LDPC decoding model construction method and LDPC decoding method
By optimizing the LDPC decoding model through grouping and positional encoding, the problems of high complexity and weak generalization ability in the long codeword decoding process are solved, achieving higher decoding accuracy and adaptability.
Patent Information
- Application Number
- CN202411080319.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-08-08
AI Technical Summary
Existing LDPC decoding methods suffer from slow convergence, high complexity, and weak generalization ability when processing long codewords. Furthermore, neural network models exhibit unstable performance on unseen data and struggle to adapt to rapidly changing communication environments.
By grouping check nodes and variable nodes, a dedicated sub-model is constructed and a positional encoding mechanism is incorporated to optimize the construction process of the LDPC decoding model, reduce computational complexity, and improve the model's generalization ability.
It improves the decoding accuracy and adaptability of the LDPC decoding model, enabling a more accurate understanding of the dependencies between nodes, reducing computational complexity and enhancing the model's generalization ability.
Smart Images

Figure CN118611683B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication technology, and in particular to an LDPC decoding model construction method and an LDPC decoding method. Background Art
[0002] In the field of modern communications, low-density parity-check codes (LDPC) are widely used in a variety of communication standards, especially in the fifth generation (5G) and future communication systems, due to their ability to achieve efficient and reliable error correction. However, traditional LDPC decoding methods based on iterative algorithms, such as belief propagation (BP), perform well on short to medium-length codewords, but face challenges such as slow convergence and high complexity when processing long codewords. In addition, its performance is highly sensitive to channel conditions and code structure, and it is difficult to adapt to rapidly changing communication environments.
[0003] In recent years, with the rise of deep learning, neural networks have been introduced into LDPC decoding in order to break through the limitations of traditional methods, but there are two major obstacles to its application: First, neural network models are prone to overfitting when processing long codes, resulting in weakened generalization ability, inability to maintain high performance on unseen data, and inability to quickly adapt to channel changes; second, the "black box" nature of neural networks and the lack of transparency in the decoding process are not conducive to system optimization and fault diagnosis, which has become a major obstacle to actual deployment.
[0004] Therefore, the existing technology still needs to be improved and developed. Summary of the invention
[0005] The present invention provides an LDPC decoding model construction method and an LDPC decoding method, which are used to construct a special sub-model by grouping and integrate a position coding mechanism, optimize the construction and decoding process of the LDPC decoding model, and improve the decoding accuracy.
[0006] The first object of the present invention is to provide a method for constructing an LDPC decoding model, which includes: grouping the check nodes according to the number of variable nodes connected to the check nodes to obtain multiple groups of check nodes, constructing a check node sub-model corresponding to each group of check nodes, and each group of check nodes shares the same check node sub-model; grouping the variable nodes according to the number of check nodes connected to the variable nodes to obtain multiple groups of variable nodes, constructing a variable node sub-model corresponding to each group of variable nodes, and the variable nodes in each group share the same variable node sub-model; constructing a variable node decoding sub-model, integrating the check node sub-model, the variable node sub-model and the variable node decoding sub-model, and integrating the position encoding mechanism into the update strategy of the check node sub-model and the variable node sub-model to obtain a graph neural network model; collecting a data set, and using the data set to train the graph neural network model to obtain an LDPC decoding model.
[0007] Preferably, the check node submodel includes a first input layer, a first neural network layer and a first output layer, the first input layer receives the current variable node scalar, the current check node scalar, received information and the current position code, the first neural network layer processes the information output by the first input layer, and uses a nonlinear transformation to obtain a check node scalar update value, the first output layer calculates and outputs a new check node scalar based on the check node scalar update value output by the first neural network layer, and the new check node scalar is equal to the sum of the check node scalar update value and the current check node scalar.
[0008] Preferably, the variable node sub-model includes a second input layer, a second neural network layer and a second output layer, the second input layer receives the current variable node scalar, the check node scalar, the received information, the Gaussian representation of the girth and the current position encoding, the second neural network layer processes the information output by the second input layer and uses a nonlinear transformation to obtain the variable node scalar update value, the second output layer outputs a new variable node scalar, and the new variable node scalar is equal to the sum of the variable node scalar update value and the current variable node scalar.
