A blind recognition and analysis method for LDPC codes in non-cooperative signal scenarios

Abstract

The present invention belongs to the field of satellite communications, and discloses a method for blind recognition and analysis of LDPC codes in a non-cooperative signal scenario, which specifically includes: first, in a non-cooperative scenario, data to be decoded after operations such as demodulation and deinterleaving are obtained, and then a neural network is constructed, and a check matrix judgment of a received signal with a high bit error rate is performed in combination with a bit error rate coding and a noise injection mechanism, so as to realize the recognition and analysis of the check matrix of a non-cooperative LDPC signal, and finally the decoding and analysis of the non-cooperative LDPC signal is performed. The present invention significantly improves the recognition capability of LDPC codes in the case of high bit error rates, is suitable for blind recognition and analysis of received high-bit error LDPC codewords in non-cooperative situations, and improves the accuracy of blind recognition of LDPC codes in the case of high bit error rates.

Description

Technical Field

[0001] The present invention belongs to the field of satellite communications, and specifically relates to a blind recognition and analysis method for LDPC codes in a non-cooperative signal scenario. Background Art

[0002] Channel coding technology plays a very important role in modern digital communications. The basic principle is to add a redundant data sequence to the information data sequence to be sent in the transmitting part of the communication system, so that a constraint can be formed between the information data sequence and the added redundant data sequence, and then the data sequence with the constraint relationship is sent. By increasing the redundancy of the data, the receiving end can detect and repair some errors, thereby improving the reliability of communication.

[0003] Among them, LDPC codes can approach the Shannon Limit due to their ability to correct random errors, and if a decoding error occurs during the decoding process, it can also be detected later. Due to the excellent performance mentioned above, LDPC codes have been widely used in modern satellite communications, optical fiber communications, deep space communications and other scenarios.

[0004] Due to the high bit error rate in non-contractual reception situations, traditional blind recognition analysis methods are not applicable to non-cooperative situations. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides a blind identification and analysis method for LDPC codes in a non-cooperative signal scenario. The method is based on KAN+Transformer, is more effective for sequences with high bit errors, and gradually improves the bit error tolerance through a noise injection mechanism to achieve more accurate check matrix judgment.

[0006] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0007] The present invention is a method for blindly identifying and analyzing LDPC codes in a non-cooperative signal scenario, and the method specifically comprises the following steps:

[0008] Step 1: Fault verification: intercept various check matrices specified in IEEE 802.16e to form a non-cooperative binary signal , the mediated non-cooperative binary signal Perform fault check processing to obtain the processed fault check matrix ;

[0009] Step 2: Bit error rate coding: After obtaining the non-cooperative binary signal After that, the non-cooperative binary signal Make a preliminary error estimation;

[0010] Step 3: Normalize the fault check matrix obtained in step 1 Perform layer normalization operation and calculate the fault check matrix The mean and variance in the feature dimension are standardized and then scaled and translated to obtain the standardized output matrix ;

[0011] Step 4: Self-attention neural network KAN + Transformer: Use Kolmogorov Arnold Networks (KAN) to build a neural network, and use the back-propagation algorithm to learn the one-dimensional function to standardize the output matrix Processing is performed and finally the normalized output matrix is ​​obtained by normalization ;

[0012] Step 5, LSTM neural network: normalize the output matrix of step 4 Input LSTM neural network, and process the normalized output matrix through LSTM neural network After that, at each time Generate a hidden state ;

[0013] Step 6, fully connected neural network: the hidden state obtained in step 5 Perform linear transformation, and then use PReLU activation function to perform nonlinear mapping on the output after linear transformation. Stack multiple layers of linear transformation and PreLU activation function to gradually transform the hidden state Transformed into the final check matrix matching result;

[0014] Step 7, noise injection: Since it is a blind recognition analysis of non-cooperative signals, the bit error rate is relatively high. Therefore, considering the noise injection mechanism, the bit error rate threshold is gradually increased during the training process of step 1 to step 6 until the recognition accuracy is lower than the threshold of 90%. The bit error rate at this time is the limit bit error rate of this method.

