A channel decoding method and system based on graph neural network

Through the channel decoding method based on graph neural network, multiple neural network models are used to update the nodes and edge vectors in the Tanner graph, the problem of insufficient dimensional curse and parallel capabilities in traditional channel decoding is solved, and a more efficient channel decoding effect is achieved.

CN115664899BActive Publication Date: 2024-11-05ZHEJIANG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202211311725.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-10-25
Publication Date
2024-11-05
Estimated Expiration
2042-10-25

AI Technical Summary

Technical Problem

Traditional neural network channel decoding is troubled by the dimensional curse problem, and its parallelism is not strong, making it difficult to effectively apply in large-scale parallel computing.

Method used

Using the channel decoding method based on graph neural network, the vectors of variable nodes, check nodes and edges are updated using multiple neural network models, and the updated vectors are mapped into the output matrix using the output transformation matrix, and finally the output is output. The convolutional neural network module performs decoding, and parameter sharing reduces the complexity of the model and the number of parameters.

Benefits of technology

It effectively weakens the dimensional curse, reduces the complexity of the model, improves the parallelism, facilitates large-scale parallel computing, and achieves more efficient channel decoding.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115664899B_ABST
    Figure CN115664899B_ABST
Patent Text Reader

Abstract

The present invention discloses a channel decoding method and system based on a graph neural network. The core of the present invention lies in a decoder represented by a graph neural network. Each iteration of an edge or point is equivalent to outputting a new point or edge vector through a multi-layer neural network for the relevant information, so that the neural network can continuously learn the structure of the graph during message transmission, thereby achieving the purpose of channel decoding. Compared with the traditional neural network decoding method, the present invention breaks through the deep learning decoding limitation of the curse of dimensionality because of parameter sharing. The parameters are reduced by an order of magnitude compared with the traditional model, and the model complexity is reduced. Because the node or edge vector can be updated in parallel, the parallel capability is enhanced. Under the same conditions, the performance is better than other traditional neural network decoding algorithms such as the belief propagation algorithm based on the neural network, which can save a lot of time and computing resources, and has broad application prospects in the fields of high-performance communication and low-power communication.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of communication and machine learning technology, and in particular to a channel decoding method and system based on graph neural network. Background Art

[0002] Communication technology has developed rapidly in the past few decades. With the rise of machine learning, new machine learning-assisted channel decoding has begun to emerge, which provides a new way of thinking for channel decoding, and is expected to replace traditional decoders and bring speed and efficiency improvements. However, traditional neural network channel decoding is plagued by the curse of dimensionality and has weak parallel capabilities. Graph Neural Network (GNN) refers to the general term for algorithms that use neural networks to learn graph structured data, extract and discover features and patterns in graph structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. Summary of the invention

[0003] The purpose of the present invention is to address the deficiencies in the prior art and provide a channel decoding method and system based on graph neural network, which weakens the curse of dimensionality, reduces the number of parameters by orders of magnitude compared to traditional models, reduces the complexity of the model, and has good parallelism in updating edge and point values, facilitating large-scale parallelism.

[0004] The technical solution adopted by the present invention is specifically as follows:

[0005] Determine the size of the check matrix based on the codeword length N of the data to be decoded and the codeword length K to be output and generate a Tanner graph, determine the variable nodes, the check nodes and the connection relationship between them; the number of variable nodes in the Tanner graph is N, and the number of check nodes is M=NK;

[0006] The data to be decoded is converted into vectors of variable nodes in the Tanner graph using the input conversion matrix T and all vectors of check nodes are set to zero vectors; the size of the input conversion matrix T is N×1, and the i-th element T of the input conversion matrix T is i Convert the data to be decoded into the vector of the i-th variable node in the Tanner graph;

[0007] Use the corresponding neural network model to pass messages to update the variable nodes, check nodes, and vectors corresponding to the edges respectively. One edge has two directions corresponding to two vectors.

