Channel decoding method and system based on boolean geometry embedding and attention decoupling

Abstract

The disclosure provides a channel decoding method and system based on Boolean geometry embedding and attention decoupling, relating to the technical field of channel error correction coding, comprising: constructing an initial observation state vector; introducing an absolute index sequence, extracting a Boolean geometry feature embedding, and constructing a global structured position feature sequence based on the Boolean geometry feature embedding; constructing a logic graph mask and a geometry graph mask, performing deep multi-modal fusion of the global structured position feature sequence and the initial observation state vector to generate network initial input features, inputting the network initial input features into a Transformer decoding network, introducing constraints of the logic graph mask and the geometry graph mask in multi-head self-attention calculation of each layer, implementing feature decoupling and dynamic gating mechanism, extracting final deep abstract feature representation, performing global joint decoding operation, and obtaining predicted decoding output result. The disclosure greatly improves the theoretical completeness and engineering robustness of the model in dealing with complex and sudden noise patterns.

Description

Technical Field

[0001] This disclosure relates to the field of channel error correction coding technology, specifically to a channel decoding method and system based on Boolean geometric embedding and attention decoupling. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Channel error correction coding technology, as a core component of modern communication systems, directly determines the reliability and efficiency of information transmission in noisy channels. Traditional communication system designs typically employ iterative decoding schemes based on the Belief Propagation (BP) algorithm or probabilistic decoding schemes based on the Successive Cancellation List (SCL) algorithm. While these classic decoding algorithms exhibit good error correction performance under specific channel conditions, as code length increases, they often suffer from problems such as dramatically increased computational complexity and large decoding delays, making it difficult to fully meet the stringent requirements of 5G and future B5G / 6G communication systems for extremely low latency and ultra-high reliability.

[0004] In recent years, breakthroughs in deep learning technology have brought new opportunities to the field of channel decoding. Researchers have successively proposed various neural decoder architectures, including decoding models based on multilayer perceptrons (MLPs), recurrent neural networks (RNNs), and graph neural networks (GNNs). These early neural decoding methods mostly involved computational graph expansion or functional simulation of the traditional backpropagation (BP) algorithm, attempting to utilize the powerful nonlinear fitting capabilities of neural networks to approximate optimal maximum likelihood decoding (MLD) performance. However, these methods face inherent topological constraints and challenges in modeling long-range dependencies in practical applications.

[0005] To overcome the challenges of topological constraints and long-distance dependency modeling, existing research has introduced the Transformer architecture and innovatively proposed the Error Correction Code Transformer (ECCT) model. The ECCT model, for the first time, transforms the soft decoding process of linear block codes into a sequence modeling problem, breaking the absolute dependence on local Tanner graph edges. Instead, it automatically learns the complex interaction relationships between variable nodes (VNs) and check nodes (CNs) through a global self-attention mechanism. This model maps the received channel soft information sequence to a high-dimensional feature space and utilizes a multi-head attention mechanism to process the global information interaction between nodes in parallel, thereby achieving a significant improvement in decoding performance in some short-to-medium code scenarios.

[0006] However, when dealing with polarization-type channel codes with highly algebraic structural features (such as Polar codes widely used in 5G control channels, and Reed-Muller codes with deep algebraic geometric foundations), the ECCT model exposes several deep-seated technical limitations: (1) Existing technologies simply copy the random index embedding of natural language processing, which cuts off the natural physical mapping of bits in the Boolean hypercube, causing the underlying "topology collapse" and forcing the network to perform inefficient data-driven fitting.

[0007] (2) Existing Transformer decoders force all attention heads to share a single parity check map (Tanner map) mask, resulting in severe feature homogenization and representation collapse in the multi-head mechanism, and they cannot take advantage of the local clustering of bits in geometric space when errors occur, forming a "topological blind zone".

[0008] (3) After introducing the decoupling mechanism of dual graphs (logical constraint graph and geometric adjacency graph), how to efficiently integrate these two heterogeneous information flows has become a new technical challenge. Simple static splicing or fixed weight addition will cause feature expression conflicts at different network depths.

[0009] (4) Existing technologies generally rely on Gaussian elimination when obtaining the system generator matrix of linear block codes. The inevitable column permutation operation will completely disrupt the natural bit sequence of polarized channel codes, causing the underlying "coordinate system collapse" and making any subsequent efforts to introduce geometric auxiliary features ineffective.

[0010] In summary, existing neural decoding techniques fail to fully exploit the unique Boolean geometric algebraic features of polarized channel codes and suffer from significant limitations such as missing underlying structural priors, homogenization of multi-head attention mechanisms, and disruption of natural bit order by conventional algebraic preprocessing. These problems lead to topological blind spots and representation collapse in the network, severely restricting the performance improvement of neural decoders in complex environments. Therefore, there is an urgent need to develop a novel neural decoding scheme that can effectively integrate Boolean geometric prior features, overcome the limitations of a single logical constraint graph, and strictly preserve the natural bit order, in order to overcome the performance bottleneck of existing technologies when processing highly structured channel codes. Summary of the Invention

