Parameter blind identification method of long code LDPC
Through the joint architecture of generative adversarial network and Transformer, using data augmentation and feature extraction technology, the problem of long-code LDPC parameter recognition is solved, and effective parameter recognition is achieved in non-cooperative communication.
Patent Information
- Application Number
- CN202510911160.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-08-22
AI Technical Summary
The parameter identification method of long code LDPC in the prior art is not yet mature, especially in non-cooperative environments or adaptive communication scenarios, it is difficult for the receiver to effectively identify the type and encoding parameters of the channel code.
The combined architecture of generative adversarial network and Transformer is adopted to generate new training samples through data augmentation, and a one-dimensional convolutional neural network and Transformer are used to identify long-code LDPC parameters to capture long-distance dependencies and global information in the data.
In the case of small sample size, effective long-code LDPC parameter recognition is achieved, simplifying the recognition steps and providing a new solution for non-cooperative communications.
Smart Images
Figure CN120528562A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of communication technology, and in particular relates to a blind identification method for parameters of a long code LDPC. Background Art
[0002] Channel coding is a crucial component in communication systems and plays a core role in multiple communication standards and protocols, covering wireless communications, satellite communications, optical communications and other fields. By applying effective channel coding, communication systems can ensure high reliability and transmission performance despite the influence of noise and interference. By adding redundant information, the receiver can more easily detect and correct errors. At the receiver, which has a complete understanding of the coding scheme used by the transmitter, the required information can be successfully decoded when the channel conditions are good enough. In non-cooperative environments or adaptive communication scenarios such as elastic optical networks where there is no prior knowledge of the encoder, the receiver needs to recover the transmitted information in a blind state from the intercepted or received signal. Therefore, it is crucial for non-cooperative or adaptive receivers to correctly identify the type of channel code used and the coding parameters.
[0003] As communications move toward higher speeds and higher capacity, long LDPC (Low Density Parity Check) sequences are widely used in various 5G communications. However, due to their long code length and large number of parameters, sequence feature extraction is difficult, and currently, no method for parameter identification of long LDPC codes has been reported. Summary of the Invention
[0004] The purpose of the present invention is to provide a method for blind identification of parameters of long code LDPC. By combining the generative adversarial network with the Transformer, it is possible to ensure that effective parameter identification tasks can be performed even with a small number of samples.
[0005] The technical solutions adopted by the present invention to solve the technical problems are as follows:
[0006] The present invention provides a method for blindly identifying parameters of a long code LDPC, which mainly includes the following steps:
[0007] First, the generative adversarial network is trained to perform sample data enhancement processing on the sample data, and then the neural network architecture built by the one-dimensional convolutional neural network and Transformer is trained for parameter identification.
[0008] Furthermore, before performing the sample data enhancement process, a real sample is first obtained, where the real sample is data obtained by demodulating a long code LDPC information sequence.
[0009] Furthermore, in the generative adversarial network, the loss function of the generator is Expressed as:
[0010]
[0011] in, Represents the discriminator's expectation of the generated data, x represents the real data, z represents the false data, G(z) represents the generated false sample, and D(G(z)) represents the probability that the discriminator judges whether the generated false sample is a real sample.
[0012] Furthermore, in the generative adversarial network, the loss function of the discriminator is Expressed as:
[0013]
[0014] in, represents the discriminator’s expectation of the real data, Represents the discriminator's expectation of the generated data, x represents real data, z represents false data, D(x) represents the real sample, G(z) represents the generated false sample, and D(G(z)) represents the probability that the discriminator judges whether the generated false sample is a real sample.
[0015] Furthermore, fine-tuning training minimizes the loss function The real data x is identified as real by the discriminator, that is, the real sample D(x) is close to 1, and the false sample G(z) is identified as false by the discriminator, that is, the probability D(G(z)) is close to 0.
