Deep learning-based belief propagation LDPC decoding method and device, medium and program product

Abstract

The invention provides a belief propagation LDPC decoding method and device based on deep learning, a medium and a program product. The method comprises the following steps: acquiring an LDPC code training sample set corresponding to a check matrix H; according to the check matrix H, determining a weight matrix and an activation function of each layer of a deep learning decoding model; based on the weight matrix and the activation function, using a belief propagation algorithm to establish a deep learning decoding model; training the deep learning decoding model by using the training sample set, and storing the trained deep learning decoding model as a decoder; and sending the LDPC code soft information which is corresponding to the check matrix H and needs to be decoded into the decoder to complete LDPC decoding. According to the method, iterative operation is completed by using deep learning in the error correction decoding process, the number of decoding iterations is reduced, and the sequence sent by the sending end is restored from the sequence containing noise and interference, so that the problem of decoding error codes caused by noise and other interference factors is solved.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular, to a belief propagation LDPC decoding method, device, medium, and program product based on deep learning. Background Art

[0002] The low-density parity-check code (LDPC, Low Density Parity Check Code), first proposed by Gallager, is a linear block error-correcting code with low decoding complexity and excellent performance, and has been proven to achieve performance close to the Shannon limit under various channels. LDPC codes can meet the requirements of mobile communication technologies for high data processing speed, high transmission speed, large-capacity transmission, and high-quality transmission. Therefore, LDPC codes are widely used in deep space communication, optical fiber communication, satellite communication, and storage fields.

[0003] The powerful capabilities of deep learning have been proven in application scenarios such as image processing, speech recognition, and natural language processing. Its recognition, classification, and fitting capabilities exceed those of traditional algorithms. Deep learning has gradually been applied to the field of channel decoding. Models such as deep neural networks (DNN), convolutional neural networks (CNN), and recurrent neural networks (RNN) are used to replace traditional decoding algorithms, which can provide stronger robustness and accuracy when dealing with complex channel conditions. To achieve efficient model training, stochastic gradient descent (SGD) is used as part of the optimization algorithm. SGD is a commonly used optimization technique that iteratively updates model parameters to reduce the loss function. To further improve the convergence speed and stability, the Adam optimizer is introduced. The Adam optimizer combines the advantages of the momentum method and adaptive learning rate, and can dynamically adjust the learning rate during training, thereby accelerating the convergence of the model and improving the final performance, effectively improving the decoding accuracy and anti-interference ability.

[0004] The decoding methods of LDPC codes can be divided into two categories: hard-decision-based decoding and soft-decision-based decoding. Hard-decision-based decoding has a relatively small computational amount and is more practical; while soft-decision decoding uses posterior probability information and through iterative operations, the performance of LDPC codes can approach the Shannon limit. The normalized min-sum (NMS, Normalized Min-Sum) decoding algorithm is obtained by introducing a correction factor into the min-sum (MS, Min-Sum) decoding algorithm to reduce the amplitude of the message. The LDPC decoding model based on deep learning is established on the combination of the DNN model and the NMS decoding algorithm, and the information transfer method of the belief propagation decoding algorithm is extended to the forward propagation of the hidden layer information of deep learning, which can reduce the computational amount and the number of iterations.

[0005] At the receiving end of an electronic communication system, problems such as decoding errors caused by noise and other interference factors are likely to occur. In view of this, in combination with the aforementioned NMS decoding algorithm and LDPC-related technologies, a belief propagation LDPC decoding method based on deep learning can be realized. Summary of the Invention

[0006] In view of the above problems, the present invention provides a belief propagation LDPC decoding method, device, medium and program product based on deep learning, which uses deep learning to complete iterative operations during the error correction decoding process, reduces the number of decoding iterations, and restores the sequence sent by the sending end from the sequence containing noise and interference, thereby solving the problem of decoding errors caused by noise and other interference factors.