[0015] Step 8. Output result: Output the received non-cooperative binary signal The probability of various check matrices is met, and the check matrix with the largest probability is considered to be the result.

[0016] A further improvement of the present invention is that: in step 1, upon receiving the demodulated non-cooperative binary signal After that, the fault check process is first performed, and the fault check matrix is ​​obtained according to the properties of the LDPC code :

[0017] ,

[0018] in, is the check matrix of LDPC code, is a non-cooperative binary signal, is a transposed, bit-error-free, non-cooperative binary signal Fault Check Matrix Correspondingly, the calculated fault check matrix is a 0 matrix, but for non-cooperative binary signals For example, due to its fault check matrix Unknown and there are bit errors, so it is necessary to check the fault matrix The faulty check matrix of the LDPC code is blindly identified by .

[0019] A further improvement of the present invention is that: in step 2, after obtaining the non-cooperative binary signal After that, for non-cooperative binary signals The preliminary noise and bit error rate estimation is as follows: Using convolutional neural network (CNN), the non-cooperative binary signal Extract the codewords for BER-encoding and preliminarily judge the non-cooperative binary signal Bit errors and bit errors, and determine non-cooperative binary signals Whether to participate in the subsequent process. If the bit error rate is judged to be greater than 20%, the subsequent process will not be participated.

[0020] A further improvement of the present invention is that: in step 3, the fault check matrix Perform layer normalization operation, which specifically includes the following steps:

[0021] Step 3.1: For each fault check matrix input , calculate the fault check matrix Mean and variance over feature dimensions:

[0022] ,

[0023] ,

[0024] in, is the mean, is the fault check matrix The number of For the A fault check matrix, is the variance;

[0025] Step 3.2: Use the mean calculated in step 3.1 and variance , for Fault Check Matrix Standardize and get standardized results :

[0026] ,

[0027] in, It is a positive number greater than 0; the purpose is to prevent the denominator from being 0;

[0028] Step 3.3: Using learnable parameters and parameters The results after standardization in step 3.2 To zoom and pan:

[0029] ,

[0030] in, is the fault check matrix The standardized output matrix output after layer normalization.

[0031] A further improvement of the present invention is that the step 4 specifically comprises the following steps:

[0032] Step 4.1. Definition for A one-dimensional function learned by the back-propagation algorithm ;

[0033] Step 4.2, Linear Projection: For each one-dimensional function, define a set of parameters to be learned by the back-propagation algorithm and , used to normalize the input output matrix Project to one dimension and get a one-dimensional projection vector :

[0034] ,

[0035] in, Represents the standardized output matrix Middle The feature vector at each position, represents transpose;

[0036] Step 4.3, nonlinear representation: standardize the output matrix Each feature vector in They are projected into one-dimensional functions respectively, and then pass through one-dimensional functions Get the intermediate representation vector :

[0037] ,

[0038] at this time, Transform to dimensional vector space, with KA's nonlinear representation, For space;

[0039] Step 4.4: Construction , , :To let the attention mechanism normalize the output matrix To calculate the similarity of different positions, each intermediate representation vector needs to be Projection to the dimension of attention , define three sets of learnable matrices:

[0040] ,

[0041] At this time, for each position :

[0042] ,

[0043] ,

[0044] ,

[0045] All , , After combining , used for subsequent scaling dot products;

[0046] Step 4.5, scaling dot product and normalization: get , , After that, , , Calculate the attention score using the formula:

[0047] ,

[0048] ,

[0049] ,

[0050] Then , and Normalize the rows to get the output matrix :

[0051] ,

[0052] in, Score for attention, Score for attention The attention weight obtained after softmax is in the range of [0,1], is the final calculated attention output vector, is a subscript, indicating different positions, Indicates The value vector of the position (Value), For the The key vector of the position (Key), For the The transpose of the query vector (Query) at the position.