[0011] To address the aforementioned issues, this disclosure proposes a channel decoding method and system based on Boolean geometric embedding and attention decoupling. By solving the problem of projecting discrete algebraic properties onto continuous features, the physical topological absolute coordinates of the Boolean hypercube are effectively injected into the network's bottom layer. A geometric adjacency graph is introduced and constructed, achieving deep decoupling of the multi-head attention mechanism. This effectively integrates Boolean geometric prior features, overcomes the limitations of a single logical constraint graph, and strictly preserves the natural bit order, thereby breaking through the performance bottleneck of existing technologies in processing highly structured channel codes.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions: Channel decoding methods based on Boolean geometric embedding and attention decoupling include: Construct a channel code based on natural bit order, obtain the received signal after transmission, and preprocess it to obtain the initial observation state vector; An absolute index sequence is introduced, a coordinate tensor is obtained based on the Boolean coordinate transformation rule, features are extracted from the coordinate tensor to obtain Boolean geometric feature embeddings, and a global structured position feature sequence is constructed based on the Boolean geometric feature embeddings. Based on algebraic logic topology and physical geometric topology, logical graph masks and geometric graph masks are constructed. The global structured position feature sequence is fused with the initial observation state vector through deep multimodal fusion to generate the initial input features of the network. The initial input features of the network are input into the Transformer decoding network. In the multi-head self-attention calculation of each layer, the constraints of logical graph masks and geometric graph masks are introduced to implement feature decoupling and dynamic gating mechanisms, extract the final deep abstract feature representation, and perform a global joint decoding operation on the deep abstract feature representation to obtain the predicted decoding output.

[0013] According to some embodiments, the present disclosure adopts the following technical solutions: A channel decoding system based on Boolean geometric embedding and attention decoupling includes: The signal acquisition and preprocessing module is used to construct a channel code based on natural bit order, acquire the received signal after transmission, and obtain the initial observation state vector after preprocessing it. The feature embedding and construction module is used to introduce an absolute index sequence, obtain a coordinate tensor based on Boolean coordinate transformation rules, extract features from the coordinate tensor to obtain Boolean geometric feature embeddings, and construct a global structured position feature sequence based on the Boolean geometric feature embeddings. The mask construction and decoding module is used to construct logical graph masks and geometric graph masks based on algebraic logical topology and physical geometric topology. It performs deep multimodal fusion of the global structured position feature sequence and the initial observation state vector to generate the initial input features of the network. The initial input features of the network are input into the Transformer decoding network. In the multi-head self-attention calculation of each layer, the constraints of logical graph masks and geometric graph masks are introduced to implement feature decoupling and dynamic gating mechanisms to extract the final deep abstract feature representation. The decision output module is used to perform a global joint decoding operation on the deep abstract feature representation to obtain the predicted decoding output result.

[0014] According to some embodiments, the present disclosure adopts the following technical solutions: A computer program product includes a computer program that, when executed by a processor, implements the channel decoding method based on Boolean geometric embedding and attention decoupling.

[0015] According to some embodiments, the present disclosure adopts the following technical solutions: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned channel decoding method based on Boolean geometric embedding and attention decoupling.

[0016] According to some embodiments, the present disclosure adopts the following technical solutions: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the channel decoding method based on Boolean geometric embedding and attention decoupling.

[0017] Compared with the prior art, the beneficial effects of this disclosure are as follows: This disclosure presents a channel decoding method based on Boolean geometric embedding and attention decoupling, which significantly improves error correction performance in complex channel environments and expands the coding gain in the medium-to-high signal-to-noise ratio (SNR) region. By introducing Boolean geometric embedding and geometric graph masking, it successfully compensates for the absolute blind spot in topology awareness of traditional single logical constraint graphs. Performance comparison test results show that, under the same polarization code configuration (such as POLAR and RM codes), this disclosure achieves stable performance superiority over standard ECCT across the entire SNR range. Particularly in the medium-to-high SNR region, this disclosure utilizes the local clustering prior of erroneous bits in the Boolean multidimensional space and, through soft information collaboration with geometric neighbors, successfully captures and corrects complex association errors in deep networks that are difficult to eliminate by pure algebraic logic. This mechanism results in a steeper downward trend in the frame error rate (FER) and bit error rate (BER) curves of this invention (i.e., better waterfall characteristics), achieving significant additional coding gain at the same target bit error rate, fundamentally breaking through the asymptotic performance bottleneck of traditional neural decoders caused by excessive feature homogenization.

[0018] This disclosure presents a channel decoding method based on Boolean geometric embedding and attention decoupling, which significantly enhances the interpretability of deep learning models and constructs an adaptive hierarchical decoding paradigm that conforms to communication priors. It pioneers a multi-view attention head decoupling and dynamic gating fusion mechanism, effectively alleviating the "black-box" limitation of deep learning models in channel decoding applications and endowing the invisible network hidden layers with explicit physical and algebraic meaning. Combined with gating weight evolution analysis, it is evident that the deep model trained in this disclosure exhibits a highly consistent "coarse-to-fine" adaptive decoding pattern that closely aligns with the underlying mathematical logic of communication: in the shallow, early layers of the network, the gating mechanism assigns dominant weights to logic heads, prioritizing the use of global parity-check equations for coarse-grained algebraic hard constraint error correction; as the network deepens, the gating mechanism dynamically and smoothly transfers feature energy to geometric heads, activating higher-order spatial awareness for local soft-smooth collaboration. The empirical data fully demonstrates that the design disclosed in this paper successfully induces the neural network to autonomously learn multi-stage, multi-perspective expert-level decoding strategies, which greatly improves the theoretical completeness and engineering robustness of the model in dealing with complex and sudden noise patterns.