[0016] Furthermore, the overall goal of the generative adversarial network is that the generator learns to generate fake data that is as close as possible to the real data distribution, while the discriminator learns to distinguish between real data and generated data.
[0017] Furthermore, the minimum-maximum adversarial loss function of the generative adversarial network is expressed as:
[0018]
[0019] in, represents the discriminator’s expectation of the real data, represents the discriminator's expectation of the variable z, p z (z) means that the variable z follows the standard normal distribution, represents the minimax game, D(x) represents the real sample, G(z) represents the generated false sample, D(G(z)) represents the probability that the discriminator judges whether the generated false sample is a real sample, D represents the discriminator, G represents the generator, x represents the real data, and z represents the false data.
[0020] Furthermore, the one-dimensional convolutional neural network performs preliminary feature extraction on the long code LDPC, reducing the computational pressure of the Transformer and the training load of the GPU.
[0021] Furthermore, the Transformer uses two encoders separately, each encoder contains a multi-head attention layer, a feedforward neural network layer, two addition and normalization layers, and the number of attention heads in the multi-head attention layer is set to four. These four attention heads learn information from four subspaces in parallel.
[0022] Furthermore, two convolutional layers are stacked before the Transformer, and a Dropout layer is added before connecting the encoder of the Transformer.
[0023] The beneficial effects of the present invention are:
[0024] The present invention provides a method for blind parameter identification of long LDPC codes. Based on a signal processing process using a generative adversarial network and a Transformer architecture, data enhancement and parameter identification are performed on long LDPC code sequences. The main signal processing and parameter identification processes for long LDPC codes in this scheme include: demodulation to obtain real samples, training a generative adversarial network to generate more signal samples, and training a one-dimensional convolutional neural network and a Transformer architecture to identify parameters of long LDPC codes. In this method, considering the potential problems of insufficient data volume and monotonous data formats during training, a generative adversarial network is used for data enhancement. New training samples are generated to supplement the existing dataset, thereby expanding the diversity of training data. Given the long length of long LDPC codes, the parameter identification algorithm draws on successful experience in natural language processing to design a Transformer model to capture long-range dependencies and global information in the data. By utilizing a combined generative adversarial network and Transformer architecture, this method simplifies the parameter identification steps for long LDPC codes, providing new insights for long LDPC code identification and promising applications in non-cooperative communications. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 This is a flow chart of a blind identification method for long code LDPC parameters provided by the present invention.
[0026] Figure 2 Flowchart for data augmentation.
[0027] Figure 3 This is the Transformer architecture diagram.
[0028] Figure 4 This is the neural network structure diagram of the generator.
[0029] Figure 5 This is the neural network structure diagram of the discriminator. DETAILED DESCRIPTION
[0030] The present invention is further described in detail below with reference to the accompanying drawings.
[0031] In a first aspect, the present invention provides a method for blindly identifying parameters of a long code LDPC.
[0032] The present invention provides a method for blindly identifying parameters of long LDPC codes, which combines a generative adversarial network (GAN) with a Transformer architecture. First, a generative adversarial network is used to enhance sample data to address the problems of insufficient data volume and single data format, generating new training samples to supplement the existing data set, thereby enhancing the diversity of training data. Second, by building a one-dimensional convolutional neural network and a Transformer neural network architecture for parameter identification of long LDPC codes, long-range dependencies and global information in the data can be effectively captured.
[0033] See also Figure 1 To illustrate, the present invention provides a method for blindly identifying parameters of a long code LDPC, and its specific implementation process is as follows:
[0034] Step S1: data enhancement module;
[0035] The data augmentation module specifically uses a generative adversarial network (GAN). This method addresses the issues of insufficient data and monotonous data formats, generating new training samples to supplement the existing dataset and increase the diversity of training data.