[0053] A further improvement of the present invention is that in step 5, the LSTM neural network includes an input gate, a forget gate, an output gate and a cell state, and the normalized output matrix As the input sequence of the LSTM neural network, the LSTM neural network is used to iteratively process the normalized output matrix , after processing the normalized output matrix through the LSTM neural network After that, at each time Generate a hidden state , specifically including the following steps:

[0054] Step 5.1, the forget gate determines how much information is retained in the memory at the previous moment. The formula is as follows:

[0055] ,

[0056] in, Indicates the current moment, Indicates the last moment, For each time The input vector is is the hidden state of the previous moment, σ is the sigmoid function, the closer the value is to 1, the more information in the memory of the previous moment is retained, and the closer it is to 0, the more information in the memory of the previous moment is forgotten. is the weight matrix of the input vector, is the hidden state weight matrix at the previous moment, is the bias vector;

[0057] Step 5.2: Input gate to control current time How much new information will be written into the cell state, and the cell state will be updated based on the output of the forget gate and the input gate:

[0058] ,

[0059] in, is the cell state, For the Gate of Forgetfulness, represents element-wise multiplication, is the cell state at the previous moment, is the weight matrix for the input gate for the input vector, is the weight matrix for the input gate of the hidden state at the previous moment, is the weight matrix for candidate cell states for the input vector, is the weight matrix for the candidate cell state for the hidden state at the previous moment, is the bias vector for the input gate, is the bias vector for the candidate cell state. The formula shows that part of the old state information is retained after being filtered by the forget gate, and the input gate is opened to the new candidate information to merge it into the new cell state. ;

[0060] Step 5.3: Output gate controls the hidden state output to the downstream layer The amount of information, and the final hidden state is the updated input vector Perform tanh and then select the gate by the output gate:

[0061] ,

[0062] in, is the output of the LSTM neural network and serves as the input of the fully connected layer. is the weight matrix for linear mapping of the current input, is the bias term.

[0063] A further improvement of the present invention is that the step 6 specifically comprises the following steps:

[0064] Step 6.1: The fully connected layer of the fully connected neural network uses a weight matrix learned by the back-propagation algorithm. and the bias term Perform a linear transformation to transform the hidden state Mapping to the new feature space:

[0065] ,

[0066] in, It is the output of the last time of the LSTM neural network. Through this linear mapping, the features are transferred from the LSTM output space to the feature subspace suitable for subsequent processing;

[0067] Step 6.2: After the linear transformation in step 6.1, use the PReLU activation function for nonlinear mapping and introduce a parameter that can be learned through the back propagation algorithm. , prompt expression ability:

[0068] ,

[0069] in, is the adjustment coefficient, when When Used for The factor by which scaling is performed;

[0070] Step 6.3: Stack multiple fully connected layers to further improve the representation capability. Assume that there are There are fully connected layers, each of which includes "linear transformation + activation function". layer, , the calculation formula is as follows:

[0071] ,

[0072] ,

[0073] in, Indicates The output of the layer linear transformation, Represents the output at the previous moment, Represents the output at this moment, , is the output result of LSTM neural network, and It is The weights and bias parameters of the layers are continuously adjusted by stacking multiple layers of fully connected layers to output the LSTM neural network. Transform and abstract it, map it to the final output dimension, and determine whether the output result conforms to the check matrix.

[0074] The beneficial effects of the present invention are:

[0075] The present invention introduces KAN+Transformer into the blind recognition task of LDPC codes. The correlation between each element in the input sequence is globally modeled, so that the model can effectively capture long-range dependencies. This is different from the traditional blind recognition method that requires explicit synchronization or prior structural information. The model of the present invention can dynamically aggregate global features without additional feature engineering or in-depth understanding of the signal structure, significantly improving the adaptability and robustness to unknown signal environments.

