A shared neural normalized min-sum decoding method and device based on residual connection

Abstract

The application provides a shared neural normalization minimum sum decoding method and device based on residual connection, which comprises the following steps: step 1, constructing a neural network input layer; step 2, constructing a neural network hidden layer; step 3, designing an adaptive residual connection mechanism; step 4, checking node-to-variable node (CN) layer iteration update, calculating the message of the node-to-variable node, and applying the adaptive residual connection according to a threshold judgment condition; step 5, variable node-to-check node (VN) layer iteration update; step 6, repeating steps 4 and 5 according to a set maximum iteration number, and outputting a final decoding result through an output layer after iteration to the maximum number; and step 7, training the network by using a cross-entropy loss function. The application effectively alleviates the gradient vanishing and degradation problems of a deep network through the adaptive residual connection mechanism, significantly improves the decoding performance of an LDPC code, and especially performs well under a low signal-to-noise ratio condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of communication technology, in particular to a shared neural normalized min-sum decoding method and device based on residual connection. BACKGROUND

[0002] LDPC (Low-Density Parity-Check) code is a kind of efficient linear error correction code, with high decoding throughput and performance close to Shannon limit, widely used in 5G mobile communication, Wi-Fi, satellite communication and storage devices such as flash memory and hard disk drive. LDPC code mainly uses belief propagation (Belief Propagation, BP) decoding algorithm, but due to its high computational complexity, it is difficult to implement on hardware. In order to reduce the complexity, the min-sum (Min Sum, MS) algorithm is proposed to sacrifice a certain bit error rate. In order to solve the decoding performance loss caused by the overestimation of the check node value in the MS algorithm, the normalized min-sum (Normalized Min Sum, NMS) and offset min-sum (Offset Min Sum, OMS) decoding algorithms are proposed.

[0003] With the remarkable achievements of deep learning in natural language processing, computer vision, autonomous driving and other fields. In recent years, combining deep learning with traditional communication has also become a research hotspot and achieved important results. The iterative process of traditional decoding algorithm is unfolded into a forward propagation neural network, the connection edge is given a trainable weight, and the weight coefficient is updated by using back propagation, which can significantly improve the decoding performance.

[0004] However, with the increase of the number of iterations and the deepening of the number of layers of the neural network, the existing neural network decoding method is prone to gradient disappearance and degradation problems, which affects the decoding performance. Residual connection is used to alleviate the training problem of deep network. However, the traditional residual connection adopts a fixed connection mode, which cannot be adaptively adjusted according to the dynamic characteristics of the decoding process. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a shared neural normalized min-sum decoding method and device based on residual connection, which solves the problems of gradient disappearance and network degradation in deep neural network decoding, and realizes adaptive control of residual connection, dynamically adjusts the connection strategy according to the network state, and improves the performance of the decoder.

[0006] The present application provides a shared neural normalized min-sum decoding method based on adaptive residual connection architecture, comprising the following steps:

[0007] Step 1, constructing a neural network input layer, the neural network input layer includes two or more input nodes, the number of input nodes is equal to the length of the LDPC (Low-Density Parity-Check) code word sequence to be decoded; the neural network input layer is configured to receive a real number vector containing channel output soft information, each element in the real number vector corresponds to a reliability measure of a bit in the code word, i.e. log-likelihood ratio (LLR, Log-Likelihood Ratio);

[0008] Step 2, constructing a neural network hidden layer, the input layer directly feeds the real number vector into the neural network hidden layer to start the decoding process of the neural network;

[0009] The neural network hidden layer includes a check node to variable node CN (Check Node) layer and a variable node to check node VN (Variable Node) layer;

[0010] The weight parameter of the check node to variable node CN layer is , the weight parameter of the variable node to check node VN layer is , and all connection edges of the check node to variable node CN layer in the current iteration share the same weight , and all connection edges of the variable node to check node VN layer share the same weight ;

[0011] Step 3, designing an adaptive residual connection mechanism, setting a threshold parameter for determining whether to enable residual connection, and defining an exponentially weighted average coefficient to control the weight distribution of the residual connection;