[0007] In a first aspect, a belief propagation LDPC decoding method based on deep learning provided by the present invention includes:

[0008] Obtain an LDPC code training sample set corresponding to the parity check matrix H;

[0009] According to the parity check matrix H, determine the weight matrices and activation functions of each layer of the deep learning decoding model;

[0010] Based on the weight matrices and activation functions, use the belief propagation algorithm to establish a deep learning decoding model;

[0011] Use the training sample set to train the deep learning decoding model, and save the trained deep learning decoding model as a decoder;

[0012] Send the soft information of the LDPC code to be decoded corresponding to the parity check matrix H into the decoder to complete LDPC decoding.

[0013] In some embodiments, the samples in the training sample set are represented as (Y, y), where Y is the input message and y is the label;

[0014] The label y is the codeword of the LDPC code. For the LDPC code (n, k), n is the codeword length and k is the information bit length. The codeword y is obtained by multiplying the information bit length k by the generator matrix G;

[0015] The input message Y is a log-likelihood ratio (LLR) value of length n, and its calculation process includes: generating received data by modulating the codeword y with BPSK and adding Gaussian white noise, and calculating the log-likelihood ratio (LLR) value of length n based on the received data.

[0016] In some embodiments, the deep learning decoding model includes an input layer, a hidden layer, and an output layer connected in sequence; the parity-check matrix H indicates the connection relationship between variable nodes and parity-check nodes. The hidden layer in the deep learning decoding model is divided into a variable layer and a parity-check layer according to the executed functions. The connection method between the variable layer and the parity-check layer is determined by the positions of "1" in the parity-check matrix H. There are two connection methods from the variable layer to the parity-check layer and from the parity-check layer to the variable layer. Weight matrices corresponding to each connection method are obtained according to the mapping relationship and the traditional NMS decoding algorithm; according to the mapping relationship of BPSK modulation, the activation function of the output layer is modified.

[0017] In some embodiments, in the hidden layer, the odd hidden layers are variable layers, and the even hidden layers are parity-check layers.

[0018] In some embodiments, establishing the deep learning decoding model using the belief propagation algorithm includes: based on the traditional NMS decoding algorithm, establishing an LDPC code on a Tanner graph. Each iteration process of BP decoding includes two steps: processing of parity-check nodes and processing of variable nodes; in each iteration, the parity-check nodes receive messages from their adjacent variable nodes, and after processing, they are passed back to the adjacent variable nodes. At this time, the variable nodes receive the messages from the adjacent parity-check nodes for processing; in the deep learning decoding model, one iteration operation is represented by going from one layer of the variable layer to the next layer of the parity-check layer, and the transmission of messages is represented by the weight matrix between the hidden layers. The weight matrix represents the process of message transmission, and the distribution of non-zero positions in the weight matrix represents the process of nodes receiving messages from adjacent nodes; at the same time, the corresponding activation functions are initialized.

[0019] In some embodiments, when training the deep learning decoding model using the training sample set, the deep learning decoding model is trained by an extended adaptive learning rate method of stochastic gradient descent, and the weight matrix of the hidden layer is updated using backpropagation. The coefficients in the weight matrix correspond to the correction factors of the traditional NMS decoding algorithm.

[0020] In some embodiments, the decoder has a hard decision unit, and the hard decision unit is used to perform a numerical hard decision to be 0 or 1 according to the probability value of 0 - 1 of the output result of the deep learning decoding model.

[0021] In a second aspect, the present invention provides an electronic device, including:

[0022] At least one processor; and a memory communicatively connected to the at least one processor;

[0023] Wherein, the memory stores instructions executable by the at least one processor, and the at least one processor, by executing the instructions stored in the memory, enables the at least one processor to execute the above method.

[0024] In a third aspect, the present invention provides a computer-readable storage medium for storing instructions which, when executed, implement the above-mentioned method.

[0025] In a fourth aspect, the present invention provides a computer program product which, when called by a computer, causes the computer to execute the above-mentioned method.

[0026] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows:

[0027] 1. The belief propagation LDPC decoding method based on deep learning provided by the present invention has a reduced number of iterations and can quickly recover the transmitted sequence from an error environment.

[0028] 2. The belief propagation LDPC decoding method based on deep learning provided by the present invention is driven by a traditional NMS decoding algorithm and is based on a deep learning decoding model, reducing the training amount. For the codewords of the same parity-check matrix H, good decoding performance can be achieved without training all the codewords. Using this feature, the number of iterations is reduced and the convergence speed is accelerated. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of a belief propagation LDPC decoding method based on deep learning provided in an embodiment of the present invention.