[0076] The present invention adds a self-attention mechanism to extract signal features, improves the overall performance of the system, and overcomes the technical defect of traditional blind recognition schemes with low recognition accuracy in situations with high bit errors.

[0077] The present invention can gradually improve the tolerance to bit errors through the noise injection mechanism, making it easier to test and approach the performance upper limit of the model under extreme conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 is a flow chart of the method of the present invention.

[0079] Figure 2 This is a schematic diagram of the recognition accuracy for different check matrices under a 10% bit error rate.

[0080] Figure 3 It is a graph showing how the training indicators change during the training process.

[0081] Figure 4 This is a comparison chart of the recognition accuracy of the present invention and the traditional likelihood method under the same bit error rate.

[0082] Figure 5 This is a comparison chart of the recognition accuracy of the present invention and the traditional likelihood method under the condition of the same number of codewords. DETAILED DESCRIPTION

[0083] The following will disclose the embodiments of the present invention with drawings. For the purpose of clear description, many practical details will be described together in the following description. However, it should be understood that these practical details should not be used to limit the present invention. That is to say, in some embodiments of the present invention, these practical details are not necessary.

[0084] like Figure 1 As shown, the present invention is a blind recognition and analysis method of LDPC codes in a non-cooperative signal scenario, comprising the following steps:

[0085] Step 1: Fault verification: intercept various check matrices specified in IEEE 802.16e to form a non-cooperative binary signal , after receiving the demodulated non-cooperative binary signal After that, the mediated non-cooperative binary signal Perform fault check processing and obtain the fault check matrix based on the properties of the LDPC code :

[0086] ,

[0087] in, is the check matrix of LDPC code, is a non-cooperative binary signal, is a transposed, bit-error-free, non-cooperative binary signal Fault Check Matrix Correspondingly, the calculated fault check matrix is a 0 matrix, but for non-cooperative binary signals For example, due to its fault check matrix Unknown and there are bit errors, so it is necessary to check the fault matrix The faulty check matrix of the LDPC code is blindly identified by .

[0088] Step 2: Use convolutional neural network to transform non-cooperative binary signals Extract the codewords for BER-encoding and preliminarily judge the non-cooperative binary signal In the judgment process, if the probability of the codeword being an error code is greater than 30%, the codeword will not participate in the subsequent process, reducing the impact of the error code on the subsequent process and improving the tolerance of this method to the error code. Define a one-dimensional convolutional network , the formula is as follows:

[0089]

[0090] in, are learnable weights, is the learnable bias, is a non-cooperative binary signal, For transposition, is sigmoid activation, so Between [0,1], is the probability that the codeword is an error code. , then the codeword does not participate in the subsequent process.

[0091] Step 3: The fault check matrix obtained in step 1 Perform layer normalization operation and calculate the fault check matrix The mean and variance in the feature dimension are standardized and then scaled and translated to obtain the standardized output matrix The layer normalization operation specifically includes the following steps:

[0092] Step 3.1: For each fault check matrix input , calculate the fault check matrix Mean and variance over feature dimensions:

[0093] ,

[0094] ,

[0095] in, is the mean, is the fault check matrix The number of For the A fault check matrix, is the variance;

[0096] Step 3.2: Use the mean calculated in step 3.1 and variance , for Fault Check Matrix Standardize and get standardized results :

[0097] ,

[0098] in, is a positive number greater than 0, which is a very small positive number. The purpose is to prevent the denominator from being 0;

[0099] Step 3.3: Using learnable parameters and parameters The results after standardization in step 3.2 To zoom and pan:

[0100] ,

[0101] in, is the fault check matrix The standardized output matrix output after layer normalization.

