained k+1 After that, the second redundant check bit is calculated using this already obtained bit; the calculation method of the second redundant bit is similar to the first one, which is obtained by multiplying the first k+1 elements of the second row of the check matrix with the known first k+1 coding bits and finally adding them; and so on, gradually obtain m redundant bits and complete the coding bit w.
[0017] Step 2.4: Send n coded bits w into the AWGN channel with a channel variance of σ 2 , the receiving end receives n signals r j ,j=1,2,…,n, converted to log-likelihood ratio Its specific definition is as follows:
[0018]
[0019] Among them, Pr(w j =0|r j ) and Pr(w j =1|r j ) are the known received signals r j In the case of j = 0 and w j =1; the log-likelihood ratio is the logarithm of the ratio of probabilities; the log-likelihood ratio represents the coded bit w j How likely is it to be 0 or 1; log likelihood ratio The larger the amplitude, the more coded bits w j The higher the certainty of being 0 or 1, the higher the certainty of being 0 or 1. If the log-likelihood ratio satisfies the condition Then Pr(w j =0|r j )>Pr(w j =1|r j ), which means that the original coded bit is more likely to be 0, and vice versa, it is more likely to be 1; thus, the LLR information under the lifting factor Z is obtained.
[0020] Step 2.5: Change the lifting factor Z, and then repeat the above steps to obtain multiple sets of LLR information under different code rates and code lengths. These LLR information constitute the training set.
[0021] Furthermore, the specific training process in step 3 is as follows:
[0022] Step 3.1: Initialize the normalization factor α of the i-th layer of the network (i) and the offset correction β (i) .
[0023] Step 3.2: Randomly select LLR data obtained by different lifting factors from the training set to improve the generalization of the decoder to different code lengths and bit rates, and load the data into the input layer of the network.
[0024] Step 3.3: The activation function of the output layer is sigmoid. After obtaining the output result of the network output layer, the cross entropy loss function is used to calculate the error. The cross entropy formula is as follows:
[0025]
[0026] Among them, v and w v are the output of the network and the vth element of the encoded information respectively, and N is the total number of data.
[0027] Step 3.4: Introduce residual connection to stop iterative decoding in advance when the decoding requirements are met, which is consistent with the early stopping mechanism of the traditional decoding algorithm; the model-driven decoder uses Adam as the optimizer.
[0028] Furthermore, the model-driven decoder hyperparameters are set as follows: the decoding network is iterated 20 times in total; the batch size for each iteration is 10,000, with 200 samples in each batch; the optimizer uses Adam; and the learning rate is 0.001.
[0029] The beneficial technical effects of the present invention are:
[0030] The present invention can optimize the traditional decoding methods NMS and OMS, and consider the normalization factor and the offset correction factor at the same time. The proposed multi-correction factor MS (MCMS) algorithm makes the decoding result of the LDPC code more accurate and optimizes the bit error rate of the communication system. The correction factor is learned using a model-driven decoder to obtain the optimal normalization factor and offset correction factor to achieve the best decoding performance. In addition, the excellent structure of the prototype graph LDPC is used to make the parameters in the neural network shareable, greatly reducing the training complexity. Finally, the training data is randomly selected during the network training process, which enhances the generalization ability of the neural network for different code lengths and code rates. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 This is an example diagram of the process of expanding a prototype diagram into a verification diagram.
[0032] Figure 2 This is a network structure diagram of the model-driven prototype LDCP decoding network of the present invention.
[0033] Figure 3 Distribution of normalized factors trained for the model-driven decoder.
[0034] Figure 4 Distribution of offset correction factors trained for the model-driven decoder.
[0035] Figure 5 Comparison of the BER performance of various decoding algorithms for the code type with base matrix BG1 in AWGN channel for improvement factor Z=3 (a) and Z=16 (b), where the number of iterations is I=25.
