Satellite spread spectrum measurement and control signal modulation identification method based on adaptive graph generation

Abstract

The invention discloses a satellite spread spectrum measurement and control signal modulation identification method based on adaptive graph generation. The method comprises the following steps: receiving and preprocessing a spread spectrum measurement and control signal; constructing a graph signal according to the pre-processed spread spectrum measurement and control signal; the graph signal comprises a node feature matrix and an adjacent matrix; the adjacency matrix is obtained by excavating correlation characteristics between I and Q signals of the spread spectrum measurement and control signal after complex convolution excavation processing and embedding excavation characteristics into an adjacency matrix form; mining and aggregating local topological information, global topological information and node feature information of the spread spectrum measurement and control graph signal from the graph signal, and generating a spread spectrum measurement and control signal representation vector according to the mined information; and identifying the modulation mode of the spread spectrum measurement and control signal according to the spread spectrum measurement and control signal representation vector. According to the invention, accurate identification of different spread spectrum measurement and control signals under the same spread spectrum system is realized in a complex and changeable electromagnetic environment.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a satellite spread spectrum measurement and control signal modulation recognition method based on adaptive graph generation. Background Art

[0002] Satellite TT&C systems are responsible for adjusting satellite orbits and flight attitudes, controlling payloads, and monitoring the operating status of various modules. They are crucial for ensuring the satellite's normal operation and mission completion. However, satellite TT&C signals are open-link signals, susceptible to interception, spoofing, and jamming, making them a key target in satellite TT&C confrontations. Satellite TT&C confrontations refer to countermeasures between opposing sides in the space-to-ground TT&C domain, competing for control of the electromagnetic spectrum. A crucial prerequisite for TT&C confrontations is the acquisition of TT&C intelligence and the reconnaissance of TT&C information. However, traditional TT&C information reconnaissance techniques struggle to accurately identify satellite TT&C information, hindering the conduct of satellite TT&C confrontations. Therefore, accurate identification of satellite TT&C signals is a crucial prerequisite in TT&C confrontations and holds significant practical significance. Modulation recognition of satellite TT&C signals involves analyzing the electromagnetic spectrum and statistical characteristics of intercepted unknown modulation signals through various signal processing methods to determine their modulation mode.

[0003] In the existing technology, modulation recognition techniques for spread spectrum measurement and control signals are mainly divided into three categories: likelihood function-based modulation recognition methods, feature extraction-based modulation recognition methods, and machine learning-based intelligent modulation recognition methods. Likelihood function-based modulation recognition methods model the modulation recognition problem as a multiple hypothesis testing problem. This method calculates the likelihood function value of the received signal and constructs a decision criterion to distinguish the modulation type. Likelihood function-based modulation methods can be divided into the average likelihood ratio test (ALRT), the generalized likelihood ratio test (GLRT), and the hybrid likelihood ratio test (HLRT). Feature extraction-based modulation recognition methods extract characteristic parameters that reflect the differences between different modulation signals, construct feature vectors, and input them into a classifier to achieve recognition. Currently, commonly used signal features include instantaneous frequency, instantaneous phase, spectrum, square spectrum, high-order cumulants, and cyclic spectrum. Machine learning-based modulation recognition methods construct a high-dimensional feature space through multi-layer nonlinear mapping, converting low-dimensional features into high-dimensional abstract representations. This eliminates the need for manual feature extraction and human intervention, enabling modulation recognition of measurement and control signals. Commonly used classifiers such as support vector machines (SVMs) and convolutional neural networks (CNNs) are widely used in signal modulation recognition.

[0004] However, modulation recognition methods based on likelihood functions have high computational complexity and strong reliance on prior information, making them unable to meet the recognition needs of increasingly complex spatial electromagnetic environments. Existing feature-based spread spectrum measurement and control signal modulation recognition methods mostly focus on identifying different spread spectrum systems, and are unable to accurately identify different spread spectrum measurement and control signals under the same spread spectrum system. At the same time, due to the strong anti-interference and anti-interception capabilities of spread spectrum measurement and control signals, it is difficult to extract highly discriminative features from spread spectrum measurement and control signals. Existing machine learning-based spread spectrum measurement and control signal modulation recognition methods do not fully mine features and have poor robustness in complex and changing electromagnetic environments.

[0005] Therefore, there is no practical solution in the existing technology to achieve accurate identification of different spread spectrum measurement and control signals under the same spread spectrum system in a complex and changeable electromagnetic environment. Summary of the Invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation.

[0007] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0008] In a first aspect, the present invention provides a method for satellite spread spectrum tracking and control signal modulation identification based on adaptive graph generation, comprising:

[0009] Receive and pre-process spread spectrum measurement and control signals;

[0010] A graph signal is constructed based on the preprocessed spread spectrum measurement and control signal; the graph signal includes a node feature matrix and an adjacency matrix; the nodes in the node feature matrix are sampling points of the preprocessed spread spectrum measurement and control time domain signal, and the adjacency matrix is ​​obtained by mining the correlation characteristics between the I and Q signals of the processed spread spectrum measurement and control signal through complex convolution, and embedding the mined features into the form of an adjacency matrix;

[0011] Mining the spread spectrum and aggregating local topology information, global topology information and node feature information of the measurement and control graph signal from the graph signal, and generating a spread spectrum measurement and control signal representation vector based on the mined information;

[0012] The modulation mode of the spread spectrum measurement and control signal is identified according to the spread spectrum measurement and control signal representation vector.

