Individual identification method under satellite modulation based on multi-task decoupled learning

Abstract

This invention discloses a satellite-based individual identification method under varying modulation conditions based on multi-task decoupling learning, primarily addressing the poor performance of existing technologies in extracting individual features under varying modulation conditions. The implementation scheme involves: receiving downlink communication link signals passing through an observation orbit and dividing them into training and test sets; establishing two knowledge extraction blocks and one shared knowledge extraction block, each composed of a complex-valued multi-scale embedding unit and an attention-gated unit, and connecting these three knowledge extraction blocks in parallel to form a feature extraction module; stacking the feature extraction modules and connecting them to a classifier to form a multi-task decoupling network; inputting the training set data into the multi-task decoupling network for training; inputting the test set data into the trained multi-task decoupling network, and outputting the modulation type and satellite identity results. This invention mitigates feature loss caused by modulation variations, can extract rich individual information under varying modulation modes, enhances the recognition performance of complex signals, and can be used for electronic reconnaissance and identification of source satellites.

Description

Technical Field

[0001] This invention belongs to the field of communications, and specifically relates to a method for individual identification under satellite modulation, which can be used for electronic reconnaissance and identification of source satellites. Background Technology

[0002] In the field of satellite communications, satellite identification has become an important research direction. Currently, satellite identification technology mainly relies on pattern analysis or feature extraction of satellite telemetry and control signals to effectively verify the satellite's identity. However, in complex electromagnetic environments, accurately identifying the source satellite of a signal is a crucial prerequisite for in-depth interpretation and analysis of communication information. Therefore, there is an urgent need for an improved satellite identification method to enhance the ability and accuracy of satellite signal analysis in variable environments.

[0003] Patent application CN202310977450.5 discloses a "satellite identification method and device based on feature fusion". Its implementation involves: first, receiving and preprocessing the IQ baseband signal of the device to be identified to obtain a stable signal segment; then, constructing a satellite identification model based on a convolutional neural network and a parallel GRU network; next, extracting IQ and temporal features from the stable signal segment; finally, combining the IQ features, temporal features, and the IQ baseband signal through convolution, and classifying the result using a fully connected layer; based on the identification result, comparing it with a legitimate satellite radio frequency fingerprint database to confirm the legitimacy of the device to be identified. This method requires a large number of correct signals to construct the fingerprint database, making the process quite complex and difficult to update. Furthermore, it does not consider the relationship between signal modulation type and individual features in individual identification, making it difficult to apply in practical satellite identification.

[0004] Patent application number 202310542492.6 discloses a "satellite communication radiation source identification method." Its implementation scheme is as follows: First, the satellite communication modulated signal is converted into digital signal data; the digital signal data is preprocessed and feature analyzed, and sub-network layers are obtained by training the data using LSTM and CNN networks; filters are added to the sub-network layer results, a channel attention mechanism is embedded, a fusion network layer is constructed, a fully connected layer is added, data graph features are fused, and training is performed; the sub-network layers, SE-Net layer, and fusion network layer are connected, the SE-Net layer and fusion network layer are trained, and the identification result is output. Although this method considers the multidimensional features of the signal, it still separates the relationship between the signal modulation type and individual characteristics, making it difficult to handle satellite identification under modulation variations. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a satellite variable modulation individual identification method based on multi-task decoupling learning, so as to reduce feature loss during modulation changes and improve the accuracy of satellite identification in variable modulation scenarios.

[0006] The technical approach to achieving the objective of this invention is as follows: by decoupling the target features into modulation features and individual features, extracting individual steady-state features under different modulation types, reducing the loss of individual features caused by modulation changes; and deeply mining the subtle features in the signal to improve the accuracy of recognition.

[0007] Based on the above ideas, the technical solution of the present invention includes the following steps:

[0008] (1) Use a broadband receiver to receive downlink communication link signals passing through a certain observation orbit, preprocess them, and then divide the preprocessed data into training set and test set in a 7:3 ratio;

[0009] (2) Construct a multi-task decoupling network:

[0010] 2a) Establish a knowledge extraction block consisting of multiple complex-valued multi-scale embedding units connected in parallel and attention-gated units connected in series;

[0011] 2b) The three knowledge extraction blocks are connected in parallel to form a feature extraction module, which is used to extract modulation knowledge, shared knowledge and individual knowledge respectively;

[0012] 2c) Stack multiple feature extraction modules in series to form a feature extraction layer;

[0013] 2d) Establish two classifiers, each composed of fully connected layers, to classify modulation types and individual features, respectively. Connect these two classifiers to the two outputs of the feature extraction layer to form a multi-task decoupling network.