[0036] Figure 6 When the base matrix is BG1, BER=10 is achieved in AWGN channel. -4 or 10 -6 Required SNR, (a) represents the code (198,132) with Z=3, (b) represents the code (1056,704) with Z=16. DETAILED DESCRIPTION
[0037] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0038] A model-driven adaptive LDPC decoding method of the present invention is specifically as follows:
[0039] 1. First, build a model-driven adaptive LDPC code decoding network.
[0040] After the prototype graph is determined, the prototype graph is expanded to construct a check matrix of the PB-LDPC code with a quasi-cyclic structure. Assume that the expansion multiple is Z. During the expansion process, the variable nodes, check nodes and edges will be copied Z times. In order to ensure that the check graph has the structure of the prototype graph and has a quasi-cyclic form, the copied edges will be cyclically permuted between variable nodes and check nodes of the same type. Nodes of the same type originate from the nodes of the prototype graph before expansion. From a matrix perspective, the construction is the process from the base matrix to the check matrix. According to the elements in the base matrix, the cyclic matrix is filled in the corresponding positions of the check matrix to form the check matrix of the PB-LDPC code. Figure 1 shows this process.
[0041] Based on the above-mentioned extended prototype graph, the final Tanner graph is obtained, and the network structure of the decoder corresponds to the information transmission relationship between the variable node and the check node in the Tanner graph. The network structure constructed in this way is obtained based on the mathematical model of the prototype graph LDPC code, and the relationship between the algorithm and the network topology is fully considered, so that its structure can be accurately interpreted and predicted. This is the model-driven deep learning method, which not only compensates for the defects caused by inaccurate models and predetermined parameters, but also has many advantages such as less training data, reduced overfitting risk, and rapid implementation. The present invention also introduces residual connections to stop iterative decoding in advance when the decoding requirements are met, which is consistent with the early stopping mechanism of the traditional decoding algorithm. And in the process of deeper network training, the gradient can be kept stable to avoid the problem of gradient disappearance or gradient explosion.
[0042] Figure 2 The structure of the decoding network is shown. Specifically, the size of the neural network input layer is N, where N is the code block length, i.e., the number of variable nodes in the Tanner graph. The size of all hidden layers in the network is E, where E represents the number of edges in the Tanner graph, i.e., the number of elements 1 in the check matrix. Each hidden layer consists of two sublayers, i v and i c , corresponding to VN update and CN update respectively, i is the serial number of the hidden layer. The neuron connection relationship between sublayers is consistent with the connection relationship between the check node and the variable node in the Tanner graph.
[0043] The network loss function is the cross entropy function, the optimizer is Adam, and the learning rate is 0.001.
[0044] The forward propagation, activation function, loss function, and operation of neurons in the hidden layer involved in the specific network structure are as follows:
[0045] In the forward propagation, the specific algorithm for transmitting messages between neurons VN and CN is to set the maximum number of iterations I, the i-th hidden layer, i = 1, 2, 3, ..., I, and e = (v, c) is the number of sublayers i v The output message of this neuron is expressed as:
[0046]
[0047] Where l is the LLR information, and
[0048] Sublayer i c The output information expression of the neuron e=(v,c) is:
[0049]
[0050] Among them, sgn is the sign function, is the normalization factor parameter of the neuron in the i-th layer e=(v,c), is the offset correction factor parameter of the neuron in the i-th layer e = (v, c), and ReLu(·) is the activation function, which is defined as follows:
[0051] ReLU(x)=max(x,0)
[0052] Compared with the traditional neural network decoder with only normalization factor or offset correction factor, the model-driven decoding proposed in the present invention considers both factors at the same time and can better compensate for the performance loss of the MS algorithm. The output information of the vth neuron in the output layer is v As follows:
[0053]
[0054] Among them, σ(·) is the sigmoid function, which is used as the activation function of the output layer and is specifically defined as follows:
[0055] σ(x)=(1+exp(-x)) -1
[0056] The cross entropy loss function of the network is:
[0057]
[0058] Among them, w v is the vth element of the encoded information, and N is the total number of input data.
