Channel coding type blind recognition method and system based on double-channel convolutional neural network

Abstract

The application discloses a channel coding type blind identification method and system based on a double-channel convolutional neural network. The method is as follows: a communication system model is established, a heterogeneous feature fusion architecture is adopted, and a double-channel convolutional neural network is constructed; a multi-scale dilated convolution layer is used in a sequence semantic branch to extract long-range spatial features of a coding structure, and an algebraic feature branch captures algebraic invariant properties of the coding based on extracted multi-dimensional artificial features; after being spliced, the sequence semantic branch and the algebraic feature branch output are input into a full connection layer, and overfitting is inhibited through Dropout; and the probability distribution of seven types of codes is output through a Softmax classification layer. The system comprises six modules of model establishment, neural network construction, sequence semantic branch, algebraic feature branch, splicing and classification, and is used for realizing the channel coding type blind identification method. The application has the advantages of high channel coding type identification efficiency, strong generalization ability, wide application range and high identification precision.

Description

Technical Field

[0001] This invention relates to the field of channel coding technology, and in particular to a blind identification method and system for channel coding types based on a dual-channel convolutional neural network. Background Technology

[0002] Channel coding techniques enhance communication reliability by introducing structured redundancy and are widely used in modern communication systems. In non-cooperative domains, the receiver is unaware of the coding scheme, necessitating blind channel coding identification techniques to obtain the current scheme. Given that practical systems require differentiated channel coding schemes to adapt to different transmission conditions, channel coding type identification becomes a crucial prerequisite for coding parameter identification.

[0003] Channel coding can be divided into three main categories: linear block codes, convolutional codes, and Turbo codes. Among them, linear block codes include subcategories such as Hamming codes, BCH codes, LDPC codes, RS codes, and Polar codes. Current research has yielded relatively mature parameter estimation techniques for specific codes such as linear block codes, BCH codes, RS codes, LDPC codes, Turbo codes, and convolutional codes, while research on blind identification of coding types is relatively scarce.

[0004] In recent years, significant progress has been made in deep learning-based channel coding type and parameter identification techniques. Compared to traditional methods, deep learning exhibits stronger nonlinear feature extraction capabilities, effectively handling complex communication data. However, most existing studies only extract artificial or overall features of codewords, achieving only coarse-grained distinction between two or three specific coding types, lacking the ability to finely identify and classify multiple coding types. Furthermore, the few algorithms available for multi-coding type identification still require further improvement in accuracy.

[0005] Currently, methods for channel coding type identification are mainly divided into two categories: traditional methods and deep learning methods. Traditional methods have many limitations, such as high computational complexity, poor adaptability to coding types, and difficulty in utilizing big data. In contrast, deep learning's powerful nonlinear feature extraction capabilities can effectively handle complex coding patterns, providing a new approach to coding type identification.

[0006] Currently, some studies have considered the similarity between code type recognition and text classification, and have introduced text classification CNN networks. For example, Qin et al. mapped the encoded sequence to word vectors and used TextCNN to achieve type recognition; based on this, Li et al. transformed the recognition problem into a natural language processing (NLP) task based on the TextCNN embedding block mechanism, and completed the classification of linear block codes, convolutional codes, and Turbo codes. However, TextCNN is difficult to extend to multi-class code joint recognition scenarios due to model capacity limitations.

[0007] Meanwhile, hybrid methods based on multidimensional feature extraction are gradually being widely used in coding type recognition. Peng et al. proposed a joint recognition architecture that integrates multidimensional feature extraction and efficient channel attention mechanism (ECA-Net), and realizes the synchronous recognition of coding type, code length and code rate through data randomness feature analysis and deep convolutional neural network. Pengcheng Laboratory

[27] proposed a finite field Fourier transform (GFFT) feature extraction method, and constructed an Euclidean distance classifier by combining artificial features such as code weight / entropy value / run length. Although the existing multidimensional feature extraction algorithms can capture the essential differences in coding, they have problems such as feature dimension explosion and generalization decay under high bit error rate.

[0008] Furthermore, to address the need for multi-scale perception, LIU et al. designed a multi-scale dilated convolutional network (MSDCNN), which expands the receptive field without increasing the number of parameters by inserting parallel dilated convolutional kernels, thereby improving recognition accuracy and achieving differentiation of seven common coding types. However, its dilated convolutional kernel design is not adapted to the algebraic periodicity of linear block codes, resulting in room for improvement in the recognition accuracy and discriminative power of LDPC codes / Polar codes / RS codes.

[0009] In summary, deep learning-based channel coding type identification has attracted attention as a novel approach. However, existing methods have not yet achieved a balance between the three objectives of model generalization, feature discriminativeness, and computational efficiency. There is an urgent need to study an innovative architecture that combines algebraic structures and deep learning to improve the accuracy of channel coding type identification. Summary of the Invention

[0010] The purpose of this invention is to provide a blind identification method and system for channel coding types based on a dual-channel convolutional neural network, which has high computational efficiency, high recognition accuracy, and strong generalization ability.

[0011] The technical solution to achieve the purpose of this invention is: a blind identification method for channel coding types based on a dual-channel convolutional neural network, comprising the following steps:

[0012] Step 1: Establish a communication system model;

[0013] Step 2: Adopt a heterogeneous feature fusion architecture to construct a dual-channel convolutional neural network DBFCNN, which includes two parallel processing branches: a sequence semantic branch and an algebraic feature branch.

[0014] Step 3: Sequence semantic branch uses multi-scale dilated convolutional layers to extract long-range spatial features of the coding structure;

[0015] Step 4: Algebraic feature branching captures the algebraic invariant properties of the encoded multidimensional artificially constructed features.

