Neural minimum and sliding window decoding method for SC-LDPC code

Abstract

The invention provides a neural minimum and sliding window decoding method for SC-LDPC codes, and the method comprises the steps: building and training a neural network, so as to determine the optimal parameters of different types of edges of a fundamental model graph in a window, and the input of the neural network is a basis matrix of a window plus a memory length; the determined optimal parameters are applied to each window of the sliding window decoding algorithm so as to decode the SC-LDPC after length coupling, and compared with a traditional method, the method has better performance and faster convergence.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and particularly to a neural minimum sum sliding window decoding method for SC-LDPC codes. Background Art

[0002] Spatial coupled low density parity check codes (SC-LDPC codes), as an error correction coding technology attracting much attention in the field of communication, the key lies in the design of spatial coupling. By connecting multiple LDPC subcodes in series to form a spatial coupling structure, this coding method can significantly improve the decoding performance, especially in the high bit error rate environment. In terms of channel coding and decoding, there is a close connection between neural networks and error correction coding technologies in mathematical methods. In recent years, neural network-assisted communication systems have gradually been introduced into the field of communication coding and decoding. A large number of literatures show that neural network-assisted coding schemes not only successfully reduce the algorithm calculation complexity, but also significantly improve the decoding performance. Currently, one of the main challenges regarding the combination of deep learning and channel coding and decoding technologies is how to make full use of this tool of deep learning to better generalize the internal relationship of data during the decoding process. This has an important impact on the application of neural networks in future channel coding and decoding, and becomes one of the key research directions at present.

[0003] For the decoding method of SC-LDPC codes, the sliding window decoding algorithm is generally adopted. The core idea of this algorithm is to gradually decode the encoded data by using the sliding window mechanism according to the special coupling structure of the SC-LDPC parity check matrix. In the traditional sliding window decoding algorithm, the belief propagation decoding algorithm is used within the window. By setting the window size as W (1 < W < L), when decoding, the leftmost base graph unit within the window is set as the target symbol, and only the target symbol is decoded during decoding. After decoding is completed, the window slides to the right to perform the decoding of the next target symbol until the decoding of the last base graph unit is completed. Then the window returns to the initial position to start the decoding of the next codeword. During the decoding process, except for the initial window, since there is message passing between each base graph unit and the adjacent ω base graph units, when decoding within the current window, some edges of the previous window will be included, but they will not be re-initialized. Instead, the information saved after the decoding of the previous window is directly used to reduce the number of iterations and the overall decoding complexity. At the same time, since only a part of the entire codeword is estimated each time, the parity check matrix cannot be used to verify the decoding accuracy. Instead, it is determined by whether the bit error rate of the current target symbol is zero or by terminating the decoding when the maximum number of iterations is reached.

[0004] Although the belief propagation algorithm is relatively efficient in most cases, its computational complexity increases significantly when dealing with large-scale sparse graphs. Each iteration involves a large number of message passing and update operations, which requires a large amount of computation. In addition, for short code lengths, there are often many short loops in the graph, which can lead to poor convergence of the algorithm and may even lead to incorrect decoding results. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an ordered statistics decoding method, device and apparatus based on a belief propagation list, which specifically include:

[0006] In a first aspect, the present invention provides a neural minimum-sum sliding window decoding method for SC-LDPC codes, comprising:

[0007] Build and train a neural network to determine the best parameters for different types of edges in the base model graph within the window. The input of the neural network is a basis matrix of the size of the window plus the memory length.

[0008] The determined optimal parameters are applied to each window of the sliding window decoding algorithm to decode the length-coupled SC-LDPC.

[0009] In a second aspect, the present invention further provides a neural minimum-sum sliding window decoding device for SC-LDPC codes, comprising:

[0010] The training module is used to build and train a neural network to determine the optimal parameters for different types of edges in the base model graph within the window. The input of the neural network is a basis matrix of the size of a window plus the memory length;

[0011] The processing module is used to apply the determined optimal parameters to each window of the sliding window decoding algorithm to decode the length-coupled SC-LDPC.

[0012] In a third aspect, the present invention further provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0013] Memory, used to store computer programs;

[0014] The processor is used to implement any method provided in the first aspect when executing a program stored in the memory.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, any one of the methods provided in the first aspect is implemented.