[0036] Specifically, if Figure 2 As shown, the received long code LDPC information sequence is first demodulated to obtain the demodulated information sequence as the real sample D(x); then the randomly generated noise sequence is used as the input of the generator G, and the setting Denotes the loss function of the generator G, G(z) denotes the generated false sample, D(G(z)) denotes the probability that the discriminator D determines whether the generated false sample is a real sample, then the loss function of the generator G is It can be expressed as follows:
[0037]
[0038] in, represents the discriminator's expectation of the generated data, x represents the real data, and z represents the fake data.
[0039] Correspondingly, the discriminator D is a two-class neural network that inputs real data x and false samples G(z). Its goal is to distinguish the two. Denotes the loss function of the discriminator D, then the loss function of the discriminator D is It can be expressed as follows:
[0040]
[0041] in, represents the discriminator's expectation of the real data.
[0042] Minimize the loss function by fine-tuning training The real data x is identified as real by the discriminator D (that is, the real sample D(x) is close to 1) and the false sample G(z) is identified as false by the discriminator D (that is, the probability D(G(z)) is close to 0).
[0043] The overall goal of a generative adversarial network is for the generator G to learn to generate fake data that is as close to the real data distribution as possible, while the discriminator D learns to distinguish between real data and generated data. The core of this is achieved by using a minimum-maximum adversarial loss function. The mathematical expression of the minimum-maximum adversarial loss function is as follows:
[0044]
[0045] in, represents the discriminator's expectation of the variable z, p z (z) indicates that the variable z follows the standard normal distribution, Represents a minimax game.
[0046] The present invention generates new training samples by generative adversarial networks to supplement existing data sets, thereby expanding the diversity of training samples and solving the problem of insufficient data volume and single data form leading to unsatisfactory training results during neural training.
[0047] Step S2: parameter identification module;
[0048] Considering the effectiveness of one-dimensional convolutional neural networks in extracting sequence features, the parameter identification module uses a one-dimensional convolutional neural network (CNN) and Transformer neural network architecture to identify the parameters of long code LDPC, which can effectively capture long-distance dependencies and global information in the data.
[0049] Specifically, the input long code LDPC information sequence is first subjected to preliminary feature extraction through a one-dimensional convolutional neural network. The preliminary feature extraction can reduce the computational pressure of the Transformer architecture and the training load of the graphics processing unit (GPU). Secondly, in the Transformer architecture designed by the present invention, two encoders are used separately, each of which includes a multi-head attention layer, a feedforward neural network layer, two addition and normalization layers, wherein the number of attention heads of the multi-head attention module is set to four, and these four attention heads learn information from four subspaces in parallel, such as Figure 3 As shown in the figure, the Transformer architecture designed in this paper can effectively capture the global dependencies of long LDPC codes, and the Transformer architecture removes the decoder module, making the model more suitable for classification problems; finally, a fully connected layer and a Softmax function are used to output the classification results.
[0050] In the present invention, since the Transformer architecture can capture long-range dependencies and global information in the data, it can better perform feature extraction and analysis on long code LDPC information sequences.
[0051] The present invention combines the generative adversarial network with the Transformer to simplify the parameter identification steps of the long code LDPC and provides a new idea for the parameter identification of the long code LDPC.
[0052] In a second aspect, the present invention provides a long code LDPC parameter blind identification system for implementing a long code LDPC parameter blind identification method provided in the first aspect of the present invention.
[0053] The present invention provides a long code LDPC parameter blind identification system, which mainly includes the following modules:
[0054] The data enhancement module is mainly used to enhance the sample data of the received long code LDPC to supplement the existing data set and enhance the diversity of training data.
[0055] The parameter identification module is mainly used to identify parameters of long code LDPC to capture long-distance dependencies and global information in the data.
[0056] Preferably, the data enhancement module can specifically use a generative adversarial network (GAN), which learns to generate fake data that is as close as possible to the real data distribution through the generator G, and learns to distinguish between real data and generated data through the discriminator D.
[0057] Preferably, the GAN utilizes a min-max adversarial loss function to achieve its overall goal.