[0030] Figure 2 It is a parity-check matrix and its corresponding Tanner graph provided in an embodiment of the present invention.

[0031] Figure 3 It is a schematic structural diagram of a deep learning decoding model provided in an embodiment of the present invention.

[0032] Figure 4 It is a schematic diagram of the weight matrix and message mapping between the input layer and the hidden layer provided in an embodiment of the present invention.

[0033] Figure 5 It is a schematic diagram of the forward propagation of the check layer provided in an embodiment of the present invention.

[0034] Figure 6a It is a performance test comparison diagram between the method of the present invention and the traditional method for illustration.

[0035] Figure 6b It is a performance test comparison diagram of the method of the present invention under different parameters for illustration.

[0036] Figure 7 It is a schematic structural diagram of an electronic device provided in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0038] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0039] As Figure 1 shown, this embodiment provides a belief propagation LDPC decoding method based on deep learning, including:

[0040] S1: Obtain an LDPC code training sample set corresponding to the parity check matrix H;

[0041] S2: According to the parity check matrix H, establish the weight matrices and activation functions of each layer of the deep learning decoding model;

[0042] S3: Based on the weight matrices and activation functions, use the belief propagation algorithm to establish a deep learning decoding model;

[0043] S4: Use the training sample set to train the deep learning decoding model, and save the trained deep learning decoding model as a decoder;

[0044] S5: Send the soft information of the LDPC code to be decoded corresponding to the parity check matrix H into the decoder to complete LDPC decoding.

[0045] The following provides a detailed description of this method. Taking the LDPC code (3,6) as an example, its parity check matrix H and Tanner are as Figure 2 shown.

[0046]

[0047] S1: Obtain an LDPC code training sample set corresponding to the parity check matrix H.

[0048] The samples in the training sample set are represented as (Y, y), where Y is the input message and y is the label;

[0049] The label y is a codeword of the LDPC code. For the LDPC code (n, k), n is the codeword length, k is the information bit length, and the codeword y is obtained by multiplying the information bit length k by the generator matrix G;

[0050] The input message Y is a log-likelihood ratio (LLR) value of length n, and its calculation process includes: generating received data by modulating the codeword y through BPSK and adding Gaussian white noise, and calculating the log-likelihood ratio (LLR) value of length n based on the received data.

[0051] Thus, for the LDPC code (3, 6), the LDPC code y with a codeword length of 6 is modulated through BPSK, Gaussian white noise is added, and the log-likelihood ratio (LLR) value Y of length 6 is calculated to form a sample (Y, y).

[0052] S2: According to the parity-check matrix H, determine the weight matrices and activation functions of each layer of the deep learning decoding model.

[0053] The deep learning decoding model includes an input layer, a hidden layer, and an output layer connected in sequence; the parity-check matrix H indicates the connection relationship between variable nodes and check nodes. The hidden layer in the deep learning decoding model is divided into a variable layer and a check layer according to the execution function. The connection method between the variable layer and the check layer is determined by the position of the "1" in the parity-check matrix H. There are two connection methods from the variable layer to the check layer and from the check layer to the variable layer. According to the mapping relationship and the traditional NMS decoding algorithm, the weight matrices corresponding to each connection method are obtained; according to the mapping relationship of BPSK modulation, the activation function of the output layer is modified.

[0054] In this embodiment, the deep learning decoding model is as Figure 3 shown, which are an input layer, a hidden layer, and an output layer in sequence. Among them, the odd hidden layers are variable layers, and the even hidden layers are check layers.

[0055] The input message is a log-likelihood ratio (LLR) value Y of length 6, denoted as L(P i ), and the output is a probability value between 0 and 1. According to its parity-check matrix H, there are 10 "1"s in total. At this time, the dimension of the hidden layer of the deep learning decoding model is 10, and the input message of the input layer is mapped according to the parity-check matrix H and expanded from dimension 6 to dimension 10.