[0016] Beneficial effects of the present invention:

[0017] The neural minimum-sum sliding window decoding method for SC-LDPC codes provided by the present invention determines the optimal parameters of different types of edges of a base mode graph in a window by building and training a neural network, wherein the input of the neural network is a basis matrix of a window plus a memory length; the determined optimal parameters are applied to each window of a sliding window decoding algorithm to decode the length-coupled SC-LDPC. The method combines a neural network with an algorithm in a sliding window decoding window for the special structure of a spatially coupled low-density parity-check code (SC-LDPC) to construct a neural MS decoder, and compared with traditional sum-product algorithm, minimum-sum algorithm, normalized minimum-sum algorithm and offset minimum-sum algorithm decoding, the method shows a certain performance gain, and has better performance and faster convergence.

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of a neural minimum-sum sliding window decoding method for SC-LDPC codes provided by the present invention;

[0020] Figure 2 A schematic diagram of the architecture of a neural network provided by the present invention;

[0021] Figure 3 A schematic diagram of a base matrix provided by the present invention;

[0022] Figure 4 A schematic diagram of sliding window decoding provided by the present invention;

[0023] Figure 5 A schematic diagram of a set of experimental data provided by the present invention;

[0024] Figure 6 Another set of experimental data schematic diagram provided by the present invention;

[0025] Figure 7 Another set of experimental data schematic diagrams provided for the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0027] In order to solve the problems existing in the prior art, the present invention uses a sliding window decoding algorithm for SC-LDPC code decoding, and adopts a hybrid minimum sum (HMS) algorithm in the window, that is, combining the NMS and OMS algorithms, and adding a normalization factor and an offset on the basis of the MS algorithm. At the same time, a neural network is used to autonomously learn the parameters in the HMS algorithm, which are applied to each window to form a neural MS decoder. And based on the coupling structure of the SC-LDPC code, when training the neural network, a base matrix of a window size plus a memory length is used as input, and the size of the parameter array of each hidden layer in the network is equal to the number of edge types of the input base matrix. The network is trained using a successive iterative greedy algorithm.

[0028] Figure 1 A flowchart of a neural minimum-sum sliding window decoding method for SC-LDPC codes provided by the present invention is shown in FIG. Figure 1 As shown, including:

[0029] S101. Build and train a neural network to determine optimal parameters for different types of edges of a base model graph within a window, wherein the input of the neural network is a base matrix of a window plus a memory length.

[0030] Optionally, the neural network is a sparse MS decoding network with a non-fully connected structure. Figure 2 As shown, the input layer and output layer of the neural network correspond to the variable node V in the base model diagram N , the number is N, and the sizes of the remaining hidden layers in the middle are the number of edges in the base model graph.

[0031] According to the structure of SC-LDPC code, the input of the neural network designed by the present invention adopts a basis matrix of the size of a window plus a memory length, which can effectively reduce the complexity and storage cost of neural network training, and can also be more suitable for windows in different situations under sliding window decoding.

[0032] For example, taking the (3, 6) SC-LDPC code as an example, the memory length ω = 2, if the window size is set to 2, the input basis matrix is ​​as follows Figure 3 shown.

[0033] Optionally, including: the neurons in each hidden layer include Z clusters, each cluster contains the corresponding total number of edges E in the base model graph b elements, represented by:

[0034]

[0035] Among them, E represents the total number of edges of the base model graph after promotion, Z represents the promotion factor, and E b represents the total number of edges of the base model graph before promotion, b i represents the i-th cluster represents the eth in the Ith iteration i Normalization factor of the edge, represents the eth in the Ith iteration j Normalization factor of the edge, represents the eth in the Ith iteration i The offset of the edge, represents the eth in the Ith iteration j The offset of the edge, e i represents any edge of the i-th cluster, e j represents any edge of the jth cluster, i and j correspond to the same position, represents all edges of the i-th cluster, represents all the edges of the j-th cluster.

[0036] The parameters are shared between the edges of the same type corresponding to each element, expressed as:

[0037]

[0038]

[0039] in, represents the eth in the Ith iteration i Normalization factor of the edge, represents the eth in the Ith iteration j Normalization factor of the edge, represents the eth in the Ith iteration i The offset of the edge, represents the eth in the Ith iteration j The offset of the edge, e i is an edge of the ith cluster, e j is an edge of the jth cluster, represents all the edges of the i-th cluster.

[0040] i and j correspond to the same position, that is, e i and e j Corresponding to edges with the same number in different clusters.