[0014] (3) Input the training set data into the multi-task decoupling network, and use the gradient descent method to train the model until the loss function tends to converge, and obtain the trained multi-task decoupling network.

[0015] (4) Input the test set data into the trained multi-task decoupling network and output the modulation type result and satellite identity result.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. The multi-task decoupling network designed in this invention employs multi-scale complex-valued embedding units to extract proprietary or shared knowledge, and achieves information sharing and coupling through multi-task attention gating units. Therefore, it can extract rich individual information even with modulation scheme variations, mitigating feature loss caused by modulation changes and enabling satellite identification with modulation scheme variations.

[0018] 2. The multi-task decoupling network constructed in this invention contains a knowledge extraction block for extracting shared knowledge, which can effectively associate modulation features during the extraction of individual features, resulting in a good synergistic effect among multiple tasks, further enhancing the ability to process complex signals and significantly improving recognition performance. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0020] Figure 2 This is a diagram of the multi-task decoupling network structure constructed in this invention;

[0021] Figure 3 for Figure 2 Structure diagram of complex-valued multi-scale embedding unit in a multi-task decoupling network;

[0022] Figure 4 for Figure 2 Attention gating unit structure diagram in multi-task decoupling network;

[0023] Figure 5 This is a simulation diagram showing the effect of using the present invention for individual satellite identification. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] Reference Figure 1 The implementation steps for this example are as follows:

[0026] Step 1: Use a broadband receiver to receive downlink communication link signals passing through a certain observation orbit, preprocess them, and divide them into training and test sets.

[0027] 1.1) Configure a broadband receiver to monitor information from a specific orbit. When a satellite enters the receiver's observable range, the receiver receives the downlink communication link signal it transmits and stores it in the workstation.

[0028] This example assumes that the receiver station has accumulated signals from N known satellites, each containing 1000 data segments, and each data segment contains 8192 complex sampling points.

[0029] 1.2) Preprocess the downlink communication link signals stored in the workstation:

[0030] 1.2.1) Let x be the downlink signal sequence of the p-th satellite in segment q. pq for:

[0031] x pq ={x pq (1),x pq (2),…,x pq (i),…,x pq (n)}

[0032] Where n represents the total number of sampling points of the signal, x pq (i) represents the i-th sampling point in the received downlink communication link signal;

[0033] 1.2.2) Perform a DC-debiasing operation on the downlink communication link signal stored in the workstation to remove the DC bias of the signal, obtaining the DC-debiased downlink communication link signal:

[0034] y pq ={y pq (1),y pq (2),…,y pq (i),…,y pq (n)}

[0035] Among them, y pq (i)=x pq (i)-μ pq μ represents the i-th sampling point of the signal sequence after DC removal. pq This represents the mean of the downlink communication link signal sequence.

[0036] 1.2.3) Normalize the downlink communication link signal after DC removal to control the signal value range within a specific interval, resulting in the normalized downlink communication link signal:

[0037] z pq ={z pq (1),z pq (2),…,z pq (i),…,z pq (n)}

[0038] in, Let y represent the i-th sampling point of the normalized signal sequence. max x represents the maximum instantaneous amplitude of the sequence after DC removal. min This represents the minimum instantaneous amplitude of the sequence after DC removal.

[0039] 1.3) Divide the preprocessed data into training and test sets in a 7:3 ratio.

[0040] Step 2: Construct a multi-task decoupling network.

[0041] Reference Figure 2 The implementation of this step is as follows:

[0042] 2.1) Construct a complex-valued multi-scale embedding unit consisting of three parallel complex convolutional layers. The first complex convolution has a kernel size of 1×3 and a stride of 2; the second complex convolution has a kernel size of 1×5 and a stride of 2; and the third complex convolution has a kernel size of 1×7 and a stride of 4. The activation function for all three complex convolutional layers is the complex ReLU function, such as... Figure 3 As shown;

[0043] 2.2) Establish an attention gating unit consisting of a fully connected layer and residuals in parallel. The fully connected layer is used to extract the weights of each knowledge vector, and the residuals are used to obtain the original knowledge vectors, such as... Figure 4 As shown;

[0044] 2.3) Multiple complex-valued multi-scale embedding units are connected in parallel and then connected in series with attention gating units to form a knowledge extraction block. The knowledge extraction block is then copied into three copies and named as the first knowledge extraction block A, the second knowledge extraction block B, and the shared knowledge extraction block S, respectively, for extracting modulation knowledge, individual knowledge, and shared knowledge.