[0009] Preferably, the update strategy of the check node submodel and the variable node submodel includes updating the variable node scalar and updating the check node scalar, the current position encoding of the variable node is created using sine encoding and cosine encoding, and the current position encoding of the check node is created using sine encoding and cosine encoding.
[0010] Preferably, for any variable node c q Current position code It is expressed as: ; In the formula, represents the sinusoidal part of the current position code, Represents the cosine part of the current position code, code is a row vector from 0 to (d / 2)-1, and Together they form a current position code of length d, where d is an even number greater than 2, D is a constant, and q∈{0, 1, 2...n-1}.
[0011] Preferably, for any check node s q Current position code It is expressed as: ; In the formula, represents the sinusoidal part of the current position code, Represents the cosine part of the current position code, code is a row vector with values from 0 to (d / 2)-1, and Together they form a current position code of length d, where d is an even number greater than 2, D is a constant, and q∈{0, 1, 2...n-1}.
[0012] Preferably, the method of collecting a data set and using the data set to train the graph neural network model to obtain an LDPC decoding model includes: collecting historical input data and historical output data of a channel LDPC code; grouping the collected data according to the signal-to-noise ratio, and randomly extracting a preset number of data from each group and putting them into the data set; dividing the data set into a training set and a test set, and using the training set to train the graph neural network model to obtain a trained graph neural network model; and inputting the test set into the trained graph neural network model for testing and tuning to obtain an LDPC decoding model.
[0013] The second object of the present invention is to provide an LDPC decoding method, which includes: constructing an LDPC decoding model, wherein the LDPC decoding model is constructed using the LDPC decoding model construction method as described above; receiving initial information, and inputting the initial information into the LDPC decoding model to obtain a decoding result output by the LDPC decoding model.
[0014] Preferably, the receiving of initial information and inputting the initial information into the LDPC decoding model to obtain a decoding result output by the LDPC decoding model includes: receiving the initial information and preprocessing the initial information; extracting features from the preprocessed initial information; and inputting the extracted features into the LDPC decoding model to obtain a decoding result output by the LDPC decoding model.
[0015] Preferably, the LDPC decoding method also includes: monitoring and recording error cases in which decoding fails during the decoding process of the LDPC decoding model; for each error case, extracting the final iteration result as a training sample, the training sample including the final value of the variable node and the final value of the corresponding check node; constructing a correction model, and using the training sample to train the correction model to obtain a trained correction model; using the trained correction model to assist the LDPC decoding model in decoding.
[0016] In this scheme, when constructing the LDPC decoding model, the nodes are grouped according to the number of connections between the check nodes and the variable nodes, and a dedicated sub-model is constructed for each group of nodes. This grouping method not only fully considers the differences in the connection conditions of different nodes in the LDPC code, but also reduces the model parameters and the computational complexity by sharing the sub-model, while enhancing the generalization ability of the model. Moreover, in order to further improve the performance of the model, when constructing the LDPC decoding model, the position encoding mechanism is integrated into the update strategy of the check node sub-model and the variable node sub-model. The position encoding enables the model to fully utilize the position information of the node in the Tanner graph. information, so as to more accurately understand the dependencies between nodes and further optimize the decoding process, so that the constructed LDPC decoding model has higher decoding accuracy; in the decoding stage, the LDPC decoding model first obtains the feature representation of the variable node through an iterative process, and then these features are sent to the variable node decoding sub-model, which outputs the posterior probability of the codeword based on the learned features. The introduction of the posterior probability provides richer information for the decoding process, so that the model can more accurately judge the value of each symbol. Finally, based on these posterior probabilities, the LDPC decoding model judges the symbol and obtains the decoded vector. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.