[0041] Since the elements in cluster Z are derived from the same base model graph, the corresponding positions belong to the same edge type, so the parameter sharing mechanism can be implemented through the above method, that is, the edges of the same type share a set of parameters. Through the parameter sharing mechanism, the number of parameters α and β in the iteration process is only related to the number of edges in the base model graph and the number of iterations, thereby saving network storage space.

[0042] See also Figure 2 , Optionally, each hidden layer in the neural network includes a first sublayer and a second sublayer,

[0043] The first sublayer corresponds to the process of variable nodes passing messages to check nodes along the associated edges, which is represented by i v , when i=1, 1 v Each neuron in is connected to the vth bit of the input layer according to the corresponding edge e = (v, c). When i>1, i v Each neuron in the layer e=(v,c) is connected to (i-1) according to e'=(v',c), v'≠v c The corresponding neurons in the layer;

[0044] The second sublayer corresponds to the process of the check node transmitting messages to the variable node, which is represented by i c ,i c The neurons in layer e=(v,c) are connected to sublayer i according to e'=(v',c), v'≠v v When the neurons associated with it reach the last hidden layer, they are connected to the vth bit of the output layer according to e=(v,c) to obtain the output result.

[0045] Optionally, given a maximum number of iterations I (i.e., the number of hidden layers), the message passing process in the neural network is:

[0046] 1. When the variable node information is received from the channel, the variable node information is initialized, which is expressed as:

[0047]

[0048] Among them, x v Represents the vth element of the codeword sequence transmission, y v represents the vth element of the received vector y, Pr(y v |x v =0) indicates that from x v to y v The transition probability, ln represents the logarithm with base e, l v Indicates the initialized variable node information.

[0049] 2. According to the initialization result, the first sub-layer i v Updates, i v The layer data processing process is expressed as:

[0050]

[0051] in, Indicates the i-th iteration i v The information of edge e in the sublayer, Indicates the i-1th iteration i c Information about edge e in the sublayer, i=1,2,...,I, e=(v,c) represents all check nodes associated with variable node v, e'=(v,c'), c'≠c represents all edges except check node c.

[0052] 3. Perform the second sub-layer i c Updates, i c The layer data processing process is expressed as:

[0053]

[0054] in, Indicates the i-th iteration i c The information of edge e in the sublayer, Indicates i v The positive and negative information of the sub-layer edge e', represents the normalization factor of the e-th edge in the i-th iteration, represents the offset of the e-th edge in the i-th iteration, e'=(v',c), v'≠v represents all edges connected to the check node c except the variable node v, ReLU(x)=max(x,0), α, β are learnable parameters, the weight α represents the normalization factor, and β represents the offset.

[0055] The second sublayer i c The updating process adopts HMS algorithm, combined with NMS and OMS algorithms, which can better compensate for the minimum and loss.

[0056] 4. Get the output neuron message, expressed as:

[0057]

[0058] Among them, v represents the vth element of the neural network output, σ represents the sigmoid function, l v Indicates the variable node information obtained by initialization. Indicates that in the last iteration I, edge e' is in I c Layer information, e'=(v,c'), c'≠c represents all edges with variable node v except check node c.

[0059] Optional, for convenience of indicating the transmission bit x v =0, we can use the sigmoid function σ(x)= ( 1+exp(-x) ) - 1 , limiting the output neuron message to between 0 and 1.

[0060] 5. Update the neural network parameters based on the output neuron message, the preset loss function and the gradient descent method.

[0061] Optional, default loss function, expressed as:

[0062]

[0063] Among them, L(o,x) represents the loss function, o represents the neural network output vector, x represents the sent codeword vector, N represents the codeword length, and o represents the output vector of the neural network. v represents the vth neuron message output by the neural network, x v represents the vth element of the codeword sequence transmission, γ v Indicates the degree of influence of data at different positions in the window on the current decoding symbol.

[0064] The loss function uses the expected cross entropy between the transmitted codeword and the output of the neural network. At the same time, when the sliding window decoding algorithm is used for decoding, the closer the data in the window is to the left, the greater the impact on the current target symbol. Therefore, the γ factor is introduced to indicate the degree of influence of data at different positions in the window on the current decoded symbol, so that the trained parameters can be better used in the sliding window decoding algorithm.