[0016] Step 5: The outputs of the sequence semantic branch and the algebraic feature branch are concatenated and then input into the fully connected layer, and the overfitting is suppressed by the Dropout layer;

[0017] Step 6: Output the probability distribution of the seven types of codes through the Softmax classification layer to complete the blind identification of the channel coding type.

[0018] Furthermore, the establishment of the communication system model described in step 1 is as follows:

[0019] Step 1.1: Due to noise, fading, and loss during transmission, the binary bit sequence w obtained after demodulation preprocessing of the intercepted sequence r is considered an error-prone version of the channel-coded sequence c from the transmitting end, equivalent to... Where e represents the error vector. Represents bitwise XOR; encoding type θ is unknown, belonging to the candidate set Θ={θ Ham ,θ Conv ,θ Turbo ,θ BCH ,θ LDPC ,θ RS ,θ Polar}, that is, the candidate set Θ includes the Hamming code θ Ham Convolutional code θ Conv Turbo code θ Turbo BCH code θ BCH LDPC code θ LDPC RS code θ RS Polar code θ Polar The encoding type determines the structural characteristics of sequence w.

[0020] Blind encoding type identification is the process of solving the mapping based on the structure of sequence w in the context of errors. L represents the length of the encoded sequence, and the candidate set of encoding types for sequence c is obtained. The decision is based on the essential difference in codeword space:

[0021]

[0022] in, This is a typical codeword space for θ-class encoding. Let P be the desired mapping, and let P be the probability that the discriminant sequence w belongs to a certain type of encoding in the candidate set.

[0023] Step 1.2: Suppose the bit error rate of sequence w is p compared to that of sequence c. Under the condition that p is unknown, establish a composite hypothesis testing model:

[0024]

[0025] Where, θ k This is the kth encoding type;

[0026] Decision-making uses generalized likelihood ratio:

[0027]

[0028] in, Let `sup(·)` be the likelihood function, and `sup(·)` be the function that takes the upper limit of the likelihood function. To estimate the results, Given a fixed p, the sequence w belongs to type θ. k The probability of;

[0029] Step 1.3: Since directly calculating the likelihood function involves high-dimensional summation and depends on the unknown p, we introduce encoding structure invariants. Define the original bitstream features mapped to a low-dimensional feature space. Satisfying the p invariance condition:

[0030]

[0031] Among them, D KL This indicates that, under the same bit error rate p, two different codes θ i and θ j The difference between the probability distributions of the generated feature f; These represent the fixed encoding type θ. i θ j The probability of feature f extracted under the given conditions; i,j traverse all encoding types that need to be distinguished;

[0032] Parsing feature mapping for each encoding type It is robust to p, and learns the feature mapping of the encoding from the encoding structure through a deep learning model, and then distinguishes different encoding types based on the features.

[0033] Furthermore, the sequence semantic branch and algebraic feature branch mentioned in step 2 are as follows:

[0034] The sequence semantic branch takes an L / g-dimensional integer sequence as input, which is converted from the original bitstream by g-bit grouping. L represents the length of the encoded sequence. It is mapped to a 128-dimensional semantic space through an embedding layer. Features are extracted through a multi-scale dilated convolutional layer with an inflation rate of {1,2,5}. Finally, a 128-dimensional feature vector is generated through global max pooling to expand the receptive field and adapt to the spatial dependencies of different encoding structures.

[0035] The algebraic feature branch takes multidimensional artificially constructed features as input, including run count, entropy, autocorrelation, coding depth, block features, linear features, and spectral features. It performs nonlinear transformation through a fully connected layer and introduces batch normalization and Dropout layers to enhance generalization.

[0036] Furthermore, the sequence semantic branch described in step 3 uses a multi-scale dilated convolutional layer to extract long-range spatial features of the coding structure, as detailed below:

[0037] Step 3.1: Addressing the long-range structural characteristics of channel coding, deep representation learning of the coded sequence is achieved through multi-scale feature extraction. The input layer is the received binary bit sequence w∈{0,1}. L Grouping by g bits and converting to an integer sequence, where L represents the sequence length, generates an L / g-dimensional discrete input vector. The range is [0, 2]. g -1]:

[0038]

[0039] in, w represents an L / g-dimensional integer vector. gτ+t It is the value of the t-th bit in the τ-th group of the original bit sequence, where τ is the output integer sequence x. int Index;

[0040] Step 3.2: The embedding mapping layer uses a trainable embedding matrix. Mapping integers to dense vectors Where the embedding dimension d e =256 was determined through grid search optimization, and the output tensor was... Preserve the temporal structure of the sequence. These represent the real parameter matrices of 256×256, 256, and (L / g)×256 dimensions respectively, generated by training the neural network.

[0041] Step 3.3: After the embedding mapping layer, a multi-scale convolutional layer is connected, which contains three parallel branches with convolutional kernel sizes of 3, 5 and 7 respectively. The other hyperparameters remain the same. Each branch is followed by a set of multi-dilation convolutional layers.

[0042] Step 3.4: The multi-dilation rate convolutional layer uses a parallel three-branch structure to capture features of different receptive fields. Multi-scale feature adaptation is achieved through differentiated dilation rates {1,2,5}. The m=1 branch extracts the local linear constraints of the block code, the m=2 branch captures the grid state transition of the convolutional code, and the m=5 branch models the long-range correlation of the interleaver of the Turbo code. The convolutional kernel parameters of each branch are shared, only the dilation rate is different.

[0043] Step 3.5: A left-filling strategy is adopted to ensure temporal invariance, and the GELU function is used as the activation function to provide a smooth nonlinear mapping;

[0044] Step 3.6: Use global max pooling to extract the strongest activation feature for each channel in the output of each branch:

[0045]

[0046] H m This represents the output feature map of the m-th dilated convolution branch; This represents a one-dimensional convolution operation, where the subscripts k = 3, 5, 7 indicate the size of the convolution kernel, and the superscript d = m indicates the dilation rate m.