[0018] Figure 1 is a flow chart of a method for constructing an LDPC decoding model provided by an embodiment of the present invention;
[0019] Figure 2 is a schematic diagram of a Tanner graph corresponding to LDPC provided in an embodiment of the present invention;
[0020] Figure 3is a schematic diagram of a situation where a check node is connected to two variable nodes according to an embodiment of the present invention;
[0021] Figure 4 is a schematic diagram of a connection between a check node and three variable nodes provided by an embodiment of the present invention;
[0022] Figure 5 is a schematic diagram of the structure of a check node sub-model provided by an embodiment of the present invention;
[0023] Figure 6 is a schematic diagram of a situation where a variable node is connected to two check nodes according to an embodiment of the present invention;
[0024] Figure 7 is a schematic diagram of a structure in which a variable node is connected to three check nodes according to an embodiment of the present invention;
[0025] Figure 8 is a schematic diagram of the structure of a variable node sub-model provided by an embodiment of the present invention;
[0026] Fig. 9 It is a flow chart of the LDPC decoding method provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0027] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0028] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In an embodiment of the present invention, a method for constructing an LDPC decoding model includes:
[0029] S101, grouping the check nodes according to the number of variable nodes connected to the check nodes to obtain multiple groups of check nodes, and constructing a check node sub-model corresponding to each group of check nodes, wherein each group of check nodes shares the same check node sub-model;
[0030] S102, grouping the variable nodes according to the number of check nodes connected to the variable nodes to obtain multiple groups of variable nodes, and constructing a variable node sub-model corresponding to each group of variable nodes, wherein the variable nodes in each group share the same variable node sub-model;
[0031] S103, construct a variable node decoding sub-model, integrate the check node sub-model, the variable node sub-model and the variable node decoding sub-model, and integrate the position encoding mechanism into the update strategy of the check node sub-model and the variable node sub-model to obtain a graph neural network model;
[0032] S104. Collect a data set, and use the data set to train a graph neural network model to obtain an LDPC decoding model.
[0033] Understandably, if Figure 2 As shown in the figure, the Tanner graph corresponding to LDPC (Low-Density Parity-Check) contains k variable nodes and m check nodes. Among them, each check node is connected to more than two variable nodes, such as s1 is connected to three variable nodes c1, c2 and c4. Similarly, each variable node is connected to more than two check nodes, such as c1 is connected to two variable nodes s1 and s2.
[0034] In this embodiment, in step S101, the check node submodel processes information related to the check node. The check node corresponds to a row in the check matrix, indicating a parity check condition that must be satisfied. The check node submodel is generally responsible for aggregating information from the variable nodes it is connected to and generating a feature representation of the check node.
[0035] For LDPC codes with regular check matrices, each check node corresponds to the same number of variable nodes. For LDPC codes with irregular check matrices, the number of variable nodes corresponding to each check node may be different. Figure 3 and Figure 4 The diagram is a diagram of the situation where the check node is connected to two variable nodes and three variable nodes respectively. The scalar V of each check node is calculated by connecting the check node submodel of the variable node and the check node to obtain the value of the updated check node scalar V for updating V. The check nodes are grouped according to the number of variable nodes connected to the check node to obtain multiple groups of check nodes, and a check node submodel corresponding to each group of check nodes is constructed. Each group of check nodes shares the same check node submodel. In other words, for check nodes connected to different numbers of variable nodes, a check node submodel corresponding to the number of connections is set, and for check nodes connected to the same number of variable nodes, the same check node submodel is set to be shared. Figure 3 and Figure 4As shown, two check nodes connected to different numbers of variable nodes are respectively set with two different check node sub-models, and the neural network parameters of different check node sub-models are set and initialized differently. Figure 3 W2 and Figure 4 The w3 in represents a neural network with different parameters, and σ represents the nonlinear transformation in the check node sub-model.
