QC-LDPC code double-branch cross-layer min-sum decoding method based on neural network optimization

Abstract

This invention discloses a dual-branch cross-layered minimum sum decoding method for QC-LDPC codes based on neural network optimization. For the input QC-LDPC code and two layer sorting methods, upper and lower branch layered minimum sum decoders based on the two sorting methods are constructed. The optimal parameters of the upper and lower branch layered minimum sum decoders are obtained through neural network optimization, and the minimum sum decoders are instantiated separately. During decoding, the received signal is simultaneously input into the upper and lower branch decoders. If decoding of either branch is successful, the decoding result is output; otherwise, the original LLR sequence value is weighted and correlated with the posterior LLR sequence value output from the first decoding of the two decoders, and then cross-input into the lower and upper branch decoders for further decoding. This invention can effectively improve the layered decoding performance of QC-LDPC codes. For typical 5G-LDPC codes, this invention shows a significant performance improvement compared to the neural network parameter-optimized layered minimum sum decoding method.

Description

Technical Field

[0001] This invention relates to the field of channel coding and decoding technology, specifically to a dual-branch cross-layered minimum sum decoding method for QC-LDPC codes based on neural network optimization. Background Technology

[0002] With the introduction of the 5G NR standard, quasi-cyclic low-density parity-check codes (QC-LDPC) have been widely used in the 5G NR standard. In recent years, deep learning and machine learning have been widely applied in the field of channel coding and decoding, and have achieved good results in optimizing the decoding performance of LDPC codes. The layered decoding structure of QC-LDPC codes is suitable for building model-driven neural networks to optimize the multiplicative weighting factors and additive correction factors in the Layered Min-Sum (LMS) decoding algorithm, thereby reducing computational complexity and achieving decoding performance close to that of the Layered Belief Propagation (LBP) algorithm. However, existing neural network decoders are mostly single-branch decoder structures, which have limited error correction capabilities for erroneous bits and slow decoding convergence speed, requiring a large number of iterations, resulting in complex neural network structures.

[0003] In layered decoding of QC-LDPC codes, the layer order can be preprocessed. Decoding with a more reasonable layer order can achieve faster decoding convergence. Static layer ordering determines the layer order before decoding and does not change it during the decoding process. It is simple to implement and does not require storing intermediate calculation results. Layer ordering preprocessing can be performed before building a model-driven neural network decoder to achieve faster convergence, reduce the number of decoding iterations, simplify the neural network structure, and greatly save the training time of the neural network decoder.

[0004] A search revealed Chinese invention patent CN115483936A, which discloses a neural network-based LDPC code decoding method for optical fiber communication systems. To mitigate the impact of short loops during iterative decoding, this invention adds weights and biases, considering optical link characteristics, to the corresponding edges of the update equations from variable nodes to check nodes in the NOMS algorithm, thereby improving decoding performance. This invention utilizes the quasi-cyclic structure of QC-LDPC codes, where a bundle of edges derived from the same edge of the exponent matrix shares the same parameters. Compared to the traditional NOMS algorithm, this invention can provide a certain signal-to-noise ratio gain in optical fiber communication links, and the trained parameters have strong generalization capabilities.

[0005] This application differs from the aforementioned applications in the following ways:

[0006] 1. Patent CN115483936A explicitly limits its application scenario to optical fiber communication systems. Its core objective is to improve the decoding performance of LDPC codes in optical fiber communication links by introducing weights and biases obtained from neural network training into the NOMS algorithm, addressing the impact of short loops on iterative decoding performance in optical fiber channel environments. It focuses on adapting to optical link characteristics and improving signal-to-noise ratio gain. In contrast, this application focuses on QC-LDPC code (especially 5G-LDPC code) decoding in wireless communication systems. Its core objective is to solve the problems of layered minimization and the tendency for decoding to get trapped in local optima under complex factor graph structures, as well as insufficient error correction capability after decoding failure. It improves the overall decoding success rate and performance through a dual-branch parallel cross-layered decoding mechanism. This invention does not model the characteristics of optical fiber channels; its technical objective is to improve the error correction performance and robustness of the QC-LDPC layered decoding algorithm itself. The application scenarios and technical objectives of the two are fundamentally different.