[0047] Step 3.7: The three features are concatenated along the channel dimension to generate a 384-dimensional semantic feature vector. Each subspace corresponds to a structural pattern at a different scale; v (1) ,v (2) ,v (5) The features are represented by inflation rates of 1, 2, and 5, respectively. Then, global max pooling is performed on the entire feature.

[0048] Step 3.8: The output features of the three multi-scale branches are concatenated with the residual path to obtain the output features of the semantic channel of the encoded sequence.

[0049] Furthermore, the algebraic feature branch described in step 4, based on the extracted multidimensional artificially constructed features, captures the algebraically invariant properties of the encoding, as detailed below:

[0050] Step 4.1: The multidimensional feature extraction module receives the preprocessed sequence as input and generates a high-dimensional discriminative feature tensor through feature space mapping. This tensor is formally represented as:

[0051]

[0052] Where T(·) is the feature transformation operator, D is the feature dimension, and X is the feature transformation operator. pre For the preprocessed sequence, F multi The extracted multidimensional features;

[0053] Step 4.2: Output the feature tensor F multi As input to the downstream neural network, the GELU activation function is used, and feature dimensionality enhancement and nonlinear mapping are achieved through fully connected layers:

[0054]

[0055] Among them, the weight matrix After Xavier initialization, bias

[0056] Step 4.3: Use batch normalization layer and 50% Dropout to enhance generalization.

[0057] Furthermore, the outputs of the sequence semantic branch and the algebraic feature branch described in step 5 are concatenated and then input into a fully connected layer. Overfitting is suppressed by a Dropout layer, as detailed below:

[0058] Step 5.1: Concatenate the multi-scale features extracted from the sequence semantic branch with the manual statistical features extracted from the algebraic feature branch to form a comprehensive feature vector;

[0059] Step 5.2: Dynamically calculate feature weights through a gating mechanism, and adaptively weight the fused features to enhance important information;

[0060] Step 5.3: Feed the weighted features into a multi-layer classifier containing fully connected layers, GELU activation function and Dropout layer, and suppress overfitting through regularization.

[0061] Furthermore, the probability distribution of the seven types of codes in step 6 is specifically the probability vector of each sample belonging to a different code type.

[0062] A blind channel coding type identification system based on a dual-channel convolutional neural network is disclosed. This system implements the aforementioned blind channel coding type identification method based on a dual-channel convolutional neural network. The system includes a model building module, a neural network construction module, a sequence semantic branching module, an algebraic feature branching module, a concatenation module, and a classification module, wherein:

[0063] The model building module is used to build a communication system model.

[0064] The neural network construction module adopts a heterogeneous feature fusion architecture to build a dual-channel convolutional neural network DBFCNN, which includes two parallel processing modules: a sequence semantic branch and an algebraic feature branch.

[0065] The sequence semantic branching module uses multi-scale dilated convolutional layers to extract long-range spatial features of the coding structure;

[0066] The algebraic feature branching module captures the algebraic invariant properties of the encoded data based on extracted multidimensional artificially constructed features.

[0067] The outputs of the concatenation module, the sequence semantic branch module, and the algebraic feature branch module are concatenated and then input into the fully connected layer, where overfitting is suppressed by the Dropout layer.

[0068] The classification module outputs the probability distribution of seven classes of codes through the Softmax classification layer.

[0069] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned blind identification method for channel coding types based on a dual-channel convolutional neural network.

[0070] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the blind channel coding type identification method based on a dual-channel convolutional neural network.

[0071] Compared with the prior art, the significant advantages of this invention are:

[0072] (1) Based on the embedding block mechanism of TextCNN, a strategy combining multi-scale convolution kernels (k=3,5,7) and differential dilation rate (m=1,2,5) is adopted to significantly expand the receptive field while maintaining a low number of parameters. Multi-scale convolution expands the receptive field without increasing the number of parameters, and multi-dilation rate convolution enables the network to capture dense local features and sparse global features at the same time by changing the sampling granularity rather than simply expanding the window.

[0073] (2) A network structure combining data-driven feature extraction and manual feature extraction is proposed to provide complementary information for the identification of different channel coding types. This improves the problem of poor robustness of the data-driven network under high bit error rate and poor accuracy of the manual feature extraction method for type identification, and better achieves coding type differentiation.

[0074] (3) A dual-channel convolutional neural network is used to extract the structured features and overall features of codewords based on the differences in different channel coding types. Then, the two features are fused to further analyze the feature information between codewords, which effectively improves the recognition accuracy of multiple channel coding types. Attached Figure Description

[0075] Figure 1 This is a flowchart of the blind channel coding type identification method based on a dual-channel convolutional neural network according to the present invention.

[0076] Figure 2 This is a schematic diagram of the structure of the coding recognition system model established in this invention.

[0077] Figure 3 This is a schematic diagram of the sequence semantic branch structure in this invention.

[0078] Figure 4 This is a schematic diagram of the algebraic characteristic branches in this invention.

[0079] Figure 5 This is a performance comparison curve between DBFCNN and other models in this embodiment of the invention.

[0080] Figure 6 This is a graph showing the recognition performance of different codeword lengths under DBFCNN in an embodiment of the present invention.

[0081] Figure 7This is a graph showing the recognition performance of different architectures under DBFCNN in an embodiment of the present invention.

[0082] Figure 8 This is a graph showing the recognition performance of different embedding dimensions under DBFCNN in an embodiment of the present invention.

[0083] Figure 9 This is a graph showing the recognition of different embedding dimensions in a single channel according to an embodiment of the present invention.

[0084] Figure 10 This is a graph showing the recognition performance of different types of channel coding under DBFCNN in an embodiment of the present invention.

[0085] Figure 11 Example 10 of the present invention -1 A schematic diagram of the confusion matrix under bit error rate.