[0056] The mapping relationship from the input layer to the first hidden layer is as Figure 4As shown in the figure, the nodes of the first hidden layer are numbered ① to ⑩ from top to bottom. The first bit of the input acts on the first, second, and third parity-check equations in the parity-check matrix H, and the sequence numbers are ①, ④, and ⑧. The second bit of the input acts on the first and second parity-check equations in the parity-check matrix H, and the sequence numbers are ② and ⑤. Similarly, the remaining input bits are mapped to obtain the connection method from the input layer to the first hidden layer, and this connection method is represented by the weight matrix W in2hid . At this time, the connection from the input layer to the first hidden layer completes the initialization of traditional decoding, that is, the extension of L(P i ).

[0057] Table 1, detailed connection table of the input layer - first hidden layer of the deep learning decoding model:

[0058] Information bit node Position in the H matrix Hidden layer connection node Hidden layer connection node number V1 (1,1)、(2,1)、(3,1) e11, e21, e31 ①④⑧ V2 (1,2)、(2,2) e12, e22 ②⑤ V3 (2,3)、(3,3) e23, e33 ⑥⑨ V4 (1,4) e14 ③ V5 (2,5) e25 ⑦ V6 (3,6) e36 ⑩

[0059] According to the parity-check node message processing formula in the iteration of the traditional NMS decoding algorithm:

[0060]

[0061] Among them, R j represents the set of variable nodes connected to the parity-check node j, and R j \i represents the set of variable nodes connected to the parity-check node j except i. At this time, the parameter to be trained is the correction factor a in the formula. The correction factor a is represented as a weight matrix in the deep learning decoding model. At this time, the weight matrix is like the connection method from the variable layer to the parity-check layer in the first iteration, and is initialized as the identity matrix of W eye , and the processing formula is written into the forward propagation.

[0062] According to the variable node message processing formula in the iteration of the traditional NMS decoding algorithm:

[0063]

[0064] Among them, C i represents the set of parity-check nodes connected to the variable node i, and C i \j represents the set of parity-check nodes connected to the variable node i except j. This set of parity-check nodes represents the connection method from the parity-check layer to the variable layer in the decoding model, that is, the weight matrix W c2v , and at this time, in the deep learning decoding model, it is represented as the connection method from the parity-check layer to the variable layer in the first iteration, and the processing formula is written into the forward propagation.

[0065] According to the decoding decision processing formula in the iteration of the traditional NMS decoding algorithm:

[0066]

[0067] Among them, Ci Denote the set of check nodes connected to variable node \(i\). This set of check nodes represents the connection mode from the variable layer to the output layer in the deep learning decoding model, that is, the weight matrix \(W\). out At this time, in the deep learning decoding model, it is manifested as the connection from the last variable layer to the output layer.

[0068] Each layer needs to transmit information completely unchanged, and the output is specified in the range of 0 - 1. The activation functions and weight matrices of each layer are shown in Table 2.

[0069] Table 2, (6,3) deep learning decoding model:

[0070]

[0071] S3: Based on the weight matrix and activation function, use the belief propagation algorithm to establish a deep learning decoding model.

[0072] Based on the traditional NMS decoding algorithm, for the LDPC code established on the Tanner graph, each iteration process of BP decoding includes two steps: the processing of check nodes and the processing of variable nodes; in each iteration, the check node receives messages from its adjacent variable nodes, and after processing, it transmits the messages back to the adjacent variable nodes. At this time, the variable node receives the messages from the adjacent check nodes for processing; while in the deep learning decoding model, one iteration operation is represented by the connection from one variable layer to the next check layer, and the message passing is represented by the weight matrix between the hidden layers. The weight matrix represents the process of message passing, and the distribution of non - zero positions in the weight matrix represents the process of a node receiving messages from adjacent nodes; at the same time, initialize the corresponding activation function.

[0073] In this embodiment, in iteration 1, the variable layer expands the input message through the input layer to obtain \(L(q\) ij ), and the check layer completes the processing of check node messages in belief propagation. Its forward propagation is completed inside the check layer. To calculate \(L(r\) ij ), that is, at this time, horizontally observe the check matrix \(H\), and obtain the forward propagation schematic diagram of the check layer as shown in Figure 5 . At this time, the horizontal calculation of three check equations is involved. Therefore, each of the three dashed boxes represents the horizontal calculation of the corresponding check equation; it mainly involves two pieces of information: the first is the sign of \(sgn(e\) ij ), and the second is the absolute value of the information \(abs(e\) ij ). The triangle is a comparator that performs the function of the min function in the iterative algorithm. According to the sign of the sign and the minimum value, the calculation of the check node message processing formula in the NMS decoding algorithm iteration can be completed, and the check node information can be updated.