[0007] 2. Patent CN115483936A employs a single-path LDPC decoding architecture. Based on the traditional NOMS algorithm, it only introduces trainable parameters to the edges corresponding to the update equations from variable nodes to check nodes. Essentially, it still involves parameter optimization within a single decoder, with the decoding process always completed within the same factor graph and iterative path. In contrast, this application constructs a dual-branch cross-layered decoding architecture: two different layer sortings are built based on the same QC-LDPC code; this forms upper and lower branch layered minimum sum decoders respectively; the two decoders are structurally independent and run in parallel; when the first decoding fails, a second round of parallel decoding is triggered through a cross-feedback mechanism. This application introduces decoding path diversity and a parallel cross-mechanism at the system architecture level, while the comparative patent only performs parameter fine-tuning within a single decoding path. The design approaches at the decoding architecture level are significantly different.

[0008] 3. The role of the neural network in patent CN115483936A is reflected in: introducing weights and biases to the update edges from variable nodes to check nodes in the NOMS algorithm; these parameters are related to the characteristics of the fiber optic link; multiple edges derived from the same exponent matrix share parameters to utilize the quasi-cyclic structure of QC-LDPC codes; the neural network is only used as a parameter optimization tool for the decoding update formula. However, the role of the neural network in this application is not limited to weight adjustment in a single update equation, but is used to: optimize the upper and lower branch layer minimums and the overall parameter configuration of the decoder; learn the optimal decoding parameters for decoders corresponding to different layer sortings; the neural network optimization in this application serves multi-decoder, multi-path decoding processes, while the neural network in the comparative patent only operates on edge weight optimization within a single decoder. The two have completely different levels of introduction and technical functions.

[0009] 4. Patent CN115483936A employs a traditional iterative update mechanism in its decoding process, where decoding information is only transmitted between nodes within the same decoder. When decoding performance is limited, it does not provide additional decoding failure recovery or multi-decoder collaboration mechanisms. In contrast, this application explicitly designs a parallel cross-decoding mechanism after decoding failure: when both the upper and lower branches fail initially, the original LLR sequence is weighted and correlated with the posterior LLR sequences output by the two decoders. The processed LLR is then used as a new input, cross-inputted into the other decoder for further decoding. This mechanism allows decoders at different levels to share and merge decoding information, effectively overcoming the performance bottleneck of single-path decoding—a technical approach not found in the comparative patent.

[0010] A search revealed Chinese invention patent CN113162632A, which discloses an intelligent QC-LDPC decoding method, decoder, and low-Earth orbit satellite communication system. The decoding method includes: S1, demodulating the received signal after QC-LDPC encoding to obtain likelihood ratio information; S2, converting the likelihood ratio information from serial to parallel to obtain N likelihood ratio sub-information, inputting the N likelihood ratio sub-information into N cascaded LSTM memory neural networks in parallel, and inputting the output data of the N LSTM memory neural networks in parallel into a fully connected network to obtain decoded data; S3, verifying the decoded data, outputting the decoded data if the verification passes, and discarding the decoded data if the verification fails. This method employs artificial intelligence technology for QC-LDPC decoding. The LSTM memory neural networks demonstrate the correlation between the input information at a given moment and the input information at previous moments, enabling self-organizing learning of the decoding algorithm, finding complex interactions between input variables, and improving decoding performance.

[0011] This application differs from the aforementioned applications in the following ways:

[0012] 1. Patent CN113162632A explicitly applies to low-Earth orbit satellite communication systems. Its core objective is to utilize an artificial intelligence model for end-to-end intelligent decoding of QC-LDPC codes. It employs an LSTM memory neural network to mine the temporal correlation between likelihood ratio sequences of received signals, achieving self-organizing learning in the decoding algorithm to improve decoding performance. In contrast, this application focuses on layered decoding of QC-LDPC codes (especially 5G-LDPC codes) in wireless communication systems. Its core objective is not to replace traditional decoding algorithms with neural networks, but rather to improve the error correction capability and decoding success rate of traditional layered decoding through a dual-branch cross-layered decoding mechanism, while maintaining minimum layer size and the basic decoding structure. The two applications differ significantly in their application focus and design goals.

[0013] 2. Patent CN113162632A employs a deep neural network-based decoding approach: it decomposes the likelihood ratio information obtained from demodulation into N sub-information; inputs these sub-information ...

[0014] 3. The decoding process of patent CN113162632A is as follows: a single neural network forward inference obtains the decoding result, and verification is only performed after the decoding result is output. If the verification fails, the decoded data is directly discarded. There is no mechanism for multi-decoder collaboration or information reuse after decoding failure. In contrast, this application explicitly designs a parallel decoding process using upper and lower dual-branch layered minimum sum decoders. Different layer ordering introduces decoding path diversity. When the first decoding fails, the original LLR is weighted and correlated with the posterior LLR output by the two decoders. The weighted LLR is then cross-inputted into the other decoder in parallel for re-decoding. This cross-re-decoding mechanism realizes information fusion and collaborative error correction between different decoding paths, significantly enhancing the error correction capability in the case of decoding failure. The contrasting patent does not involve similar multi-branch decoding or cross-information feedback mechanisms.