[0013] In a second aspect, the present invention provides a satellite spread spectrum tracking and control signal modulation identification device based on adaptive graph generation, comprising:

[0014] Receiving module, used for receiving and pre-processing spread spectrum measurement and control signals;

[0015] A construction module is configured to construct a graph signal based on the preprocessed spread spectrum measurement and control signal; the graph signal includes a node feature matrix and an adjacency matrix; the nodes in the node feature matrix are sampling points of the preprocessed spread spectrum measurement and control time domain signal, and the adjacency matrix is ​​obtained by mining the correlation characteristics between the I and Q signals of the processed spread spectrum measurement and control signal through complex convolution, and embedding the mined features into the form of an adjacency matrix;

[0016] a generation module, configured to mine and aggregate local topology information, global topology information, and node feature information of the spread spectrum measurement and control graph signal from the graph signal, and generate a spread spectrum measurement and control signal representation vector based on the mined information;

[0017] An identification module is used to identify the modulation mode of the spread spectrum measurement and control signal according to the spread spectrum measurement and control signal representation vector.

[0018] In a third aspect, the present invention provides an electronic device comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0019] Memory for storing computer programs;

[0020] The processor is configured to implement the steps of the satellite spread spectrum measurement and control signal modulation identification method based on adaptive graph generation when executing the program stored in the memory.

[0021] The satellite spread spectrum measurement and control signal modulation recognition method based on adaptive graph generation provided by the present invention converts the satellite spread spectrum measurement and control signal recognition problem into a whole-graph classification problem in the graph domain, fully considers the correlation between the I / Q paths of the spread spectrum measurement and control signal by using complex convolution, enhances the interpretability of the spread spectrum measurement and control signal on the basis of retaining the time domain characteristics, obtains potential features that can characterize the spread spectrum measurement and control graph signal by mining and aggregating the local topology, global topology and node feature information of the spread spectrum measurement and control signal from the graph signal of the spread spectrum measurement and control signal, generates a graph signal representation vector (spread spectrum measurement and control signal representation vector) with higher discriminability, improves the recognition accuracy of the spread spectrum measurement and control signal, and realizes accurate recognition of different spread spectrum measurement and control signals under the same spread spectrum system in a complex and changeable electromagnetic environment, with strong robustness.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a satellite spread spectrum measurement and control signal modulation identification method based on adaptive graph generation provided by the present invention;

[0024] FIG2( a ) is an architecture diagram of a CAVG-MPoolGCN model provided by the present invention;

[0025] Figure 2(b) is the architecture diagram of the multi-channel pooling in Figure 2(a);

[0026] Figure 3 Schematically shows the signal time domain diagrams of different types of satellite spread spectrum tracking and control signals;

[0027] Figure 4 This is the comparison result of the robustness (measured by recognition accuracy) of the CAVG-MPoolGCN in the present invention and several control models using different graph generation methods under different signal-to-noise ratio conditions;

[0028] Figure 5 This is the robustness comparison result of the CAVG-MPoolGCN in the present invention and several control models using different graph pooling methods under different signal-to-noise ratio conditions;

[0029] Figure 6 This is the robustness comparison result between the CAVG-MPoolGCN in the present invention and the traditional SVM (support vector machine) method under different signal-to-noise ratio conditions;

[0030] Figure 7(a) shows the robustness experimental results of CAVG-MPoolGCN under different timing errors in the present invention;

[0031] Figure 7(b) shows the robustness experimental results of CAVG-MPoolGCN under different phase offsets in the present invention;

[0032] Figure 7(c) shows the robustness experimental results of CAVG-MPoolGCN in the present invention under different frequency offsets. DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0034] To accurately identify different spread spectrum tracking signals under the same spread spectrum system in complex and changing electromagnetic environments, the present invention provides a satellite spread spectrum tracking signal modulation recognition method based on adaptive graph generation. This method combines spread spectrum tracking signal modulation recognition with deep learning to construct a spread spectrum tracking signal recognition model based on adaptive graph generation, CAVG-MPoolGCN. This model can automatically extract signal features and has strong robustness, making it suitable for intelligent recognition of spread spectrum tracking signals in complex and changing electromagnetic environments. Figure 1 As shown, the satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation provided by the present invention includes the following steps:

[0035] S10. Receive and pre-process the spread spectrum measurement and control signal.

[0036] Specifically, the real part and the imaginary part of the spread spectrum measurement and control signal are extracted, and the data format thereof is converted into a 1×N×2 format, where N is the signal length, and the extracted real part and the imaginary part are jointly normalized.

[0037] S20. Construct a graph signal based on the preprocessed spread spectrum measurement and control signal; the graph signal includes a node feature matrix and an adjacency matrix; the nodes in the node feature matrix are sampling points of the preprocessed spread spectrum measurement and control time domain signal, and the adjacency matrix is ​​obtained by mining the correlation characteristics between the I and Q signals of the processed spread spectrum measurement and control signal through complex convolution, and embedding the mined features into the form of an adjacency matrix.

[0038] Specifically, constructing a graph signal according to the pre-processed spread spectrum measurement and control signal includes:

[0039] (1) The pre-processed spread spectrum measurement and control signal is input into multiple parallel one-dimensional complex convolutional layers, and the features output by each one-dimensional complex convolutional layer are nonlinearized using the complex ReLu activation function to obtain the convolution output features; wherein the convolution kernel size of multiple one-dimensional complex convolutional layers increases from 2;

[0040] (2) The convolution output features are embedded into the adjacency matrix, and the node feature matrix is ​​constructed according to the preprocessed spread spectrum measurement and control signal to obtain the graph signal of the spread spectrum measurement and control signal.