[0086] Figure 12 Example 10 of the present invention -3 A schematic diagram of the confusion matrix under bit error rate.

[0087] Figure 13 Example 10 of the present invention -5 A schematic diagram of the confusion matrix under bit error rate. Detailed Implementation

[0088] like Figure 1 As shown, the present invention provides a blind method for channel coding type identification based on a dual-channel convolutional neural network, comprising the following steps:

[0089] Step 1: Establish a communication system model;

[0090] Step 2: Adopt a heterogeneous feature fusion architecture to construct a dual-channel convolutional neural network DBFCNN, which includes two parallel processing branches: a sequence semantic branch and an algebraic feature branch.

[0091] Step 3: Sequence semantic branch uses multi-scale dilated convolutional layers to extract long-range spatial features of the coding structure;

[0092] Step 4: Algebraic feature branching captures the algebraic invariant properties of the encoded multidimensional artificially constructed features.

[0093] Step 5: The outputs of the sequence semantic branch and the algebraic feature branch are concatenated and then input into the fully connected layer, and the overfitting is suppressed by the Dropout layer;

[0094] Step 6: Output the probability distribution of the seven types of codes through the Softmax classification layer to complete the blind identification of the channel coding type.

[0095] As a specific example, step 1 establishes a communication system model, such as... Figure 2As shown, the details are as follows:

[0096] Step 1.1: Due to noise, fading, and loss during transmission, the binary bit sequence w obtained after demodulation and other preprocessing of the intercepted sequence r can be regarded as an error-laden version of the channel-coded sequence c from the transmitting end, equivalent to... Where e represents the error vector. Represents bitwise XOR; encoding type θ is unknown, belonging to the candidate set Θ={θ Ham ,θ Conv ,θ Turbo ,θ BCH ,θ LDPC ,θ RS ,θ Polor}, that is, the candidate set Θ includes the Hamming code θ Ham Convolutional code θ Conv Turbo code θ Turbo BCH code θ BCh LDPC code θ LDPC RS code θ RS Polar code θ Polar The encoding type determines the structural characteristics of sequence w.

[0097] Blind encoding type identification is the process of solving for the mapping based on the structure of sequence w in a noisy environment. L represents the length of the encoded sequence, and the candidate set of encoding types for sequence c is obtained. The decision is based on the essential difference in codeword space:

[0098]

[0099] in, The typical codeword space for θ-class encoding is shown in Table 1.

[0100] Table 1 Typical codeword space for channel coding

[0101]

[0102] in, Let P be the desired mapping, and let P be the probability that the discriminant sequence w belongs to a certain type of encoding in the candidate set.

[0103] Step 1.2: Suppose the bit error rate of sequence w is p compared to that of sequence c. Under the condition that p is unknown, establish a composite hypothesis testing model:

[0104]

[0105] Where, θ k This is the kth encoding type;

[0106] Decision-making uses generalized likelihood ratio:

[0107]

[0108] in, Let `sup(·)` be the likelihood function, and `sup(·)` be the function that takes the upper limit of the likelihood function. To estimate the results, Given a fixed p, the sequence w belongs to type θ. k The probability of;

[0109] Step 1.3: Since directly calculating the likelihood function involves high-dimensional summation and depends on the unknown p, we introduce encoding structure invariants. Define the original bitstream features mapped to a low-dimensional feature space. Satisfying the p invariance condition:

[0110]

[0111] Among them, D KL This indicates that, under the same bit error rate p, two different codes θ i and θ j The difference between the probability distributions of the generated feature f; These represent the fixed encoding type θ. i θ j The probability of feature f extracted under the given conditions; i,j traverse all encoding types that need to be distinguished;

[0112] That is, the parsing feature mapping for each encoding type It exhibits a certain degree of robustness to p, and feature mappings can be learned from the encoding structure through deep learning models. Then, different encoding types can be identified based on this feature.

[0113] As a specific example, the sequence semantic branch and algebraic feature branch mentioned in step 2 are as follows:

[0114] The sequence semantic branch takes an L / g-dimensional integer sequence as input, which is converted from the original bitstream by g-bit grouping. L represents the length of the encoded sequence. It is mapped to a 128-dimensional semantic space through an embedding layer. Features are extracted through a multi-scale dilated convolutional layer with an inflation rate of {1,2,5}. Finally, a 128-dimensional feature vector is generated through global max pooling to expand the receptive field and adapt to the spatial dependencies of different encoding structures.

[0115] The algebraic feature branch takes multidimensional artificially constructed features as input, including run count, entropy, autocorrelation, coding depth, block features, linear features, and spectral features. It performs nonlinear transformation through a fully connected layer and introduces batch normalization and Dropout layers to enhance generalization.

[0116] As a specific example, the sequence semantic branch described in step 3 uses a multi-scale dilated convolutional layer to extract long-range spatial features of the coding structure, such as... Figure 3 As shown, the details are as follows:

[0117] Step 3.1: Addressing the long-range structural characteristics of channel coding, such as the sparse parity correlation of LDPC codes and the interleaver dependency of Turbo codes, deep representation learning of the coded sequence is achieved through multi-scale feature extraction. The input layer is the received binary bit sequence w∈{0,1}. L Grouping by g bits and converting to an integer sequence, where L represents the sequence length, generates an L / g-dimensional discrete input vector. The range is [0, 2]. g -1]:

[0118]

[0119] in, w represents an L / g-dimensional integer vector. gτ+t It is the value of the t-th bit in the τ-th group of the original bit sequence, where τ is the output integer sequence x. int Index;

[0120] Step 3.2: The embedding mapping layer uses a trainable embedding matrix. Mapping integers to dense vectors Where the embedding dimension d e =256 was determined through grid search optimization, and the output tensor was... Preserve the temporal structure of the sequence. These represent the real parameter matrices of 256×256, 256, and (L / g)×256 dimensions respectively, generated by training the neural network.