[0015] A search revealed Chinese invention patent CN110212920A, which discloses a deep learning-based LDPC code decoding algorithm. Combining deep learning technology, it studies the construction of large-length random LDPC codes and structured LDPC codes, improved belief propagation decoding on cyclic Tanner graphs, and its low-complexity simplified decoding algorithm. Employing an optimal approach, it verifies the distribution of short cycles containing each non-zero element in the check matrix, combining the optimized distribution of short cycle numbers in the current column with the standard deviation distribution of the short cycle numbers in each column to construct a random LDPC code with good characteristics. The application also studies a cycle elimination algorithm, simulates the performance of QC-LDPC codes constructed using the substitution elimination method, and constructs a QC-LDPC code with a better large-length through improvements. Furthermore, it improves the BP decoding algorithm and proposes an improved minimum-sum decoding algorithm. Based on deep learning technology and the cycle statistical characteristics of Tanner graphs, this application first constructs a random LDPC code with good characteristics, and then constructs a QC-LDPC code with a better large-length through simulation. Its algorithm is superior, and its error correction performance is better.

[0016] This application differs from the aforementioned applications in the following ways:

[0017] 1. Patent CN110212920A focuses on the overall performance optimization of LDPC codes. Its core technical issues include: the adverse effects of short cycles in the Tanner graph on decoding performance; optimization of the construction methods for random LDPC codes and QC-LDPC codes; improving decoding performance by increasing the encirclement length and reducing short cycles; and improvements to the BP and minimum-sum decoding algorithms based on these improvements. Therefore, the research object of this comparative patent is not limited to the decoding process itself, but covers the overall design issues of code construction, Tanner graph optimization, and decoding algorithm improvement. This application does not involve the construction of LDPC codes or the optimization of Tanner graph structures; the technical issues and research objects addressed by the two are fundamentally different.

[0018] 2. The deep learning technology in patent CN110212920A is mainly used to: assist in analyzing the short-loop statistical characteristics of Tanner graphs; guide the construction optimization of random LDPC codes and QC-LDPC codes; and improve decoding performance by combining with improved BP or minimum sum algorithms. In this scheme, deep learning is highly coupled with code construction and Tanner graph structure design, and is not simply used for decoder parameter optimization. However, in this application, the role of the neural network is clearly limited to optimizing the parameters of the upper and lower branch decoders. Therefore, there are substantial differences between the two in terms of the target of the neural network, the layers introduced, and the technical purpose. Summary of the Invention

[0019] To address the shortcomings mentioned in the background section, the present invention aims to provide a neural network-optimized method for decoding QC-LDPC codes with dual-branch cross-layered minimum sums.

[0020] A neural network-optimized method for dual-branch cross-layered minimum sum decoding of QC-LDPC codes includes the following steps:

[0021] S101, for the input QC-LDPC code, the maximum number of iterations. And two specific layer ordering methods: minimum layer and decoder. , Construct a hierarchical minimum sum decoder with upper and lower branches. , .

[0022] S102. Parameterized upper and lower branch hierarchical minimum sum decoder based on multiplicative weighting and additive correction mechanism of check node operation. , Multiplicative weighting factors at each level , With additive correction factor , This yields a minimum sum decoder for the upper and lower branch neural networks. , Initialize the multiplicative weighting factors of each layer in the hierarchical minimum sum decoder of the upper and lower branch neural networks. , The additive correction factor is 1. , It is 0.

[0023] S103. Under the target operating signal-to-noise ratio condition, for the transmission and reception Monte Carlo process of QC-LDPC code through additive white Gaussian noise channel and BPSK modulation, generate QC-LDPC code received signal vector-transmit codeword dataset, and use cross-entropy loss function and backpropagation algorithm of neural network to perform hierarchical minimum sum decoder of upper and lower branch neural networks. , Independent training was performed to obtain a hierarchical minimum sum decoder for the lower and upper branches of the neural network. , The optimal multiplicative weighting factor , With the optimal additive correction factor , Through optimal multiplicative weighting factors , With the optimal additive correction factor , The upper and lower branch decoders are instantiated to establish a hierarchical minimum sum decoder for the lower and upper branch neural networks with optimized neural network parameters. , .