[0045] 2.4) Connect the first knowledge extraction block A, the second knowledge extraction block B, and the shared knowledge extraction block S in parallel to form a feature extraction module;

[0046] 2.5) Stack the six feature extraction modules in series to form a feature extraction layer. This layer contains two outputs, which are used for modulation type classification and individual feature classification, respectively.

[0047] 2.6) Construct a classifier consisting of fully connected layers. The number of intermediate neurons in the fully connected layers shall be no less than 128. The number of output neurons in the two fully connected layers shall be N1 and N2, respectively, where N1 is the number of modulation categories and N2 is the number of satellite individuals.

[0048] 2.7) Connect the two classifiers to the two outputs of the feature extraction layer respectively to form a multi-task decoupling network.

[0049] Step 3: Input the training set data into the multi-task decoupling network and train it using gradient descent.

[0050] 3.1) Set the learning rate to 0.0001, the batch size to 16, the maximum number of training iterations to 50, the optimizer to Adam, and the cross-entropy function to be used as the network's loss function;

[0051] 3.2) Input the training set data into the multi-task decoupling network for forward propagation and calculate the predicted output:

[0052] 3.2.1) Given the normalized downlink communication link signal input sequence x;

[0053] x = z pq ={[I1,Q1] T [I2,Q2] T ,…,[I i Q i ] T ,…,[I n Q n ] T}

[0054] Among them [I] i Q i ] T The column vector x represents the combination of the real part and the imaginary part of the i-th sample point of the preprocessed downlink communication link signal. i , [·] T Indicates taking the transpose, I i =real(z) pq (i)), real(·) denotes taking the real part, Q i =imag(z pq (i)), imag(·) denotes taking the imaginary part;

[0055] 3.2.2) Construct the complex-valued convolution kernel: W = [W I W Q ] T , where W I With W Q Let these represent the real and imaginary parts of the complex convolution, respectively:

[0056] 3.2.3) Convolve the input signal sequence x with a complex-valued convolution kernel W, that is, use parallel complex-valued convolution kernels of three different scales (1×3, 1×5, and 1×7) to perform multi-scale feature extraction on the signal, and obtain the knowledge vector f. θ (x):

[0057]

[0058] in, This represents the convolution operation.

[0059] 3.2.4) The signal input sequence x is processed by the i-th complex-valued multi-scale embedding unit connected in parallel in the first knowledge extraction block A to extract the knowledge vector E. A,i and m A The knowledge vectors are concatenated to obtain the concatenated feature vector E. A :

[0060]

[0061] in, m A E represents the number of complex-valued multi-scale embedding units in the first knowledge extraction block A, d represents the dimension of information embedding, and E represents the number of complex-valued multi-scale embedding units in the first knowledge extraction block A. A,i =f θ (x), i∈(1,m) A );

[0062] 3.2.5) The signal input sequence x is processed by the j-th complex-valued multi-scale embedding unit in parallel within the second knowledge extraction block B to extract the knowledge vector E. B,j and m B The knowledge vectors are concatenated to obtain the concatenated feature vector E. B :

[0063]

[0064] in, m B E represents the number of complex-valued multi-scale embedding units in the second knowledge extraction block B; d represents the dimension of information embedding. B,j =f θ (x), j∈(1,m) B ).