[0041] In detail, the CAVG-MPoolGCN proposed in the present invention includes a pre-trained adaptive graph generation module for adaptively outputting an adjacency matrix based on the pre-processed spread spectrum measurement and control signal. Specifically, the adaptive graph generation module is composed of multiple parallel one-dimensional complex convolutions, and each convolution layer uses an independent convolution kernel; the size of the convolution kernel increases from 2, and the total number is controlled by the hyperparameter m; after the complex convolution, the complex ReLu activation function is applied to nonlinearize the output features of the complex convolution layer; the output of each convolution layer is embedded in the adjacency matrix, wherein the output features of the i-th convolution layer are diagonally embedded at the position offset i+1 along the main diagonal, the I channel features are embedded in the upper right corner of the adjacency matrix, and the Q channel features are embedded in the lower left corner of the adjacency matrix, thereby obtaining the adjacency matrix of the spread spectrum measurement and control graph signal. In addition, the value of each sampling point in the spread spectrum measurement and control time domain signal constitutes a node feature matrix, thereby obtaining the graph signal of the spread spectrum measurement and control signal.

[0042] S30, mining and aggregating local topology information, global topology information and node feature information of the spread spectrum measurement and control graph signal from the graph signal, and generating a spread spectrum measurement and control signal representation vector based on the mined information.

[0043] Specifically, the CAVG-MPoolGCN proposed in the present invention also includes a pre-trained graph representation learning module, so the pre-trained graph representation learning module can be used to extract features representing the local topology information, global topology information and node feature information of the spread spectrum measurement and control graph signal from the graph signal of the spread spectrum measurement and control graph signal.

[0044] The figure shows that the learning module includes multiple cascaded MPoolGCN modules, where the input of the first MPoolGCN module is the graph signal, and the last MPoolGCN module outputs features containing local topology information, global topology information and node feature information of the spread spectrum measurement and control graph signal; each MPoolGCN module includes at least one GCN layer and a multi-channel pooling module; the multi-channel pooling module is used to perform multi-channel pooling on the features output by the GCN layer to which it is connected, and perform cross-channel convolution and pooling result aggregation on the pooling results; here, multi-channel pooling includes: pooling based on local topology, pooling based on global topology and pooling based on node features.

[0045] As shown in Figure 2(a), the entire graph signal representation learning module is composed of multiple MPoolGCN stacks (3 in Figure 2). Each MPoolGCN module is formed by stacking at least one GCN (Graph Convolutional Network) and a multi-channel pooling (MPool) module. In multiple cascaded MPoolGCN blocks, the first MPoolGCN block contains two layers of GCN and a multi-channel pooling module. The first layer of GCN learns the node representation vector and converts the dimension of the node feature vector to the latent space vector dimension. The second layer of GCN is used to encode the graph signal and aggregate node information. Except for the first MPoolGCN block, each of the remaining MPoolGCN blocks is composed of a layer of GCN and a multi-channel pooling module. The layer of GCN is used to aggregate node information and update the representation vector of the nodes in the graph.

[0046] As shown in Figure 2(b), the multi-channel pooling module consists of three parts: multi-channel pooling, cross-channel convolution, and pooling result aggregation. Multi-channel pooling performs graph pooling using three channels: local topology-based pooling, global topology-based pooling, and node feature-based pooling. These channels are referred to as channels 1, 2, and 3, respectively. These channels are described in detail below.

[0047] Channel 1 is a local topology-based pooling method, which aims to extract the local topology features of the image signal and output a fine-grained pooling map. In this channel 1, the local topology-based pooling is implemented in the following way:

[0048] i. Determine the first node set:

[0049] n i =sum(A[i,:])

[0050] s1=[n1,n2,…n N ];

[0051] Idx1=rank(s1,m)

[0052] In the above formula, A is the adjacency matrix of the graph signal, i corresponds to the i-th node, n i is the degree of the i-th node, s1 represents the vector of degrees of all N nodes, m is the number of nodes to be retained, rank(s1,m) means sorting all the degrees in s1 from large to small and returning the subscripts Idx1 of the first m nodes; the first node set is the set consisting of the nodes whose subscripts belong to Idx1;

[0053] ii. Pooling the graph signal according to the first node set to obtain a first fine-grained pooled graph.

[0054] Specifically, node feature vectors for the nodes in the first node set are obtained from the graph signal's node feature matrix X, and a new node feature matrix is ​​constructed based on the obtained node feature vectors. Similarly, a new adjacency matrix is ​​constructed based on the elements related to the nodes in the first node set contained in the graph signal's adjacency matrix A. The new node feature matrix and the new adjacency matrix together constitute the first fine-grained pooling graph.

[0055] Channel 2 is a global topology-based pooling method that divides the nodes in the spread spectrum measurement and control graph signal into multiple clusters. Each cluster is represented by a representative node, thereby outputting a coarse-grained pooling graph. In this channel 2, the global topology-based pooling is implemented in the following way:

[0056] i. Calculate the node cluster probability matrix:

[0057]

[0058] Among them, A is the adjacency matrix of the graph signal, X is the node feature matrix of the graph signal, N is the number of nodes, and p is the preset number of node clusters. Represents the real number field, GNN embed (·) is a graph neural network model used to learn the node cluster probability matrix. It is part of CAVG-MPoolGCN and is trained together with CAVG-MPoolGCN. softmax(·) represents the softmax function, and S is the node cluster probability matrix.

[0059] ii. Calculate the coarse-grained pooling graph based on the node cluster probability matrix:

[0060]

[0061] Among them, GNN embed (·) is a graph neural network model used to learn node features. It is part of CAVG-MPoolGCN and is trained together with CAVG-MPoolGCN. Z is GNN embed (·) The learned node features, q is the dimension of the node features, S T is the transpose of S, X coarse is the node feature matrix of the calculated coarse-grained pooling graph, A coarse is the adjacency matrix of the calculated coarse-grained pooling graph.