[0074] Taking the first check equation as an example, in Figure 5In the uppermost dashed-line box, the forward propagation connection relationship of the verification layer is shown in Table 3, and the relationship between the verification layer and the variable layer is shown in Table 4.

[0075] Table 3, Forward Propagation Connection Relationship of the Verification Layer:

[0076] Node information Horizontally adjacent points (R(j)\i) Comparator input E11 E12 E14 E12 E14 E12 E11 E14 E11 E14 E14 E11 E12 E11 E12

[0077] In iteration 1, from the verification layer to the variable layer, the message processing of the variable nodes in the belief propagation is completed. To calculate L(q ij ), that is, at this time, by observing the parity-check matrix H vertically, the initial value of the weight matrix W c2v can be obtained. The weight matrix represents the connection structure, and the corresponding connection relationships are as follows:

[0078]

[0079] Table 4, Relationship between the Verification Layer and the Variable Layer:

[0080] Variable node information Vertically adjacent nodes (C(i)\j) Connected check node e11 E21 E31 E21 E31 e12 E22 E22 e14 \ \ e21 E11 E31 E11 E31 e22 E12 E12 e23 E33 E33 e25 \ \ e31 E21 E11 E21 E11 e33 E23 E23 e36 \ \

[0081] S4: Use the training sample set to train the deep learning decoding model, and save the trained deep learning decoding model as a decoder. Usually, being trained well means achieving the training effect or reaching the preset maximum number of training times.

[0082] When using the training sample set to train the deep learning decoding model, the deep learning decoding model is trained by the extended adaptive learning rate method of stochastic gradient descent, and the weight matrix of the hidden layer is updated using backpropagation. The coefficients in the weight matrix correspond to the correction factors of the traditional NMS decoding algorithm, which are used to change the amplitude of the message and thus improve the performance.

[0083] In this embodiment, according to the above deep learning decoding network construction method, a decoding network is built for the codeword type to be decoded. The signal-to-noise ratio range is set from -1 dB to 3 dB with a step size of 1, a total of five noise environments. For each noise environment and the codeword type to be decoded, a total of 5000 codewords are used to calculate the soft information after adding noise, obtaining a training sample set for the current noise environment with a size of 5000. The learning rate Lr = 0.0001 is set for model training, the number of training epochs = 5, and a training sample set with the same size is regenerated for each training; after training is completed, the trained deep learning decoding model is saved as a decoder.

[0084] S5: Send the soft information of the LDPC code to be decoded corresponding to the parity-check matrix H into the decoder to complete the LDPC decoding.

[0085] The decoder has a hard decision unit, and the hard decision unit is used to perform a numerical hard decision to be 0 or 1 according to the probability value of 0 - 1 output by the deep learning decoding model.

[0086] In this embodiment, the soft information of the LDPC code receiving the same parity-check matrix H is read by the decoder for decoding. At this time, the decoding output of the deep learning decoding model is a real value in the range (0, 1). Through hard decision: It is converted into 01 bits to complete the decoding.

[0087] An example result of the performance test for the present invention is as Figure 6a 、 Figure 6b shown. The present invention only needs to use low signal-to-noise ratio data for network training, and the training data is greatly reduced compared with traditional neural networks. The decoding performance has a 1dB improvement compared with the traditional NMS decoding algorithm, and the number of iterations is fixed at ten times, reducing the number of iterations under low signal-to-noise ratio and reducing the decoding delay.

[0088] In summary, the beneficial effects of the present invention are:

[0089] 1. The belief propagation LDPC decoding method based on deep learning provided by the present invention has a reduced number of iterations and can quickly recover the transmitting end sequence from an error code environment.

[0090] 2. The belief propagation LDPC decoding method based on deep learning provided by the present invention is driven by the traditional NMS decoding algorithm and is based on the deep learning decoding model, reducing the training amount. For the codewords of the same parity-check matrix H, good decoding performance can be achieved without training all the codewords. Using this feature reduces the number of iterations and speeds up the convergence rate.