[0008] Among them, the corresponding neural network model at least includes:

[0009] The first neural network model outputs an update vector of the check node according to the constructed input matrix of the check node;

[0010] The second neural network model outputs an update vector of the variable node according to the constructed input matrix of the variable node;

[0011] The third neural network model outputs an update vector of the first direction edge according to the constructed input matrix of the first direction edge;

[0012] The fourth neural network model outputs an update vector of the second direction edge according to the constructed input matrix of the second direction edge;

[0013] The output conversion matrix O is used to map the vector conversion of the variable node in the updated Tanner graph to the output matrix, wherein the size of the output conversion matrix O is N×1, the size of the output matrix is ​​N×N, and the i-th element O of the output conversion matrix O is i Convert the vector of the i-th check node in the updated Tanner graph into a row or column of the output matrix respectively;

[0014] The output matrix is ​​decoded by the output convolutional neural network module to obtain the codeword and complete the decoding;

[0015] The input conversion matrix T, output conversion matrix O, first neural network model, second neural network model, third neural network model, fourth neural network model and output convolutional neural network module are obtained by training using training data to minimize the loss function between the codeword decoded by the output neural network module and the true value.

[0016] Furthermore, the input matrix of the second neural network model required for updating the corresponding vectors of the variable nodes is at least composed of the vectors of the variable nodes themselves and the vectors of the edges pointing to themselves; the input matrix of the first neural network model required for updating the corresponding vectors of the check nodes is at least composed of the vectors of the check nodes themselves and the vectors of the edges pointing to themselves; the input matrix of the third neural network model required for updating the corresponding vectors of the first-direction edges is at least composed of the vectors of the first-direction edges themselves and the vectors of the points connecting the two ends of the edge; the input matrix of the fourth neural network model required for updating the corresponding vectors of the second-direction edges is at least composed of the vectors of the second-direction edges themselves and the vectors of the points connecting the two ends of the edge.

[0017] Furthermore, the input matrix of the first neural network model required to update the corresponding vector of the check node and the input matrix of the second neural network model required to update the corresponding vector of the variable node also include additional parameters of the point itself; the input matrix of the third neural network model required to update the corresponding vector of the edge in the first direction and the input matrix of the fourth neural network model required to update the corresponding vector of the edge in the second direction also include additional parameters of the edge itself.

[0018] Furthermore, the output convolutional neural network module is composed of a convolutional layer, a fully connected layer and an activation function layer.

[0019] A channel decoding system based on a graph neural network includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the channel decoding method based on the graph neural network is implemented.

[0020] A storage medium containing computer executable instructions, which implement the channel decoding method based on graph neural network when executed by a computer processor.

[0021] The beneficial effect of the present invention is that the channel decoding method based on graph neural network provided by the present invention has parameter sharing when updating values, so the parameters are reduced by an order of magnitude compared with the traditional model, the model complexity is reduced, the curse of dimensionality is weakened, and because the vector values ​​of the edges can be updated together, the vector values ​​of the points can also be updated together, the parallelism is very good, which is convenient for large-scale computer parallelism. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the Tanner graph required for message passing, which includes variable nodes and check nodes.

[0023] Figure 2 This is a flow chart of the channel decoding method based on graph neural network.

[0024] Figure 3 It is the simulation verification and comparison result. DETAILED DESCRIPTION

[0025] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0026] The present invention provides a channel decoding method based on a graph neural network. The core of the present invention lies in a decoder represented by a graph neural network. Each iteration of an edge or point is equivalent to outputting a new point or edge vector through a multi-layer neural network for the relevant information, so that the neural network can continuously learn the structure of the graph in the message transmission, thereby achieving the purpose of channel decoding. Specifically, by determining the size of the check matrix based on the codeword length N of the data to be decoded and the codeword length K to be output and generating a Tanner graph, the data to be decoded is converted into a vector of a variable node in the Tanner graph using the input conversion matrix T; then the neural network model is used for message transmission to update the variable node, the check node, and the vector corresponding to the edge respectively, and the output conversion matrix O is used to map the vector of the variable node in the updated Tanner graph to the output matrix, and finally the output matrix is ​​decoded by the output convolutional neural network module to obtain the codeword. A channel decoding method based on a graph neural network provided by the present invention has parameter sharing when updating the value, and the parameters are reduced by an order of magnitude compared with the traditional model, the model complexity is reduced, and the curse of dimensionality is weakened. Moreover, since the vector values ​​of the edges can be updated together, the vector values ​​of the points can also be updated together, and the parallelism is very good, which is convenient for large-scale parallelism of computers.