[0059] 2. Build the training data set and train the network.
[0060] In order to construct the data set, the present invention first expands the base graph. Here, the base graph BG1 in the 5G standard is selected, and its matrix dimension is 46×68. All subsequent expansions use this base graph. The base graph is expanded using the four lifting factors Z=3, 6, 10, and 16 to obtain four check matrices. If the check matrix is:
[0061]
[0062] Among them, h m,k+m-1 is the element in the mth row and k+m-1th column of the H matrix, then the bit sequence w is encoded using the following formula.
[0063]
[0064] Among them, the coded bits w1,w2,…,w k is known, and is composed of the information bit sequence u1,u2,…,u kAssign the value to obtain. For the first redundant check bit w k+1 , which is obtained by multiplying the first k elements of the first row of the check matrix with the first k known coded bits and then adding them together. Then the coded data w is passed through the AWGN channel to obtain the LLR data.
[0065]
[0066] Among them, σ 2 is the channel variance, Pr(w j =0|r j ) and Pr(w j =1|r j ) are the known received signals r j In the case of j = 0 and w j =1, n is the total number of input data. The coded data and LLR data generated by the four lifting factors are saved separately, and the network data set is obtained.
[0067] 3. A decoder driven by a training data set and a label training model constructed using the LLR of the received signal and its corresponding bit sequence information label.
[0068] The training process of the network decoder is as follows:
[0069] Step 1: Randomly select one data from the four sets of received signal LLR data for training.
[0070] Step 2: Calculate the loss value of the network decoded data and the label bit sequence through the loss function, and then use the Adam optimizer to adjust the network parameters according to the loss value.
[0071] Step 3: Use another set of received signal LLRs that have not been learned as a verification set to compare whether the bit error rate of the verification result reaches the expected bit error rate.
[0072] Step 4: Introduce residual connections to stop iterative decoding in advance when the decoding requirements are met, which is consistent with the early stopping mechanism of traditional decoding algorithms. In addition, it can keep the gradient stable during deeper network training and avoid the problem of gradient vanishing or gradient exploding.
[0073] Step 5: Record the network parameters and end the training.
[0074] 4. Verify the performance of the model-driven adaptive LDPC decoder.
[0075] The received signal LLR information is decoded into a bit sequence by a model-driven decoder, and the decoded bit information is compared with the label bit information to obtain the signal bit error rate. In order to compare the improvement of decoding performance by the model-driven adaptive LDPC code decoder, the present invention compares the optimization method with other traditional decoding methods. These include the Sum-Product (SP) decoding algorithm, the Minimum Sum (MS) decoding algorithm, the Normalized Minimum Sum (NMS) decoding algorithm, and the Offset Min-Sum (OMS) decoding algorithm.
[0076] The model-driven prototype graph LDPC decoder training optimized in the present invention is carried out in Windows 10 system, the programming language is Python 3.9, and the deep learning framework is Tensorflow 2.6. The algorithm runs on the GPU, and the graphics card model is NVIDIA 4080.