[0012] Step 4, check node to variable node CN layer iteration update, calculating the check node to variable node message, and applying adaptive residual connection according to the threshold judgment condition;

[0013] Step 5, variable node to check node VN layer iteration update, calculating the variable node to check node message, and applying adaptive residual connection according to the threshold judgment condition;

[0014] Step 6, iteration and output layer processing, repeating steps 4 and 5 according to the set maximum number of iterations, and outputting the final decoding result through the output layer after iteration to the maximum number, and the output layer is processed using an activation function;

[0015] Step 7, network training optimization, training the network using a cross-entropy loss function, and updating the shared weight parameters through a backpropagation algorithm.

[0016] In step 1, the sequence of code words of the LDPC code to be decoded received by the input layer of the neural network is the soft information sequence obtained after the transmitted code word is modulated by binary phase shift keying and transmitted in an additive white Gaussian noise channel.

[0017] In step 3, the adaptive residual connection mechanism comprises:

[0018] In the check node to variable node CN layer update: the output of the variable node to check node VN layer in the t-2th iteration is combined with the output of the variable node to check node VN layer in the t-1th iteration by exponential weighted average, and the weighted coefficient is When , the residual connection is enabled; when , the residual connection is not used. is a threshold value.

[0019] In the variable node to check node VN layer update: the output of the check node to variable node CN layer in the t-1th iteration is combined with the output of the check node to variable node CN layer in the tth iteration by exponential weighted average, and the weighted coefficient is When , the residual connection is enabled; when , the residual connection is not used.

[0020] In step 4, for the check node to variable node CN layer, the message of the check node CN to the variable node VN is calculated, and the update formula is:

[0021] (1),

[0022] Wherein, is the shared weight of the node to variable node CN layer connection edge at the tth iteration; represents the message of the check node m to the variable node n at the tth iteration, represents the message of the variable node to the check node m at the t-1th iteration; represents the set of all variable nodes connected to the check node m but excluding the variable node n; the function f is a mathematical mapping, is the input of the function f, and the result represents the message value sent by the check node to the variable node in the tth iteration;

[0023] Then the adaptive residual connection mechanism is applied, and the update formula is:

[0024] (2).

[0025] In step 5, for the variable node to check node VN layer, the variable node VN to check node CN message is calculated, and the update formula is:

[0026] (3),

[0027] wherein, is the shared weight of the variable node to check node VN layer connection edge at the tth iteration, represents the message of the variable node n to the check node m at the tth iteration, represents the message of the check node to the variable node n at the tth iteration, represents the set of all check nodes connected to the variable node n but excluding the check node m; represents a set of check nodes connected to the variable node n; represents the channel log-likelihood ratio corresponding to the nth bit; the function g is a mathematical mapping, is the input of the function g, and the result represents the message value sent by the variable node to the check node in the nth iteration;

[0028] Then, the adaptive residual connection mechanism is applied, and the update formula is:

[0029] (4).

[0030] In step 6, the calculation formula of the final decoding output is:

[0031] (5),

[0032] wherein, T is the maximum iteration number of decoding, represents the message of the check node m to the variable node n at the Tth iteration; represents the hard decision output of the variable node n after reaching the maximum iteration number; using as the activation function, wherein x is the input of the activation function;

[0033] The final decoding decision is:

[0034] (6),

[0035] wherein, represents the final binary decision result for the nth symbol position.

[0036] Step 7 includes: then training the network by using the cross-entropy Cross-Entropy loss function, updating the weight coefficients in the neural network through back propagation, and the loss function formula is:

[0037] ​ (7),

[0038] wherein N is the number of output layer neurons, i.e., the code word length, s is the transmitted code word, o is the decoded sequence of network output, is the cross-entropy loss function calculated based on the transmitted code word s and the decoded sequence of network output o; represents the nth bit in the transmitted code word sequence.

[0039] The application also provides a shared neural normalized minimum sum decoding device based on an adaptive residual connection architecture, which is realized by the method and is realized by a computer processor, and comprises logical function modules and can execute the method; the device comprises an input layer module, a shared weight processing module, an adaptive residual connection control module, an iteration control module and a training optimization module.

