Neural min-sum and sliding window decoding method for sc-lpd codes

Abstract

The application provides a neural minimum sum and sliding window decoding method for an SC-LDPC code. A neural network is built and trained to determine optimal parameters of different types of edges of a base module graph in a window. An input of the neural network is a base matrix with a window and a memory length. The determined optimal parameters are applied to each window of a sliding window decoding algorithm to decode the SC-LDPC after length coupling. Compared with a traditional method, the method has better performance and faster convergence.

Description

TECHNICAL FIELD

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

[0002] As a communication field's error correction coding technology, the key of space-coupled low-density parity-check code (SC-LDPC code) is the design of space coupling. By connecting multiple LDPC sub-codes to form a space-coupled structure, this coding method can significantly improve the decoding performance, especially in high error rate environments. In the field of channel coding and decoding, neural networks and error correction codes are closely related in mathematical methods. In recent years, neural network assisted communication systems have been gradually introduced into the field of communication coding and decoding. A large number of documents show that neural network assisted coding schemes not only successfully reduce the algorithmic complexity, but also significantly improve the decoding performance. At present, one of the main challenges of the combination of deep learning and channel coding and decoding technology is how to make full use of deep learning as a tool to better summarize the internal relationship of the data in the decoding process. This has an important influence on the application of neural networks in future channel coding and decoding, and has become one of the key directions of current research.

[0003] For the decoding method of SC-LDPC code, the sliding window decoding algorithm is generally used. The core idea of this algorithm is to use the sliding window mechanism to decode the coded data step by step according to the special coupling structure of the SC-LDPC check matrix. The traditional sliding window decoding algorithm uses the belief propagation decoding algorithm in the window. By setting the window size to W (1 < W < L), the leftmost base module graph unit in the window is set as the target symbol during decoding. Only the target symbol is decoded, and after decoding, the window slides to the right to decode the next target symbol. Until the decoding of the last base module graph unit is completed, the window returns to the initial position and starts decoding the next code word. During the decoding process, in addition to the initial window, since there is message passing between each base module graph unit and the adjacent ω base module graph units, when decoding in the current window, some edges of the previous window will be included, but they will not be reinitialized, but directly use the information saved after the decoding of the previous window to reduce the number of iterations and the complexity of the overall decoding. At the same time, since only a part of the entire code word is estimated each time, the check matrix cannot be used to verify the accuracy of the decoding, but the decoding is terminated by judging whether the error rate of the current target symbol is zero or by reaching the maximum number of iterations.

[0004] Although the belief propagation algorithm is efficient in most cases, its computational complexity increases significantly when dealing with large-scale sparse graphs, involving a large number of message passing and updating operations in each iteration, and the amount of calculation is large. And for short code length, there are often many short loops in the graph, which will cause the convergence of the algorithm to be poor, and even may cause false decoding results. SUMMARY

[0005] In order to solve the above problems existing in the prior art, the application provides an ordered statistic decoding method, device and equipment based on a belief propagation list, which specifically comprises:

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

[0007] The neural network is built and trained to determine the optimal parameters of different types of edges in the base graph within the window, and the input of the neural network is a base matrix with a window plus a 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 application further provides a neural minimum-sum sliding window decoding device for SC-LDPC codes, comprising:

[0010] The training module is configured to build and train the neural network to determine the optimal parameters of different types of edges in the base graph within the window, and the input of the neural network is a base matrix with a window plus a memory length;

[0011] The processing module is configured 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 application 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 complete mutual communication through the communication bus;

[0013] The memory is configured to store a computer program;

[0014] The processor is configured to execute the program stored on the memory to implement any method provided in the first aspect.

[0015] In a fourth aspect, the application provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and the computer program is executed by the processor to implement any method provided in the first aspect.

[0016] The application has the following beneficial effects:

[0017] The application provides a neural minimum sum and sliding window decoding method for SC-LDPC codes, wherein a neural network is built and trained to determine optimal parameters of different types of edges of a base graph in a window, and the input of the neural network is a base matrix with a window and a memory length; the determined optimal parameters are applied to each window of a sliding window decoding algorithm to decode the SC-LDPC after length coupling, the method combines the neural network with the algorithm in the window of the sliding window decoding, constructs a neural MS decoder, and compared with the traditional sum-product algorithm, the minimum sum algorithm, the normalized minimum sum algorithm and the offset minimum sum algorithm, the neural MS decoder has a certain performance gain and better performance and faster convergence.