[0062] Channel 3 is node feature-based pooling, which focuses on capturing the node feature information of the graph signal and also generates a fine-grained pooled graph after pooling. In this channel 3, node feature-based pooling is achieved through the following methods:

[0063] i. Calculate the weight vector:

[0064]

[0065] Among them, X (l) is the node feature matrix X of the graph signal output by the l-th layer GCN, readout(·) is the readout function, Z is the readout global feature representation of the graph signal, d is its dimension size, represents the real number domain, W is the learnable parameter matrix of the neural network used to learn the importance weights of node features in X. This neural network is part of CAVG-MPoolGCN and is trained together with CAVG-MPoolGCN. σ(·) is the ReLu activation function. is the weight vector;

[0066] ii. Determine the second node set:

[0067]

[0068] Among them, X[i] is the node feature vector of the i-th node in the node feature matrix X of the graph signal, n i is the node feature importance score of the i-th node, s2 is a vector consisting of N node feature importance scores, m is the number of nodes to be retained, rank(s2,m) means sorting all the degrees in s2 from large to small and returning the subscripts Idx2 of the first m nodes; the second node set is the set consisting of nodes whose subscripts belong to Idx2;

[0069] iii. Pooling the graph signal according to the second node set to obtain a second fine-grained pooled graph.

[0070] Specifically, node feature vectors for the nodes in the second node set are obtained from the graph signal's node feature matrix X, and a new node feature matrix is ​​constructed based on the obtained node feature vectors. Similarly, a new adjacency matrix is ​​constructed based on the elements related to the nodes in the second node set contained in the graph signal's adjacency matrix A. The new node feature matrix and the new adjacency matrix together constitute the second fine-grained pooling graph.

[0071] As mentioned earlier, the multi-channel pooling module not only performs multi-channel pooling on the features output by the connected GCN layer, but also performs cross-channel convolution and pooling result aggregation on the pooling results.

[0072] Specifically, the cross-channel convolution and pooling result aggregation of the pooling results are achieved in the following way:

[0073] i. Perform cross-channel convolution on the node feature matrices of the first fine-grained pooling map and the coarse-grained pooling map to obtain the first node embedding matrix:

[0074] X1=σ([H fine1 +A cross ·H corase ]·W);

[0075] Among them, H fine1 is the node embedding matrix composed of the node feature matrix and adjacency matrix of the first fine-grained pooling graph, H corase is the node embedding matrix composed of the node feature matrix and adjacency matrix of the coarse-grained pooling graph, A cross [i]=S[i], i∈Idx1∪Idx2, S[i] is the node cluster probability vector of the i-th node in the node cluster probability matrix; wherein, the node feature matrix and the adjacency matrix can jointly generate a node embedding matrix through aggregation and transformation operations of a graph neural network (GNN). For the specific generation method, please refer to the relevant prior art, which will not be elaborated in the present invention.

[0076] ii. Perform cross-channel convolution on the node feature matrices of the second fine-grained pooling map and the coarse-grained pooling map to obtain the second node embedding matrix:

[0077] X2=σ([H fine2 +A cross ·H corase ]·W);

[0078] Among them, H fine2 is the node embedding matrix of the second fine-grained pooling graph;

[0079] iii. Aggregate pooling results:

[0080]

[0081] Where X1[i,:] is the embedding vector of the i-th node in the first node embedding matrix X1, X2[i,:] is the embedding vector of the i-th node in the second node embedding matrix X2, Idx=Idx1∪Idx2, A[Idx,:] is the element with the subscript Idx in the adjacency matrix A of the graph signal, X P is the node feature matrix contained in the aggregation result, is the adjacency matrix contained in the aggregation result, and K is the number of nodes retained in the aggregation result.

[0082] Therefore, the present invention uses cross-channel convolution to convolve the two fine-grained pooling maps with the coarse-grained pooling map, achieving the fusion of different channel features. Finally, the graph pooling method based on the multi-channel mechanism aggregates the pooling results to obtain a pooled map after multi-channel pooling.

[0083] Then, each MPoolGCN module obtains its pooled graph after multi-channel pooling. After each MPoolGCN module, a Readout function is used to read the whole graph representation vector of the graph signal from the pooled graph. Finally, all the whole graph representation vectors are added together to obtain the final representation vector, which is the generation of the spread spectrum measurement and control signal representation vector based on the mined information as mentioned in step S30.

[0084] S40. Identify the modulation mode of the spread spectrum measurement and control signal according to the spread spectrum measurement and control signal representation vector.

[0085] Specifically, CAVG-MPoolGCN also includes a pre-trained two-layer MLP classifier. The spread-spectrum measurement and control signal representation vector is input into the pre-trained two-layer MLP classifier. The classifier then uses a Softmax function to normalize the output, producing the probability distribution of each modulation mode for the spread-spectrum measurement and control signal. The modulation mode with the highest probability value is then determined as the modulation mode for the spread-spectrum measurement and control signal.

[0086] It can be understood that in order to enable CAVG-MPoolGCN to extract more accurate signal features and accurately classify the spread spectrum measurement and control signal representation vector, it is necessary to pre-train CAVG-MPoolGCN in advance, so as to use the trained CAVG-MPoolGCN to implement a series of operations in steps S20 to S40.

[0087] When training CAVG-MPoolGCN, a simulated spread spectrum measurement and control signal dataset can be used. The parameters of this dataset include: direct sequence spread spectrum (DSSS) and direct sequence / frequency hopping (DS / FH) spread spectrum; six signal types: DS-BPSK, DS-QPSK, DS-UQPSK, DS / FH-BPSK, DS / FH-QPSK, and DS / FH-UQPSK; a symbol rate of 1 MHz; a frequency hopping rate of 2500 kHz; a sampling frequency of 630 MHz; a carrier frequency of 70 MHz; a frequency hopping frequency set of [5 MHz, 20 MHz, 35 MHz, and 50 MHz]; a spreading code length of 63; a direct sequence spread spectrum code rate of 63 MHz; a UQPSK imbalance coefficient of 0.3; and a signal-to-noise ratio of -10 dB:20 dB with a 2 dB interval.