[0065] 3.2.6) The signal input sequence x is processed by the k-th complex-valued multi-scale embedding unit in parallel within the shared knowledge extraction block S to extract the knowledge vector E. S,k and m S The knowledge vectors are concatenated to obtain the concatenated feature vector E. S :

[0066]

[0067] in, m S E represents the number of complex-valued multi-scale embedding units in the shared knowledge extraction block S; d represents the dimension of information embedding; E S,k =f θ (x), k∈(1,m) S );

[0068] 3.2.7) The above two eigenvectors E A With ES Perform concatenation, and then concatenate the feature vectors [E] A E S The attention value corresponding to each knowledge vector is obtained by using the gating unit input as the first knowledge extraction block A. These attention values ​​are then concatenated to obtain the attention weight vector g of the knowledge vector. A :

[0069]

[0070] in, E α This represents the α-th knowledge vector in the input feature vector, where α∈(1,m) A +m S ), ||·|| p Let p be the p-norm of the eigenvector, where p = 2;

[0071] 3.2.8) Transfer the attention weight vector g A The gating unit input [E] of the first knowledge extraction block A A E S ] Perform matrix multiplication to obtain the modulation knowledge output G of the first knowledge extraction block A. A :

[0072] G A =g A ×[E A E S ]

[0073] 3.2.9) The above two eigenvectors E B With E S Perform concatenation, and then concatenate the feature vectors [E] B E S The attention value corresponding to each knowledge vector is obtained by using the gating unit input as the second knowledge extraction block B. These attention values ​​are then concatenated to obtain the attention weight vector g of the knowledge vector. B :

[0074]

[0075] in, E α This represents the β-th knowledge vector in the input feature vector, where β∈(1,m). B +m S );

[0076] 3.2.10) Transfer the attention weight vector g B Input [E] to the second knowledge extraction block B gate unit B ES ] Perform matrix multiplication to obtain the individual knowledge output G of the second knowledge extraction block B. B :

[0077] G B =g B ×[E B E S ]

[0078] 3.2.11) The above three eigenvectors E A E B With E S Perform concatenation, and then concatenate the feature vectors [E] A E B E S The attention value corresponding to each knowledge vector is obtained by using the gating unit input as the input to the shared knowledge extraction block S. These attention values ​​are then concatenated to obtain the attention weight vector g of the knowledge vector. S :

[0079]

[0080] in, E γ This represents the γ-th knowledge vector in the input feature vector, where γ∈(1,m) A +m B +m S );

[0081] 3.2.12) Transfer the attention weight vector g S Input to the shared knowledge extraction block S gated unit [E] A E B E S Perform matrix multiplication to obtain the shared knowledge output G of the shared knowledge extraction block S. S :

[0082] G S =g S ×[E A E B E S ]

[0083] 3.2.13) Multiple feature extraction modules are stacked in series to form a feature extraction layer to output modulation features and individual features. Specifically, the signal input sequence x is used as the input of the first stacked feature extraction module, and the first knowledge extraction block A1 of the first layer outputs modulation knowledge G. A1 As input to the first knowledge extraction block A2 of the second layer; the output individual knowledge G of the second knowledge extraction block B1 of the first layer. B1As input to the second knowledge extraction block B2 of the second layer; the shared knowledge G output by the first layer shared extraction block S1 is used. S1 This serves as the input to the second layer shared extraction block S2; and so on, after six layers of forward propagation, the final output modulation feature G is obtained. A and individual characteristics G B ;

[0084] 3.2.14) The modulation features are processed by a classifier to obtain the modulation recognition result y. mod Individual features are processed by a classifier to obtain the individual identification result y. id :

[0085] y mod =f φ (G A )

[0086] y id =f φ (G B )

[0087] Among them, y mod ∈(1,N1), where N1 is the total number of modulation categories, y id ∈(1,N2), where N2 is the total number of individual satellites;

[0088] 3.3) Calculate the loss value (Loss) between the predicted class and the true class output by the network using the loss function;

[0089]

[0090] Where N represents the batch size in training, i represents the sequence number of the training sample in the training batch, M1 represents the number of satellite signal modulation categories in the training set, M2 represents the number of individual satellite signal categories in the training set, and c represents the sequence number of the training sample in the downlink signal category of the individual satellite.

[0091] y ic Let y represent the symbol function for modulation identification. If the true modulation category of the i-th sample is c, then y ic The value is 1 if it is not 1, otherwise the value is 0.

[0092] z ic The sign function representing individual identification is z if the true individual category of the i-th sample is c. ic It takes the value 1, otherwise it takes the value 0.