[0036] The check node submodel includes a first input layer, a first neural network layer and a first output layer. The first input layer receives the current variable node scalar, the current check node scalar, the received information and the current position code. The first neural network layer processes the information output by the first input layer and uses a nonlinear transformation to obtain the check node scalar update value. The first output layer calculates and outputs a new check node scalar based on the check node scalar update value output by the first neural network layer. The new check node scalar is equal to the sum of the check node scalar update value and the check node current scalar. In other words, the input of the check node submodel is the variable node scalar U i , check node scalar V i , receive information y and current position code p. i 、V i , y and p are taken as input, processed by the first neural network layer, and a nonlinear transformation σ is applied to obtain the check node scalar update value ΔV i , the output of the check node sub-model is the new check node scalar V i+1 , V i+1 ΔV is the scalar update value for the check node i and the current check node scalar V i The sum of the check node scalar update value ΔV i and the new check node scalar V i+1 It is expressed as: Where W 校验 is the variable of the first neural network layer, and the received information y is the information received by the LDPC decoding model.
[0037] It should be noted that the current variable node scalar U i It is a tensor. Although it is called a scalar, its actual dimension is a set constant.
[0038] In this embodiment, the current variable node scalar U i It is the information passed from the current variable node to the current check node, indicating a measure or confidence level of the current state of the current variable node. iIndicates the state of the current check node before the update, which may be the accumulated information in the previous iteration or some form of confidence. The received information y is the original information received from the channel or other external source, or the information transmitted by other check nodes. The current position code p is used to introduce position information in the check node sub-model. The current position code can help the check node sub-model understand the importance or difference of information in different positions or orders.
[0039] In this embodiment, the nonlinear transformation σ refers to a method of mapping values in the input space to the output space, which does not follow a linear relationship. Specifically, the nonlinear transformation does not satisfy the additivity and homogeneity of the linear transformation, that is, for any input a and b, and scalar k, the nonlinear transformation σ does not satisfy σ(a+b)=σ(a)+σ(b) and σ(ka)=kσ(a).
[0040] See also Figure 5 As shown, illustratively, the variable W of the first neural network in the check node submodel 校验 ={W1, W2}, during the operation of the check node sub-model, the input variable node scalar set {U i} and the current position code are added element-wise to obtain the result X i Then, X i Through W1 and nonlinear transformation σ, we get Z i At the same time, the received information y is passed through W2 and nonlinear transformation σ to obtain Z2, and then Z i Perform element-wise multiplication with Z2 to obtain Z, and then split Z into two parts Z a and Z b , and Za and Z b and the check node scalar V i Perform element-wise multiplication to get V a and V b Finally, V a and V b Add together to get the new check node scalar V i+1 .
[0041] In this embodiment, in step S102, the variable node submodel is responsible for processing information related to the variable node. The variable node corresponds to the column in the check matrix of the LDPC code, representing the coded bit. In the graph neural network, the variable node submodel includes a nonlinear transformation of the node features and an information exchange mechanism with adjacent check nodes.
[0042] Similarly, for LDPC codes with regular check matrices, each variable node corresponds to the same number of check nodes, and for LDPC codes with irregular check matrices, the number of check nodes corresponding to each variable node may be different. Figure 6 and Figure 7 The following is a schematic diagram of the case where a variable node is connected to two check nodes and three check nodes respectively. Each variable node scalar U is calculated through the variable node sub-model to obtain the updated value of the variable node scalar U for updating U. The variable nodes are grouped according to the number of check nodes connected to the variable nodes to obtain multiple groups of variable nodes, and a variable node sub-model corresponding to each group of variable nodes is constructed. The variable nodes in each group share the same variable node sub-model. In other words, for variable nodes connected to different numbers of check nodes, the variable node sub-model corresponding to the number of connections is used, and for variable nodes connected to the same number of check nodes, the same variable node sub-model is shared. Figure 6 and Figure 7 As shown, two variable nodes connected with different numbers of check nodes are respectively set with two different variable node sub-models. Figure 6 The l2 and Figure 7 The l3 in represents a neural network with different parameters, and γ represents the nonlinear transformation in the variable node sub-model.