[0024] S104. During decoding, the received signal is simultaneously input into the upper and lower branch neural network layered minimum sum decoder. , If either branch decodes successfully, the decoding result is output; otherwise, the original LLR sequence value is weighted and correlated with the posterior LLR sequence value output by the first decoding of the two-way decoder, and then fed into the lower and upper branch neural network layered minimum sum decoder for decoding again.

[0025] Further, step S101 specifically includes:

[0026] QC-LDPC codes are obtained through a size of Mother matrix Quasi-cyclic expansion yields a size of The verification matrix Where m represents the number of rows in the parent matrix, n represents the number of columns in the parent matrix, M represents the number of rows in the parity check matrix, and N represents the number of columns in the parity check matrix, and... , , Indicates the magnitude of the expansion factor, and the parity-check matrix. From top to bottom each Each row constitutes one layer, with a total of m layers;

[0027] Two specific layer ordering methods for input layer minimum and decoder: , A specific layer ordering of the layered minimum sum decoder is defined as the input QC-LDPC mother matrix. Row permutation sort The row is sorted by permutation. yes A permutation. Given a layer order. The following layered minimum sum decoder It is given Layered minimum sum decoder under matrix, Represents the relationship between the parent matrix and the mother matrix. Sort by layer The row permutation matrix performed, that is The first of the matrix Take from The first of the matrix OK.

[0028] Furthermore, step S101 specifically includes:

[0029] Given a mother matrix The one-time iteration calculation of the layered minimum sum decoder is based on the input layer of the received signal and the... It consists of Z-row decoding units, each of which comprises a check node layer and a variable node layer. A single Z-row decoding unit performs... Matrix The row check corresponding to The second least-sum confidence propagation decoding message passing, processed by a single Z-row decoding unit. Matrix Row verification is performed by the parent matrix The result is formed by expanding the given parent matrix by one row. The following layered minimum and decoder Each Z-line decoding unit is denoted as Press it Reordering , One iteration of the calculation is performed by the received signal input layer and Composition, total layer, The next iteration calculation is performed by the received signal input layer and... indivual Connecting components.

[0030] Two layer sorting methods for input QC-LDPC codes , The following options are available:

[0031] This can be obtained through the following operation: [The parent matrix is ​​then used for this purpose.] Mapped to a Tanner graph, the parent matrix The rows and columns correspond to the check nodes and variable nodes in the Tanner diagram, respectively. First, all check nodes in the parent matrix Tanner diagram are sorted in ascending order according to the sum of the degrees of the connected variable nodes; that is, the smaller the sum of the degrees of the connected variable nodes, the higher the ranking. For check nodes with the same priority (equal sums of the degrees of the connected variable nodes), the sum of the number of times each check node is connected to a higher-priority check node is calculated; the larger the sum, the higher the priority, and the higher the ranking. The target ranking of the parent matrix rows is obtained from the ranking of the check nodes in the parent matrix. .

[0032] The following steps are used to calculate the parent matrix. The row weight of each row in the matrix is ​​used to sort the rows of the parent matrix in descending order of row weight, i.e., the rows with larger row weights appear earlier in the sorting order. .

[0033] Further, step S102 specifically includes:

[0034] A parameterized hierarchical minimum sum decoder based on a multiplicative weighted and additive correction mechanism using check node operations. The multiplicative weighting factors and additive correction factors of each level are obtained This refers to using a multiplicative weighting and additive correction mechanism in the message update formula at the verification node layer, and... Multiplicative weighting factors at each level With additive correction factor Set as learnable parameters to obtain , Can represent The set of multiplicative weighting factors used by each verification node layer, i.e. , Can represent The set of additive correction factors used by each check node layer, i.e. ,in, , . The message update formula for the middle check node layer is expressed as:

[0035] ;

[0036] In the formula, , , , Indicates in Matrix and verification nodes The set of all connected variable nodes, here Represents the parity check matrix Sort by layer The reordered matrix Indicates in Remove variable nodes The set of variable nodes, The message representing the variable node. Indicates the first During the nth iteration calculation Learnable multiplicative weighting factors for layer verification node layers. Indicates the first During the nth iteration calculation Learnable additive correction factor for layer verification node layer.