[0019] This disclosure presents a channel decoding method based on Boolean geometric embedding and attention decoupling, which overcomes the topological representation bottleneck with minimal marginal computational overhead, possessing extremely high hardware computational efficiency and deployment value. While introducing a multimodal heterogeneous graph feature extraction and decoupling collaborative mechanism, this disclosure demonstrates outstanding nonlinear gains in model lightweighting and complexity control. Combined with model complexity test data, it is shown that under various mainstream code length and code rate configurations, compared to the standard ECCT model, the total number of network parameters increases by only about 7% to 9%, and the number of floating-point operations per frame (FLOPs per frame), which determines the actual baseband processing latency, increases by only about 4%. This disclosure, through the natural order alignment of the underlying data structure and efficient Boolean coordinate bit operations, perfectly unlocks the ability to extract deep geometric topological features of polarized codes with extremely low marginal computational overhead, achieving a leapfrog improvement in convergence efficiency and error correction accuracy. Considering the logic gate resource consumption of silicon-based hardware and the return on communication reliability, the architecture of this disclosure achieves an extremely high cost-effectiveness ratio (Cost-E). (iciency), which is highly compatible with the engineering evolution requirements of 5G-Advanced and future 6G communication systems for ultra-reliable low-latency communication (URLLC) baseband chips. Attached Figure Description

[0020] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0021] Figure 1 This is a flowchart illustrating the channel decoding method based on Boolean geometric embedding and attention decoupling according to an embodiment of this disclosure. Figure 2 This is a performance comparison chart of POLAR codes with N = 64 and K = 32 according to embodiments of this disclosure; Figure 3 This is a performance comparison chart of POLAR codes with N = 64 and K = 48 according to embodiments of this disclosure; Figure 4 This is a performance comparison chart of POLAR codes with N = 128 and K = 64 according to embodiments of this disclosure; Figure 5 This is a performance comparison chart of POLAR codes with N = 128 and K = 86 according to embodiments of this disclosure; Figure 6 This is a performance comparison chart of N = 64 K = 22 RM codes according to embodiments of this disclosure; Figure 7 This is a performance comparison chart of N = 64 K = 42 RM codes according to embodiments of this disclosure; Figure 8 This is a performance comparison chart of N = 128 K = 64 RM codes according to embodiments of this disclosure; Figure 9 This is a comparison chart of the number of parameters and computational complexity of embodiments of this disclosure; Figure 10 This is a schematic diagram of the alpha (α) values ​​of the POLAR code with N = 64 and K = 32 according to an embodiment of this disclosure; Figure 11 This is a schematic diagram of the alpha values ​​of the POLAR code with N = 64 and K = 48 according to an embodiment of this disclosure; Figure 12 This is a schematic diagram of the alpha values ​​of the N = 128 K = 64 POLAR code according to an embodiment of this disclosure; Figure 13 This is a schematic diagram of the alpha values ​​of the POLAR code with N = 128 and K = 86 according to an embodiment of this disclosure; Figure 14 This is a schematic diagram of the alpha values ​​of the N = 64 K = 22 RM code according to an embodiment of this disclosure; Figure 15 This is a schematic diagram of the alpha values ​​of the N = 64 K = 42 RM code according to an embodiment of this disclosure; Figure 16 This is a schematic diagram of the alpha value of the N = 128 K = 64 RM code according to an embodiment of this disclosure; Figure 17 This is a schematic diagram of the mask matrix of N = 128 K = 64 RM codes according to an embodiment of this disclosure. Detailed Implementation

[0022] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Example 1 One embodiment of this disclosure provides a channel decoding method based on Boolean geometric embedding and attention decoupling, the method steps of which include: Step 1: Construct a channel code based on natural bit order, obtain the received signal after transmission, and preprocess it to obtain the initial observation state vector; Step 2: Introduce an absolute index sequence, obtain a coordinate tensor based on Boolean coordinate transformation rules, extract features from the coordinate tensor to obtain Boolean geometric feature embeddings, and construct a global structured position feature sequence based on the Boolean geometric feature embeddings; Step 3: Construct logic graph masks and geometric graph masks based on algebraic logic topology and physical geometric topology; Step 4: Perform deep multimodal fusion of the global structured location feature sequence and the initial observation state vector to generate the initial input features of the network; input the initial input features of the network into the Transformer decoding network, and introduce constraints of logical graph mask and geometric graph mask in the multi-head self-attention calculation of each layer to implement feature decoupling and dynamic gating mechanism, and extract the final deep abstract feature representation. Step 5: Perform a global joint decoding operation on the deep abstract feature representation to obtain the predicted decoding output.

[0026] As one embodiment, the channel decoding method based on Boolean geometric embedding and attention decoupling disclosed herein addresses the shortcomings of existing Transformer-based deep learning neural decoders when processing polarization-type structured channel codes (such as Reed-Muller codes and Polar codes). These shortcomings stem from the lack of algebraic geometric constraints, resulting in slow convergence, limited feature representation dimensions, and performance bottlenecks in deep networks. The specific implementation process is as follows: Step 1: Constructing a channel code based on natural bit order and preprocessing the signal; Obtain the parameters N (code length) and k (information bit length) of the target channel code (such as Reed-Muller code or Polar code). When constructing the generator matrix and parity check matrix, abandon the traditional Gaussian elimination method which may lead to column permutation operations, and directly perform structured construction based on algebraic geometric constraints to strictly preserve the natural bit order of codeword variable nodes.