[0088] It should be noted that the present invention does not limit the size of the dataset used for training CAVG-MPoolGCN and the signal modulation and spread spectrum types it contains. In practice, the dataset can be constructed according to needs. The dataset given above is only an illustrative example.

[0089] In addition, the present invention does not require the method for obtaining signal samples in the data set. Actual spread spectrum measurement and control signals can be collected, or they can be obtained through simulation generation. After obtaining a large number of signal samples, they are divided into a training set and a test set. The training set is used for training, and the test set is used to test the performance of the preliminarily trained CAVG-MPoolGCN. For example, 100 spread spectrum measurement and control modulation signals of each signal-to-noise ratio can be set in the training set, and 30 different spread spectrum measurement and control modulation signals of each signal-to-noise ratio can be set in the test set.

[0090] Then, you can use the dataset for training, including:

[0091] (1) Initialize the network parameters, set the initial learning rate, rounds, batch size, loss function and optimizer;

[0092] The specific parameter settings are as follows: the pooling retention ratio in all MPoolGCN layers in the recognition model is set to 0.5, the feature representation of the entire image is obtained by the average aggregation function, the parameters are initialized using the Xavier Normal distribution, the hidden layer dimensions are 128 and 256, the embedding layer dimensions are 64, 128, and 256, the number of stacked layers of MPoolGCN is 3, the optimization method is the Adam optimizer, the batch size is 8, the training rounds are 130 rounds, and the learning rate is set to 0.001.

[0093] (2) The signal sample is input into the CAVG-MPoolGCN in training, so that it predicts the modulation mode corresponding to the output signal sample. Then, the loss value is calculated based on the predicted modulation mode and the actual modulation mode of the signal sample. The network parameters of the CAVG-MPoolGCN are continuously updated through the loss value calculated each time, and the training loss change is dynamically monitored. If the loss value converges and no longer decreases, the training process is terminated and the optimal model is saved. Otherwise, the iterative training is repeated.

[0094] Preferably, a cross entropy loss function can be used to measure the difference between the network prediction output and the training set label, and its formula is:

[0095]

[0096] Where x j is the signal sample input to CAVG-MPoolGCN, y j is the label of the signal sample, which is used to indicate the actual modulation mode corresponding to the signal sample. N represents the number of signal samples in a training batch, and C represents the number of categories of spread spectrum measurement and control signals. is the output x predicted by CAVG-MPoolGCN jIn the probability distribution corresponding to all C modulation modes, j Label y j The probability of the same modulation mode, x j,c x is the output of CAVG-MPoolGCN j In the probability distribution corresponding to all C modulation modes, x j The probability corresponding to the cth modulation mode, L(x,y) is the loss value.

[0097] The effects of the present invention can be further illustrated by the following simulation experiments.

[0098] Specifically, the present invention is compared with different graph generation methods, graph pooling methods and traditional methods, which are described in detail as follows.

[0099] (1) Comparison with different graph generation methods

[0100] Visibility Graph (VG): This algorithm is a classic method for converting time series data into a graph structure. This method captures the visibility relationship between time series data points, that is, if all points between points a and b are below the line connecting them, then a and b are connected. This method maps complex nonlinear signals into a graph with a clear topology.

[0101] Horizontal Visibility Graph (HVG): The principle of the HVG method is similar to that of the VG method. It makes some improvements based on the VG method. That is, if all points between points a and b are below the horizontal line connecting the two points a and b, then a and b are connected, thereby mapping the time domain signal to the graph domain.

[0102] Adaptive Visibility Graph (AVG) algorithm based on one-dimensional convolution: Convolution is used to extract the characteristic relationship between nodes and their adjacent nodes, and the adjacency matrix of the graph signal is constructed based on this, thereby mapping the time domain signal to the graph domain to generate the corresponding graph signal.

[0103] (2) Comparison with different graph pooling methods

[0104] DiffPool: This algorithm uses a GNN model to learn an assignment matrix to determine the cluster to which each node belongs. It then collapses each cluster into a single node to implement graph pooling. This pooling method takes into account the global topology of the graph.

[0105] SAGPool: This algorithm is a hierarchical pooling method that uses the attention mechanism to implement graph pooling; this pooling method takes into account the node features and topological structure of the graph.

[0106] (3) Comparison with traditional methods

[0107] SVM method: By extracting the signal's autocorrelation peak-to-average ratio, autocorrelation second-order moment, spectral bandwidth, power spectrum cancellation gain, time-frequency matrix normalization column mode and time-frequency matrix normalization row mode, combined with a multi-classification support vector machine, the recognition of direct sequence spread spectrum and hybrid spread spectrum signals is achieved.

[0108] The comparison results are as follows Figures 4-6 As shown, Figure 4 This is the comparison result of the robustness (measured by recognition accuracy) of the CAVG-MPoolGCN in the present invention and several control models using different graph generation methods under different signal-to-noise ratio conditions; Figure 5 This is the robustness comparison result of the CAVG-MPoolGCN in the present invention and several control models using different graph pooling methods under different signal-to-noise ratio conditions; Figure 6 This is the robustness comparison result between the CAVG-MPoolGCN in the present invention and the traditional SVM (support vector machine) method under different signal-to-noise ratio conditions.