[0102] Step 4: Use the Kolmogorov Arnold Networks (KAN) to construct a neural network and use the back-propagation algorithm to learn the one-dimensional function to standardize the output matrix Processing is performed and finally the normalized output matrix is ​​obtained by normalization , specifically including the following steps:

[0103] Step 4.1, define a one-dimensional function. Any multidimensional continuous function revealed by the Kolmogorov-Arnold representation theorem can be approximated by a finite number of "one-dimensional functions" and linear combinations / composites. Therefore, first define for A one-dimensional function learned by the back-propagation algorithm ;

[0104] Step 4.2, Linear Projection: For each one-dimensional function, define a set of parameters to be learned by the back-propagation algorithm and , used to normalize the input output matrix Project to one dimension and get a one-dimensional projection vector :

[0105] ,

[0106] in, Represents the standardized output matrix Middle The feature vector at each position, represents transpose;

[0107] Step 4.3, nonlinear representation: standardize the output matrix Each feature vector in They are projected into one-dimensional functions respectively, and then pass through one-dimensional functions Get the intermediate representation vector :

[0108] ,

[0109] at this time, Transform to dimensional vector space, with KA's nonlinear representation, For space;

[0110] Step 4.4: Construction , , :To let the attention mechanism normalize the output matrix To calculate the similarity of different positions, each intermediate representation vector needs to be Projection to the dimension of attention , define three sets of learnable matrices:

[0111] ,

[0112] At this time, for each position :

[0113] ,

[0114] ,

[0115] ,

[0116] All , , After combining , used for subsequent scaling dot products;

[0117] Step 4.5, scaling dot product and normalization: get , , After that, , , Calculate the attention score using the formula:

[0118] ,

[0119] ,

[0120] ,

[0121] Then , and Normalize the rows to get the output matrix :

[0122] ,

[0123] in, Score for attention, Score for attention The attention weight obtained after softmax is in the range of [0,1], is the final calculated attention output vector, is a subscript, indicating different positions, Indicates The value vector of the position (Value), For the The key vector of the position (Key), For the The transpose of the query vector (Query) at the position.

[0124] Step 5, LSTM neural network: normalize the output matrix of step 4 Input LSTM neural network, and process the normalized output matrix through LSTM neural network After that, at each time Generate a hidden state .

[0125] The LSTM neural network includes an input gate, a forget gate, an output gate, and a cell state. The normalized output matrix As the input sequence of the LSTM neural network, the LSTM neural network is used to iteratively process the normalized output matrix The core of LSTM is to "read, write" and "forget" information through the gating mechanism. It consists of four key components: forget gate, input gate, output gate and cell state. After the normalized output matrix is ​​processed by the LSTM neural network After that, at each time Generate a hidden state , specifically including the following steps:

[0126] Step 5.1, the forget gate determines how much information is retained in the memory at the previous moment. The formula is as follows:

[0127] ,

[0128] in, Indicates the current moment, Indicates the last moment, For each time The input vector is is the hidden state of the previous moment, σ is the sigmoid function, the closer the value is to 1, the more information in the memory of the previous moment is retained, and the closer it is to 0, the more information in the memory of the previous moment is forgotten. is the weight matrix of the input vector, is the hidden state weight matrix at the previous moment, is the bias vector;

[0129] Step 5.2: Input gate to control current time How much new information will be written into the cell state, and the cell state will be updated based on the output of the forget gate and the input gate:

[0130] ,

[0131] in, is the cell state, For the Gate of Forgetfulness, represents element-wise multiplication, is the cell state at the previous moment, is the weight matrix for the input gate for the input vector, is the weight matrix for the input gate of the hidden state at the previous moment, is the weight matrix for candidate cell states for the input vector, is the weight matrix for the candidate cell state for the hidden state at the previous moment, is the bias vector for the input gate, is the bias vector for the candidate cell state. The formula shows that part of the old state information is retained after being filtered by the forget gate, and the input gate is opened to the new candidate information to merge it into the new cell state. ;

[0132] Step 5.3: Output gate controls the hidden state output to the downstream layer The amount of information, and the final hidden state is the updated input vector Perform tanh and then select the gate by the output gate:

[0133] ,

[0134] in, is the output of the LSTM neural network and serves as the input of the fully connected layer. is the weight matrix for linear mapping of the current input, is the bias term.