[0027] Specifically, for RM codes, the generating matrix constraints are constructed by traversing all monomial combinations with a degree not exceeding r (determined by inversely from the value of k). Variable node indexes. (in It is uniquely mapped to m-dimensional Boolean hypercube coordinates. Its coordinate components The mapping rules are as follows:

[0028] in, j ∈{0, 1, ..., m 1}. For Polar codes, the full kernel matrix GN is generated recursively and constructed using Kronecker product operations:

[0029] By combining channel polarization reliability (such as the Batachalia parameter) sorting, information bits and frozen bits are separated to directly construct the generator matrix and check constraints, ensuring the absolute alignment of the bit sequence with the Boolean geometric space from the underlying data structure.

[0030] Furthermore, the received signal after BPSK modulation and transmission through a noisy channel at the transmitting end is obtained. y ∈RN. For the received signal y Feature extraction preprocessing: Extracting soft observation amplitude sequences from physical channels Simultaneously, algebraic verification features are proposed, firstly mapping the received signal into a binary bit field through hard sign decision. Subsequently, the adjoint expression was calculated in the Galois domain GF(2). Finally, it is remapped back to the bipolar feature space (+1 / 1) Obtain the syndrome sequence s ∈RM, where M=N k represents the number of verification nodes. The amplitude sequence and the adjoint sequence are concatenated along the feature dimension to construct an initial observation state vector containing all nodes in the graph. .

[0031] Step 2: Introduce an absolute index sequence, obtain a coordinate tensor based on Boolean coordinate transformation rules, extract features from the coordinate tensor to obtain Boolean geometric feature embeddings, and construct a global structured position feature sequence based on the Boolean geometric feature embeddings; This step, through a feature extraction network, completely abandons discrete index embedding and transforms discrete physical topological coordinates into continuous high-dimensional spatial features. Specifically, it includes the following processes: (1) Variable node embedding process: The one-dimensional absolute index sequence of the input variable nodes is used to obtain a coordinate tensor based on the Boolean coordinate transformation rule in step 1. This discrete tensor is then input into a nonlinear feature extraction network, sequentially passing through an up-dimensional linear mapping layer (to...). Dimensionality), GELU activation function, first layer LayerNorm, dimension reduction linear mapping layer (to The system uses a second layer (LayerNorm) to output the final geometric embedding features of the variable nodes. That is, Boolean geometric feature embedding.

[0032] (2) Verification node embedding and fusion process: Based on the parity check matrix H, the features of variable nodes connected to the check node are aggregated using the scatter-add operation. After being averaged by dividing by the node degree, the check node features are output through a multilayer perceptron incorporating LayerNorm and ReLU. .

[0033] Furthermore, the spatial centroids of the Boolean coordinates of all variable nodes connected to the verification node are calculated, and a binarization threshold of 0.5 is applied. These centroids are then projected back onto the hypercube vertex set to achieve geometric alignment. The specific implementation process is as follows: 1) Establishing topological connectivity: based on the parity check matrix of the target channel code. (in To verify the total number of nodes, (total number of variable nodes), for the th Verification nodes ( Extracting the matrix No. The column index of an element with a value of 1 in the row is used to determine the set of all variable nodes connected to that check node. And obtain the degree of the verification node. .

[0034] 2) Get the absolute coordinates of variable nodes: Get the set Each variable node exist The Vibull hypercube (in which The corresponding absolute coordinate vector in ) Calculating the barycenter of a multidimensional space: In the real number field, for a set... The arithmetic mean of the coordinate vectors of all variable nodes in the check node is calculated along each dimension to obtain the check node's coordinates. 2D space centroid vector For the first vector... Dimensions ( The calculation formula is as follows: This operation establishes the continuous "pseudo-coordinates" of the verification node in Euclidean space.

[0035] 3) Binarization Projection and Geometric Space Alignment: To ensure that the calculated check node coordinates can share the same Boolean geometric metric space as the variable nodes (for subsequent Hamming distance calculation), the centroid vector is aligned... Apply to each dimension The binarization threshold is used to hard-determine (project) the result back to the vertex set of the Boolean hypercube. Specifically, the verification node... Final geometric alignment coordinates The projection rule is: if Then the mapping ;like Then the mapping .

[0036] Finally, the variable node features Features of verification nodes Concatenate the sequences to generate a globally structured location feature sequence. .

[0037] Step 3: Construct logic graph masks and geometric graph masks based on algebraic logic topology and physical geometric topology; For a global graph structure containing variable nodes and check nodes, this step simultaneously proposes algebraic logical topology and physical geometric topology, constructing two physically orthogonal self-attention masks: (1) Construction of Logic Mask: Parity check constraint based on target channel code (i.e., parity check matrix) Sure Construct a logical graph adjacency matrix. It is initialized as a self-loop identity matrix. For each check node c, it and all variable nodes connected to it are considered as a fully connected subgraph. Any pair of nodes belonging to the same subgraph is set ( i , j The adjacency weight between nodes is 1. In this logical subgraph, all variable nodes involved in the same parity check equation form fully connected cliques to accelerate the transmission of local algebraic messages. This mask characterizes the macroscopic global algebraic check equation system, giving the network the ability to learn hard logic of parity constraints.