[0058] Preferably, the parameter identification module can specifically adopt a one-dimensional convolutional neural network (CNN) and a Transformer neural network architecture, first performing preliminary feature extraction on the long code LDPC through a one-dimensional convolutional neural network, and then performing parameter identification on the long code LDPC through the Transformer architecture.
[0059] The present invention provides a blind parameter identification system for long code LDPC. By enhancing the diversity of training data, the system can effectively perform parameter identification tasks with a small number of samples. The system can effectively capture long-range dependencies and global information in the data, and better complete the parameter identification of long code LDPC.
[0060] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following description is given by taking LDPC (17664, 14592) modulated by DP-16QAM and LDPC (16440, 14392) after check bit puncture as examples in conjunction with the accompanying drawings.
[0062] (1) First, a binary sequence is randomly generated as the source information. After two LDPC channel codes are applied, 16QAM modulation is performed, and the signal-to-noise ratio is set to 10dB through an AWGN channel. Soft decision demodulation is used at the receiving end and normalization is performed so that the results are between [-1, 1]. The length of each sample data is 17664, which is the length of an LDPC codeword. 1000 LDPC (17664, 14592) sample data are selected as the real samples of the generative adversarial network. At the same time, 1000 data enhancement samples can be obtained. These 2000 sample data are used as the test set of LDPC (17664, 14592). The number of samples in the test set and training set of the two long codes LDPC is 6000 and 2000 respectively.
[0063] (2) The neural network structure of the generator G designed in this embodiment is as follows Figure 4 As shown in the figure, it contains five fully connected layers and a hidden layer in the middle. Except for the first fully connected layer, the output lengths of the remaining fully connected layers increase layer by layer, and the final output length is equal to the code length 17664. Figure 4 In the equation, x1, x2…x 1000 represents a random noise vector with an input length of 1000, w1, w2, w3, w4, w5 represent the five fully connected layers of the generator G, h1, h2, h3…h 512 Indicates that the output vector length of the first fully connected layer is 512, h1, h2, h3…h 1024 Indicates that the output vector length of the second fully connected layer is 1024, h1, h2, h3…h 2048 Indicates that the output vector length of the third fully connected layer is 2048, h1, h2, h3…h 4096 Indicates that the output vector length of the fourth fully connected layer is 4096, f1, f2…f 17664 This means the final output vector length is 17664.
[0064] The neural network structure of the discriminator D designed in this embodiment is as follows Figure 5 As shown in Figure 1, it consists of multiple fully connected layers. The input to the discriminator D is either a real sample or a generated fake sample. Each sample has a data length of 17664. The intermediate hidden layer still uses the LeakyReLU function with a slope α of 0.2. Considering the long code length and the complexity of the discriminator D task, six fully connected layers are stacked. The input sample data length is 17664, decreasing layer by layer. The final output layer is the score of whether the input sample data is a real sample. Figure 5 In the equation, x1, x2…x 17664 Indicates that the input data length is 17664, w1, w2, w3, w4, w5, w6 represent the six fully connected layers of the discriminator D, h1, h2, h3…h 4096 Indicates that the output vector length of the first fully connected layer is 4096, h1, h2, h3…h 2048 Indicates that the output vector length of the second fully connected layer is 2048, h1, h2, h3…h 1024 Indicates that the output vector length of the third fully connected layer is 1024, h1, h2, h3…h 512 Indicates that the output vector length of the fourth fully connected layer is 512, h1, h2, h3…h 256 It indicates that the output vector length of the fifth fully connected layer is 256, f represents the final output, and after the Sigmoid activation function, the discrimination result is mapped to the range of [0,1], indicating the probability that the data is real data.