[0040] The input layer module is used for receiving an LDPC code soft information sequence to be decoded.

[0041] The shared weight processing module comprises a CN layer processing unit and a VN layer processing unit, which execute steps 4 and 5, respectively.

[0042] The adaptive residual connection control module dynamically controls the enabling and disabling of residual connection according to the threshold judgment mechanism in the method.

[0043] The iteration control module manages the alternating iteration process of the CN layer and the VN layer and executes step 6.

[0044] The training optimization module realizes step 7 in the method and optimizes the network parameters using the cross-entropy loss function.

[0045] The application also provides an electronic device comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0046] The application also provides a storage medium storing a computer program or instructions, and when the computer program or instructions are run on a computer, the steps of the method are executed.

[0047] Beneficial effects: the application effectively alleviates the gradient vanishing and degradation problems of a deep network through an adaptive residual connection mechanism, significantly improves the decoding performance of an LDPC code, especially under a low signal-to-noise ratio condition; the shared weight strategy is adopted to make the CN layer and the VN layer share weight parameters in all iterations, which greatly reduces the storage and calculation overhead; meanwhile, the threshold judgment mechanism is introduced to adaptively determine whether to enable residual connection according to the current state of the network, which improves the flexibility and adaptability of the network compared with a fixed connection mode. Attached Figure Description

[0048] Figure 1 It is a hidden layer design diagram.

[0049] Figure 2 This is a schematic diagram of the overall structure of a shared neural network based on an adaptive residual connection architecture.

[0050] Figure 3 This is a schematic diagram comparing the bit error rate at different thresholds.

[0051] Figure 4 This is a schematic diagram showing the comparison of error rates of various decoding algorithms for LDPC codes with a code length of 576.

[0052] Figure 5 This is a schematic diagram comparing the bit error rates of various decoding algorithms at different iteration counts. Detailed Implementation

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0054] This invention provides a shared neural normalized minimum sum decoding method based on residual connections, comprising:

[0055] Step 1: Select an LDPC code with a code rate of 0.75 and a code length of 576 from the IEEE 802.16e standard. Use mini-batch gradient descent to train the network. Each batch contains 100 codewords, which are evenly distributed between 1 and 5 dB to ensure that each signal-to-noise ratio is allocated the same number of codewords.

[0056] Step 2: Construct the neural network input layer to receive the codeword sequence of the LDPC code to be decoded. Specifically, the input sequence is a soft information sequence obtained by transmitting the transmitted codeword through an additive white Gaussian noise (AWGN) channel after binary phase shift keying (BPSK) modulation.

[0057] Set the sending codeword After binary phase-shift keying modulation and transmission through an additive white Gaussian noise channel, the received codeword at the receiver is: The input layer receives a sequence of soft information. As the initial input to the network.

[0058] Step 3: Construct the hidden layers of the neural network, including the CN layer from the verification node to the variable node and the VN layer from the variable node to the verification node.

[0059] CN layer design: As shown in Figure 1 , the CN layer is responsible for computing the message passing from check nodes to variable nodes. Each CN layer contains multiple neurons, each of which corresponds to an edge in the check matrix of the LDPC code. All the connecting edges of a CN layer share the same weight parameter in the same iteration .

[0060] VN layer design: As shown in Figure 1 , the VN layer is responsible for computing the message passing from variable nodes to check nodes. Similarly, all the connecting edges of a VN layer share the same weight parameter in the same iteration .

[0061] Step 4, adaptive residual connection mechanism design, which includes the following key elements: define the exponential weighted average coefficient and , control the weight distribution of residual connection; set a unified threshold parameter , to determine whether to enable residual connection.