[0065] Optionally, the neural network parameters are updated according to a preset loss function and a gradient descent method, including: estimating the loss function value based on a preset initial normalization factor α, an offset β and a preset loss function; and calculating the gradient according to the estimation result. Based on the calculation results, the values ​​of the parameters α and β in the neural network are updated using the gradient descent method.

[0066] Specifically, in the training process of the neural network, in order to improve the generalization ability of the neural MS decoder for different code lengths, the present invention uses multiple sets of boosted codes derived from the same base model graph. In each iteration, different samples are randomly selected for training, so that the trained neural MS decoder can be applicable to codes of various code lengths. In addition, a layer-by-layer greedy training method is used to construct and optimize the network, so as to reduce the possibility of gradient vanishing problems during training. The training process is simplified by gradually increasing the depth of the network, so that the network is gradually expanded into a multi-layer network that can be iterated multiple times. After each expansion, the newly added layers are trained on the basis of the existing layers, and only the parameters of the most recently added layer are learnable.

[0067] The training process is shown in the following table.

[0068]

[0069]

[0070] S102: Apply the determined optimal parameters to each window of the sliding window decoding algorithm to decode the length-coupled SC-LDPC.

[0071] As Figure 4 shown, by setting the window size to W, where 1 < W < L, during current decoding, the leftmost base graph unit is set as the target symbol. Using the hybrid min-sum algorithm, the parameters trained by the neural network are brought into the corresponding edge update process within the window to complete decoding, and only the symbols contained within the current window are decoded. After decoding is completed, the window slides to the right for decoding of the next window. This process continues until decoding of the last base graph unit is completed, and the window returns to its initial position to start decoding of the next codeword.

[0072] The neural min-sum sliding window decoding method for SC-LDPC codes provided by the present invention determines the optimal parameters for different types of edges of the base graph within the window by building and training a neural network. The input of the neural network is a base matrix of the size of a window plus the memory length; the determined optimal parameters are applied to each window of the sliding window decoding algorithm to decode the SC-LDPC after length coupling. This method combines the neural network with the algorithm within the sliding window decoding window for the special structure of spatially coupled low-density parity-check codes (SC-LDPC), constructs a neural MS decoder, and shows certain performance gains compared with the traditional sum-product algorithm (SPA), min-sum algorithm (MSA), normalized min-sum algorithm (NMSA), and offset min-sum algorithm (OMSA) decoding, having better performance and faster convergence.

[0073] To prove the beneficial effects of the present invention, the present invention also provides a set of experimental data as follows:

[0074] First, a neural decoding network suitable for SC-LDPC was constructed, and the parameters trained by it were used in the sliding window decoding algorithm of SC-LDPC codes with the same structure. The bit error rate (BER) and block error rate (BLER) performance under BPSK modulation over an additive white Gaussian noise (AWGN) channel for different code lengths were simulated.

[0075] In the actual training of the neural decoding network, the base code used was a (3, 6) regular SC-LDPC code with a window size of 4, forward memory, and backward memory lengths, a code rate of 1 / 2, and a code type structure of The model is of type, memory length ω=1, and the size of the input check matrix is ​​5 rows and 12 columns, with a total of 30 non-zero elements. The base codes with M values ​​of 60, 120, 160, and 200 are selected for random selection in each iteration. Since too many hidden layers will increase the complexity of network training and even cause the gradient disappearance problem, the number of hidden layers is set to 25. In each iteration, the parameters are trained for more than 10,000 batches, each containing 50 samples. At the same time, the ADAM optimizer with an initial learning rate of 0.001 is used for training to update the normalization factor and offset factor. At initialization, α=0.8, β=0.0 are set.

[0076] After training the neural decoding network, the obtained parameter distribution is as follows Figure 5 As shown, the normalization factors are mostly concentrated between 0.7 and 1.0, the offsets are concentrated between -0.1 and 0.2, and the values ​​vary with the number of iterations.

[0077] When the trained parameters are used for sliding window decoding of (3, 6) SC-LDPC codes with the same structure as the training, four codes with different M values ​​(60, 120, 160, 200) are tested. The coupling length is set to L = 50, the window size is W = 4, and the number of iterations within the window is I max =25, and on this basis, each generation is iterated 10 times (i.e. the total number of iterations is 250 times). In addition, SPA, MSA, NMSA and OMSA algorithms with the same parameter conditions are selected as controls. Among them, since NMSA with α=0.8 and OMSA with β=0.15 are regarded as the benchmarks for comparison in many existing works, the same settings are also used this time.