[0077] The training parameters are as follows: learning rate is 0.001, number of training times is 10000, and batch size is 200. The initial values of normalization factor and offset correction factor are set to 0.8 and 0.2 respectively. The neural network is constructed according to BG1. The initial log-likelihood ratio training data is obtained by BPSK modulated signal through AWGN channel. The training results are shown in Figure 3 and Figure 4 As shown in FIG. 1 , the distribution of the trained normalization factor and the offset correction factor are shown in FIG. 1 . Figure 5The decoding performance of five decoding algorithms, SP, MS, NMS, OMS, and MCMS, for (198, 132) and (1056, 704) BG1 codes with code rate 2 / 3 (with boosting factors Z=3 and Z=16, respectively) is shown. The maximum number of iterations is set to I=25. As shown in the figure, the model-driven MCMS decoder performs better at shorter information bit lengths. This may be because for short codes, the performance is more affected by the correlation of message passing between nodes, while for longer codewords, the performance is more constrained by the decoding threshold. The MS algorithm simplifies the tanh operation, so the corresponding decoding threshold will also deteriorate. Moreover, for long codes, the correlation of message passing between nodes will be weaker due to the larger circle length of the check matrix. For longer information bit lengths, the performance of the model-driven MCMS decoder is better than that of the OMS and NMS algorithms, but there is still a certain gap between it and the SP algorithm. For shorter information bit lengths, the optimized decoder shows better performance. Compared with the OMS algorithm and the NMS algorithm, it has a significant performance improvement. In addition, when the signal-to-noise ratio is greater than 5dB, the model-driven MCMS even exceeds the SP algorithm. The reason for exceeding the BP algorithm is that when the code length is short, the model-driven MCMS algorithm can effectively reduce the negative impact of the correlation between the messages transmitted by the nodes.
[0078] Figure 6 It shows that in AWGN channel, various decoders can achieve BER=10 at a certain number of iterations. -4 or 10 -6 Required SNR. Figure 6 (a) When Z = 3, BER = 10 -4 The required signal-to-noise ratio, the performance of the model-driven MCMS decoder is more than 0.5dB better than the traditional MS algorithm and is similar to the performance of the SP decoding algorithm. Figure 6 (a) also shows that BER=10 is achieved -6 Required SNR,In this high SNR region, the model-driven MCMS decoder can outperform the SP decoder even in early iterations, because the model-driven decoding network converges faster. Figure 6 (b) shows that when Z = 16, BER = 10 is achieved. -6 The required SNR. It can be seen that the model-driven MCMS decoder is still better than the MS, NMS, and OMS algorithms, and slightly inferior to the SP decoding. At high SNR, it can be observed that the performance of the model-driven MCMS decoder and the SP decoding algorithm are similar, which shows the superiority of the model-driven MCMS decoder in the high SNR area.
[0079] In summary, the model-driven adaptive LDPC decoding method proposed in the present invention improves the traditional decoding algorithm, reduces the decoding complexity and enhances the decoding accuracy. A single model can have good decoding performance for codes with different code lengths and code rates after one training, so that better bit error rate performance can be obtained in complex and changeable transmission systems.
Claims
1. A model-driven adaptive LDPC decoding method, characterized in that: The following steps are involved: Step 1: Based on the base graph of the prototype LDPC code and the improved decoding algorithm, a model-driven decoder with a connection structure consistent with the Tanner graph is constructed; Step 2: Use multiple lifting factors to expand the same PB-LDPC base graph into multiple check matrices to obtain multiple bit sequences with different code rates and code lengths. The bit sequences are modulated by BPSK and then pass through an additive white Gaussian noise channel. The final LLR is used as the data set for training the network. Step 3: Preprocess the LLR data set obtained in step 2 and input it into the model-driven decoder constructed in step 1, and train the two correction factors simultaneously; after the training, the final model-driven decoder is obtained.
2. The method of adaptive LDPC decoding based on model driving according to claim 1, characterized in that: The model-driven decoder is constructed based on the base matrix of the prototype graph LDPC code. The input layer of the decoding network receives the log-likelihood ratio information through the channel. The number of hidden layers depends on the number of iterations of the iterative decoding algorithm. Each iteration corresponds to a hidden layer. Each hidden layer has two sublayers, which are used for the check node to transmit messages to the variable node and the variable node to transmit messages to the check node. The neuron connection between the sublayers is determined by the connection relationship of the Tanner graph of the check matrix expanded from the base matrix. The output layer converts the log-likelihood ratio into bits.