[0018] The application will be further described in detail below in combination with the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

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

[0020] Figure 2 An architecture diagram of a neural network provided by the application;

[0021] Figure 3 A base matrix diagram provided by the application;

[0022] Figure 4 A sliding window decoding diagram provided by the application;

[0023] Figure 5 A set of experimental data diagrams provided by the application;

[0024] Figure 6 Another set of experimental data diagrams provided by the application;

[0025] Figure 7 Still another set of experimental data diagrams provided by the application. DETAILED DESCRIPTION

[0026] The application will be further described in detail below in combination with the drawings and embodiments, but the embodiments of the application are not limited thereto.

[0027] To address the problems existing in the prior art, this invention employs a sliding window decoding algorithm for SC-LDPC code decoding. Within the window, a Hybrid Minimum Sum (HMS) algorithm is used, combining the NMS and OMS algorithms, and adding a normalization factor and offset to the MS algorithm. Simultaneously, a neural network autonomously learns the parameters from the HMS algorithm and applies them to each window, forming a neural MS decoder. Furthermore, based on the coupled structure of the SC-LDPC code, during neural network training, a basis matrix of the window size plus the memory length is used as input. The parameter array size of each hidden layer in the network is equal to the number of edge types in the input basis matrix. The network is trained using an iterative greedy algorithm.

[0028] Figure 1 This invention provides a flowchart illustrating a neural minimum-sum sliding window decoding method for SC-LDPC codes, as shown below. Figure 1 As shown, it includes:

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

[0030] Optionally, the neural network can be a sparse MS decoding network with a partially connected structure. For example... Figure 2 As shown, the variable nodes V in the base schema diagram corresponding to the input and output layers of the neural network are... N The number of layers is N, and the size of the remaining hidden layers is equal to the number of edges in the base schema graph.

[0031] For the structure of SC-LDPC codes, this invention designs the input of the neural network to use a window plus a basis matrix of memory length, which can effectively reduce the complexity of neural network training and storage cost, and is also more applicable to different window conditions in sliding window decoding.

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

[0033] Optionally, this includes: neurons in each hidden layer comprising Z-clusters, each cluster containing the corresponding total number of edges E in the fundamental model graph. b Each element is represented as:

[0034]

[0035] Where E represents the total number of edges in the lifted fundamental graph, and Z represents the lifting factor. b b represents the total number of edges in the fundamental graph before promotion. i Represents the i-th cluster denotes the e-th edge in the i-th iteration, i a normalization factor of the e-th edge in the i-th iteration, denotes the e-th edge in the i-th iteration, j a normalization factor of the e-th edge in the i-th iteration, denotes the e-th edge in the i-th iteration, i an offset of the e-th edge in the i-th iteration, denotes the e-th edge in the i-th iteration, j an offset of the e-th edge in the i-th iteration, e i denotes an arbitrary edge of the i-th cluster, j denotes an arbitrary edge of the j-th cluster, i and j correspond to the same position, denotes all edges of the i-th cluster, denotes all edges of the j-th cluster.

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

[0037]

[0038]

[0039] wherein, denotes the e-th edge in the i-th iteration, i a normalization factor of the e-th edge in the i-th iteration, denotes the e-th edge in the i-th iteration, j a normalization factor of the e-th edge in the i-th iteration, denotes the e-th edge in the i-th iteration, i an offset of the e-th edge in the i-th iteration, denotes the e-th edge in the i-th iteration, j an offset of the e-th edge in the i-th iteration, e i denotes an arbitrary edge of the i-th cluster, j denotes an arbitrary edge of the j-th cluster, denotes all edges of the i-th cluster.

[0040] i and j correspond to the same position, i.e. e i and e j correspond to edges of the same number in different clusters.

[0041] Since the elements in the Z cluster are derived from the same base model graph, the corresponding positions belong to the same edge type, and thus the parameter sharing mechanism can be realized by the above method, i.e. edges of the same type share a set of parameters. Through the parameter sharing mechanism, the number of parameters a and b 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] Further referring to Figure 2 Optionally, each hidden layer in the neural network comprises a first sub-layer and a second sub-layer,

[0043] where the first sub-layer corresponds to the process of variable nodes passing messages to check nodes along the associated edges, denoted as i v When i = 1, 1 v Each neuron in the input layer is connected to the vth bit of the input vector x = (x1, x2,..., xN), and each neuron in the i v Each neuron in the i c th hidden layer is connected to the corresponding neuron in the (i-1)

[0044] The second sub-layer corresponds to the process of check nodes passing messages to variable nodes, denoted as i c , i c Each neuron in the i v th hidden layer is connected to the corresponding neuron in the i

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

[0046] 1. When receiving variable node information from the channel, the variable node information is initialized, denoted as:

[0047]

[0048] where x v represents the vth element of the transmitted codeword sequence, y v represents the vth element of the received vector y, Pr(y v | x v = 0) represents the transition probability from x v to y v , ln represents the logarithm with base e, and l v represents the initialized variable node information.