[0027] The neural network model, input conversion matrix T, output conversion matrix O and output convolutional neural network module used in the present invention together constitute an overall combined neural network module, which is obtained by training with training data to minimize the loss function between the codeword obtained by decoding the output neural network module and the true value. Specifically, Figure 2 As shown, the training method is as follows:

[0028] Step 1: Prepare training data, and use a specific coding method, such as the BCH channel coding method, to map a codeword of length K to a codeword of length N, that is, to add an error correction redundant part, which is used as the data to be decoded (length N) and the corresponding decoded true value (length K). Then add noise through the AWGN channel to obtain a signal with a certain signal-to-noise ratio, which is convenient for the next step of processing;

[0029] Step 2: Define the graph neural network model and initialize its parameters, including the vectors of edges and points and additional parameters; the graph neural network model of the present invention is constructed based on the Tanner graph, which is a graph that represents the connection relationship between check nodes and variable nodes. Each row in the check matrix represents a linear constraint, corresponding to a check node in the Tanner graph, and each column in the check matrix represents a variable, corresponding to a variable node in the Tanner graph. If a variable participates in a linear constraint, there is a connecting edge between the variable node and the check node in the Tanner graph. According to the Tanner graph, we can construct Figure 1The decoder represented by defines the input endpoint as v i , the constraint end is u i , the vector of each endpoint is or The edge vector is or (related to direction), the additional parameter of the node is a vector or The extra parameter for edges is a vector or In the present invention, the size of the check matrix is ​​determined based on the codeword length N of the data to be decoded and the codeword length K to be output, and a Tanner graph is generated to determine the variable nodes, the check nodes and the connection relationship therebetween; the number of variable nodes in the generated Tanner graph is N, and the number of check nodes is M=NK.

[0030] The graph neural network model of the present invention is specifically a neural network model, an input conversion matrix T, an output conversion matrix O and an output convolutional neural network module, which together constitute an integrated combined neural network module; wherein the input conversion matrix T is a vector that converts the data to be decoded into the variable nodes in the Tanner graph, the size of the input conversion matrix T is N×1, and the i-th element T of the input conversion matrix T i Convert the data to be decoded into the i-th variable node v in the Tanner graph i Vector At the same time, the point vector corresponding to U is set to the zero vector; the output conversion matrix O maps the vector conversion of the variable node in the updated Tanner graph to the output matrix, wherein the size of the output conversion matrix O is N×1, the size of the output matrix is ​​N×N, and the i-th element O of the output conversion matrix O is i The vector of the i-th check node in the updated Tanner graph is converted into a row or column of an output matrix respectively; the neural network model at least includes:

[0031] The first neural network model is based on the constructed check node u i The input matrix is ​​the update vector of the check node.

[0032] The second neural network model is constructed based on the variable node v i The input matrix is ​​the update vector of the variable node.

[0033] The third neural network model outputs an update vector of the first direction edge according to the constructed input matrix of the first direction edge;

[0034] The fourth neural network model outputs an update vector of the second direction edge according to the constructed input matrix of the second direction edge;

[0035] The output convolutional neural network module decodes the output matrix to obtain the codeword. The output convolutional neural network module consists of a convolutional layer, a fully connected layer and an activation function layer, and a traditional convolutional neural network can be used.

[0036] The neural network (second neural network model, first neural network model) parameter matrices required to define the update vectors of nodes in V and U are W v With W u , the neural network (third neural network model, fourth neural network model) parameter matrices required to update the vector of an edge are W vu With W uv (related to direction), the output convolutional neural network parameter matrix is ​​W out ; and initialize the weight parameter matrix W u , W v , W uv , W vu , the input / output transformation matrix T and O and the neural network parameters W of the output convolutional neural network module out .