[0043] In this embodiment, the variable node sub-model includes a second input layer, a second neural network layer, and a second output layer. The second input layer receives the current variable node scalar, the check node scalar, the received information, the Gaussian representation of the girth, and the current position encoding. The second neural network layer processes the information output by the second input layer and uses a nonlinear transformation to obtain the variable node scalar update value. The second output layer outputs a new variable node scalar, which is equal to the sum of the variable node scalar update value and the current variable node scalar. That is, the input of the variable node sub-model is the current variable node scalar U i , current check node scalar V i , receive information y, Gaussian representation of girth r and current position code p. i 、V i , y, r and p are used as inputs, processed by the second neural network layer, and a nonlinear transformation σ is used to obtain the variable node scalar update value ΔU i , the output of the variable node sub-model is the new check node scalar U i+1 , U i+1 Update the value ΔU for the variable node scalar i and the current variable node scalar U i The sum of the variable node scalar update value ΔU i and the new check node scalar U i+1 It is expressed as: Where W 变量 are the variables of the second neural network layer.
[0044] In this embodiment, the length of the girth of the variable node is assigned equivariant information, and a Gaussian kernel is used for calculation to obtain a Gaussian representation r of the girth.
[0045] See also Figure 8 As shown, illustratively, the variable W of the second neural network of the variable node sub-model is variable = {W3, W4, W5}, and during the operation of the variable node sub-model, the input check node scalar set {V i} and the current position code are added element-wise to obtain the result X i Then, X i Through W3 and nonlinear transformation σ, we get Y i At the same time, the received information y is passed through W4 and nonlinear transformation σ to obtain Y2, and then Y i Perform element-wise multiplication with Y2 to obtain Y, and then split Y into two parts Y a and Y b , and Y a and Y b Respectively with the variable node scalar U i And the Gaussian representation r of the girth is transformed and multiplied element-wise to obtain U a and U b , and finally a and U b Add together to get the new check node scalar U i+1 , where U a and U b Respectively expressed as: U a =Y a ☉σ(W5 ˙ U i );U b =Y b ☉σ(W5 ˙ r).
[0046] In this embodiment, in step S103, the variable node decoding sub-model is mainly responsible for receiving information from the check node and the initial information of the channel, performing decoding operations, and generating new external information to pass to the check node. After the check sub-model and the variable node model complete the specified number of iterations, the U of each variable node is obtained. final and the V of the check node final During decoding, the variable node decoding sub-model is used to decode each variable node. The input of the variable node decoding sub-model is the U final and the scalar collection of check nodes connected to it {V final}, the output of the variable node decoding sub-model is the codeword posterior probability prediction pred, where pred is the probability of the decoding result being 0 and 1. The decoding result can be obtained by making a hard decision on the symbol based on pred.
[0047] In this embodiment, a belief propagation (BP) algorithm, a minimum sum (Min-Sum) algorithm, and a normalized minimum sum (Normalized Min-Sum) algorithm may be used to construct a variable node decoding sub-model.
[0048] In step S103 , the update strategy of the check node sub-model and the variable node sub-model includes updating the variable node scalar U and updating the check node scalar V.
[0049] The variable node scalar U is updated as follows:
[0050] For an LDPC code with a code length of n, the information received by the LDPC decoding model is recorded as the received information y=[y0,y1,...y n-1 ], when the variable node sub-model is used for the ith time to update the variable node scalar U, the input of the variable node sub-model is the set of check scalars connected to the variable node {V i} (the scalar V of all check nodes connected to the current node i ), receiving information y, current position code p, Gaussian representation r of the girth to which the current node belongs, and current variable node scalar U i The current position code p of the current variable node is created using sine code and cosine code. Assuming that the current position code is d bits, d is an even number, then for any variable node c q Current position code It is expressed as: ; In the formula, represents the sinusoidal part of the current position code, Represents the cosine part of the current position code, code is a row vector with values from 0 to (d / 2)-1, and Together they form a current position code of length d, where d is an even number greater than 2 (such that d / 2 is an integer greater than 1), D is a constant, and q∈{0, 1, 2...n-1}.