3. The method of adaptive LDPC decoding based on model driving according to claim 1, characterized in that: The step 2 is specifically as follows: Step 2.1: Assume that the expansion multiple is Z. During the expansion process, the variable nodes, check nodes, and edges will be replicated Z times. In order to ensure that the check graph has the structure of the prototype graph and has a quasi-cyclic form, the replicated edges will be cyclically replaced between the variable nodes and check nodes of the same type. Nodes of the same type originate from the nodes of the prototype graph before expansion; Step 2.2: Directly encode the check matrix H through the lower triangular check structure. For a check matrix H without any check structure, add and permute its rows to make it have a lower triangular coding structure. The resulting matrix H is as follows: Among them, h m,k+m-1 is the element in the mth row and k+m-1th column of the H matrix; Step 2.3: Assume that the first k bits of the encoded bit sequence w are information bits, which are directly assigned by u; the last m bits of the bit sequence w are redundant bits, which are gradually obtained by the check matrix H and the existing bits in the sequence w. The formula is as follows: Among them, the coded bits w1,w2,…,w k is known, and is composed of the information bit sequence u1,u2,…,u k Assignment is obtained; for the first redundant check bit w k+1 , which is obtained by multiplying the first k elements of the first row of the check matrix with the first k known coded bits and then adding them together; the first redundant bit w is obtained k+1 After that, the second redundant check bit is calculated using the bit that has been obtained. The calculation method of the second redundant check bit is similar to the first one. It is obtained by multiplying the first k+1 elements of the second row of the check matrix with the first k+1 known coded bits and then adding them together. In this way, m redundant bits are gradually obtained and the coded bit w is completed. Step 2.4: Send n coded bits w into the AWGN channel with a channel variance of σ 2 , the receiving end receives n signals r j ,j=1,2,…,n, converted to log-likelihood ratio Its specific definition is as follows: Among them, Pr(w j =0|r j ) and Pr(w j =1|r j ) are the known received signals r j In the case of j = 0 and w j =1; the log-likelihood ratio is the logarithm of the ratio of probabilities; the log-likelihood ratio represents the coded bit w j How likely is it to be 0 or 1; log likelihood ratio The larger the amplitude, the more coded bits w j The higher the certainty of being 0 or 1, the higher the certainty of being 0 or 1. If the log-likelihood ratio satisfies the condition Then Pr(w j =0|r j )>Pr(w j =1|r j ), which means that the original coded bit is more likely to be 0, and vice versa. Thus, the LLR information under the lifting factor Z is obtained; Step 2.5: Change the lifting factor Z, and then repeat the above steps to obtain multiple sets of LLR information under different code rates and code lengths. These LLR information constitute the training set.
4. The method of adaptive LDPC decoding based on model driving according to claim 1, characterized in that: The specific training process in step 3 is as follows: Step 3.1: Initialize the normalization factor α of the i-th layer of the network (i) and the offset correction β (i) ; Step 3.2: Randomly select LLR data obtained by different lifting factors from the training set to improve the generalization of the decoder to different code lengths and bit rates, and load the data into the input layer of the network; Step 3.3: The activation function of the output layer is sigmoid. After obtaining the output result of the network output layer, the cross entropy loss function is used to calculate the error. The cross entropy formula is as follows: Among them, v and w v are the output of the network and the vth element of the encoded information, respectively, and N is the total number of data; Step 3.4: Introduce residual connection to stop iterative decoding in advance when the decoding requirements are met, which is consistent with the early stopping mechanism of the traditional decoding algorithm; the model-driven decoder uses Adam as the optimizer.
5. The method of adaptive LDPC decoding based on model driving according to claim 1, characterized in that: The model-driven decoder hyperparameters are set as follows: the decoding network is iterated 20 times in total; the batch size for each iteration is 10,000, and each batch has 200 samples; the optimizer uses Adam; and the learning rate is 0.001.
Citation Information
Cited By
QC-LDPC code double-branch cross layered minimum sum decoding method based on neural network optimization
CN121485702A