[0109] from Figure 4 、 Figure 5 、 Figure 6 It can be seen that under all signal-to-noise ratios, the model CAVG-MPoolGCN proposed in this invention has good recognition performance; when the signal-to-noise ratio is greater than 5dB, the recognition accuracy can reach 100%; under low signal-to-noise ratio conditions where the signal-to-noise ratio is less than 0dB, the recognition accuracy of the method proposed in this invention can also reach more than 75%. At the same time, the recognition performance of the algorithm proposed in this invention is far superior to different graph generation methods, pooling methods, and traditional spread spectrum measurement and control recognition methods, effectively verifying that the adaptive visual graph generation method based on complex convolution proposed in this invention can enhance the interpretability of the generated graph signal while retaining the time domain features, and the multi-channel pooling method adopted in this invention can capture and aggregate local topology, global topology, and node feature information, fully explore the potential features of the spread spectrum measurement and control graph signal, and better and more accurately realize the modulation recognition of the spread spectrum measurement and control signal.

[0110] Figure 7 shows the robustness test results of the present invention under different timing errors, phase offsets, and frequency offsets. As can be seen from the figure, the proposed method is highly robust to different timing errors, phase offsets, and carrier frequency offsets, and can, to a certain extent, meet the requirements for spread-spectrum measurement and control signal recognition in complex electromagnetic environments.

[0111] Corresponding to the above-mentioned method for identifying modulation of satellite spread spectrum tracking and control signals based on adaptive graph generation, the present invention further provides a device for identifying modulation of satellite spread spectrum tracking and control signals based on adaptive graph generation, comprising:

[0112] Receiving module, used for receiving and pre-processing spread spectrum measurement and control signals;

[0113] A construction module is configured to construct a graph signal based on the preprocessed spread spectrum measurement and control signal; the graph signal includes a node feature matrix and an adjacency matrix; the nodes in the node feature matrix are sampling points of the preprocessed spread spectrum measurement and control time domain signal, and the adjacency matrix is ​​obtained by mining the correlation characteristics between the I and Q signals of the processed spread spectrum measurement and control signal through complex convolution, and embedding the mined features into the form of an adjacency matrix;

[0114] a generation module, configured to mine local topology information, global topology information, and node feature information of the spread spectrum measurement and control graph signal from the graph signal, and generate a spread spectrum measurement and control signal representation vector based on the mined information;

[0115] An identification module is used to identify the modulation mode of the spread spectrum measurement and control signal according to the spread spectrum measurement and control signal representation vector.

[0116] Optionally, a construction module is provided for constructing a graph signal according to the preprocessed spread spectrum measurement and control signal, including:

[0117] The preprocessed spread spectrum measurement and control signal is input into multiple parallel one-dimensional complex convolutional layers, and the features output by each one-dimensional complex convolutional layer are nonlinearized using a complex ReLu activation function to obtain convolution output features; wherein the convolution kernel size of the multiple one-dimensional complex convolutional layers increases from 2;

[0118] The convolution output features are embedded into an adjacency matrix, and a node feature matrix is ​​constructed according to the preprocessed spread spectrum measurement and control signal to obtain the graph signal.

[0119] Optionally, the generation module mines local topology information, global topology information and node feature information of the spread spectrum measurement and control graph signal from the graph signal, including:

[0120] Extracting features representing local topology information, global topology information, and node feature information of the spread spectrum measurement and control graph signal from the graph signal using a pre-trained graph representation learning module;

[0121] The graph representation learning module includes multiple MPoolGCN modules; the multiple MPoolGCN modules are cascaded, wherein the input of the first MPoolGCN module is the graph signal, and the output of the last MPoolGCN module includes features of local topology information, global topology information and node feature information of the spread spectrum measurement and control graph signal;

[0122] The MPoolGCN module includes at least one GCN layer and a multi-channel pooling module; the multi-channel pooling module is used to perform multi-channel pooling on the features output by the GCN layer to which it is connected, and perform cross-channel convolution and pooling result aggregation on the pooling results; the multi-channel pooling includes: pooling based on local topology, pooling based on global topology, and pooling based on node features.

[0123] Optionally, pooling based on local topology is implemented as follows:

[0124] i. Determine the first node set:

[0125] n i =sum(A[i,:])

[0126] s1=[n1,n2,…n N ];

[0127] Idx1=rank(s1,m)

[0128] In the above formula, A is the adjacency matrix of the graph signal, i corresponds to the i-th node, n i is the degree of the i-th node, s1 represents the vector of degrees of all N nodes, m is the number of nodes to be retained, rank(s1,m) means sorting all the degrees in s1 from large to small and returning the subscripts Idx1 of the first m nodes; the first node set is the set consisting of the nodes whose subscripts belong to Idx1;

[0129] ii. Perform pooling processing on the graph signal according to the first node set to obtain a first fine-grained pooling graph.

[0130] Optionally, pooling based on the global topology is implemented as follows:

[0131] i. Calculate the node cluster probability matrix:

[0132]

[0133] Where A is the adjacency matrix of the graph signal, X is the node feature matrix of the graph signal, N is the number of nodes, and p is the number of preset node clusters. Represents the real number field, GNN embed (·) is a graph neural network model used to learn the node cluster probability matrix, softmax(·) represents the softmax function, and S is the node cluster probability matrix;

[0134] ii. Calculate the coarse-grained pooling graph based on the node cluster probability matrix:

[0135]

[0136] Among them, GNN embed (·) is the graph neural network model used to learn node features, and Z is GNN embed (·) The learned node features, q is the dimension of the node features, S T is the transpose of S, X coarse is the node feature matrix of the calculated coarse-grained pooling graph, A coarse is the adjacency matrix of the calculated coarse-grained pooling graph.