[0093] This represents the probability that the predicted modulation category of the i-th sample is c after passing through the multi-task decoupling network;

[0094] This represents the probability that the predicted individual category of the i-th sample is c after passing through the multi-task decoupling network;

[0095] 3.4) Use the chain rule to backpropagate the loss value to each layer, calculate the gradient of the weights and biases of each layer to update the network parameters;

[0096] 3.4.1) Let the loss function be L. If the output of a certain layer is a, and the weighted input of that layer is z, then a = f(z). 3.4.2) Calculate the gradient of the loss function with respect to the weighted input. for:

[0097]

[0098] 3.4.3) Calculate the gradient of the loss with respect to the weights. gradient of bias

[0099]

[0100] 3.4.4) Update the network weights and biases using the calculated gradients:

[0101]

[0102] Where η is the learning rate;

[0103] 3.5) Repeat steps 3.2) to 3.4) until the maximum number of iterations is reached or the loss function converges to a small value; the trained multi-task decoupling network is obtained.

[0104] Step 4: Input the test set data into the trained multi-task decoupling network, and output the modulation type result and satellite identity result. Specifically, use the signal input sequence x as the input to the feature extraction layer, and output the modulation feature G of the satellite signal. A and individual characteristics of the signal G B The modulation features are used as input to the signal modulation classification layer, and the modulation results are output. Individual features are used as input to the satellite individual identification classification layer, and the individual identification results are output.

[0105] The effects of this invention will be further illustrated below with simulation experiments:

[0106] 1. Simulation experimental conditions:

[0107] The hardware platform for the simulation experiment of this invention is as follows: the processor is an i9-11900F CPU with a main frequency of 2.5GHz, the memory is 32GB, and the graphics card is an NVIDIA GeForce RTX 3090 GPU.

[0108] The software platform for the simulation experiment of this invention is: Windows 10 operating system, Python 3.8 and PyTorch 0.12.0.

[0109] The input data used in the simulation experiment of this invention is simulated S-band satellite downlink IQ signals. The intermediate frequency of each satellite downlink IQ signal is 70MHz, and it is sampled at a sampling rate of 56Mbps, resulting in a sample signal sequence size of 8192×2.

[0110] The prior condition for the experiment was that, before time t, the receiver had successfully accumulated signals from 10 known satellites, each containing 1000 data segments, and each data segment containing 8192 sampling points. These satellite individuals were simulated in 3, 3, 2, and 2 different modulation schemes: PCM-BPSK-PM, PCM-QPSK-PM, PCM-BPSK-QPSK-PM, and PCM-FM. Labels 1, 2, and 3 represent different satellite individuals under the same modulation type, and similarly, labels 4, 5, 6, 7, 8, 9, and 10 also represent different satellite individuals under the same modulation type. The main carrier frequency was 14MHz, the subcarrier symbol rate was 1MHz, and the simulation was conducted under a 20dB condition. After preprocessing, each IQ signal was labeled with a satellite individual category, including 10 satellite categories, and the data format was h5py.

[0111] 2. Simulation content and result analysis:

[0112] Under the above conditions, this invention is used to identify 10 types of source satellites for downlink communication links, resulting in a confusion matrix as follows: Figure 5 In this matrix, rows represent predicted categories, and columns represent true categories. The value in the i-th row and j-th column represents the probability that the i-th satellite is predicted to be of category j. By observing the values ​​on the diagonal, the ability of this invention to correctly identify satellites can be evaluated.

[0113] from Figure 5 As can be seen, the average recognition accuracy of the present invention reached 96.7%, indicating that the present invention can effectively identify satellite identities.

[0114] The embodiments described above are merely some, not all, embodiments of the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention; however, these modifications and changes based on the spirit of the present invention are still within the scope of protection of the claims of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should also fall within the scope of protection of the present invention.