[0135] Step 6, fully connected neural network: the hidden state obtained in step 5 Perform linear transformation, and then use PReLU activation function to perform nonlinear mapping on the output after linear transformation. Stack multiple layers of linear transformation and PreLU activation function to gradually transform the hidden state Transformed into the final check matrix matching result;

[0136] The step 6 specifically includes the following steps:

[0137] Step 6.1: The fully connected layer of the fully connected neural network uses a weight matrix learned by the back-propagation algorithm. and the bias term Perform a linear transformation to transform the hidden state Mapping to the new feature space:

[0138] ,

[0139] in, It is the output of the last time of the LSTM neural network. Through this linear mapping, the features are transferred from the LSTM output space to the feature subspace suitable for subsequent processing;

[0140] Step 6.2: After the linear transformation in step 6.1, use the PReLU activation function for nonlinear mapping and introduce a parameter that can be learned through the back propagation algorithm. , prompt expression ability:

[0141] ,

[0142] in, is the adjustment coefficient, when When Used for The factor by which scaling is performed;

[0143] Step 6.3: Stack multiple fully connected layers to further improve the representation capability. Assume that there are There are fully connected layers, each of which includes "linear transformation + activation function". layer, , the calculation formula is as follows:

[0144] ,

[0145] ,

[0146] in, Indicates The output of the layer linear transformation, Represents the output at the previous moment, Represents the output at this moment, , is the output result of LSTM neural network, and It is The weights and bias parameters of the layers are continuously adjusted by stacking multiple layers of fully connected layers to output the LSTM neural network. Transform and abstract it, map it to the final output dimension, and determine whether the output result conforms to the check matrix, such as Figure 2 As shown, the horizontal and vertical axes are the code rate of the check matrix. It can be seen that when the bit error rate is 10%, the recognition result of the present invention is basically consistent with the real check matrix.

[0147] Step 7, noise injection: Since it is a blind recognition analysis of non-cooperative signals, the bit error rate is relatively high. Therefore, the noise injection mechanism is considered. During the training process of steps 1 to 6, the threshold of the bit error rate is gradually increased to enhance the training performance of the system until the recognition accuracy is lower than the threshold of 90%. The bit error rate at this time is the limit bit error rate of this method. Figure 3 As shown, the horizontal axis represents the number of training times, the vertical axis represents the maximum bit error rate, and the training index refers to loss. As the number of training times increases, the maximum bit error rate of the present invention also increases. When the maximum bit error rate reaches 18%, the recognition accuracy of the present invention can no longer reach the threshold of 90%, so it can be considered that the bit error rate of 18% is the maximum bit error rate of this method;

[0148] Step 8. Output result: Output the received non-cooperative binary signal The probability of various check matrices is met, and the check matrix with the largest probability is considered to be the result.

[0149] Figure 4 , Figure 5 This is a comparison chart of the recognition effect of the present invention and the traditional likelihood method. The check matrix specified by IEEE 802.16e is used. When the code length is 576, the , , , , , The check matrices corresponding to the six code rates construct the transmission signal and add bit errors to it. Figure 4 It can be seen that when the bit error rate is fixed at 14%, the recognition accuracy of the present invention is higher than that of the traditional likelihood method, and as the number of codewords increases, the recognition accuracy also increases; Figure 5 It can be seen that when the number of codeword inputs is the same, as the bit error rate increases, the recognition accuracy of the present invention is better than that of the traditional likelihood method, wherein the illustration of the present invention is KAT and the traditional likelihood method is Likehood.