[0121] Step 3.3: After the embedding mapping layer, a multi-scale convolutional layer is connected, which contains three parallel branches with convolutional kernel sizes of 3, 5 and 7 respectively. The other hyperparameters remain the same. Each branch is followed by a set of multi-dilation convolutional layers.

[0122] Step 3.4: The multi-dilation rate convolutional layer uses a parallel three-branch structure to capture features of different receptive fields. Multi-scale feature adaptation is achieved through differentiated dilation rates {1,2,5}. The m=1 branch extracts the local linear constraints of the block code, the m=2 branch captures the grid state transition of the convolutional code, and the m=5 branch models the long-range correlation of the interleaver of the Turbo code. The convolutional kernel parameters of each branch are shared, only the dilation rate is different.

[0123] Step 3.5: A left-filling strategy is adopted to ensure temporal invariance, and the GELU function is used as the activation function to provide a smooth nonlinear mapping;

[0124] Step 3.6: Use global max pooling to extract the strongest activation feature for each channel in the output of each branch:

[0125]

[0126] H m This represents the output feature map of the m-th dilated convolution branch; This represents a one-dimensional convolution operation, where the subscripts k = 3, 5, 7 indicate the size of the convolution kernel, and the superscript d = m indicates the dilation rate m.

[0127] Step 3.7: The three features are concatenated along the channel dimension to generate a 384-dimensional semantic feature vector. Each subspace corresponds to a structural pattern at a different scale; v (1) ,v (2) ,v (5) The features are represented by inflation rates of 1, 2, and 5, respectively. Then, global max pooling is performed on the entire feature.

[0128] Step 3.8: The output features of the three multi-scale branches are concatenated with the residual path to obtain the output features of the semantic channel of the encoded sequence.

[0129] As a specific example, the algebraic feature branch described in step 4, based on extracted multidimensional artificially constructed features, captures the algebraically invariant properties of the encoding, such as... Figure 4 As shown, the details are as follows:

[0130] Step 4.1: The multidimensional feature extraction module receives the preprocessed sequence as input and generates a high-dimensional discriminative feature tensor through feature space mapping. The definitions and encoding discriminative properties of each artificial feature are shown in Table 2. This tensor can be formally represented as:

[0131]

[0132] Where T(·) is the feature transformation operator, and D is the feature dimension; X pre For the preprocessed sequence, F multi The extracted multidimensional features;

[0133] Table 2 Artificial Features

[0134]

[0135] Step 4.2: Output the feature tensor F multi As input to the downstream neural network, the GELU activation function is used, and feature dimensionality enhancement and nonlinear mapping are achieved through fully connected layers:

[0136]

[0137] Among them, the weight matrix After Xavier initialization, bias

[0138] Step 4.3: Use batch normalization layer and 50% Dropout to enhance generalization.

[0139] As a specific example, the outputs of the sequence semantic branch and the algebraic feature branch described in step 5 are concatenated and then input into a fully connected layer. Overfitting is suppressed by a Dropout layer, as detailed below:

[0140] Step 5.1: Concatenate the multi-scale features extracted from the sequence semantic branch with the manual statistical features extracted from the algebraic feature branch to form a comprehensive feature vector;

[0141] Step 5.2: Dynamically calculate feature weights through a gating mechanism, and adaptively weight the fused features to enhance important information;

[0142] Step 5.3: Feed the weighted features into a multi-layer classifier (containing a fully connected layer, a GELU activation function, and a Dropout layer) and use regularization to effectively suppress overfitting.

[0143] As a specific example, the probability distribution of the seven types of codes in step 6 is specifically the probability vector of each sample belonging to a different code type.

[0144] This invention also provides a blind channel coding type identification system based on a dual-channel convolutional neural network. This system is used to implement the aforementioned blind channel coding type identification method based on a dual-channel convolutional neural network. The system includes a model building module, a neural network construction module, a sequence semantic branching module, an algebraic feature branching module, a concatenation module, and a classification module, wherein:

[0145] The model building module is used to build a communication system model.

[0146] The neural network construction module adopts a heterogeneous feature fusion architecture to build a dual-channel convolutional neural network DBFCNN, which includes two parallel processing modules: a sequence semantic branch and an algebraic feature branch.

[0147] The sequence semantic branching module uses multi-scale dilated convolutional layers to extract long-range spatial features of the coding structure;

[0148] The algebraic feature branching module captures the algebraic invariant properties of the encoded data based on extracted multidimensional artificially constructed features.

[0149] The outputs of the concatenation module, the sequence semantic branch module, and the algebraic feature branch module are concatenated and then input into the fully connected layer, where overfitting is suppressed by the Dropout layer.

[0150] The classification module outputs the probability distribution of seven classes of codes through the Softmax classification layer.

[0151] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned blind identification method for channel coding type based on a dual-channel convolutional neural network.

[0152] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the blind identification method for channel coding types based on a dual-channel convolutional neural network.

[0153] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0154] Example

[0155] This embodiment employs the PyTorch deep learning framework and utilizes an NVIDIA GeForce PTX 4060 GPU for training acceleration. To meet the data requirements of the deep learning model, simulation datasets for seven types of channel coding with different parameters were generated based on MATLAB 2022a: Hamming codes, BCH codes, LDPC codes, RS codes, Polar codes, convolutional codes, and Turbo codes. Each parameter's coding generated 2000 sample frames. Key parameter configurations are detailed in Table 3. Random bit errors were injected using an additive white Gaussian noise (AWGN) channel model, with a bit error rate of 10%. -5 ~10 -1 Twenty points are evenly selected within the range to generate noisy received sequences, which are then used as input after preprocessing. This results in a dataset containing 1,120,000 samples. The training and test sets are divided in an 8:2 ratio, and the cross-entropy loss function is used. With the Adam optimizer, initial learning rate 10 -3 Model training was performed with decay factors beta_1 = 0.9 and beta_2 = 0.999, batch size B = 256, and maximum iteration rounds T = 150.