[0037] Further, step S104 specifically includes:

[0038] S104. During decoding, the received signal is decoded using a dual-branch parallel decoding mechanism. Therefore, the received signal must be simultaneously connected to both the lower and upper branch neural network hierarchical minimum sum decoders. , The input layer of the upper and lower branches receives the following data: The original LLR sequence value of each channel is Record the branch iterations The posterior LLR sequence value after the next decoding is Record the a posteriori LLR sequence value after T iterations of branch decoding. superscript This represents the output layer. The received signal is simultaneously processed by the upper and lower branches through parallel input, initiating the first layered minimum sum decoding of each branch. If at least one of the decoding results from the upper and lower branches passes the target QC-LDPC code parity-check matrix... If the decoding fails the verification, decoding terminates, and the output of any branch that passes the decoding verification is selected as the final decoding result; otherwise, dual-branch secondary decoding is initiated, and the LLR sequence is input through the original channel. The a posteriori LLR sequence of the first decoded output of the two-way decoder , Weighted correlation is performed to obtain weighted LLR sequences. , Then, the inputs are fed in parallel and cross-referenced to the lower and upper branch neural network hierarchical minimum sum decoder. , Decoding. For the two decoding outputs, the decoding result with the smallest Euclidean distance to the channel received signal vector is selected as the final decoding output.

[0039] Step S104 further includes:

[0040] and Search within [0,1] based on the performance of the validation dataset and select the value that performs best on the validation set.

[0041] The beneficial effects of this invention are:

[0042] This invention significantly improves the performance of the Neural Layered Min-Sum (NLMS) decoding algorithm by optimizing the dual-branch cross-decoding strategy through neural networks, and effectively enhances the error correction capability of the NLMS decoder for erroneous bits, making it suitable for decoding QC-LDPC codes. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0045] Figure 2 This is a flowchart illustrating step S104 in this invention;

[0046] Figure 3 This is a diagram of the neural network structure of the present invention;

[0047] Figure 4 This is a comparison chart of frame error rates obtained by decoding using the present invention and existing NLMS decoding methods. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] like Figure 1 , 2As shown in Figure 3, this embodiment is a method for decoding a dual-branch cross-layered minimum sum of QC-LDPC codes based on neural network optimization, including the following steps:

[0050] S101: Maximum number of iterations for the input QC-LDPC code. And two specific layer ordering methods: minimum layer and decoder. , Construct a hierarchical minimum sum decoder with upper and lower branches. , .

[0051] QC-LDPC codes are obtained through a size of Mother matrix Quasi-cyclic expansion yields a size of The verification matrix Among them are , , Indicates the magnitude of the expansion factor, and the parity-check matrix. From top to bottom each The rows form a layer, totaling layer.

[0052] Two specific layer ordering methods for input layer minimum and decoder: , A specific layer ordering of the layered minimum sum decoder is defined as the input QC-LDPC mother matrix. Row permutation sort The row is sorted by permutation. yes A permutation. Given a layer order. The following layered minimum sum decoder It is given Layered minimum sum decoder under matrix, Represents the relationship between the parent matrix and the mother matrix. Sort by layer The row permutation matrix performed, that is The first of the matrix Take from The first of the matrix OK.

[0053] Given a mother matrix The one-time iteration calculation of the layered minimum sum decoder is based on the input layer of the received signal and the... It consists of Z-row decoding units, each of which comprises a check node layer and a variable node layer. A single Z-row decoding unit performs... Matrix The row check corresponding to The second least-sum confidence propagation decoding message passing, processed by a single Z-row decoding unit. Matrix Row verification is performed by the parent matrix The result is formed by expanding the given parent matrix by one row. The following layered minimum and decoder Each Z-line decoding unit is denoted as Press it Reordering , One iteration of the calculation is performed by the received signal input layer and Composition, total layer, The next iteration calculation is performed by the received signal input layer and... indivual Connecting components.

[0054] Two layer sorting methods for input QC-LDPC codes , The following options are available:

[0055] This can be obtained through the following operation: [The parent matrix is ​​then used for this purpose.] Mapped to a Tanner graph, the parent matrix The rows and columns correspond to the check nodes and variable nodes in the Tanner diagram, respectively. First, all check nodes in the parent matrix Tanner diagram are sorted in ascending order according to the sum of the degrees of the connected variable nodes; that is, the smaller the sum of the degrees of the connected variable nodes, the higher the ranking. For check nodes with the same priority (equal sums of the degrees of the connected variable nodes), the sum of the number of times each check node is connected to a higher-priority check node is calculated; the larger the sum, the higher the priority, and the higher the ranking. The target ranking of the parent matrix rows is obtained from the ranking of the check nodes in the parent matrix. .