[0150] 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 principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for blind identification and analysis of LDPC codes in a non-cooperative signal scenario, characterized in that: The LDPC code blind recognition and analysis method specifically comprises the following steps: Step 1: Fault verification: intercept various check matrices specified in IEEE 802.16e to form a non-cooperative binary signal , the mediated non-cooperative binary signal Perform fault check processing to obtain the processed fault check matrix ; Step 2: Bit error rate coding: After obtaining the non-cooperative binary signal After that, the non-cooperative binary signal Make a preliminary error estimation; Step 3: Normalize the fault check matrix obtained in step 1 Perform layer normalization operation and calculate the fault check matrix The mean and variance in the feature dimension are standardized and then scaled and translated to obtain the standardized output matrix ; Step 4: Self-attention neural network KAN + Transformer: Use the Kolmogorov-Arnold representation theorem to construct a neural network, and use the back-propagation algorithm to learn the one-dimensional function to standardize the output matrix Processing is performed and finally the normalized output matrix is ​​obtained by normalization ; Step 5, LSTM neural network: normalize the output matrix of step 4 Input LSTM neural network, and process the normalized output matrix through LSTM neural network After that, at each time Generate a hidden state ; Step 6, fully connected neural network: the hidden state obtained in step 5 Perform linear transformation, and then use PReLU activation function to perform nonlinear mapping on the output after linear transformation. Stack multiple layers of linear transformation and PreLU activation function to gradually transform the hidden state Transformed into the final check matrix matching result; Step 7, noise injection: during the training process from step 1 to step 6, the bit error rate threshold is gradually increased until the recognition accuracy is lower than the 90% threshold. The bit error rate at this time is the limit bit error rate. Step 8. Output result: Output the received non-cooperative binary signal The probability of various check matrices is met, and the check matrix with the largest probability is considered to be the result.

2. The method for blind identification and analysis of LDPC codes in a non-cooperative signal scenario according to claim 1, characterized in that: In step 1, upon receiving the demodulated non-cooperative binary signal After that, the fault check process is first performed, and the fault check matrix is ​​obtained according to the properties of the LDPC code : ， in, is the check matrix of LDPC code, is a non-cooperative binary signal, is the transpose, for non-cooperative binary signals , fault check matrix Unknown and there are bit errors, it is necessary to check the fault matrix The faulty check matrix of the LDPC code is blindly identified by .

3. The method for blind identification and analysis of LDPC codes in a non-cooperative signal scenario according to claim 1, characterized in that: In step 2, after obtaining the non-cooperative binary signal After that, for non-cooperative binary signals The preliminary error code estimation is as follows: using convolutional neural network, the non-cooperative binary signal Extract the codewords for BER-encoding and preliminarily judge the non-cooperative binary signal In the calculation, the probability that the code word is an error code is greater than 30%, and the code word does not participate in the subsequent process.

4. The method for blind identification and analysis of LDPC codes in a non-cooperative signal scenario according to claim 1, characterized in that: In step 3, the fault check matrix Perform layer normalization operation, which specifically includes the following steps: Step 3.1: For each fault check matrix input , calculate the fault check matrix Mean and variance over feature dimensions: ， ， in, is the mean, is the fault check matrix The number of For the A fault check matrix, is the variance; Step 3.2: Use the mean calculated in step 3.1 and variance , for Fault Check Matrix Standardize and get standardized results : ， in, is a positive number greater than 0; Step 3.3: Using learnable parameters and parameters The results after standardization in step 3.2 To zoom and pan: ， in, is the fault check matrix The standardized output matrix output after layer normalization.

5. The method for blind identification and analysis of LDPC codes in a non-cooperative signal scenario according to claim 1, characterized in that: The step 4 specifically includes the following steps: Step 4.