[0156] To prevent overfitting, an early stopping mechanism is implemented: training automatically stops if the validation loss does not improve further after 15 epochs. Simultaneously, an automatic learning rate optimization mechanism is employed: if the validation loss does not improve further after 5 epochs, the learning rate is reduced to half its original value, with a minimum of 10. -6 .

[0157] Table 3 Channel Coding Parameters

[0158] Convolutional codes (4,[15,17])(5,[23,35])(6,[53,75])(7,[133,174]) Hamming code [7,4][15,11][31,26][63,57] LDPC code [648,1 / 2][648,2 / 3][1296,1 / 2][1296,2 / 3] Turbo code (4,[15,17])(5,[23,35])(6,[53,75])(7,[133,174]) BCH code [63,47][127,95][204,188][255,223] RS code [15,7][15,11][255,139][255,251] Polar codes [64,32][128,40][256,80][512,256]

[0159] (1) Comparison of recognition performance of different models:

[0160] The performance of DBFCNN was compared with that of various network architectures at different bit error rates, including: MSDCNN, which uses dual multi-scale modules and dilated convolutional structures; TextCNN, which processes encoded sequences as text; and a fusion network that starts with dual convolutional layers followed by five multi-scale Inception modules. Figure 5 As shown, experimental results indicate that the DBFCNN architecture performs well in 10... -5 Up to 10 -1 It exhibits significant performance advantages across a wide range of bit error rates, especially at 10. -5 ~10 -3 It performs best under low error rate conditions, with an accuracy of over 99%.

[0161] Among the comparison networks, MSDCNN performed best, but its recognition accuracy was still significantly lower than that of the DBFCNN architecture. When the bit error rate was 10... -5 10 -4 10 -3 and 10 -2 At 10, DBFCNN's recognition accuracy improved by 3.75%, 3.86%, 5.05%, and 12.91% compared to MSDCNN, respectively, only falling behind MSDCNN at high bit error rates. This performance advantage mainly stems from DBFCNN's dual-channel feature fusion mechanism, which effectively enhances the model's feature representation ability and robustness under complex channel conditions. Experimental data further demonstrates that at 10 -2 In the high error rate region mentioned above, DBFCNN still maintains an average performance advantage of 82.30%.

[0162] (2) The impact of semantic partition length on performance:

[0163] Figure 6 The comparison of recognition results is shown when the semantic segmentation length g = {4, 5, 8, 10, 16} bits. Experimental results show that the network achieves optimal recognition performance when the word length is set to 8 bits. This is because an 8-bit word length transforms a 5000-bit input sequence into a reasonable length of 625 words, avoiding the problems of excessively long sequences caused by 4-bit words and excessively short sequences caused by 16-bit words. Furthermore, the 8-bit structure is equivalent to 1 byte, matching the computer architecture and enabling efficient execution of CPU / GPU SIMD instructions while reducing memory access overhead. In addition, an 8-bit window can cover the core features of typical channel coding: the 5-7 bit constraint length of convolutional codes, the byte-based interleaver structure of Turbo codes, the 8-bit symbol system of RS / BCH codes, and the byte boundary check matrix design of LDPC codes.

[0164] While BER>6×10-2 In high-error-rate environments, the recognition accuracy of 8-bit words is slightly lower than that of 4-bit words. This is primarily because shorter words divide the sequence into more words, each containing less information; therefore, a single word error has a relatively smaller impact on the overall features. In contrast, 8-bit words contain only 625 words, each carrying more information, making their errors more significant for the feature vector. Furthermore, since a word represents an integer from 0 to 255, a single bit error in an 8-bit word leads to a completely incorrect value for the entire word, while a single bit error in a 4-bit word only affects those four words. Therefore, under high error rates, 4-bit words offer better recognition accuracy. However, overall, 8-bit words remain the optimal choice.

[0165] (3) The impact of network architecture on performance:

[0166] Figure 7 , Figure 8 and Figure 9 The results show a comparison of recognition performance under different network architectures and embedding dimension configurations. Experimental data shows that the dual-channel convolutional neural network DBFCNN achieves optimal performance with an embedding dimension of 128, achieving a test accuracy of 92.95%, a 4.29 percentage point improvement compared to the optimal single-channel architecture. Notably, when the embedding dimension increases from 32 to 512, the recognition accuracy fluctuation range of DBFCNN is only 2.73%, lower than the 5.25% fluctuation range of the optimal single-channel architecture, proving that the dual-channel design can effectively suppress performance fluctuations caused by changes in embedding dimension. This stability stems from the enhancement effect of the dual-channel feature fusion mechanism on the representation space, which reaches the optimal balance point at an embedding dimension of 128, enabling the model to obtain optimal feature discrimination ability while maintaining parameter efficiency.

[0167] (4) Identification of different encoding types:

[0168] like Figure 10 As shown, the recognition performance of different types of channel coding varies with the bit error rate (BER). Hamming codes, as the simplest linear block codes, are easier to recognize at lower BERs. However, the higher the coding complexity of a linear block code, the weaker its recognition performance under the same conditions. In contrast, convolutional codes, Turbo codes, and Polar codes significantly outperform other linear block codes at high BERs. Polar codes, at BER=10... -1 Even under high error rates, it maintains a recognition accuracy of >80%. In 10... -5 ~10 -3 In low to medium error rate environments, all encoding types can achieve a recognition accuracy of >95%, and approach the 99% limit as the BER continues to decrease.