[0049] 2. According to the initialization result, the first sub-layer i v is updated, i v The data processing process of the i

[0050]

[0051] where, represents the information of edge e in the i v th iteration of the i th sub-layer, i = 1, 2,..., I, c represents the information of edge e in the i e = (v, c) represents all check nodes associated with variable node v, and e' = (v, c'), c'≠ c represents all edges except the check node c.

[0052] 3. Update the second sub-layer i c . c The data processing process of the second sub-layer is represented as:

[0053]

[0054] wherein, represents the information of edge e in the i c th iteration of the i v th sub-layer, represents the positive or negative case of the information of edge e' in the i c th sub-layer, represents the normalization factor of the e th edge in the i c 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), and a, b are learnable parameters, wherein a represents the normalization factor and b represents the offset.

[0055] The update process of the second sub-layer i c uses the HMS algorithm, which combines the NMS and OMS algorithms and can better compensate for the minimum sum loss.

[0056] 4. Obtain the output neuron message, which is represented as:

[0057]

[0058] wherein, v o v represents the v c th element of the output of the neural network, and v σ represents the sigmoid function, ( l ) represents the initialized variable node information, represents the information of edge e' in the i c th iteration of the i v th sub-layer, and e' = (v, c'), c'≠ c represents all edges connected to the variable node v except the check node c.

[0059] Optionally, in order to facilitate the representation of the probability that the transmission bit x v = 0, the sigmoid function σ(x) = 1 + exp(-x) ( - ) 1 is used to limit the output neuron message to between 0 and 1.

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

[0061] Optionally, the preset loss function is represented as:

[0062]

[0063] Wherein, L(o,x) represents the loss function, o represents the neural network output vector, x represents the transmitted code word vector, N represents the code word length, o v represents the vth neuron message of the neural network output, x v represents the vth element of the code word sequence transmission, γ v represents the influence degree of the data at different positions in the window on the current decoding symbol.

[0064] The loss function adopts the expected cross entropy between the transmission code word and the neural network output, and since the sliding window decoding algorithm is used for decoding, the data at the left end in the window has greater influence on the current target symbol, therefore, the γ factor is introduced to represent the influence degree of the data at different positions in the window on the current decoding symbol, so that the trained parameters are better used in the sliding window decoding algorithm.

[0065] Optionally, the neural network parameters are updated according to the preset loss function and the gradient descent method, including: estimating the loss function value based on the preset initial normalization factor α, the offset β and the preset loss function; calculating the gradient Based on the calculation result, the values of the parameters α and β in the neural network are updated by 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 to different code lengths, the present application uses multiple groups of lifting codes derived from the same base graph. Different samples are randomly selected for training each time, so that the trained neural MS decoder can be applied to codes of multiple code lengths. Moreover, a layer-by-layer greedy training method is used to construct and optimize the network, so as to reduce the possibility of gradient disappearance problem in the training process. 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 capable of multiple iterations. After each expansion, the new layer is trained on the basis of the existing layers, and only the parameters of the newly 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 SC-LDPC after length coupling.

[0071] As shown in the formula (1), by setting the window size as W, 1 < W < L, the leftmost base pattern unit is set as the target symbol during the current decoding, the parameter trained by the neural network is brought into the corresponding edge update process in the window to complete the decoding by using the hybrid minimum sum algorithm, and only the symbols contained in the current window are decoded. Figure 4

[0072] The neural minimum sum sliding window decoding method for the SC-LDPC code provided by the application determines the optimal parameters of different types of edges of the base pattern in the window by building and training a neural network, and the input of the neural network is a base matrix with a memory length size; the determined optimal parameters are applied to each window of the sliding window decoding algorithm to decode the SC-LDPC after length coupling, and the method combines the neural network with the algorithm in the sliding window to construct a neural MS decoder, compared with the traditional sum-product algorithm (SPA), minimum sum algorithm (MSA), normalized minimum sum algorithm (NMSA) and offset minimum sum algorithm (OMSA) decoding, the neural MS decoder presents a certain performance gain, and has better performance and faster convergence.