[0051] The Gaussian representation of the girth r is also d bits, and the Gaussian representation of the girth r is expressed as: ; where code_r is a row vector with values from 0 to d, D_r is a constant set according to the girth in the H check matrix, and ring is the length of the ring containing the current variable node and the check nodes around the check node.
[0052] Finally, the variable node scalar update value ΔU is calculated by the variable node sub-model i, Variable node scalar update value ΔU i and Ui Add together to get the new variable node scalar U i+1 .
[0053] The update of the scalar V of the check node is as follows:
[0054] For an LDPC code with a code length of n, the information received by the LDPC decoding model is recorded as the received information y=[y0,y1,...y n-1 ], when the check node sub-model is used for the ith time to update the check node scalar V, the input of the check node sub-model is the current variable node scalar V i , current check node scalar V i , receive information y and current position code p, assuming that the current position code is d bits, d is an even number, then for any check node s q Current position code It is expressed as: ; In the formula, represents the sinusoidal part of the current position code, Represents the cosine part of the current position code, where code is a row vector with values ranging from 0 to (d / 2)-1. and Together they form a current position code of length d, where d is an even number greater than 2 (so that d / 2 is an integer greater than 1), and D is a constant that is set according to requirements, q∈{0, 1, 2...n-1}.
[0055] Finally, the check node sub-model calculates the check node scalar update value ΔV i, Check node scalar update value ΔV i and V i Add together to get the new check node scalar V i+1 .
[0056] In this embodiment, in step S104, a data set is collected, and the data set is used to train a graph neural network model to obtain an LDPC decoding model, including: collecting historical input data and historical output data of the channel LDPC code; grouping the collected data according to the signal-to-noise ratio, and randomly extracting a preset number of data from each group and putting them into the data set; dividing the data set into a training set and a test set, and using the training set to train the graph neural network model to obtain a trained graph neural network model; inputting the test set into the trained graph neural network model for testing and tuning to obtain an LDPC decoding model.
[0057] In this embodiment, the input data includes the original binary information bit sequence (or message bits), which are LDPC encoded to generate coded bits. The output data refers to the signal received by the receiving end after being transmitted through the channel (which may be interfered by noise). The collected data is grouped according to its corresponding signal-to-noise ratio (SNR) value. Each group contains the data transmitted and received under specific SNR conditions.
[0058] In this embodiment, a preset amount of data is randomly selected from each group and put into the data set, which helps to ensure the diversity and representativeness of the data set, while controlling the size of the data set to suit training requirements.
[0059] In this embodiment, the training set is 80% of the data set, and the test set is 20% of the data set.
[0060] During the testing process, if you find that the model performs poorly on the test set (such as overfitting), you can try to use regularization techniques (such as L1 / L2 regularization, Dropout, etc.) to tune the model. Regularization is a technique to reduce model overfitting and help improve the generalization ability of the model.
[0061] In this embodiment, other methods (such as stratified sampling) may also be used to obtain the data set.
[0062] In this embodiment, when constructing the LDPC decoding model, the nodes are grouped according to the number of nodes connected to the check nodes and the variable nodes, and a dedicated sub-model is constructed for each group of nodes. This grouping method not only fully considers the differences in the connection conditions of different nodes in the LDPC code, but also reduces the model parameters by sharing sub-models, reduces the computational complexity, and enhances the generalization ability of the model. Moreover, in order to further improve the performance of the model, when constructing the LDPC decoding model, the position coding mechanism is integrated into the update strategy of the check node sub-model and the variable node sub-model. The position coding enables the model to make full use of the position information of the nodes in the Tanner graph, thereby more accurately understanding the dependency between the nodes, further optimizing the decoding process, and thus making the constructed LDPC decoding model have higher decoding accuracy.
[0063] See also Fig. 9 As shown, the embodiment of the present invention further provides an LDPC decoding method, including:
[0064] S201, constructing an LDPC decoding model;
[0065] S202: Receive initial information, and input the initial information into an LDPC decoding model to obtain a decoding result output by the LDPC decoding model.