1. Definition for A one-dimensional function learned by the back-propagation algorithm ; Step 4.2, Linear Projection: For each one-dimensional function, define a set of parameters to be learned by the back-propagation algorithm and , used to normalize the input output matrix Project to one dimension and get a one-dimensional projection vector : ， in, Represents the standardized output matrix Middle The feature vector at each position, represents transpose; Step 4.3, nonlinear representation: standardize the output matrix Each feature vector in They are projected into one-dimensional functions respectively, and then pass through one-dimensional functions Get the intermediate representation vector : ， at this time, Transform to dimensional vector space, For space; Step 4.4: Construction , , :Each intermediate representation vector Projection to the dimension of attention , define three sets of learnable matrices: ， At this time, for each position : ， ， ， All , , After combining ; Step 4.5, scaling dot product and normalization: get , , After that, , , Calculate the attention score using the formula: ， , ， Then , and Normalize the rows to get the output matrix : ， in, Score for attention, Score for attention The attention weight obtained after softmax is in the range of [0,1], is the final calculated attention output vector, Indicates The value vector of the positions, For the The key vector for each position, For the The transpose of the query vector at position .

6. The method for blind identification and analysis of LDPC codes in a non-cooperative signal scenario according to claim 1, characterized in that: In step 5, the LSTM neural network includes an input gate, a forget gate, an output gate, and a cell state, and the normalized output matrix As the input sequence of the LSTM neural network, the LSTM neural network is used to iteratively process the normalized output matrix , after processing the normalized output matrix through the LSTM neural network After that, at each time Generate a hidden state , specifically including the following steps: Step 5.1, the forget gate determines how much information is retained in the memory at the previous moment. The formula is as follows: ， in, Indicates the current moment, Indicates the last moment, For each time The input vector is is the hidden state of the previous moment, σ is the sigmoid function, the closer the value is to 1, the more information in the memory of the previous moment is retained, and the closer it is to 0, the more information in the memory of the previous moment is forgotten. is the weight matrix of the input vector, is the hidden state weight matrix at the previous moment, is the bias vector; Step 5.2: Input gate to control current time How much new information will be written into the cell state, and the cell state will be updated based on the output of the forget gate and the input gate: ， in, is the cell state, For the Gate of Forgetfulness, represents element-wise multiplication, is the cell state at the previous moment, is the weight matrix for the input gate for the input vector, is the weight matrix for the input gate of the hidden state at the previous moment, is the weight matrix for candidate cell states for the input vector, is the weight matrix for the candidate cell state for the hidden state at the previous moment, is the bias vector for the input gate, is the bias vector for the candidate cell state. The formula shows that part of the old state information is retained after being filtered by the forget gate, and the input gate is opened to the new candidate information to merge it into the new cell state. ; Step 5.3: Output gate controls the hidden state output to the downstream layer The amount of information, and the final hidden state is the updated input vector Perform tanh and then select the gate by the output gate: ， in, is the output of the LSTM neural network and serves as the input of the fully connected layer. is the weight matrix for linear mapping of the current input, is the bias term.

7. The method for blind identification and analysis of LDPC codes in a non-cooperative signal scenario according to claim 1, characterized in that: The step 6 specifically includes the following steps: Step 6.1: The fully connected layer of the fully connected neural network uses a weight matrix learned by the back-propagation algorithm. and the bias term Perform a linear transformation to transform the hidden state Mapping to the new feature space: ， in, The output of the last time of the LSTM neural network; Step 6.2: After the linear transformation in step 6.1, use the PReLU activation function for nonlinear mapping and introduce a parameter that can be learned through the back propagation algorithm. , prompt expression ability: ， in, is the adjustment coefficient, when When Used for The factor by which scaling is performed; Step 6.3: Stack multiple fully connected layers to improve representation capabilities. Assume that there are Layer fully connected layer, for the layer, , the calculation formula is as follows: ， ， in, Indicates The output of the layer linear transformation, Represents the output at the previous moment, Represents the output at this moment, , is the output result of LSTM neural network, and It is The weights and bias parameters of the layers are continuously adjusted by stacking multiple layers of fully connected layers to output the LSTM neural network. Transform and abstract it, map it to the final output dimension, and determine whether the output result conforms to the check matrix.

Images