[0073] In order to prove the beneficial effects of the application, a set of experimental data is provided, as follows:

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

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

[0076] After the neural decoding network is trained, the obtained parameter distribution is as shown in Figure 5 The normalization factor is mostly concentrated between 0.7 and 1.0, and the offset is concentrated between -0.1 and 0.2. The values change with the number of iterations.

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

[0078] The test results are shown in the following figures, Figure 6 , Figure 7 which are the BER and BLER performance graphs, respectively. In the case of a coupling length of 50, the performances of the SPA, MSA, and OMSA algorithms 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 performances of the five algorithms are similar; when the signal-to-noise ratio is greater than 2.5 dB, the advantage of the neural MS decoder gradually emerges.

[0079] As shown in Figure 6 , in the BER performance graph, for the shorter codes with M values of 60 and 120 shown in (a) and (b) of Figure 6 , the gain of the neural MS decoder is more obvious. Figure 6 In (a), 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 decoder, and a gain of about 1 dB compared to the other three algorithms. 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 Fig. (b) is a 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, and when the signal-to-noise ratio is greater than 4dB, the gain gradually increases and finally remains at 1.5dB; Figure 6 Fig. (c) corresponds to a code with M=160, when the signal-to-noise ratio is greater than 2.5dB, the neural MS decoder has a gain of about 0.3dB compared with the NMS algorithm; Figure 6 Fig. (d) is a 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.

[0080] As shown in Fig. (a), when the signal-to-noise ratio is greater than 2.5dB, the neural MS decoder has a gain compared with the NMSA decoding, the gain gradually increases, and when the signal-to-noise ratio is 5dB, the gain remains at about 1.5dB; Figure 7 Fig. (b) shows that the neural MS decoder finally has a gain of about 1.5dB compared with the NMSA decoding; Figure 7 Fig. (c) shows that when the signal-to-noise ratio is greater than 2.5dB, the neural MS decoder has a gain of about 0.3dB compared with the NMSA decoding; Figure 7 Fig. (d) shows that the neural MS decoder has a gain of about 0.5dB compared with the NMSA decoding. Figure 7 Figure 7 For the special structure of the spatially coupled low-density parity-check code (SC-LDPC), the neural network is combined with the sliding window decoding algorithm in the window to construct a neural MS decoder, which has a certain performance gain compared with the traditional SPA, MSA, NMSA and OMSA decoding.

[0081] The application further provides a structure of an electronic device, which comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus,

[0082] The memory is used for storing a computer program.

[0083] The processor is used for executing the program stored on the memory to realize the steps provided in the method embodiments.

[0084] The communication interface is used for communication between the electronic device and other devices.

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

[0086] ​The method provided by the embodiment of the present application can be applied to an electronic device. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. Herein, no limitation is made, and any electronic device that can implement the present application falls within the protection scope of the present application.

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

[0088] For the electronic device / storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the specific content, beneficial effects, and the like are described in the method embodiment part.

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

[0090] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For ordinary skilled persons in the art to which the present application belongs, a number of simple deductions or replacements can be made without departing from the concept of the present application, and all of them should be regarded as falling within the protection scope of the present application.

Claims

1. A neural min-sum sliding window decoding method for SC-LDPC codes, characterized in that, The shared parameters between edges of the same type corresponding to each of the elements are represented as: The neural network is built and trained to determine the optimal parameters of different types of edges of the base mode graph in a window, and an input of the neural network is a base matrix with a window plus a memory length; wherein the neural network is a sparse MS decoding network with a non-complete connection structure, and an input layer and an output layer of the neural network correspond to variable nodes in a base mode graph , the number of N , and the size of the remaining intermediate hidden layers are all the number of edges in the base mode graph; The neurons in each of the hidden layers include Z clusters, each cluster containing a corresponding total number of edges in the base motif elements, denoted as: ， in, This represents the total number of edges in the improved fundamental graph. Indicates the boosting factor. This represents the total number of edges in the primitive graph before promotion. Indicates the first A cluster; The determined optimal parameters are applied to each window of the sliding window decoding algorithm to decode the length-coupled SC-LDPC. ， ， wherein, denotes the normalized factor of the edge in the th iteration, denotes the normalized factor of the edge in the th iteration, denotes the offset of the edge in the th iteration, denotes the offset of the edge in the th iteration, , denotes any edge of the th cluster, denotes any edge of the th cluster, denotes all edges of the th cluster, denotes all edges of the th cluster; Each of the hidden layers in the neural network comprises a first sub-layer and a second sub-layer, 2. The method of claim 1, wherein, The process of message passing in the neural network is: wherein the first sub-layer corresponds to a process of passing messages from a variable node to a check node along an associated edge, denoted as When , each neuron in the first sub-layer passes a message to a corresponding neuron in the second sub-layer according to a corresponding edge connected to the input layer, each neuron in the first sub-layer passes a message to a corresponding neuron in the second sub-layer according to a corresponding edge When , each neuron in the second sub-layer passes a message to a corresponding neuron in the third sub-layer according to a corresponding edge according to connected to the corresponding neuron in the third sub-layer; The second sub-layer corresponds to a process of transmitting messages from check nodes to variable nodes, denoted as , neurons in the layer According to connection sub-layer neurons associated therewith, when reaching the last hidden layer, according to connection of the first bit of the output layer, obtaining an output result.