[0066] In this embodiment, the LDPC decoding model is constructed by using the LDPC decoding model construction method as described above, which will not be described in detail here.
[0067] In this embodiment, initial information is received, and the initial information is input into an LDPC decoding model to obtain a decoding result output by the LDPC decoding model, including: receiving the initial information and preprocessing the initial information; extracting features from the preprocessed initial information; and inputting the extracted features into the LDPC decoding model to obtain a decoding result output by the LDPC decoding model.
[0068] In this embodiment, the preprocessing of the initial information includes signal demodulation, sampling, quantization, format conversion and other processing.
[0069] In this embodiment, features are extracted from the pre-processed initial information, such as by calculating the log-likelihood ratio (LLR) of each received symbol.
[0070] In this embodiment, the LDPC decoding method also includes: monitoring and recording error cases in which decoding fails during the decoding process of the LDPC decoding model; for each error case, extracting the final iteration result as a training sample, the training sample includes the final value of the variable node and the final value of the corresponding check node; constructing a correction model, and using the training sample to train the correction model to obtain a trained correction model; using the trained correction model to assist the LDPC decoding model in decoding.
[0071] In this embodiment, the correction model can be based on machine learning (such as neural networks, decision trees, etc.) or other types of models (such as rule-based systems). The goal of the correction model is to verify the final value V of the node final (and possibly other contextual information) predicts a correction vector that will be consistent with the final value U of the variable node final In the LDPC decoding process, the trained correction model is used to assist in decoding. When the decoding model encounters a situation that is difficult to judge, the correction model can be called to intervene and correct it to improve the decoding accuracy.
[0072] In this embodiment, when constructing the LDPC decoding model, the nodes are grouped according to the number of nodes connected to the check nodes and the variable nodes, and a dedicated sub-model is constructed for each group of nodes. This grouping method not only fully considers the differences in the connection conditions of different nodes in the LDPC code, but also reduces the model parameters and the computational complexity by sharing the sub-model, while enhancing the generalization ability of the model. Moreover, in order to further improve the performance of the model, when constructing the LDPC decoding model, the position encoding mechanism is integrated into the update strategy of the check node sub-model and the variable node sub-model. The position encoding enables the model to fully utilize the position information of the node in the Tanner graph. information, so as to more accurately understand the dependencies between nodes and further optimize the decoding process, so that the constructed LDPC decoding model has higher decoding accuracy; in the decoding stage, the LDPC decoding model first obtains the feature representation of the variable node through an iterative process, and then these features are sent to the variable node decoding sub-model, which outputs the posterior probability of the codeword based on the learned features. The introduction of the posterior probability provides richer information for the decoding process, so that the model can more accurately judge the value of each symbol. Finally, based on these posterior probabilities, the LDPC decoding model judges the symbol and obtains the decoded vector.
[0073] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. All equivalent structural changes made by using the contents of the present invention specification and drawings under the inventive concept of the present invention, or directly / indirectly applied in other related technical fields are included in the patent protection scope of the present invention.
Claims
1. A method for constructing an LDPC decoding model, characterized in that: The LDPC decoding model construction method comprises: The check nodes are grouped according to the number of variable nodes connected to the check nodes to obtain multiple groups of check nodes, and a check node sub-model corresponding to each group of check nodes is constructed, and the check nodes in each group share the same check node sub-model, wherein the check node sub-model includes a first input layer, a first neural network layer, and a first output layer, wherein the first input layer receives a current variable node scalar, a current check node scalar, received information, and a current position code, the first neural network layer processes the information output by the first input layer, and uses a nonlinear transformation to obtain a check node scalar update value, and the first output layer calculates and outputs a new check node scalar according to the check node scalar update value output by the first neural network layer, and the new check node scalar is equal to the sum of the check node scalar update value and the current check node scalar; The variable nodes are grouped according to the number of check nodes connected to the variable nodes to obtain multiple groups of variable nodes, and a variable node sub-model corresponding to each group of variable nodes is constructed, and the variable nodes in each group share the same variable node sub-model; the variable node sub-model includes a second input layer, a second neural network layer, and a second output layer, the second input layer receives the current variable node scalar, the check node scalar, the received information, the Gaussian representation of the girth, and the current position encoding, the second neural network layer processes the information output by the second input layer, and uses a nonlinear transformation to obtain the variable node scalar update value, and the second output layer outputs a new variable node scalar, and the new variable node scalar is equal to the sum of the variable node scalar update value and the current variable node scalar; Construct a variable node decoding sub-model, the input of which is the final value of the variable node and the set of check node scalars connected to it, and the output of which is the codeword posterior probability prediction pred, where pred is the probability that the decoding result is 0 and 1, integrate the check node sub-model, the variable node sub-model and the variable node decoding sub-model, and integrate the position encoding mechanism into the update strategy of the check node sub-model and the variable node sub-model to obtain a graph neural network model; Collect a data set, and use the data set to train the graph neural network model to obtain an LDPC decoding model.