[0056] The following operation can be performed to obtain the parent matrix: The row weight of each row in the matrix is ​​used to sort the rows of the parent matrix in descending order of row weight, i.e., the rows with larger row weights appear earlier in the sorting order. .

[0057] S102: Parameterized upper and lower branch hierarchical minimum sum decoder based on multiplicative weighting and additive correction mechanism of check node operation. , Multiplicative weighting factors at each level , With additive correction factor , This yields a minimum sum decoder for the upper and lower branch neural networks. , Initialize the multiplicative weighting factors of each layer in the hierarchical minimum sum decoder of the upper and lower branch neural networks. , The additive correction factor is 1. , It is 0.

[0058] First, construct the input layer of the neural network. The input layer utilizes the received signal. Initialize the LLR value received by the channel:

[0059] ;

[0060] In the formula, This represents the noise variance under additive white Gaussian noise channel conditions. Indicates the code length. This represents the codeword after being encoded using QC-LDPC.

[0061] Then, the hidden layers of the neural network are constructed. In one iteration, the hidden layers include... indivual - Line decoding unit . No. Layered deployment of verification nodes, each layer containing The verification node, the first Layered deployment of variable nodes, each layer contains 1 variable node The input layer is treated as the 0th layer of variable nodes.

[0062] The message sent from the variable node layer to the verification node layer is represented as follows:

[0063] ;

[0064] In the formula, express Matrix number Layer and variable nodes The set of connected verification nodes, initialized. It is 0.

[0065] The message sent from the verification node layer to the variable node layer is represented as follows:

[0066] ;

[0067] In the formula, , , , Indicates in Matrix and verification nodes The set of all connected variable nodes, here Represents the parity check matrix Sort by layer The reordered matrix Indicates in Remove variable nodes , The message representing the variable node. Indicates the first During the nth iteration calculation Learnable multiplicative weighting factors for layer verification node layers. Indicates the first During the nth iteration calculation Learnable additive correction factor for layer verification node layer.

[0068] The formula for the posterior LLR value is expressed as follows:

[0069] ;

[0070] An activation function is used after the output layer of the model. During the training phase, the activation function is expressed by the following formula:

[0071] ;

[0072] The activation function used during the testing phase is expressed by the following formula:

[0073] ;

[0074] The above illustrates the neural network layer structure in one iteration of computation. If the maximum number of decoding iterations is... The network structure remains unchanged, with the input layer and softmax layer remaining the same, and simple repeated hidden layers. Second-rate.

[0075] The cross-entropy loss can be expressed as follows:

[0076] .

[0077] S103: Under the target operating signal-to-noise ratio condition, a QC-LDPC code received signal vector-transmit codeword dataset is generated for the transmission and reception Monte Carlo process of QC-LDPC codes through an additive white Gaussian noise channel and BPSK modulation. Then, a layered minimum sum decoder for the upper and lower branch neural networks is developed using the cross-entropy loss function and the backpropagation algorithm of the neural network. , Perform independent training to obtain the hierarchical minimum sum decoder of the upper and lower branch neural networks. , The optimal multiplicative weighting factor , With the optimal additive correction factor , Through optimal multiplicative weighting factors , With the optimal additive correction factor , The upper and lower branch decoders are instantiated to establish a hierarchical minimum sum decoder for the lower and upper branch neural networks with optimized neural network parameters. , .

[0078] S104: During decoding, the received signal is decoded using a dual-branch parallel decoding mechanism. Therefore, the received signal must be simultaneously connected to both the lower and upper branch neural network hierarchical minimum sum decoders. , The input layer of the upper and lower branches receives the following data: The original LLR sequence value of each channel is Record the branch iterations The posterior LLR sequence value after the next decoding is Record the a posteriori LLR sequence value after T iterations of branch decoding. superscript This represents the output layer. The received signal is simultaneously processed by the upper and lower branches through parallel input, initiating the first layered minimum sum decoding of each branch. If at least one of the decoding results from the upper and lower branches passes the target QC-LDPC code parity-check matrix... If the decoding fails the verification, decoding terminates, and the output of any branch that passes the decoding verification is selected as the final decoding result; otherwise, dual-branch secondary decoding is initiated, and the LLR sequence is input through the original channel. The a posteriori LLR sequence of the first decoded output of the two-way decoder , Weighted correlation is performed to obtain weighted LLR sequences. , Then, the inputs are fed in parallel and cross-referenced to the lower and upper branch neural network hierarchical minimum sum decoder. , Decoding. For the two decoding outputs, the decoding result with the smallest Euclidean distance to the channel received signal vector is selected as the final decoding output.