[0078] The test results are shown in the figure below. Figure 6 , Figure 7 The BER and BLER performance diagrams are shown respectively. When the coupling length is 50, the performance of the three algorithms, SPA, MSA and OMSA, in the control group are similar, and the performance of the NMSA algorithm is slightly better. When the signal-to-noise ratio is low, the performance of the five algorithms is similar; when the signal-to-noise ratio is greater than 2.5dB, the advantages of the neural MS decoder gradually emerge.

[0079] like Figure 6 As shown in the BER performance diagram, for Figure 6 For the shorter codes with M of 60 and 120 shown in (a) and (b), the gain of the neural MS decoder is more obvious. Figure 6 The middle (a) figure corresponds to the code with M=60. When the signal-to-noise ratio is between 3 and 5 dB, the neural MS decoder has a gain of about 0.5 dB compared to the NMSA decoding. Compared with the other three algorithms, it has a gain of about 1 dB. When the signal-to-noise ratio is greater than 5 dB, the neural MS decoder has a gain of about 1 dB compared to the NMS algorithm. Figure 6 The middle (b) figure shows the code with M=120. When the signal-to-noise ratio is between 3-4dB, the neural MS decoder has a gain of about 0.5dB compared with the NMS algorithm with better performance. When the signal-to-noise ratio is greater than 4dB, the gain gradually increases and finally remains at 1.5dB. Figure 6 The middle (c) figure corresponds to the code with M=160. When the signal-to-noise ratio is greater than 2.5dB, the neural MS decoder maintains a gain of about 0.3dB compared to the NMS algorithm; Figure 6 The middle (d) figure shows the code with M=200. When the signal-to-noise ratio is greater than 2.5dB, the neural MS decoder has a gain of about 0.5dB compared with the NMS algorithm. It can be seen that the neural network plays a role in optimizing the parameters.

[0080] like Figure 7 As shown in the BLER performance diagram, Figure 7 In the middle (a) figure, when the signal-to-noise ratio is greater than 2.5dB, the neural MS decoder begins to have a gain compared to the NMSA decoder with better performance, and the gain gradually increases until the signal-to-noise ratio reaches 5dB, when the gain remains at around 1.5dB; Figure 7 In the middle (b) figure, the neural MS decoder finally maintains a gain of about 1.5dB with the NMSA decoding; Figure 7 In the middle (c) figure, after the signal-to-noise ratio is greater than 2.5dB, the neural MS decoder and NMSA decoding always maintain a gain of about 0.3dB; Figure 7 In the middle figure (d), the neural MS decoder always maintains a gain of about 0.5dB with the NMSA decoding.

[0081] Aiming at the special structure of spatially coupled low-density parity-check code (SC-LDPC), the neural network is combined with the sliding window decoding algorithm to construct a neural MS decoder, which shows certain performance gain compared with traditional SPA, MSA, NMSA and OMSA decoding.

[0082] The present invention also provides a structure of an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus.

[0083] Memory, used to store computer programs;

[0084] The processor is used to implement the steps provided in the above method embodiment when executing the program stored in the memory.

[0085] The communication interface is used for communication between the above electronic device and other devices.

[0086] The method provided in the embodiment of the present invention can be applied to electronic devices. Specifically, the electronic device can be: a desktop computer, a portable computer, an intelligent mobile terminal, a server, etc. This is not limited here, and any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0087] The present invention also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps provided in the above method embodiment are implemented.

[0088] As for the electronic device / storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the specific content and beneficial effects and other related matters can be referred to the partial description of the method embodiment.

[0089] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0090] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A neural minimum-sum sliding window decoding method for SC-LDPC codes, characterized in that: include: Building and training a neural network to determine optimal parameters for different types of edges of a base model graph within a window, wherein the input of the neural network is a basis matrix of the size of a window plus a memory length; The determined optimal parameters are applied to each window of the sliding window decoding algorithm to decode the length-coupled SC-LDPC.

2. The method according to claim 1, characterized in that The neural network is a sparse MS decoding network with a non-completely connected structure, comprising: The input layer and output layer of the neural network correspond to the variable nodes V in the base model graph N , the number is N, and the sizes of the remaining hidden layers in the middle are the number of edges in the base model graph.