[0169] To better understand the recognition performance of different encoding types Figure 11, Figure 12 and Figure 13 The confusion matrices at several specific bit error rates are presented to compare the performance differences under high, low, and extremely low bit error rates. Figure 11 In high-error-rate scenarios, Polar codes maintain an accuracy of 60% thanks to their unique frozen bit pattern formed by channel polarization, significantly outperforming other codes. Hamming codes achieve 51.5% accuracy by relying on a fixed block structure, while other code types suffer from severe cross-contamination. Figure 12 In a low-error-rate environment, convolutional codes, Hamming codes, and polar codes all achieved 100% accuracy. RS codes were mainly misclassified as Hamming codes due to the similarity of their algebraic grouping structures, while LDPC codes experienced a slight misclassification of 0.8% due to overlapping sparse parity features. Figure 13 Under low error conditions, the average accuracy of all codes exceeds 99%, with only minor confusions such as BCH→Hamming code and RS→convolution / Hamming code remaining, reflecting the approximation of the cyclic code generator polynomial and the cross-interference of Galois field operations.

[0170] In summary, this invention provides a blind channel coding type identification method based on a dual-channel convolutional neural network (DBFCNN) to achieve blind type identification of unknown channel coding in non-cooperative systems without requiring any prior knowledge. Experimental results show that compared to TextCNN, Inception, and MSDCNN, DBFCNN improves the identification accuracy under both high and low bit error rates, with overall recognition performance 5-10% higher than the previous best network. When the bit error rate is below 10%, DBFCNN achieves even better results. -3 At that time, the accuracy of the test set can reach more than 98%, which effectively improves the recognition accuracy of various channel coding types.

[0171] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A blind method for channel coding type identification based on a dual-channel convolutional neural network, characterized in that, Includes the following steps: Step 1: Establish a communication system model; Step 2: Adopt a heterogeneous feature fusion architecture to construct a dual-channel convolutional neural network DBFCNN, which includes two parallel processing branches: a sequence semantic branch and an algebraic feature branch. Step 3: Sequence semantic branch uses multi-scale dilated convolutional layers to extract long-range spatial features of the coding structure; Step 4: Algebraic feature branching captures the algebraic invariant properties of the encoded multidimensional artificially constructed features. Step 5: The outputs of the sequence semantic branch and the algebraic feature branch are concatenated and then input into the fully connected layer, and the overfitting is suppressed by the Dropout layer; Step 6: Output the probability distribution of the seven types of codes through the Softmax classification layer to complete the blind identification of the channel coding type.

2. The blind channel coding type identification method based on a dual-channel convolutional neural network according to claim 1, characterized in that, Step 1, which involves establishing a communication system model, is as follows: Step 1.1: Due to noise, fading, and loss during transmission, the binary bit sequence w obtained after demodulation preprocessing of the intercepted sequence r is considered an error-prone version of the channel-coded sequence c from the transmitting end, equivalent to... Where e represents the error vector. Represents bitwise XOR; encoding type θ is unknown, belonging to the candidate set Θ={θ Ham ,θ Conv ,θ Turbo ,θ BCH ,θ LDPC ,θ RS ,θ Polar }, that is, the candidate set Θ includes the Hamming code θ Ham Convolutional code θ Conv Turbo code θ Turbo BCH code θ BCH LDPC code θ LDPC RS code θ RS Polar code θ Polar The encoding type determines the structural characteristics of sequence w. Blind encoding type identification is the process of solving the mapping based on the structure of sequence w in the context of errors. L represents the length of the encoded sequence, and the candidate set of encoding types for sequence c is obtained. The decision is based on the essential difference in codeword space: in, This is a typical codeword space for θ-class encoding. Let P be the desired mapping, and let P be the probability that the discriminant sequence w belongs to a certain type of encoding in the candidate set. Step 1.2: Suppose the bit error rate of sequence w is p compared to that of sequence c. Under the condition that p is unknown, establish a composite hypothesis testing model: Where, θ k This is the kth encoding type; Decision-making uses generalized likelihood ratio: in, Let `sup(·)` be the likelihood function, and `sup(·)` be the function that takes the upper limit of the likelihood function. To estimate the results, Given a fixed p, the sequence w belongs to type θ. k The probability of; Step 1.3: Since directly calculating the likelihood function involves high-dimensional summation and depends on the unknown p, we introduce encoding structure invariants. Define the original bitstream features mapped to a low-dimensional feature space. Satisfying the p invariance condition: Among them, D KL This indicates that, under the same bit error rate p, two different codes θ i and θ j The difference between the probability distributions of the generated feature f; These represent the fixed encoding type θ. i θ j The probability of feature f extracted under the given conditions; i,j traverse all encoding types that need to be distinguished; Parsing feature mapping for each encoding type It is robust to p, and learns the feature mapping of the encoding from the encoding structure through a deep learning model, and then distinguishes different encoding types based on the features.

3. The blind channel coding type identification method based on a dual-channel convolutional neural network according to claim 1, characterized in that, The sequence semantic branch and algebraic feature branch mentioned in step 2 are as follows: The sequence semantic branch takes an L / g-dimensional integer sequence as input, which is converted from the original bitstream by g-bit grouping. L represents the length of the encoded sequence. It is mapped to a 128-dimensional semantic space through an embedding layer. Features are extracted through a multi-scale dilated convolutional layer with an inflation rate of {1,2,5}. Finally, a 128-dimensional feature vector is generated through global max pooling to expand the receptive field and adapt to the spatial dependencies of different encoding structures. The algebraic feature branch takes multidimensional artificially constructed features as input, including run count, entropy, autocorrelation, coding depth, block features, linear features, and spectral features. It performs nonlinear transformation through a fully connected layer and introduces batch normalization and Dropout layers to enhance generalization.