[0037] Step 3: Train the graph neural network model. Input the data in step 1 into the combined neural network module above and perform a certain number of cycles. During this process, message passing is continuously performed. Whenever a codeword is decoded during training, the loss function is evaluated, back propagated, and the parameters are updated. Specifically, this step is as follows:

[0038] (3.1) First, each element T in the input transformation matrix T i Convert the input codeword to V i The corresponding dimension of U, and set the point vector corresponding to U to the zero vector;

[0039] (3.2) Then construct the input matrix to start updating the vector of the point. By weighting and splicing the vector of the point itself, the vector of the edge pointing to itself, the point's own parameters and weight parameters into an input matrix, input a corresponding multi-layer convolutional neural network model respectively, and then output the result, update the vector of the corresponding point, and update the vector of the point itself. All points must be updated once. In general, the input matrix of the second neural network model required to update the vector corresponding to the variable node is at least composed of the vector of the variable node itself and the vector of the edge pointing to itself; the input matrix of the first neural network model required to update the vector corresponding to the check node is at least composed of the vector of the check node itself and the vector parameters of the edge pointing to itself, but not limited to this; the additional parameters of the point itself are optional and can also be randomly generated.

[0040] The second neural network model that updates all nodes belonging to V should use the same parameter matrix W v, the first neural network to update all nodes belonging to U should use the same parameter matrix W u , such sharing can reduce parameters and reduce complexity;

[0041] (3.3) Start to build the input matrix to update the edge vector. By using the edge vector, the vector of the points at both ends of the edge, the edge parameters and the weight parameters, etc., weighted and spliced ​​into an input matrix, input a corresponding multi-layer convolutional neural network model, and then output the result, update the corresponding edge vector, and update the edge vector. Since the edge has a direction, all edges must be updated twice. For an edge The vector representing the edge from the node of V to the node of U is defined as the first direction edge vector in this embodiment. When updating the vector of this type of edge, the same parameter matrix W needs to be used. vu Similarly, when updating the edge value from the node of U to the node of V (the second direction edge vector), the same parameter matrix W needs to be used uv , such sharing can reduce parameters and complexity; in general, the input matrix of the third neural network model required to update the corresponding vector of the first direction edge is at least composed of the vector of the first direction edge itself and the vector of the points at both ends of the edge; the input matrix of the fourth neural network model required to update the corresponding vector of the second direction edge is at least composed of the vector of the second direction edge itself and the vector of the points at both ends of the edge, but not limited to this, feature information such as additional parameters of the edge itself can be added according to actual conditions to improve accuracy. Similarly, the additional parameters of the edge itself are optional and can also be randomly generated.

[0042] (3.4) Use each element O of the output transformation matrix O i The updated v i The corresponding output vector is mapped to the final output convolutional neural network module, and then passes through the designed convolution and fully connected layers, and finally through an activation function layer to output the final codeword result;

[0043] (3.5) Compare the output result with the expected result, i.e. the true value, and calculate the loss function, such as the mean square error value, etc. Back propagate the entire network, and use the stochastic gradient descent method to update the gradient to obtain the updated parameter value;

[0044] (3.6) Repeat the above steps (3.1) to (3.5) for a certain number of iterations to end the training process.

[0045] After the training is completed, the trained neural network module can be used for channel decoding based on graph neural network. The specific method is as follows:

[0046] Determine the size of the check matrix based on the codeword length N of the data to be decoded and the codeword length K to be output and generate a Tanner graph, determine the variable nodes, the check nodes and the connection relationship between them;

[0047] The data to be decoded is converted into the vector of the variable node in the Tanner graph using the input transformation matrix T and the vectors of the check nodes are all set to zero vectors;

[0048] The corresponding neural network model is used for message passing to update the variable nodes, check nodes, and the vectors corresponding to the edges respectively, and the output conversion matrix O is used to convert the vectors of the variable nodes in the updated Tanner graph into the output matrix, wherein the size of the output conversion matrix O is N×1, the size of the output matrix is ​​N×N, and the i-th element O of the output conversion matrix O is i Convert the vector of the i-th check node in the updated Tanner graph into a row or column of the output matrix respectively;

[0049] The output matrix is ​​decoded by the output convolutional neural network module to obtain the codeword and complete the decoding.

[0050] In this embodiment, the bit error rate is tested under different signal-to-noise ratios, and compared with the method of direct transmission without channel coding and channel decoding and the traditional neural network decoding method. Here, the belief propagation (BP) algorithm based on deep learning is selected to compare with the method based on graph neural network of the present invention. Here, N=63, K=NM=45, and the signal-to-noise ratio ranges from 3dB to 8dB. The bit error rate simulation results are shown in Figure 3 It can be seen that the decoding method based on graph neural network of the present invention obtains a lower bit error rate with fewer iterations than the neural network BP decoding algorithm.