[0038] The parity check equation is as follows: For a matrix The OK( If the row is in the ), The element in the column is 1 (i.e. This indicates that... These variable nodes (i.e., the codewords) (each bit) participated in the first A parity check equation. The mathematical expression of this equation is that the modulo-2 addition (i.e., the XOR operation) of these bits must result in 0: ).

[0039] (2) Geometric Mask Construction: Based on the aligned Boolean coordinates established in step 2, calculate the L1 distance (i.e., Hamming distance) between all nodes in the entire graph, and construct the spatial metric matrix. Set the absolute distance threshold τ=1.01, and construct the geometric graph adjacency matrix using an indicator function. This mask-based forced attention mechanism focuses on local algebraic topological dependencies in Boolean space, ensuring high-frequency cooperative interactions between neighboring nodes with a Hamming distance of 1.

[0040] Specifically, in step 2, the following was obtained: A set of Boolean coordinates of variable nodes (in ),as well as The set of alignment coordinates obtained by centroid calculation and binarization projection of each verification node (in Aligned Boolean coordinates refer to unifying the coordinates of these two types of heterogeneous nodes to a single coordinate system. In a 3D geometric space, let the total number of nodes in the entire graph be . Define a globally aligned Boolean coordinate matrix. This is a concatenation of the variable node coordinates and the check node coordinates along the node dimension. (Concatenation):

[0041] For any i in the entire graph Nodes ( The corresponding row vector in this matrix is ​​denoted as . This representation establishes that all nodes in the entire graph are in the same... Absolute position within a 3D hypercube.

[0042] Furthermore, the mathematical representation of the spatial metric matrix and the indicator function is as follows: Based on the above aligned Boolean coordinate matrix Calculate any two nodes in the entire graph and Between Distance (equivalent to Hamming distance in Boolean space), constructing a spatial metric matrix. , its first The formula for calculating each element is: Set absolute distance threshold (This threshold ensures that only Hamming distances are equal to 0.) or (The node pairs are activated).

[0043] Indicator Function The standard mathematical definition of graph connectivity weights used to convert continuous or discrete distance scalars into binary values ​​is:

[0044] By applying the indicator function element-by-element to the spatial metric matrix This allows the construction of the geometric graph adjacency matrix. The indicator function operation is topologically equivalent to extracting a local neighborhood of radius 1 in the Boolean hypercube (i.e., retaining only its own edges and first-order nearest neighbor edges), thereby forcing the attention mechanism to focus on local algebraic topological dependencies in Boolean space, ensuring that neighboring nodes with a Hamming distance of 1 can engage in high-frequency cooperative interactions.

[0045] Finally, the adjacency matrix of the logical graph is... Adjacency matrix of geometric graph Convert to a Boolean mask tensor and use Boolean values ​​to effectively block the calculation of invalid attention scores between unrelated nodes.

[0046] Step 4: Perform deep multimodal fusion of the global structured location feature sequence and the initial observation state vector to generate the initial input features of the network. Input the initial input features of the network into the Transformer decoding network. In the multi-head self-attention calculation of each layer, introduce the constraints of the logical graph mask and the geometric graph mask to implement the feature decoupling and dynamic gating mechanism and extract the final deep abstract feature representation. Specifically, firstly, the global structured location feature sequence P With the initial observed state vector Performing deep multimodal fusion includes: The initial observation state vector is obtained through linear mapping. Project to dmodel Dimension, and global structured location feature sequence P After concatenation along the feature dimension, the input is fed into a feature fusion network (FusionLayer) consisting of linear layers, LayerNorm, and ReLU for dimensionality reduction and nonlinear mapping, generating the initial input features for the network. .

[0047] Furthermore, the initial input features are fed into a Transformer decoding network containing an L-layer encoder. In the multi-head attention computation of each layer, feature decoupling and dynamic gating mechanisms are implemented: (1) Multi-view attention head decoupling mechanism: The total number is hAttention heads are divided into "Logic Heads" and "Geometric Heads." This is achieved by constructing a combined mask tensor. The first half of the logic header is forced to be constrained only by the logic graph mask, and is dedicated to extracting global parity correlation; the second half of the geometry header is forced to be constrained only by the geometry graph mask, and is dedicated to mining soft features of local Boolean neighborhoods.

[0048] Specifically, the mask polarity definition and single-image mask generation process include: In standard self-attention mechanisms, a mask is used to block the flow of information between unrelated nodes. The polarity of a Boolean mask is defined as follows: (or False) indicates that information exchange is allowed. (or True) indicates blocking information exchange. Based on the constructed logical graph adjacency matrix. Adjacency matrix of geometric graph (in (where the total number of nodes in the graph is 1). First, take the logical NOT operation on it to generate the corresponding two-dimensional mask matrix: Logical mask matrix (i.e., positions without edges in the original adjacency matrix are set to 1, and positions with edges are set to 0):

[0049] Geometric mask matrix:

[0050] Furthermore, the block construction process of the Combined Mask Tensor is as follows: Let the total number of heads in the multi-head attention mechanism be . (For even numbers), to achieve deep decoupling of multi-view feature extraction, a global combined mask tensor spanning all attention heads is constructed. The construction of this tensor strictly follows the block concatenation rule based on the head index: a. Allocate the logical header subspace (previous) Size): For tensors The first in the head dimension slices (of which) ), forcibly applying the logical mask matrix Copy and assign to the slice, that is .