[0062] Connection mode design as shown in Figure 2 , from left to right, the nth corresponds to the channel log-likelihood ratio through the input layer to get the 0th iteration, that is, the initialized variable node n to check node m message ; then pass into the first iteration of CN layer to update the check node information, output the iteration message from check node m to variable node n , and pass through the VN layer to update the variable node information, output the message of variable node n to check node m . When updating the CN layer in the second iteration, the input of the CN layer is not only the output of the VN layer in the last iteration , but also introduces the initialized through the residual connection, which is combined using exponential weighted average, and the output of the second iteration CN layer is ; when updating the VN layer in the second iteration, the input of the VN layer is not only the output of the CN layer in the last iteration , but also introduces the output of the first iteration CN layer through the residual connection, which is combined using exponential weighted average, and the output of the second iteration CN layer is . This process continues until the maximum number of iterations , is the posterior log-likelihood ratio after the Tth iteration, represents the hard decision output of variable node n after the maximum number of iterations.

[0063] Adaptive control strategy: by comparing the size relationship between the weighted coefficient and the threshold value, dynamically decide whether to enable residual connection. When the residual connection is enabled when the residual connection is not used when the residual connection is enabled when the residual connection is not used when

[0064] Step 5, update the check node to variable node CN layer, whose neurons calculate the check node CN to variable node VN message, update formula as follows:

[0065] (1),

[0066] wherein, is the shared weight of the CN layer connection edge at the tthiteration, denotes the message from the check node m to the variable node n at the tthiteration, denotes the message from the variable node to the check node m at the t-1thiteration, denotes the set of all variable nodes connected to the check node m but excluding the variable node n.

[0067] Then apply the adaptive residual connection mechanism, update formula as follows:

[0068] (2),

[0069] wherein, is the weighted coefficient of the exponential weighted average, is the threshold value.

[0070] Step 6, update the variable node to check node VN layer, whose neurons calculate the variable node VN to check node CN message, update formula as follows:

[0071] (3),

[0072] wherein, is the shared weight of the VN layer connection edge at the tthiteration, denotes the message from the variable node n to the check node m at the tthiteration, denotes the message from the check node to the variable node n at the tthiteration, denotes the set of all check nodes connected to the variable node n but excluding the check node m.

[0073] Then apply the adaptive residual connection mechanism, update formula as follows:

[0074] (4),

[0075] wherein, is the weighting coefficient for the exponentially weighted average.

[0076] Step 7, the iterative execution stage repeats step 4 (CN layer update) and step 5 (VN layer update) according to the set maximum number of iterations T. In each iteration, the network adaptively decides whether to enable the residual connection according to the current weighting coefficient and threshold parameter. When the maximum number of iterations is reached, the final decoding result is calculated by the output layer:

[0077] (5),

[0078] where, denotes the message from check node m to variable node n at the Tth iteration, denotes the set of check nodes connected to variable node n. Use as the activation function.

[0079] The final decoding decision is:

[0080] (6).

[0081] Step S8: After determining the maximum number of iterations T, the structure of the entire neural network is also determined, and then the network is trained using the Cross-Entropy loss function. The weight coefficients in the neural network are updated through backpropagation. The loss function formula is as follows:

[0082] (7),

[0083] where N is the number of output layer neurons, i.e. the code word length, s is the transmitted code word, and o is the decoding sequence output by the network.

[0084] The LDPC code neural network normalization minimum sum decoding method of the application is based on the neural network expansion implementation of the traditional normalization minimum sum algorithm. The neural normalization minimum sum (NNMS) algorithm expands the iteration process of the traditional normalization minimum sum algorithm into a multi-layer neural network structure, each connection edge has an independent weight parameter; the shared neural normalization minimum sum (SNNMS) algorithm introduces a weight sharing mechanism on the basis of the NNMS, so that all connection edges in the same iteration layer share the same weight parameter, effectively reducing the number of network parameters; the application further proposes a residual connection architecture shared neural normalization minimum sum (ResSNNMS) algorithm, which introduces a residual connection mechanism on the basis of the SNNMS, and relieves the gradient disappearance problem of the deep network through cross-iteration information transmission; on this basis, the application finally proposes an adaptive residual connection architecture shared neural normalization minimum sum (AResSNNMS) algorithm, which introduces an adaptive residual connection control mechanism, dynamically adjusts the enabling strategy of the residual connection according to the network state, and further optimizes the decoding performance.