[0051] Corresponding to the aforementioned embodiment of the channel decoding method based on graph neural network, the present invention also provides an embodiment of a channel decoding system based on graph neural network.

[0052] An embodiment of the present invention provides a graph neural network-based channel decoding system, including one or more processors for implementing the graph neural network-based channel decoding method in the above embodiment.

[0053] The embodiment of the channel decoding system based on graph neural network of the present invention can be applied to any device with data processing capability, and the any device with data processing capability can be a device or apparatus such as a computer.

[0054] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0055] An embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the channel decoding method based on graph neural network in the above embodiment is implemented.

[0056] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capability and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0057] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A channel decoding method based on graph neural network, characterized in that: Specifically: Determine the size of the check matrix based on the codeword length N of the data to be decoded and the codeword length K to be output and generate a Tanner graph, determine the variable nodes, the check nodes and the connection relationship between them; the number of variable nodes in the Tanner graph is N, and the number of check nodes is M=NK; The data to be decoded is converted into the vector of the variable node in the Tanner graph using the input transformation matrix T and the vectors of the check nodes are all set to zero vectors; The size of the input transformation matrix T is N×1, and the i-th element T of the input transformation matrix T i Convert the data to be decoded into the vector of the i-th variable node in the Tanner graph; Use the corresponding neural network model to pass messages to update the variable nodes, check nodes, and vectors corresponding to the edges respectively. One edge has two directions corresponding to two vectors. The corresponding neural network model at least includes: The first neural network model outputs an update vector of the check node according to the constructed input matrix of the check node; The second neural network model outputs an update vector of the variable node according to the constructed input matrix of the variable node; The third neural network model outputs an update vector of the first direction edge according to the constructed input matrix of the first direction edge; The fourth neural network model outputs an update vector of the second direction edge according to the constructed input matrix of the second direction edge; The output conversion matrix O is used to map the vector conversion of the variable node in the updated Tanner graph to the output matrix, wherein the size of the output conversion matrix O is N×1, the size of the output matrix is ​​N×N, and the i-th element O of the output conversion matrix O is i Convert the vector of the i-th check node in the updated Tanner graph into a row or column of the output matrix respectively; The output matrix is ​​decoded by the output convolutional neural network module to obtain the codeword and complete the decoding; The input conversion matrix T, output conversion matrix O, first neural network model, second neural network model, third neural network model, fourth neural network model and output convolutional neural network module are obtained by training using training data to minimize the loss function between the codeword decoded by the output neural network module and the true value.

2. The method according to claim 1, characterized in that The input matrix of the second neural network model required for updating the corresponding vectors of the variable nodes is at least composed of the vectors of the variable nodes themselves and the vectors of the edges pointing to themselves; the input matrix of the first neural network model required for updating the corresponding vectors of the check nodes is at least composed of the vectors of the check nodes themselves and the vectors of the edges pointing to themselves; the input matrix of the third neural network model required for updating the corresponding vectors of the first-direction edges is at least composed of the vectors of the first-direction edges themselves and the vectors of the points connecting the two ends of the edge; the input matrix of the fourth neural network model required for updating the corresponding vectors of the second-direction edges is at least composed of the vectors of the second-direction edges themselves and the vectors of the points connecting the two ends of the edge.

3. The method according to claim 1, characterized in that The input matrix of the first neural network model required to update the corresponding vector of the check node and the input matrix of the second neural network model required to update the corresponding vector of the variable node also include additional parameters of the point itself; the input matrix of the third neural network model required to update the corresponding vector of the edge in the first direction and the input matrix of the fourth neural network model required to update the corresponding vector of the edge in the second direction also include additional parameters of the edge itself.

4. The method according to claim 1, characterized in that: The output convolutional neural network module consists of a convolutional layer, a fully connected layer and an activation function layer.

5. A channel decoding system based on a graph neural network, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the channel decoding method based on the graph neural network as described in any one of claims 1 to 4 is implemented.

6. A storage medium comprising computer executable instructions, which, when executed by a computer processor, implement the channel decoding method based on graph neural network as described in any one of claims 1 to 4.

Citation Information

Patent Citations

  • Deep learning-based RCM method

    CN110430013A

  • RS code belief propagation decoding method based on deep learning

    CN110730008A