[0051] b. Allocate the geometric head subspace (after) Size): For tensors The first in the head dimension slices (of which) ), forcibly applying the geometric mask matrix Copy and assign to the slice, that is .

[0052] (2) Energy-conserving adaptive gating fusion: Each network layer is configured with learnable dynamic gating parameters αl∈[0, 1] (activated by Sigmoid). To eliminate the risk of gradient vanishing in deep networks caused by decoupling truncation, a dynamic gating vector g∈Rh is constructed: a weighting coefficient 2αl is assigned to the logical head, and a weighting coefficient 2(1) is assigned to the geometric head. αl). The output tensor of each attention head is scaled and weighted element-wise by the gate vector, and then updated by the feedforward neural network to output the feature Xl+1 of this layer.

[0053] Specifically, in the first layer( In the Transformer decoding network, an independent learnable scalar parameter is configured. First, the Sigmoid activation function is used to map it to dynamic gating parameters.

[0054]

[0055] To mitigate the risk of gradient vanishing in deep networks that may be caused by multi-head decoupling and truncation, this disclosure proposes an "energy-conserving" weighted strategy. The construction dimension is... A one-dimensional dynamic gating vector (total number of attention heads) , its first The assignment rules for each element are as follows:

[0056] This design ensures that the mathematical expectation of all attention head weights remains strictly 1 (i.e., While changing the feature distribution, it maintains the physical conservation of the total feature energy flowing into the next layer of the network, thus ensuring the stable convergence of deep networks.

[0057] After processing by an L-layer decoupled gated Transformer, the final deep abstract feature representation is extracted. .

[0058] Step 5: Perform a global joint decoding operation on the deep abstract feature representation to obtain the predicted decoding output.

[0059] Specifically, for deep abstract feature representation Perform a global joint decoding operation: (1) Global layer normalization: for nodes containing all nodes in the graph Perform LayerNorm normalization to smooth the statistical variance of node features; (2) Feature Dimension Compression: A shared linear projection layer is used to compress the feature dimension of each node. dmodel The abstract features are compressed into a one-dimensional scalar, and redundant representation dimensions are removed to obtain a global sequence representing the soft information of the entire graph. ; (3) Global joint remapping: A fully connected classification mapping layer is fed into the entire graph topology. This layer fully utilizes the adjoint constraint information accumulated by the verification nodes to assist the final decision of the variable nodes, directly reducing the dimensionality of the N + M sequence to a core target prediction sequence of length N. .

[0060] Finally, physical layer decisions from soft input to hard output are performed using symbolic functions:

[0061] During the end-to-end training phase, the binary cross-entropy loss (BCEwithLogitsLoss) between the predicted sequence and the actual transmitted codewords is used as the objective function, and the backpropagation algorithm is used to perform global parameter optimization for the entire decoupled gating network.

[0062] Theoretical proof The effectiveness of the Boolean hypercube sensing mechanism employed in this disclosure is rigorously demonstrated from a theoretical perspective. Through proofs in two dimensions—the overlap of local parity check equations (from an algebraic geometric perspective) and the shortest path of the Boolean hypercube (from a metric space perspective)—the following core conclusion is established: the correlation of error correction information between two code points is their binary Hamming distance d. H ( i , j It is a monotonically decreasing function.

[0063] 1. Symbol and Background Definition In the encoding system of Polar Codes and Reed-Muller (RM) codes, let the code length be N = 2^n. Define the indices of any two code points (physical bits or polarization channel bits) as i,j ∈ {0, 1, ..., N}. 1}.

[0064] Definition 1 (Binary Hamming Distance between Code Points). Define the code point index. i and j of n The binary representations are as follows: and The binary Hamming distance between two code points ( Binary Hamming Distance ) is defined as:

[0065] 2. Algebraic proof based on the overlap of local parity check equations In belief propagation (BP) and neural network-based message passing decoding, the strength of information exchange between code points essentially depends on the number of local parity check equations in which they participate.

[0066] Let RM(r,n) be an RM code of length N=2n (a polar code can be considered as its sub-code under a specific channel polarity). According to the dual code property, the weight of the minimum weight parity check equation for RM(r,n) is 2r+1. Geometrically, each such parity check equation corresponds exactly to an (r+1)-dimensional affine subcube (A) in F2n. ineSubcube).

[0067] Theorem 1 (Code Point Check Overlap Theorem) states that for any two code points... i , j Let the binary Hamming distance between ∈F2n be d = dH( i , j Then it also contains code points. i and j The number of minimum weight parity check equations Np ( i , j The following combination formulas are satisfied:

[0068] prove . To make a (r + 1)-dimensional subcube contain vertices i and j The sub-cube must be i and j It spans across all dimensions with different coordinates. Because i and j There is d The sub-cube must contain these dimensions differently. d Each dimension.

[0069] If d > r + 1, then there is no (r + 1)-dimensional subcube that can simultaneously accommodate both points, so the number of check equations containing both is 0.