[0091] Based on the same technical concept, an embodiment of the present invention also provides an electronic device, which can implement the process of the belief propagation LDPC decoding method based on deep learning provided in the above embodiments of the present invention. In one embodiment, the electronic device can be a server, a terminal device or other electronic devices. As Figure 7 shown, the electronic device may include:

[0092] At least one processor, and a memory connected to at least one processor. In the embodiments of the present invention, the specific connection medium between the processor and the memory is not limited. Figure 7 Taking the example that the processor and the memory are connected by a bus. The bus is represented by a thick line in Figure 7 . The connection methods between other components are only for illustrative purposes and are not limited thereto. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 7 only a thick line is used to represent it in

[0093] In an embodiment of the present invention, a memory stores instructions executable by at least one processor. By executing the instructions stored in the memory, the at least one processor can execute a belief propagation LDPC decoding method based on deep learning described above. The processor can implement Figure 7 the functions of each module in the device shown.

[0094] Among them, the processor is the control center of the device. It can connect various parts of the entire control device through various interfaces and lines. By running or executing the instructions stored in the memory and calling the data stored in the memory, various functions of the device and process data, so as to monitor the device as a whole.

[0095] In an alternative design, the processor may include one or more processing units. The processor may integrate an application processor and a modem processor. Among them, the application processor mainly processes the operating system, user interface, application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above modem processor may not be integrated into the processor. In some embodiments, the processor and the memory can be implemented on the same chip. In some embodiments, they can also be implemented separately on independent chips.

[0096] The processor can be a general-purpose processor, such as a CPU, a digital signal processor, an application-specific integrated circuit, a field programmable gate array, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of a belief propagation LDPC decoding method based on deep learning disclosed in combination with the embodiments of the present invention can be directly embodied as being completed by a hardware processor, or completed by a combination of hardware and software modules in the processor.

[0097] A memory, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The memory may include at least one type of storage medium, for example, it may include flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disc, and so on. The memory is any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory in the embodiments of the present invention may also be a circuit or any other device capable of implementing a storage function, for storing program instructions and / or data.

[0098] By designing and programming the processor, the code corresponding to a deep learning-based belief propagation LDPC decoding method introduced in the foregoing embodiments can be solidified into the chip, so that the chip can execute the steps of the method in the above embodiments when running. How to design and program the processor is a well-known technology to those skilled in the art and will not be elaborated here.

[0099] Based on the same inventive concept, the embodiments of the present invention also provide a storage medium storing computer instructions, which, when run on a computer, cause the computer to execute a deep learning-based belief propagation LDPC decoding method discussed above.

[0100] In some optional embodiments, the various aspects of a deep learning-based belief propagation LDPC decoding method of the present invention can also be implemented in the form of a program product, which includes program code that, when the program product runs on a device, causes the control device to execute the steps in a deep learning-based belief propagation LDPC decoding method according to various exemplary embodiments of the present invention described above in this specification.

[0101] It should be noted that although several units or subunits of the device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present invention, the features and functions of two or more of the above-described units may be embodied in one unit. Conversely, the features and functions of one unit described above may be further divided and embodied by multiple units. Additionally, although the operations of the method of the present invention are described in a specific order in the drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the illustrated operations must be performed to achieve the desired result. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step for execution, and / or one step may be decomposed into multiple steps for execution.

[0102] Those skilled in the art should understand that the embodiments of the present invention may be provided as a method, a system, or a computer program product. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0103] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a server, such that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one or more of the flows or multiple flows and / or blocks Figure 1 one or more of the blocks or multiple blocks.

[0104] Program code for performing the operations of the present invention may be written using any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., as well as conventional procedural programming languages such as the "C" language or similar programming languages. The program code may execute entirely on the user computing device, partially on the user device, execute as a stand-alone software package, execute partially on the user computing device and partially on a remote computing device, or execute entirely on a remote computing device or server.

[0105] In the case of a remote computing device, the remote computing device can be connected to the user computing device via any kind of network including a local area network (LAN) or a wide area network (WAN), or, it can be connected to an external computing device (e.g., by using an Internet service provider to connect via the Internet).