2. The LDPC decoding model construction method according to claim 1, characterized in that: The update strategy of the check node submodel and the variable node submodel includes the update of the variable node scalar and the update of the check node scalar. The current position code of the variable node is created using sine code and cosine code, and the current position code of the check node is created using sine code and cosine code.
3. The LDPC decoding model construction method according to claim 2, characterized in that: For any variable node c q Current position code It is expressed as: ; In the formula, represents the sinusoidal part of the current position code, Represents the cosine part of the current position code, code is a row vector with values from 0 to (d / 2)-1, and Together they form a current position code of length d, where d is an even number greater than 2, D is a constant, and q∈{0, 1, 2...n-1}.
4. The LDPC decoding model construction method according to claim 3, characterized in that: For any check node s q Current position code It is expressed as: ; In the formula, represents the sinusoidal part of the current position code, Represents the cosine part of the current position code, code is a row vector with values from 0 to (d / 2)-1, and Together they form a current position code of length d, where d is an even number greater than 2, D is a constant, and q∈{0, 1, 2...n-1}.
5. The LDPC decoding model construction method according to claim 1, characterized in that: The collecting of the data set and using the data set to train the graph neural network model to obtain the LDPC decoding model include: Collect historical input data and historical output data of channel LDPC codes; The collected data are grouped according to the signal-to-noise ratio, and a preset number of data are randomly selected from each group and put into the data set; Divide the data set into a training set and a test set, and use the training set to train the graph neural network model to obtain a trained graph neural network model; The test set is input into the trained graph neural network model for testing and tuning to obtain the LDPC decoding model.
6. A LDPC decoding method, characterized in that: The LDPC decoding method comprises: Constructing an LDPC decoding model, wherein the LDPC decoding model is constructed by using the LDPC decoding model construction method according to any one of claims 1 to 5; Initial information is received, and the initial information is input into the LDPC decoding model to obtain a decoding result output by the LDPC decoding model.
7. The LDPC decoding method according to claim 6, wherein: The receiving initial information and inputting the initial information into the LDPC decoding model to obtain a decoding result output by the LDPC decoding model includes: Receive initial information and pre-process the initial information; Extract features from the preprocessed initial information; The extracted features are input into the LDPC decoding model to obtain a decoding result output by the LDPC decoding model.
8. The LDPC decoding method according to claim 6, wherein: Also includes: Monitor and record error cases of decoding failure during the decoding process of the LDPC decoding model; For each error case, extract the last iteration result as a training sample, wherein the training sample includes the final value of the variable node and the final value of the corresponding check node; Constructing a correction model, and using the training samples to train the correction model to obtain a trained correction model; The trained correction model is used to assist the LDPC decoding model in decoding.
Citation Information
Patent Citations
Neural networks for forward error correction decoding
CN110741553A
Self-learning normalized offset minimum sum decoding method for LDPC codes
CN111565051A
Methods and systems for intelligent collection and analysis of vehicle data
US20190025813A1
Neural networks for forward error correction decoding
US20210142158A1
Scheduling method for LDPC decoding
US20210152282A1