[0079] and Search within [0,1] based on the performance of the validation dataset and select the value that performs best on the validation set.

[0080] Simulation content:

[0081] The proposed decoding network was trained and tested using the PyTorch framework. The LDPC codes were designed according to the 5G NR standard. As the mother matrix Mother matrix Size is The code length is The number of information bits is The maximum number of decoding iterations is , The selected values ​​are [5, 28, 46, 38, 41, 43, 26, 36, 32, 30, 44, 24, 42, 40, 34, 37, 45, 35, 27, 33, 39, 23, 31, 29, 20, 19, 18, 14, 21, 25, 17, 22, 8, 13, 16, 11, 15, 6, 12, 10, 7, 9, 4, 1, 3, 2]. The selected values ​​are [1, 2, 3, 4, 9, 7, 10, 6, 12, 8, 11, 13, 15, 16, 14, 17, 18, 19, 20, 21, 22, 25, 23, 24, 26, 27, 29, 30, 31, 32, 33, 34, 35, 36, 37, 39, 40, 42, 44, 45, 28, 38, 41, 43, 46, 5]. Under the condition of 0-3dB signal-to-noise ratio, the received signal vector-transmit codeword dataset of 5G-LDPC code is generated for the transmission and reception Monte Carlo process of 5G-LDPC code through BPSK modulation and additive white Gaussian noise channel. The network was trained using mini-batch stochastic gradient descent, with each small batch containing 64 training data points. The optimizer used was the Adam adaptive moment estimator with a learning rate of 0.01. The upper and lower branch decoders were trained independently using the cross-entropy loss function.

[0082] Decoding was performed using the method of this invention and the existing NLMS decoding method, and the respective frame error rates were obtained as follows: Figure 4 As shown, the decoding method of the present invention has a significant performance improvement compared to the single-branch neural network optimization layered minimum sum decoding.