[0070] If d ≤ r + 1, the sub-cube needs to freely choose (r + 1 - d) dimensions from the remaining (n - d) dimensions to form a complete (r + 1) dimensional space. According to the principles of combinatorics, the number of freely chosen solutions is... .

[0071] Corollary 1. For a non-trivial RM code (satisfying the order $r < n - 1$), the combination number function is a strictly monotonically decreasing function with respect to the Hamming distance d . That is, the smaller the Hamming distance $d_H( i , j )$, the number of parity-check equations shared by the code positions i and j increases in a combinatorial series, and the information interaction and correlation between the two in the decoding process are stronger.

[0072] When using graph neural network models such as the Self-Attention mechanism for channel decoding, the ability of the network to capture node relationships is restricted by the topological shortest path.

[0073] Theorem 2 (Equivalence of the shortest path of hypercube message passing). Polar The physical connectivity of the RM code and the RM code forms a Boolean hypercube graph $H_n=(V, E)$, where the vertex set $V = F_2^n$, and the edge set $E$ is composed of pairs of points with a Hamming distance of 1 . For any two bits i , j on this topological graph, the shortest path geodesic distance ( GeodesicDistance ) is strictly equal to their binary Hamming distance $d_H(i,j)$.

[0074] Proof . In the $H_n$ graph, when information is transmitted along an edge by one step, only 1 bit in the binary index can be flipped. Since the code positions $i$ and $j$ differ in $d_H(i,j)$ bits, at least $d_H(i,j)$ state flips are required to reach vertex $j$ from vertex $i$. Therefore, the graph-theoretic shortest path $D_{H_n}(i,j)\equiv d_H(i,j)$.

[0075] In summary, whether from the parity-check overlap in algebraic geometry or from the shortest path distance of the graph topology, the Hamming distance $d$ H (i,j) of the binary index of the code positions perfectly quantifies the error-correcting physical correlation between the code positions. Introducing this metric into the attention mask or position encoding of the neural network can provide the optimal inductive bias for the model that conforms to the mechanism of channel coding.

[0076] Example 2 In an embodiment of the present disclosure, a channel decoding system based on Boolean geometric embedding and attention decoupling is provided, including: The signal acquisition and preprocessing module is used to construct a channel code based on natural bit order, acquire the received signal after transmission, and obtain the initial observation state vector after preprocessing it. The feature embedding and construction module is used to introduce an absolute index sequence, obtain a coordinate tensor based on Boolean coordinate transformation rules, extract features from the coordinate tensor to obtain Boolean geometric feature embeddings, and construct a global structured position feature sequence based on the Boolean geometric feature embeddings. The mask construction and decoding module is used to construct logical graph masks and geometric graph masks based on algebraic logical topology and physical geometric topology. It performs deep multimodal fusion of the global structured position feature sequence and the initial observation state vector to generate the initial input features of the network. The initial input features of the network are input into the Transformer decoding network. In the multi-head self-attention calculation of each layer, the constraints of logical graph masks and geometric graph masks are introduced to implement feature decoupling and dynamic gating mechanisms to extract the final deep abstract feature representation. The decision output module is used to perform a global joint decoding operation on the deep abstract feature representation to obtain the predicted decoding output result.

[0077] Example 3 One embodiment of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the channel decoding method based on Boolean geometric embedding and attention decoupling.

[0078] Example 4 One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the channel decoding method based on Boolean geometric embedding and attention decoupling.

[0079] Example 5 One embodiment of this disclosure provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the channel decoding method based on Boolean geometric embedding and attention decoupling.

[0080] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A channel decoding method based on Boolean geometric embedding and attention decoupling, characterized in that, include: Construct a channel code based on natural bit order, obtain the received signal after transmission, and preprocess it to obtain the initial observation state vector; An absolute index sequence is introduced, a coordinate tensor is obtained based on the Boolean coordinate transformation rule, features are extracted from the coordinate tensor to obtain Boolean geometric feature embeddings, and a global structured position feature sequence is constructed based on the Boolean geometric feature embeddings. Based on algebraic logic topology and physical geometric topology, logical graph masks and geometric graph masks are constructed. The global structured position feature sequence is fused with the initial observation state vector through deep multimodal fusion to generate the initial input features of the network. The initial input features of the network are input into the Transformer decoding network. In the multi-head self-attention calculation of each layer, the constraints of logical graph masks and geometric graph masks are introduced to implement feature decoupling and dynamic gating mechanisms, extract the final deep abstract feature representation, and perform a global joint decoding operation on the deep abstract feature representation to obtain the predicted decoding output.

2. The channel decoding method based on Boolean geometric embedding and attention decoupling as described in claim 1, characterized in that, The process of constructing a channel code based on natural bit order, obtaining the received signal after transmission, and preprocessing it to obtain an initial observation state vector includes: Obtain the parameters of the target channel code and construct the generator matrix and parity check matrix; Acquire the received signal after BPSK modulation at the transmitting end and transmission through a noisy channel; The soft observation amplitude sequence of the physical channel is extracted from the received signal, and the algebraic check feature is extracted at the same time. First, the received signal is mapped to the binary bit field by symbolic hard decision, then the syndrome is calculated in the Galois field, and finally it is remapped back to the bipolar feature space to obtain the syndrome sequence. The amplitude sequence and the syndrome sequence are concatenated in the feature dimension to construct the initial observation state vector containing all nodes in the graph.