[0115] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

Claims

1. A satellite-based individual identification method under variable modulation based on multi-task decoupling learning, characterized in that: The steps include: (1) Use a broadband receiver to receive downlink communication link signals passing through a certain observation orbit, preprocess them, and then divide the preprocessed data into training set and test set in a 7:3 ratio; (2) Construct a multi-task decoupling network: 2a) Establish a knowledge extraction block consisting of multiple complex-valued multi-scale embedding units connected in parallel and then serially connected attention gating units; 2b) The three knowledge extraction blocks are connected in parallel to form a feature extraction module, which is used to extract modulation knowledge, shared knowledge and individual knowledge respectively; 2c) Stack multiple feature extraction modules in series to form a feature extraction layer; 2d) Establish two classifiers, each composed of fully connected layers, to classify modulation types and individual features, respectively. Connect these two classifiers to the two outputs of the feature extraction layer to form a multi-task decoupling network. (3) Input the training set data into the multi-task decoupling network and train the model using gradient descent until the loss function converges to obtain the trained multi-task decoupling network; specifically, inputting the training set data into the multi-task decoupling network involves calculating the predicted output through forward propagation: 3.2.1) Given the normalized downlink communication link signal input sequence x; x=z pq ={[I1,Q1] T ,[I2,Q2] T ,…,[I i ,Q i ] T ,…,[I n ,Q n ] T } Where [I] i Q i ] T The column vector x represents the combination of the real part and the imaginary part of the i-th sample point of the preprocessed downlink communication link signal. i , [·] T Indicates taking the transpose, I i =real(z) pq (i)), real(·) denotes taking the real part, Q i =imag(z pq (i)), imag(·) denotes taking the imaginary part; 3.2.2) Construct the complex-valued convolution kernel: W = [W I W Q ] T , where W I With W Q Let these represent the real and imaginary parts of the complex convolution, respectively: 3.2.3) Convolve the input signal sequence x with a complex-valued convolution kernel W, that is, use parallel complex-valued convolution kernels of three different scales (1×3, 1×5, and 1×7) to perform multi-scale feature extraction on the signal, and obtain the knowledge vector f. θ (x): in, This represents the convolution operation; 3.2.4) The signal input sequence x is processed by the i-th complex-valued multi-scale embedding unit connected in parallel in the first knowledge extraction block A to extract the knowledge vector E. A,i and m A The knowledge vectors are concatenated to obtain the concatenated feature vector E. A : in, m A E represents the number of complex-valued multi-scale embedding units in the first knowledge extraction block A, d represents the dimension of information embedding, and E represents the number of complex-valued multi-scale embedding units in the first knowledge extraction block A. A,i =f θ (x), i∈(1,m) A ); 3.2.5) The signal input sequence x is processed by the j-th complex-valued multi-scale embedding unit in parallel within the second knowledge extraction block B to extract the knowledge vector E. B,j and m B The knowledge vectors are concatenated to obtain the concatenated feature vector E. B : in, m B E represents the number of complex-valued multi-scale embedding units in the second knowledge extraction block B; d represents the dimension of information embedding. B,j =f θ (x), j∈(1,m) B ); 3.2.6) The signal input sequence x is processed by the k-th complex-valued multi-scale embedding unit in parallel within the shared knowledge extraction block S to extract the knowledge vector E. S,k and m S The knowledge vectors are concatenated to obtain the concatenated feature vector E. S : in, m S E represents the number of complex-valued multi-scale embedding units in the shared knowledge extraction block S; d represents the dimension of information embedding; E S,k =f θ (x), k∈(1,m) S ); 3.2.7) The above two eigenvectors E A With E S Perform concatenation, and then concatenate the feature vectors [E] A E S The attention value corresponding to each knowledge vector is obtained by using the gating unit input as the first knowledge extraction block A. These attention values ​​are then concatenated to obtain the attention weight vector g of the knowledge vector. A : in, E α This represents the α-th knowledge vector in the input feature vector, where α∈(1,m) A +m S ), ||·|| p Let p be the p-norm of the eigenvector, where p = 2; 3.2.8) Transfer the attention weight vector g A The gating unit input [E] of the first knowledge extraction block A A E S ] Perform matrix multiplication to obtain the modulation knowledge output G of the first knowledge extraction block A. A : G A =g A ×[E A ,E S ] 3.2.9) The above two eigenvectors E B With E S Perform concatenation, and then concatenate the feature vectors [E] B E S The attention value corresponding to each knowledge vector is obtained by using the gating unit input as the second knowledge extraction block B. These attention values ​​are then concatenated to obtain the attention weight vector g of the knowledge vector. B : in, E α This represents the β-th knowledge vector in the input feature vector, where β∈(1,m) B +m S ); 3.2.10) Transfer the attention weight vector g B Input [E] to the second knowledge extraction block B gate unit B E S ] Perform matrix multiplication to obtain the individual knowledge output G of the second knowledge extraction block B. B : G B =g B ×[E B ,E S ] 3.2.11) The above three eigenvectors E A E B With E S Perform concatenation, and then concatenate the feature vectors [E] A E B E S As the input to the gating unit of the shared knowledge extraction block S, the attention value ω corresponding to each knowledge vector is obtained. Eγ And these attention values ​​are concatenated to obtain the attention weight vector g of the knowledge vector. S : in, E γ This represents the γ-th knowledge vector in the input feature vector, where γ∈(1,m) A +m B +m S ); 3.2.12) Transfer the attention weight vector g S Input to the shared knowledge extraction block S gated unit [E] A E B E S Perform matrix multiplication to obtain the shared knowledge output G of the shared knowledge extraction block S. S : G S =g S ×[E A ,E B ,E S ] 3.2.13) Multiple feature extraction modules are stacked in series to form a feature extraction layer to output modulation features and individual features. Specifically, the signal input sequence x is used as the input of the first stacked feature extraction module, and the first knowledge extraction block A1 of the first layer outputs modulation knowledge G. A1 As input to the first knowledge extraction block A2 of the second layer; the output individual knowledge G of the second knowledge extraction block B1 of the first layer. B1 As input to the second knowledge extraction block B2 of the second layer; the shared knowledge G output by the first layer shared extraction block S1 is used. S1 This serves as the input to the second layer shared extraction block S2; and so on, after six layers of forward propagation, the final output modulation feature G is obtained. A and individual characteristics G B ; (4) Input the test set data into the trained multi-task decoupling network and output the modulation type result and satellite identity result.