[0137] Optionally, pooling based on node features is implemented as follows:

[0138] i. Calculate the weight vector:

[0139]

[0140] Among them, X (l) is the node feature matrix X of the graph signal output by the l-th layer GCN, readout(·) is the readout function, Z is the readout global feature representation of the graph signal, d is its dimension size, represents the real number domain, W is the learnable parameter matrix of the neural network used to learn the importance weights of node features in X, σ(·) is the ReLu activation function, is the weight vector;

[0141] ii. Determine the second node set:

[0142]

[0143] Where X[i] is the node feature vector of the i-th node in the node feature matrix X of the graph signal, n i is the node feature importance score of the i-th node, s2 is a vector consisting of N node feature importance scores, m is the number of nodes to be retained, rank(s2,m) means sorting all the degrees in s2 from large to small and returning the subscripts Idx2 of the first m nodes; the second node set is the set consisting of the nodes whose subscripts belong to Idx2;

[0144] iii. Perform pooling processing on the graph signal according to the second node set to obtain a second fine-grained pooling graph.

[0145] Optionally, perform cross-channel convolution and pooling aggregation on the pooling results, which can be achieved by:

[0146] i. Perform cross-channel convolution on the node feature matrices of the first fine-grained pooling map and the coarse-grained pooling map to obtain a first node embedding matrix:

[0147] X1=σ([H fine1 +A cross ·H corase ]·W);

[0148] Among them, H fine1 is the node embedding matrix of the first fine-grained pooling graph, H corase is the node embedding matrix of the coarse-grained pooling graph, A cross [i]=S[i], i∈Idx1∪Idx2, S[i] is the node cluster probability vector of the i-th node in the node cluster probability matrix;

[0149] ii. Perform cross-channel convolution on the node feature matrices of the second fine-grained pooling map and the coarse-grained pooling map to obtain a second node embedding matrix:

[0150] X2=σ([H fine2 +A cross ·H corase ]·W);

[0151] Among them, H fine2 is the node embedding matrix of the second fine-grained pooling graph;

[0152] iii. Aggregate pooling results:

[0153]

[0154] Where X1[i,:] is the embedding vector of the i-th node in the first node embedding matrix, X2[i,:] is the embedding vector of the i-th node in the second node embedding matrix, Idx=Idx1∪Idx2, A[Idx,:] is the element with subscript Idx in the adjacency matrix A of the graph signal, is the adjacency matrix of the aggregation result, X P is the node feature matrix of the aggregation result, and K is the number of nodes retained in the aggregation result.

[0155] Optionally, the identification module is specifically used to: input the spread spectrum measurement and control signal representation vector into a pre-trained two-layer MLP classifier, and output the probability distribution of the spread spectrum measurement and control signal corresponding to each type of modulation mode through normalization of the Softmax function, and determine the modulation mode with the largest probability value as the modulation mode of the spread spectrum measurement and control signal.

[0156] It should be noted that, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For relevant details, please refer to the partial description of the method embodiment, and both the device and the method can achieve the same or similar beneficial effects.

[0157] It should be noted that the terms "first," "second," and the like are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in sequences other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention.

[0158] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0159] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other variations of the disclosed embodiments by viewing the drawings and the disclosed content. In the description of the present invention, the word "comprising" does not exclude other components or steps, "one" or "a" does not exclude multiple situations, and "multiple" means two or more, unless otherwise clearly and specifically defined. In addition, certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0160] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, devices (equipment), or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or a combination of software and hardware embodiments, all of which are collectively referred to herein as "modules" or "systems." Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The computer program may be stored / distributed in a suitable medium, provided together with other hardware or as part of the hardware, or may be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems.

[0161] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, apparatus (devices) and computer program products of the embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0162] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0163] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0164] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for satellite spread spectrum tracking and control signal modulation recognition based on adaptive graph generation, characterized in that: include: Receive and pre-process spread spectrum measurement and control signals; Constructing a graph signal according to the pre-processed spread spectrum measurement and control signal; The graph signal includes a node feature matrix and an adjacency matrix; The nodes in the node feature matrix are sampling points of the pre-processed spread spectrum measurement and control time domain signal, and the adjacency matrix is ​​obtained by mining the correlation characteristics between the I and Q signals of the processed spread spectrum measurement and control signal through complex convolution, and embedding the mined features into the form of an adjacency matrix; Mining the spread spectrum and aggregating local topology information, global topology information and node feature information of the measurement and control graph signal from the graph signal, and generating a spread spectrum measurement and control signal representation vector based on the mined information; The modulation mode of the spread spectrum measurement and control signal is identified according to the spread spectrum measurement and control signal representation vector.

2. The satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation according to claim 1 is characterized in that: The constructing of the image signal according to the pre-processed spread spectrum measurement and control signal includes: The preprocessed spread spectrum measurement and control signal is input into multiple parallel one-dimensional complex convolutional layers, and the features output by each one-dimensional complex convolutional layer are nonlinearized using a complex ReLu activation function to obtain convolution output features; wherein the convolution kernel size of the multiple one-dimensional complex convolutional layers increases from 2; The convolution output features are embedded into an adjacency matrix, and a node feature matrix is ​​constructed according to the preprocessed spread spectrum measurement and control signal to obtain the graph signal.

3. The satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation according to claim 1, characterized in that: The mining and aggregating of local topology information, global topology information and node feature information of the spread spectrum measurement and control graph signal from the graph signal includes: Extracting features representing local topology information, global topology information, and node feature information of the spread spectrum measurement and control graph signal from the graph signal using a pre-trained graph representation learning module; The graph representation learning module includes multiple MPoolGCN modules; the multiple MPoolGCN modules are cascaded, wherein the input of the first MPoolGCN module is the graph signal, and the output of the last MPoolGCN module includes features of local topology information, global topology information and node feature information of the spread spectrum measurement and control graph signal; The MPoolGCN module includes at least one GCN layer and a multi-channel pooling module; the multi-channel pooling module is used to perform multi-channel pooling on the features output by the GCN layer to which it is connected, and perform cross-channel convolution and pooling result aggregation on the pooling results; the multi-channel pooling includes: pooling based on local topology, pooling based on global topology, and pooling based on node features.