3. The method according to claim 2, characterized in that include: Each neuron in the hidden layer includes Z clusters, each cluster containing the corresponding total number of edges E in the base model graph b elements, represented by: Among them, E represents the total number of edges of the base model graph after promotion, Z represents the promotion factor, and E b represents the total number of edges of the base model graph before promotion, b i represents the i-th cluster; The parameters shared between the edges of the same type corresponding to each of the elements are expressed as: in, represents the eth in the Ith iteration i Normalization factor of the edge, represents the eth in the Ith iteration j Normalization factor of the edge, represents the eth in the Ith iteration i The offset of the edge, represents the eth in the Ith iteration j The offset of the edge, e i represents any edge of the i-th cluster, e j represents any edge of the jth cluster, represents all edges of the i-th cluster, represents all the edges of the j-th cluster.

4. The method according to claim 3, characterized in that Each of the hidden layers in the neural network includes a first sublayer and a second sublayer, The first sublayer corresponds to the process of variable nodes passing messages to check nodes along the associated edges, which is represented by i v , when i=1, 1 v Each neuron in is connected to the vth bit of the input layer according to the corresponding edge e = (v, c). When i>1, i v Each neuron in the layer e=(v,c) is connected to (i-1) according to e'=(v',c), v'≠v c The corresponding neurons in the layer; The second sub-layer corresponds to the process of the check node transmitting the message to the variable node, which is represented by i c ,i c The neurons in layer e=(v,c) are connected to sublayer i according to e'=(v',c), v'≠v v When the neurons associated with it reach the last hidden layer, they are connected to the vth bit of the output layer according to e=(v,c) to obtain the output result.

5. The method according to claim 4, characterized in that The process of message passing in the neural network is: When variable node information is received from the channel, the variable node information is initialized, which is expressed as: Among them, x v Represents the vth element of the codeword sequence transmission, y v represents the vth element of the received vector y, Pr(y v |x v =0) indicates that from x v to y v The transition probability, ln represents the logarithm with base e, l v Indicates the variable node information obtained by initialization; According to the initialization result, the first sub-layer i v Updates, i v The layer data processing process is expressed as: in, Indicates the i-th iteration i v The information of edge e in the sublayer, Indicates the i-1th iteration i c Information about edge e in the sublayer, i=1,2,...,I, e=(v,c) represents all check nodes associated with variable node v, e'=(v,c'), c'≠c represents all edges except check node c; Perform the second sublayer i c Updates, i c The layer data processing process is expressed as: in, Indicates the i-th iteration i c The information of edge e in the sublayer, Indicates i v The positive and negative information of the sub-layer edge e', represents the normalization factor of the e-th edge in the i-th iteration, represents the offset of the e-th edge in the i-th iteration, e'=(v',c), v'≠v represents all edges connected to the check node c except the variable node v, ReLU(x)=max(x,0), α, β are learnable parameters, the weight α represents the normalization factor, and β represents the offset; The output neuron message is obtained, expressed as: Among them, v represents the vth element of the neural network output, σ represents the sigmoid function, l v Indicates the variable node information obtained by initialization. Indicates that in the last iteration I, edge e' is in I c Layer information, e'=(v,c'), c'≠c represents all edges with variable node v except check node c; The neural network parameters are updated according to the output neuron message, a preset loss function and a gradient descent method.

6. The method according to claim 5, characterized in that: The preset loss function is expressed as: Among them, L(o,x) represents the loss function, o represents the neural network output vector, x represents the sent codeword vector, N represents the codeword length, and o represents the output vector of the neural network. v represents the vth neuron message output by the neural network, x v represents the vth element of the codeword sequence transmission, γ v Indicates the degree of influence of data at different positions in the window on the current decoding symbol.

7. The method according to claim 5, characterized in that: The updating of neural network parameters according to a preset loss function and gradient descent method includes: Based on a preset initial normalization factor α, an offset β and the preset loss function, estimating a loss function value; Based on the estimated results, calculate the gradient Based on the calculation results, the values ​​of the parameters α and β in the neural network are updated using the gradient descent method.

8. A neural minimum-sum sliding window decoding device for SC-LDPC codes, characterized in that: include: A training module is used to build and train a neural network to determine the optimal parameters of different types of edges of the base model graph in the window, and the input of the neural network is a base matrix of the size of a window plus a memory length; The processing module is used to apply the determined optimal parameters to each window of the sliding window decoding algorithm to decode the length-coupled SC-LDPC.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, for implementing any of the methods described in claims 1-7 when executing a program stored in a memory.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images