2. The method according to claim 1, characterized in that, In step (1), the downlink communication link signal is preprocessed by first performing a DC removal operation on the input signal to remove the DC bias of the signal; then, the DC bias of the removed signal is normalized to control the signal value range within a specific interval, ensuring that different signals are compared and processed under the same standard.

3. The method according to claim 1, characterized in that, Each complex-valued multi-scale embedding unit constituting the knowledge extraction block in step (2a) is composed of three complex convolutions in parallel. The first complex convolution has a kernel size of 1×3 and a stride of 2; the second complex convolution has a kernel size of 1×5 and a stride of 2; and the third complex convolution has a kernel size of 1×7 and a stride of 4. The activation function of all three complex convolution layers is the complex ReLU function.

4. The method according to claim 1, characterized in that, In step (2b), an attention mechanism is introduced to form a gating unit. Its input is the knowledge vector to be fused, and its output is the knowledge vector with fused attention. The gating unit consists of a fully connected layer and a residual in parallel. The fully connected layer is used to extract the weight of each knowledge vector, and the residual is used to obtain the original knowledge vector.

5. The method according to claim 1, characterized in that, In step (2c), the fully connected layer that constitutes the classifier takes the output of the feature extraction layer as input and outputs the modulation type result and the satellite identity result. The number of intermediate neurons in the fully connected layer is set to be no less than 128, and the number of output neurons in the two fully connected layers are N1 and N2, respectively, where N1 is the number of modulation categories and N2 is the number of satellite individuals.

6. The method according to claim 1, characterized in that, In step (3), the training set data is input into the multi-task decoupling network, and gradient descent is used to train it, as follows: 3a) Set the learning rate to 0.0001, the batch size to 16, the maximum number of training iterations to 50, the optimizer to Adam, and the cross-entropy function to be used as the network loss function; 3b) The input is propagated forward through the network to calculate the predicted output; 3c) Use the loss function to calculate the loss value between the predicted class and the true class of the network output; Where N represents the batch size in training, i represents the sequence number of the training sample in the training batch, M1 represents the number of satellite signal modulation categories in the training set, M2 represents the number of individual satellite signal categories in the training set, and c represents the sequence number of the training sample in the downlink signal category of the individual satellite. y ic Let y represent the symbol function for modulation identification. If the true modulation category of the i-th sample is c, then y ic The value is 1 if it is not 1, otherwise the value is 0. z ic The sign function representing individual identification is z if the true individual category of the i-th sample is c. ic Its value is 1 if it is not 1, otherwise its value is 0. This represents the probability that the predicted modulation category of the i-th sample is c after passing through the multi-task decoupling network; This represents the probability that the predicted individual category of the i-th sample is c after passing through the multi-task decoupling network; 3d) Use the chain rule to backpropagate the loss value to each layer, calculate the gradient of the weights and biases of each layer to update the network parameters; 3e) Repeat steps 3b) to 3d) until the maximum number of iterations is reached or the loss function converges to a small value; the trained multi-task decoupling network is obtained.

Images