4. The satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation according to claim 3 is characterized in that: The pooling based on local topology is achieved in the following way: i. Determine the first node set: In the above formula, A is the adjacency matrix of the graph signal, i corresponds to the i-th node, n i is the degree of the i-th node, s1 represents the vector of degrees of all N nodes, m is the number of nodes to be retained, rank(s1,m) means sorting all the degrees in s1 from large to small and returning the subscripts Idx1 of the first m nodes; the first node set is the set consisting of the nodes whose subscripts belong to Idx1; ii. Perform pooling processing on the graph signal according to the first node set to obtain a first fine-grained pooling graph.

5. The satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation according to claim 4 is characterized in that: The pooling based on the global topology is achieved in the following way: i. Calculate the node cluster probability matrix: Where A is the adjacency matrix of the graph signal, X is the node feature matrix of the graph signal, N is the number of nodes, and p is the number of preset node clusters. Represents the real number field, GNN embed (·) is a graph neural network model used to learn the node cluster probability matrix, softmax(·) represents the softmax activation function, and S is the node cluster probability matrix; ii. Calculate the coarse-grained pooling graph based on the node cluster probability matrix: Among them, GNN embed (·) is the graph neural network model used to learn node features, and Z is GNN embed (·) The learned node features, q is the dimension of the node features, S T is the transpose of S, X coarse is the node feature matrix of the calculated coarse-grained pooling graph, A coarse is the adjacency matrix of the calculated coarse-grained pooling graph.

6. The satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation according to claim 5, characterized in that: The node feature-based pooling is achieved in the following way: i. Calculate the weight vector: Among them, X (l) is the node feature matrix X of the graph signal output by the l-th layer GCN, readout(·) is the readout function, Z is the readout global feature representation of the graph signal, d is its dimension size, represents the real number domain, W is the learnable parameter matrix of the neural network used to learn the importance weights of node features in X, σ(·) is the ReLu activation function, is the weight vector; ii. Determine the second node set: Where X[i] is the node feature vector of the i-th node in the node feature matrix X of the graph signal, n i is the node feature importance score of the i-th node, s2 is a vector consisting of N node feature importance scores, m is the number of nodes to be retained, rank(s2,m) means sorting all the degrees in s2 from large to small and returning the subscripts Idx2 of the first m nodes; the second node set is the set consisting of the nodes whose subscripts belong to Idx2; iii. Perform pooling processing on the graph signal according to the second node set to obtain a second fine-grained pooling graph.

7. The satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation according to claim 6, characterized in that: The cross-channel convolution and pooling result aggregation of the pooling results are achieved in the following way: i. Perform cross-channel convolution on the node feature matrices of the first fine-grained pooling map and the coarse-grained pooling map to obtain a first node embedding matrix: X1=σ([H fine1 +A cross ·H corase ]·W); Among them, H fine1 is the node embedding matrix composed of the node feature matrix and adjacency matrix of the first fine-grained pooling graph, H corase is the node embedding matrix composed of the node feature matrix and adjacency matrix of the coarse-grained pooling graph, A cross [i]=S[i], i∈Idx1∪Idx2, S[i] is the node cluster probability vector of the i-th node in the node cluster probability matrix; ii. Perform cross-channel convolution on the node feature matrices of the second fine-grained pooling map and the coarse-grained pooling map to obtain a second node embedding matrix: X2=σ([H fine2 +A cross ·H corase ]W); Among them, H fine2 is the node embedding matrix of the second fine-grained pooling graph; iii. Aggregate pooling results: Where X1[i,:] is the embedding vector of the i-th node in the first node embedding matrix, X2[i,:] is the embedding vector of the i-th node in the second node embedding matrix, Idx=Idx1∪Idx2, A[Idx,:] is the element with subscript Idx in the node feature matrix of the graph signal, is the adjacency matrix contained in the aggregation result, X P is the node feature matrix contained in the aggregation result, and K is the number of nodes retained in the aggregation result.

8. The satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation according to claim 1, characterized in that: Identifying a modulation mode of the spread spectrum measurement and control signal according to the spread spectrum measurement and control signal representation vector includes: The spread spectrum measurement and control signal representation vector is input into a pre-trained two-layer MLP classifier, and the probability distribution of the spread spectrum measurement and control signal corresponding to each type of modulation mode is output through normalization by the Softmax function, and the modulation mode with the largest probability value is determined as the modulation mode of the spread spectrum measurement and control signal.

9. A satellite spread spectrum tracking and control signal modulation identification device based on adaptive graph generation, characterized in that: include: Receiving module, used for receiving and pre-processing spread spectrum measurement and control signals; A construction module, used for constructing a graph signal according to the pre-processed spread spectrum measurement and control signal; The graph signal includes a node feature matrix and an adjacency matrix; the nodes in the node feature matrix are sampling points of the pre-processed spread spectrum measurement and control time domain signal, and the adjacency matrix is ​​obtained by mining the correlation characteristics between the I and Q signals of the processed spread spectrum measurement and control signal through complex convolution, and embedding the mined features into the form of an adjacency matrix; a generation module, configured to mine and aggregate local topology information, global topology information, and node feature information of the spread spectrum measurement and control graph signal from the graph signal, and generate a spread spectrum measurement and control signal representation vector based on the mined information; An identification module is used to identify the modulation mode of the spread spectrum measurement and control signal according to the spread spectrum measurement and control signal representation vector.

10. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; The processor is configured to implement the steps of the satellite spread spectrum tracking and control signal modulation identification method based on adaptive graph generation as described in any one of claims 1 to 7 when executing the computer program stored in the memory.

Images