3. The method of claim 2, wherein, When receiving variable node information from a channel, the variable node information is initialized, represented as: An output neuron message is obtained, represented as: ， wherein, represents the v-th element of the transmitted codeword sequence, v represents the v-th element of the received vector represents the v-th element of the transmitted codeword sequence, represents the v-th element of the received vector represents the transition probability from to represents the transition probability from represents the natural logarithm with base e, represents the initialized variable node information; According to the initialization result, the first sub-layer is updated, The layer data processing process is represented as: ， wherein, represents the th iteration information of edge e in the sublayer, represents the th iteration information of edge e in the sublayer, , , represents all check nodes associated with variable node , represents all edges except check node . A second sub-layer is performed for updating, The layer data processing procedure is represented as: ， wherein, denotes the iteration information of edge e in sublayer, denotes sublayer edge information of positive or negative case, denotes the iteration normalization factor of the iteration normalization factor of the offset of the denotes all edges connected to check node except variable node , is a learnable parameter, weight denotes normalization factor, denotes offset;​ According to the output neuron message, a preset loss function, and a gradient descent method, neural network parameters are updated. ， wherein, represents the i-th element of the output of the neural network, represents a sigmoid function, represents the initialized variable node information, represents the information of the edges in the last iteration layer, represents all edges except the check node variable node​​​​ The preset loss function is represented as:

4. The method of claim 3, wherein, The updating of the neural network parameters according to the output neuron message, the preset loss function, and the gradient descent method comprises: ， wherein, represents the loss function, represents the neural network output vector, represents the transmitted codeword vector, represents the codeword length, represents the message of the v th neuron of the neural network output, represents the v th element of the codeword sequence transmission, represents the degree of influence of the data at different positions in the window on the current decoded symbol.

5. The method of claim 3, wherein, The shared parameters between edges of the same type corresponding to each of the elements are represented as: based on the output neuron message, a preset initial normalization factor , an offset and the preset loss function, estimate a loss function value; calculating a gradient based on the estimation result updating parameters in the neural network using a gradient descent method based on the calculation result and values.

6. A neural minimum-sum sliding-window decoding device for SC-LDPC codes, characterized in that, The processing module is configured to apply the determined optimal parameters to each window of the sliding window decoding algorithm to decode the length-coupled SC-LDPC. The training module is configured to build and train a neural network to determine optimal parameters of different types of edges in a base graph in a window, wherein an input of the neural network is a base matrix with a size of a window plus a memory length; wherein the neural network is a sparse MS decoding network with a non-complete connection structure, and an input layer and an output layer of the neural network correspond to variable nodes in a base graph , the number of N , and the size of the remaining intermediate hidden layers are all equal to the number of edges in the base graph; The neurons in each of the hidden layers include Z clusters, each cluster containing a corresponding total number of edges in the base motif elements, denoted as: ， wherein, denotes the total number of edges of the base graph after lifting, denotes the lifting factor, denotes the total number of edges of the base graph before lifting, denotes the th cluster, cluster; The device comprises a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory are in communication with each other through the communication bus; ， ， wherein, denotes the normalized factor of the edge in the th iteration, denotes the normalized factor of the edge in the th iteration, denotes the offset of the edge in the th iteration, denotes the offset of the edge in the th iteration, , denotes any edge of the th cluster, denotes any edge of the th cluster, denotes all edges of the th cluster, denotes all edges of the th cluster; The memory is configured to store a computer program.

7. An electronic device, comprising: The processor is configured to execute the program stored in the memory to implement the method in any one of claims 1-5. The computer program stored in the computer-readable storage medium is executed by the processor to implement the method in any one of claims 1-5. ​ 8. A computer-readable storage medium, characterized in that, ​

Images