[0085] In order to verify the effectiveness of the adaptive residual connection architecture based shared neural normalization minimum sum decoding method proposed in the application, Figure 3 The AResSNNMS algorithm performance under different threshold parameters is compared and analyzed in the application, wherein BER (Bit Error Rate) represents the bit error rate; E b / N0 represents the ratio of the energy of each binary bit to the noise power spectral density, and the unit is decibel (dB). The decoding performance of the AResSNNMS algorithm is lower than that of the ResSNNMS algorithm when This is because when the message reliability is sufficient, the introduction of historical messages will increase the complexity of decoding, so the threshold is selected through simulation, and subsequent simulations are based on this threshold.

[0086] Figure 4 The different schemes in the neural network normalization minimum sum algorithm are compared and analyzed, wherein BER (Bit Error Rate) represents the bit error rate; E b / N0 represents the ratio of the energy of each binary bit to the noise power spectral density, and the unit is decibel (dB). From Figure 4 It can be seen that the decoding effect of NNMS, SNNMS, ResSNNMS and AResSNNMS conforms to the theoretical analysis, and the decoding performance is further improved after using the adaptive residual connection mechanism. When BER=10 -4At this time, the NNMS algorithm needs about 4.68dB, the SNNMS needs about 4.22dB, the ResSNNMS algorithm needs about 4.01dB, and the AResSNNMS algorithm needs about 3.88dB. It can be seen that only the residual connection mechanism in the neural network decoding process is improved and the adaptive control is introduced, which can be improved by 0.8dB, 0.34dB and 0.13dB compared with the foregoing algorithms.

[0087] Figure 5 The decoding performance of the three algorithms under different iteration conditions is compared, wherein BER (Bit Error Rate) represents the bit error rate; E b / N0 represents the ratio of the energy of each binary bit to the noise power spectral density, and the unit is decibel (dB). It can be seen from the figure that the decoding results of the three algorithms are more and more accurate with the increase of the iteration number, and when BER=10 –6 When the iteration number is 15 times, the AResSNNMS algorithm is improved by 0.44dB and 0.20dB compared with the SNNMS and ResSNNMS algorithms.

[0088] According to the above analysis, the gradient disappearance and degradation phenomenon of the deep network are relieved, and better decoding performance is obtained under a certain complexity.

[0089] The embodiment of the present application also provides a residual connection-based shared neural normalized minimum sum decoding device, which can be realized by a computer processor and is used for executing the decoding method. The device comprises the following modules:

[0090] The input layer module is responsible for receiving the LDPC code soft information sequence to be decoded, and is the data inlet of the entire decoding device. The module receives the soft information sequence obtained by transmitting the code word after binary phase shift keying modulation in an additive white Gaussian noise channel. The module performs necessary preprocessing on the received soft information, including data normalization and format conversion, to ensure that the input data meet the numerical stability requirements of subsequent neural network processing. Meanwhile, the module also provides a data buffering function to support batch processing mode decoding operation, thereby improving the overall decoding throughput. In software implementation, the module stores the input data through an array or tensor structure and provides a standardized data interface for other modules to call.

[0091] Shared weight processing module: responsible for implementing the main calculation logic of LDPC code decoding. This module contains two core components, CN layer processing unit and VN layer processing unit. The CN layer processing unit is responsible for performing the message calculation from check node to variable node, implementing the message update operation described in formula (1); the VN layer processing unit is responsible for performing the message calculation from variable node to check node, implementing the message update operation described in formula (3). The two processing units significantly reduce the number of network parameters through the shared weight strategy, improve the calculation efficiency and storage efficiency. In software implementation, this module realizes efficient parallel processing through matrix operation and vectorization calculation.

[0092] Adaptive residual connection control module: responsible for dynamically controlling the enabling strategy of residual connection according to the current state of the network. According to the threshold judgment mechanism, when the weighted coefficient is greater than the threshold, the residual connection is closed, otherwise the residual connection is used. This module implements the residual connection function of CN layer and VN layer, respectively performing the exponential weighted average operation described in formula (2) and formula (4). Through the cross-iteration information transmission, it provides a direct propagation path for the gradient, effectively alleviating the gradient vanishing and network degradation problems in deep network. The adaptive control mechanism of this module can dynamically adjust the residual connection strategy according to the message reliability, avoiding unnecessary interference of historical information when the message is fully reliable.