3. The channel decoding method based on Boolean geometric embedding and attention decoupling as described in claim 1, characterized in that, The process involves introducing an absolute index sequence, obtaining a coordinate tensor based on Boolean coordinate transformation rules, extracting features from the coordinate tensor to obtain Boolean geometric feature embeddings, and constructing a global structured position feature sequence based on the Boolean geometric feature embeddings, including: The coordinate tensor is obtained from the one-dimensional absolute index sequence of the input variable nodes based on the Boolean coordinate transformation rule. The coordinate tensor is input into a nonlinear feature extraction network, which passes through a dimension-upgrading linear mapping layer, a GELU activation function, a first LayerNorm layer, a dimension-reducing linear mapping layer, and a second LayerNorm layer in sequence, and outputs the final variable node geometric embedding features, resulting in Boolean geometric feature embeddings. Based on the parity check matrix, the features of variable nodes connected to the check node are aggregated by scattering accumulation operation, and after being averaged by the node degree, the check node features are output through a multilayer perceptron containing LayerNorm and ReLU. Calculate the spatial centroid of the Boolean coordinates of all variable nodes connected to the verification node, apply a binarization threshold of 0.5, and project them back to the hypercube vertex set to achieve geometric alignment. Finally, the Boolean geometric features are embedded and the verification node features are concatenated along the sequence dimension to generate a globally structured position feature sequence.

4. The channel decoding method based on Boolean geometric embedding and attention decoupling as described in claim 1, characterized in that, Constructing logical graph masks and geometric graph masks based on algebraic logic topology and physical geometric topology, including: Based on the parity check constraint of the target channel code, a logical graph adjacency matrix is ​​constructed and initialized as a self-loop identity matrix. For each check node, it and all the variable nodes connected to it are regarded as a fully connected subgraph. Based on the established aligned Boolean coordinates, calculate the Hamming distance between all nodes in the entire graph, construct a spatial metric matrix, set an absolute distance threshold, and construct a geometric graph adjacency matrix using an indicator function; Convert the adjacency matrices of the logical graph and the geometric graph into Boolean mask tensors to obtain the logical graph mask and the geometric graph mask.

5. The channel decoding method based on Boolean geometric embedding and attention decoupling as described in claim 1, characterized in that, The step of performing deep multimodal fusion of the global structured location feature sequence and the initial observation state vector to generate the initial input features of the network includes: The initial observed state vector is projected onto d through a linear mapping. model The feature dimension is concatenated with the global structured location feature sequence along the feature dimension and then input into a feature fusion network consisting of linear layers, LayerNorm, and ReLU for dimensionality reduction and nonlinear mapping to generate the initial input features of the network.

6. The channel decoding method based on Boolean geometric embedding and attention decoupling as described in claim 1, characterized in that, The process involves inputting the initial input features into the Transformer decoding network, and in the multi-head self-attention computation of each layer, introducing constraints from logistic graph masks and geometric graph masks to implement feature decoupling and dynamic gating mechanisms, thereby extracting the final deep abstract feature representation, including: The initial input features of the network are input into a Transformer decoding network containing an L-layer encoder. The total number of attention heads h are divided into logical heads and geometric heads. By constructing a combined mask tensor, the first half of the logical heads is forced to be constrained only by the logical graph mask to extract global parity correlation; the second half of the geometric heads is forced to be constrained only by the geometric graph mask to mine soft features of local Boolean neighborhoods. Each layer of the network is configured with learnable dynamic gating parameters, and a dynamic gating vector is constructed. Different weighting coefficients are assigned to the logical head and the geometric head. The output tensor of each attention head is scaled and weighted element by element by the gating vector, and then updated by the feedforward neural network to output the features of this layer. After being processed by an L-layer decoupled gating Transformer, the final deep abstract feature representation is extracted.

7. A channel decoding system based on Boolean geometric embedding and attention decoupling, characterized in that, include: The signal acquisition and preprocessing module is used to construct a channel code based on natural bit order, acquire the received signal after transmission, and obtain the initial observation state vector after preprocessing it. The feature embedding and construction module is used to introduce an absolute index sequence, obtain a coordinate tensor based on Boolean coordinate transformation rules, extract features from the coordinate tensor to obtain Boolean geometric feature embeddings, and construct a global structured position feature sequence based on the Boolean geometric feature embeddings. The mask construction and decoding module is used to construct logical graph masks and geometric graph masks based on algebraic logical topology and physical geometric topology. It performs deep multimodal fusion of the global structured position feature sequence and the initial observation state vector to generate the initial input features of the network. The initial input features of the network are input into the Transformer decoding network. In the multi-head self-attention calculation of each layer, the constraints of logical graph masks and geometric graph masks are introduced to implement feature decoupling and dynamic gating mechanisms to extract the final deep abstract feature representation. The decision output module is used to perform a global joint decoding operation on the deep abstract feature representation to obtain the predicted decoding output result.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the channel decoding method based on Boolean geometric embedding and attention decoupling as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the channel decoding method based on Boolean geometric embedding and attention decoupling as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the channel decoding method based on Boolean geometric embedding and attention decoupling as described in any one of claims 1-6.

Images