[0083] The foregoing has shown and described the basic principles, main features, and performance improvements of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for decoding QC-LDPC codes with dual-branch cross-layered minimum sums based on neural network optimization, characterized in that, The method includes the following steps: S101, for the input QC-LDPC code, the maximum number of iterations. And two specific layer ordering methods: minimum layer and decoder. , Construct a hierarchical minimum sum decoder with upper and lower branches. , ; Step S101 specifically includes: The input QC-LDPC code is obtained through a size of Mother matrix Quasi-cyclic expansion yields a size of The verification matrix Where m represents the number of rows in the parent matrix, n represents the number of columns in the parent matrix, M represents the number of rows in the parity check matrix, and N represents the number of columns in the parity check matrix, and... , , Indicates the magnitude of the expansion factor, and the parity-check matrix. From top to bottom each Each row constitutes one layer, with a total of m layers; Two specific layer ordering methods for input layer minimum and decoder: , A specific layer ordering of the layered minimum sum decoder is defined as the input QC-LDPC mother matrix. Row permutation sort The row is sorted by permutation. yes A permutation, given a layer order The following layered minimum sum decoder It is given Layered minimum sum decoder under matrix, Represents the parent matrix Sort by layer The row permutation matrix performed, i.e. The first of the matrix , Take from The first of the matrix , OK; Given a mother matrix The one-time iteration calculation of the layered minimum sum decoder is based on the input layer of the received signal and the... It consists of Z-row decoding units, each of which comprises a check node layer and a variable node layer. A single Z-row decoding unit performs... Matrix The row check corresponding to The second least-sum confidence propagation decoding message passing, processed by a single Z-row decoding unit. Matrix Row verification is performed by the parent matrix The result of a single-row expansion; given a parent matrix The following layered minimum and decoder indivual - Line decoding unit is denoted as Press it Reordering , One iteration of the calculation is performed by the received signal input layer and Composition, total layer, The next iteration calculation is performed by the received signal input layer and... indivual Connecting components; S102. Parameterized upper and lower branch hierarchical minimum sum decoder based on multiplicative weighting and additive correction mechanism of check node operation. , Multiplicative weighting factors at each level , With additive correction factor , This yields a minimum sum decoder for the upper and lower branch neural networks. , Initialize the multiplicative weighting factors of each layer in the hierarchical minimum sum decoder of the upper and lower branch neural networks. , The additive correction factor is 1. , =0; In step S102, the parameterized hierarchical minimum sum decoder is based on the multiplicative weighted and additive correction mechanism of the check node operation. The multiplicative weighting factors and additive correction factors of each level are obtained The specific method is as follows: A multiplicative weighting and additive correction mechanism is used in the message update formula of the verification node layer, and... Multiplicative weighting factors at each level With additive correction factor Set as learnable parameters to obtain , express The set of multiplicative weighting factors used by each verification node layer, i.e. , express The set of additive correction factors used by each check node layer, i.e. ,in, , ; The message update formula for the middle check node layer is expressed as: ； In the formula, , , , Indicates in Matrix and verification nodes The set of all connected variable nodes. Represents the parity check matrix Sort by layer The rearranged matrix Indicates in Remove variable nodes The set of variable nodes, The message representing the variable node, Indicates the first During the nth iteration calculation Learnable multiplicative weighting factors for layer verification node layers. Indicates the first During the nth iteration calculation Learnable additive correction factor for layer verification node layer; S103. Under the target operating signal-to-noise ratio condition, for the transmission and reception Monte Carlo process of QC-LDPC encoded QC-LDPC codewords after BPSK modulation and additive white Gaussian noise channel, generate the received signal vector-transmit codeword dataset of QC-LDPC codewords, and use the cross-entropy loss function and the backpropagation algorithm of the neural network to perform hierarchical minimum sum decoder of the upper and lower branch neural networks. , Perform independent training to obtain the hierarchical minimum sum decoder of the upper and lower branch neural networks. , The optimal multiplicative weighting factor , With the optimal additive correction factor , ; through optimal multiplicative weighting factors , With the optimal additive correction factor , The upper and lower branch decoders are instantiated to establish a hierarchical minimum sum decoder for the upper and lower branch neural networks with optimized neural network parameters. , ; S104. During decoding, the received signal is simultaneously input into the upper and lower branch neural network layered minimum sum decoder. , If either branch decodes successfully, the decoding result is output; otherwise, the original LLR sequence value is weighted and correlated with the posterior LLR sequence value output by the first decoding of the two-way decoder, and then fed into the lower and upper branch neural network layered minimum sum decoder in parallel for decoding again. Step S104 specifically includes: During decoding, the received signal is decoded using a dual-branch parallel decoding mechanism. Therefore, the received signal must be simultaneously connected to both the upper and lower branch neural network hierarchical minimum sum decoders. , The input layer of the upper and lower branches receives the following data: The original LLR sequence value of each channel is Record the branch iterations The posterior LLR sequence value after the next decoding is Record the a posteriori LLR sequence value after T iterations of branch decoding. superscript This represents the output layer; the received signal simultaneously initiates the first layered minimum sum decoding of each branch through the parallel input upper and lower branches. If at least one of the decoding results of the upper and lower branches can pass the target QC-LDPC code parity check matrix... If the decoding fails the verification, decoding terminates, and the output of any branch that passes the decoding verification is selected as the final decoding result; otherwise, dual-branch secondary decoding is initiated, and the LLR sequence is input through the original channel. The a posteriori LLR sequence of the first decoded output of the two-way decoder , Weighted correlation is performed to obtain weighted LLR sequences. , Then, the inputs are fed in parallel and cross-referenced to the lower and upper branch neural network hierarchical minimum sum decoder. , Decoding; For the two decoding outputs, select the decoding result with the smallest Euclidean distance to the channel received signal vector as the final decoding output.

2. The method for decoding QC-LDPC codes with dual-branch cross-layered minimum sums based on neural network optimization according to claim 1, characterized in that, Step S104 further includes: and Search within [0,1] based on the performance of the validation dataset and select the value that performs best on the validation set.

3. The method for decoding QC-LDPC codes with dual-branch cross-layered minimum sums based on neural network optimization according to claim 2, characterized in that, The two specific layer orderings of the minimum layer and the decoder. , The method to obtain it is as follows: The parent matrix Mapped to a Tanner graph, the parent matrix The rows and columns correspond to the check nodes and variable nodes in the Tanner diagram, respectively. First, all check nodes in the parent matrix Tanner diagram are sorted in ascending order based on the sum of the degrees of the variable nodes they are connected to. For check nodes with the same priority, the sum of the number of times each check node is connected to a higher-priority check node is calculated; the larger the sum, the higher the priority, and the earlier it is sorted. The target sorting of the parent matrix rows is obtained from the sorting of the check nodes in the parent matrix. ; Calculate the parent matrix The row weight of each row in the matrix is ​​used to sort the rows of the parent matrix in descending order of row weight, with larger row weights appearing earlier in the sorted order. .

Images