[0093] Iteration control module: responsible for managing the execution flow and timing control of the entire decoding process, ensuring that the CN layer and VN layer perform ordered alternating iteration according to the predetermined strategy. According to the preset maximum number of iterations, the execution of the iteration process is controlled, the working timing of the CN layer processing unit and the VN layer processing unit in the shared weight processing module is coordinated, and the correctness and consistency of message transmission are ensured. When the maximum number of iterations is reached, the module performs the final decoding output calculation, implements the output operation described in formula (5), and applies the activation function for the final hard decision processing.

[0094] Training optimization module: realizes the training and optimization function of neural network parameters, ensures that the shared weight parameters can obtain the optimal value through learning. This module implements the cross-entropy loss function calculation, and performs the loss calculation described in formula (6). The training process uses the stochastic mini-batch gradient descent method to update the network parameters, supports adaptive learning rate adjustment and multiple optimization algorithms. This module plays a key role in the training stage, and through the learning of a large amount of training data, the network parameters converge to the optimal state, laying a foundation for the improvement of decoding performance.

[0095] The decoding device proposed in this embodiment has the following technical advantages: each functional module is realized by software, which is convenient for parameter adjustment and algorithm optimization, and supports fast adaptation of different LDPC code configurations. The adaptive residual connection control module dynamically adjusts the connection strategy according to the network state through program logic, and can maintain excellent performance under different channel conditions.

[0096] The application provides a residual connection-based shared neural normalized minimum sum decoding method and device. There are many methods and approaches to realize the technical scheme, and the above description is only the preferred embodiment of the application. It should be pointed out that, for ordinary skilled persons in the technical field, some improvements and refinements can be made without departing from the principle of the application, and these improvements and refinements should also be regarded as the protection scope of the application. The components not explicitly described in the embodiment can be realized by using the existing technology.

Claims

1. A method of shared neural normalized min-sum decoding based on adaptive residual connections, characterized in that, The method comprises the following steps: Step 1, constructing a neural network input layer, the neural network input layer comprising two or more input nodes, the number of input nodes being equal to the length of a LDPC code word sequence to be decoded; the neural network input layer being configured to receive a real number vector comprising channel output soft information, each element of the real number vector corresponding to a reliability measure, i.e. a log-likelihood ratio, of a bit in the code word; Step 2, constructing a neural network hidden layer, the input layer feeding the real number vector directly into the neural network hidden layer to start the decoding process of the neural network; the neural network hidden layer comprising a check node to variable node CN layer and a variable node to check node VN layer; The weight parameter of the check node to variable node CN layer is The weight parameter of the variable node to check node VN layer is All the connecting edges of the check node to variable node CN layer share the same weight in the current iteration All the connecting edges of the variable node to check node VN layer share the same weight ; Step 3, designing an adaptive residual connection mechanism, setting a threshold parameter for determining whether to enable residual connection, and defining an exponential weighted average coefficient to control the weight distribution of the residual connection; Step 4, iterative updating of the check node to variable node CN layer, calculating the check node to variable node message, and applying the adaptive residual connection according to the threshold judgment condition; Step 5, iterative updating of the variable node to check node VN layer, calculating the variable node to check node message, and applying the adaptive residual connection according to the threshold judgment condition; Step 6, iteratively performing and output layer processing, repeating steps 4 and 5 according to the set maximum number of iterations, and outputting the final decoding result through the output layer after the maximum number of iterations, the output layer being processed using an activation function; Step 7, network training and optimization, training the network using a cross-entropy loss function, and updating the shared weight parameters through a backpropagation algorithm; In step 3, the adaptive residual connection mechanism comprises: At check node to variable node CN layer update: the output of the (t-2)th iteration variable node to check node VN layer is combined with the output of the (t-1)th iteration variable node to check node VN layer by exponential weighted average, with the weighted coefficient When the residual connection is enabled; when the residual connection is not used, is the threshold value; At the variable node to check node VN layer update: the output of the check node to variable node CN layer of the t-1th iteration is combined with the output of the check node to variable node CN layer of the tth iteration by an exponential weighted average with a weighting factor of The residual connection is enabled when and disabled when .