[0106] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device that implements the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are performed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0108] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A belief propagation LDPC decoding method based on deep learning, characterized in that, Including: Obtaining an LDPC code training sample set corresponding to a parity-check matrix H; Establishing weight matrices and activation functions for each layer of a deep learning decoding model according to the parity-check matrix H; Based on the weight matrices and activation functions, establishing a deep learning decoding model using the belief propagation algorithm; Training the deep learning decoding model using the training sample set, and saving the trained deep learning decoding model as a decoder; Feeding the soft information of the LDPC code to be decoded corresponding to the parity-check matrix H into the decoder to complete LDPC decoding.

2. The belief propagation LDPC decoding method based on deep learning according to claim 1, characterized in that, Samples in the training sample set are represented as (Y, y), where Y is the input message and y is the label; The label y is the codeword of the LDPC code. For an LDPC code (n, k), n is the codeword length and k is the information bit length. The codeword y is obtained by multiplying the information bit length k by the generator matrix G; The input message Y is a log-likelihood ratio (LLR) value of length n. Its calculation process includes: generating received data by modulating the codeword y through BPSK and adding Gaussian white noise, and calculating the log-likelihood ratio (LLR) value of length n based on the received data.

3. The belief propagation LDPC decoding method based on deep learning according to claim 1, characterized in that The deep learning decoding model includes an input layer, a hidden layer, and an output layer connected in sequence; the parity-check matrix H indicates the connection relationship between variable nodes and check nodes. The hidden layer in the deep learning decoding model is divided into a variable layer and a check layer according to the execution function. The connection method between the variable layer and the check layer is determined by the position of "1" distribution in the parity-check matrix H. There are two connection methods from the variable layer to the check layer and from the check layer to the variable layer. The weight matrices corresponding to each connection method are obtained according to the mapping relationship and the traditional NMS decoding algorithm; the activation function of the output layer is modified according to the mapping relationship of BPSK modulation.

4. The belief propagation LDPC decoding method based on deep learning according to claim 3, characterized in that In the hidden layer, odd-numbered hidden layers are variable layers and even-numbered hidden layers are check layers.

5. The belief propagation LDPC decoding method based on deep learning according to claim 3, characterized in that, The establishment of the deep learning decoding model using the belief propagation algorithm includes: based on the traditional NMS decoding algorithm, establishing an LDPC code on a Tanner graph. Each iteration process of BP decoding includes two steps: processing of check nodes and processing of variable nodes; in each iteration, a check node receives messages from its adjacent variable nodes, and after processing, it sends them back to the adjacent variable nodes. At this time, the variable node receives the messages from the adjacent check nodes for processing; in the deep learning decoding model, one iteration operation is represented by going from one layer of the variable layer to the next layer of the check layer, and the message passing is represented by the weight matrix between the hidden layers. The weight matrix represents the process of message passing, and the distribution of non-zero bits in the weight matrix represents the process of a node receiving messages from adjacent nodes; at the same time, the corresponding activation function is initialized.

6. The belief propagation LDPC decoding method based on deep learning according to claim 1, wherein When training the deep learning decoding model using the training sample set, the deep learning decoding model is trained by an extended adaptive learning rate method of stochastic gradient descent, and the weight matrix of the hidden layer is updated using backpropagation. The coefficients in the weight matrix correspond to the correction factors of the traditional NMS decoding algorithm.

7. The belief propagation LDPC decoding method based on deep learning according to claim 1, wherein The decoder has a hard decision unit, and the hard decision unit is configured to perform a numerical hard decision to be 0 or 1 according to the probability value of 0-1 output by the deep learning decoding model.

8. An electronic device, characterized in that, It includes: At least one processor; And a memory communicatively connected to the at least one processor; Wherein, the memory stores instructions executable by the at least one processor, and the at least one processor, by executing the instructions stored in the memory, causes the at least one processor to execute the method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store instructions, and when the instructions are executed, the method according to any one of claims 1-7 is implemented.

10. A computer program product, characterized in that, When the computer program product is called by a computer, the computer is caused to execute the method according to any one of claims 1-7.

Images