4. The blind channel coding type identification method based on a dual-channel convolutional neural network according to claim 1, characterized in that, Step 3 describes a sequence semantic branch that uses multi-scale dilated convolutional layers to extract long-range spatial features of the coding structure, as detailed below: Step 3.1: Addressing the long-range structural characteristics of channel coding, deep representation learning of the coded sequence is achieved through multi-scale feature extraction. The input layer is the received binary bit sequence w∈{0,1}. L Grouping by g bits and converting to an integer sequence, where L represents the sequence length, generates an L / g-dimensional discrete input vector. The range is [0, 2]. g -1]: in, w represents an L / g-dimensional integer vector. gτ+t It is the value of the t-th bit in the τ-th group of the original bit sequence, where τ is the output integer sequence x. int The index; Step 3.2: The embedding mapping layer uses a trainable embedding matrix. Mapping integers to dense vectors Where the embedding dimension d e =256 was determined through grid search optimization, and the output tensor was... Preserve the temporal structure of the sequence. These represent the real parameter matrices of 256×256, 256, and (L / g)×256 dimensions respectively, generated by training the neural network. Step 3.3: After the embedding mapping layer, a multi-scale convolutional layer is connected, which contains three parallel branches with convolutional kernel sizes of 3, 5 and 7 respectively. The other hyperparameters remain the same. Each branch is followed by a set of multi-dilation convolutional layers. Step 3.4: The multi-dilation rate convolutional layer uses a parallel three-branch structure to capture features of different receptive fields. Multi-scale feature adaptation is achieved through differentiated dilation rates {1,2,5}. The m=1 branch extracts the local linear constraints of the block code, the m=2 branch captures the grid state transition of the convolutional code, and the m=5 branch models the long-range correlation of the interleaver of the Turbo code. The convolutional kernel parameters of each branch are shared, only the dilation rate is different. Step 3.5: A left-filling strategy is adopted to ensure temporal invariance, and the GELU function is used as the activation function to provide a smooth nonlinear mapping; Step 3.6: Use global max pooling to extract the strongest activation feature for each channel in the output of each branch: H m This represents the output feature map of the m-th dilated convolution branch; This represents a one-dimensional convolution operation, where the subscripts k = 3, 5, 7 indicate the size of the convolution kernel, and the superscript d = m indicates the dilation rate m. Step 3.7: The three features are concatenated along the channel dimension to generate a 384-dimensional semantic feature vector. Each subspace corresponds to a structural pattern at a different scale; v (1) ,v (2) ,v (5) The features are represented by inflation rates of 1, 2, and 5, respectively. Then, global max pooling is performed on the entire feature. Step 3.8: The output features of the three multi-scale branches are concatenated with the residual path to obtain the output features of the semantic channel of the encoded sequence.

5. The blind channel coding type identification method based on a dual-channel convolutional neural network according to claim 1, characterized in that, The algebraic feature branch described in step 4 is based on the extracted multidimensional artificially constructed features, capturing the algebraic invariant properties of the encoding, as detailed below: Step 4.1: The multidimensional feature extraction module receives the preprocessed sequence as input and generates a high-dimensional discriminative feature tensor through feature space mapping. This tensor is formally represented as: Where T(·) is the feature transformation operator, D is the feature dimension, and X is the feature transformation operator. pre For the preprocessed sequence, F multi The extracted multidimensional features; Step 4.2: Output the feature tensor F multi As input to the downstream neural network, the GELU activation function is used, and feature dimensionality enhancement and nonlinear mapping are achieved through fully connected layers: Among them, the weight matrix After Xavier initialization, bias Step 4.3: Use batch normalization layer and 50% Dropout to enhance generalization.

6. The blind channel coding type identification method based on a dual-channel convolutional neural network according to claim 1, characterized in that, The outputs of the sequence semantic branch and algebraic feature branch described in step 5 are concatenated and then input into a fully connected layer. Overfitting is suppressed by a Dropout layer, as detailed below: Step 5.1: Concatenate the multi-scale features extracted from the sequence semantic branch with the manual statistical features extracted from the algebraic feature branch to form a comprehensive feature vector; Step 5.2: Dynamically calculate feature weights through a gating mechanism, and adaptively weight the fused features to enhance important information; Step 5.3: Feed the weighted features into a multi-layer classifier containing fully connected layers, GELU activation function and Dropout layer, and suppress overfitting through regularization.

7. The blind channel coding type identification method based on a dual-channel convolutional neural network according to claim 1, characterized in that, The probability distribution of the seven types of codes in step 6 is specifically the probability vector of each sample belonging to a different code type.

8. A blind channel coding type identification system based on a dual-channel convolutional neural network, characterized in that, This system is used to implement the blind channel coding type identification method based on a dual-channel convolutional neural network as described in any one of claims 1 to 7. The system includes a model building module, a neural network construction module, a sequence semantic branching module, an algebraic feature branching module, a concatenation module, and a classification module, wherein: The model building module is used to build a communication system model. The neural network construction module adopts a heterogeneous feature fusion architecture to build a dual-channel convolutional neural network DBFCNN, which includes two parallel processing modules: a sequence semantic branch and an algebraic feature branch. The sequence semantic branching module uses multi-scale dilated convolutional layers to extract long-range spatial features of the coding structure; The algebraic feature branching module captures the algebraic invariant properties of the encoded data based on extracted multidimensional artificially constructed features. The outputs of the concatenation module, the sequence semantic branch module, and the algebraic feature branch module are concatenated and then input into the fully connected layer, where overfitting is suppressed by the Dropout layer. The classification module outputs the probability distribution of seven classes of codes through the Softmax classification layer.

9. A mobile terminal, 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 program, it implements the blind channel coding type identification method based on a dual-channel convolutional neural network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the blind channel coding type identification method based on a dual-channel convolutional neural network as described in any one of claims 1 to 7.

Images