2. The method of claim 1, wherein, In step 1, the code word sequence of the LDPC code to be decoded received by the neural network input layer is a soft information sequence obtained after a transmitted code word is modulated by binary phase shift keying and transmitted in an additive white Gaussian noise channel.

3. The method of claim 2, wherein, In step 4, for the check node to variable node CN layer, the check node CN to variable node VN message is calculated, and the update formula is: （1）， wherein, is the shared weight of the edge connecting the node to the variable node CN layer at the tth iteration; denotes the message from check node m to variable node n at the tth iteration, denotes the message from variable node n to check node m at the t-1th iteration; denotes the message from check node m to variable node n at the tth iteration, denotes the set of all variable nodes connected to check node m but not including variable node n; the function f is a mathematical mapping, is the input of the function f and the result denotes the message value sent by the check node to the variable node at the tth iteration; Then the adaptive residual connection mechanism is applied, and the update formula is: （2）。 4. The method of claim 3, wherein, In step 5, for the variable node to check node VN layer, the variable node VN to check node CN message is calculated, and the update formula is: （3）， in, Let V be the shared weight of the connection edges from the variable node to the verification node VN layer at iteration t. This represents the message from variable node n to check node m during the t-th iteration. This represents the check node at the t-th iteration. Message to variable node n, This represents the set of all other check nodes connected to variable node n, excluding check node m. This represents a set of check nodes connected to variable node n; It is expressed as the channel log-likelihood ratio corresponding to the nth bit; the function g is a mathematical mapping. It is the input of function g, and the result indicates the variable node at the th . The message value sent to the verification node in the next iteration; Then the adaptive residual connection mechanism is applied, and the update formula is: （4）。 5. The method of claim 4, wherein, In step 6, the calculation formula of the final decoding output is: （5）， where T is the maximum number of iterations for decoding, represents the message from check node m to variable node n at the Tth iteration; represents the hard decision output of variable node n after reaching the maximum number of iterations; uses as the activation function, where x is the input of the activation function. The final decoding decision is: （6）， wherein denotes the final binary decision result for the n-th symbol position.

6. The method of claim 5, wherein, Step 7 comprises: then training the network using a cross-entropy Cross-Entropy loss function, updating the weight coefficients in the neural network through backpropagation, and the loss function formula is: （7）， where N is the number of output layer neurons, i.e., the code word length, s is the transmitted code word, and o is the decoded sequence of the network output, is a cross-entropy loss function computed based on the transmitted code word s and the decoded sequence o of the network output; represents the nth bit in the transmitted code word sequence.

7. A shared neural normalized min-sum decoding device based on an adaptive residual connection architecture, implemented by the method of any one of claims 1-6. The device is implemented by a computer processor and comprises logical function modules, and can execute the method; the device comprises an input layer module, a shared weight processing module, an adaptive residual connection control module, an iteration control module and a training optimization module; The input layer module is used to receive a soft information sequence of a LDPC code to be decoded; The shared weight processing module comprises a CN layer processing unit and a VN layer processing unit, which execute steps 4 and 5, respectively; The adaptive residual connection control module: according to the threshold judgment mechanism in the method, dynamically controls the enablement and shutdown of the residual connection; The iteration control module: manages the alternating iteration process of the CN layer and the VN layer, and executes step 6; The training optimization module: implements step 7 in the method, and optimizes the network parameters using a cross-entropy loss function.

8. An electronic device, comprising: A processor and a memory are included, and the memory stores program codes, which, when executed by the processor, cause the processor to execute the steps of the method according to any one of claims 1 to 6.

9. A storage medium, characterized by A computer program or instructions are stored, which, when running on a computer, execute the steps of the method according to any one of claims 1 to 6.

Images