[0065] In this embodiment, during training, Adam is selected as the optimizer, the learning rate is 0.0001, and the decay rates of the first-order moment estimate and the second-order moment estimate are set to 0.5 and 0.9, respectively, to maintain the stability of the training. The outer loop and the inner loop are defined, the outer loop is set to 1000 training cycles, and the inner loop obtains real samples for each batch. The generator G is executed once per training, and the discriminator D is executed 3 times per training, and the neural network structures of the generator G and the discriminator D are trained and optimized respectively. The batch size of the read data is set to 64, and the training cycle of the generative adversarial network is set to 1000.
[0066] (3) Parameter identification: Two convolutional layers are stacked before the Transformer architecture. The convolution kernel size of each layer is 7, the padding is 3, and the stride is 2. The number of input channels of the first convolutional layer is 1, and the number of output channels of the first convolutional layer is 64. The number of input channels of the second convolutional layer is 64, and the number of output channels of the second convolutional layer is 128. Both the first and second convolutional layers use ReLU as the activation function. A Dropout layer is added before connecting to the Transformer encoder, and the probability is set to 0.1. The Adam optimization algorithm is used for training, with an initial learning rate of 0.0001, a training cycle of 50, and a batch size of 64 when reading the dataset.
[0067] This embodiment performs data enhancement and parameter identification on two long LDPC codes, providing a solution for neural network training when samples are scarce. At the same time, it can also effectively perform data enhancement and parameter identification on the long LDPC code based on a generative adversarial network.
[0068] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.
Claims
1. A method for blind identification of parameters of long code LDPC, characterized in that: The following steps are involved: First, the generative adversarial network is trained to perform sample data enhancement processing on the sample data, and then the neural network architecture built by the one-dimensional convolutional neural network and Transformer is trained for parameter identification.
2. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: Before performing sample data enhancement processing, real samples are first obtained, where the real samples are data obtained by demodulating a long code LDPC information sequence.
3. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: In the generative adversarial network, the loss function of the generator is Expressed as: in, Represents the discriminator's expectation of the generated data, x represents the real data, z represents the false data, G(z) represents the generated false sample, and D(G(z)) represents the probability that the discriminator judges whether the generated false sample is a real sample.
4. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: In the generative adversarial network, the loss function of the discriminator is Expressed as: in, represents the discriminator’s expectation of the real data, Represents the discriminator's expectation of the generated data, x represents real data, z represents false data, D(x) represents the real sample, G(z) represents the generated false sample, and D(G(z)) represents the probability that the discriminator judges whether the generated false sample is a real sample.
5. The method for blind identification of parameters of long code LDPC according to claim 4, characterized in that: Fine-tuning training minimizes the loss function The real data x is identified as real by the discriminator, that is, the real sample D(x) is close to 1, and the false sample G(z) is identified as false by the discriminator, that is, the probability D(G(z)) is close to 0.
6. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: The overall goal of the generative adversarial network is that the generator learns to generate fake data that is as close as possible to the real data distribution, while the discriminator learns to distinguish between real data and generated data.
7. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: The minimum-maximum adversarial loss function of the generative adversarial network is expressed as: in, represents the discriminator’s expectation of the real data, represents the discriminator's expectation of the variable z, p z (z) indicates that the variable z follows the standard normal distribution, represents the minimax game, D(x) represents the real sample, G(z) represents the generated false sample, D(G(z)) represents the probability that the discriminator judges whether the generated false sample is a real sample, D represents the discriminator, G represents the generator, x represents the real data, and z represents the false data.
8. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: The one-dimensional convolutional neural network performs preliminary feature extraction on the long code LDPC, reducing the computational pressure of the Transformer and the training load of the GPU.
9. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: The Transformer uses two encoders separately, each encoder contains a multi-head attention layer, a feedforward neural network layer, two addition and normalization layers, and the number of attention heads in the multi-head attention layer is set to four. These four attention heads learn information from four subspaces in parallel.
10. The method for blind identification of parameters of long code LDPC according to claim 1, characterized in that: Two convolutional layers are stacked before the Transformer, and a Dropout layer is added before connecting the Transformer encoder.