A communication signal anti-jamming method and system based on graph signal processing

Abstract

The application discloses a communication signal anti-suppression interference method and system based on graph signal processing, specifically comprising the following steps: in the training stage, creating a signal data set, the data set comprising a training set and a verification set; building a graph signal processing convolutional recurrent network; creating a custom network training objective function; using the obtained signal data set to train and verify the graph signal processing convolutional recurrent network, and obtaining the trained graph signal processing convolutional recurrent network; in the application stage, obtaining a time-frequency aliasing signal to be deinterfered through a single antenna; and using the trained graph signal processing convolutional recurrent network to recover a useful signal from the complex number form aliasing signal. The system comprises a training module and an application module, and is used for realizing the communication signal anti-suppression interference method based on the graph signal processing. The application can realize weak signal enhancement under high suppression interference in the case of single antenna reception, and improves the wireless communication anti-interference capability.

Description

Technical Field

[0001] This invention relates to the field of communication anti-interference technology, and in particular to a method and system for communication signal anti-suppression interference based on graph signal processing. Background Technology

[0002] The rapid development of next-generation wireless communication technologies has led to a significant increase in data transmission demands. In wireless communication, due to limited frequency band resources, multiple wireless devices share the same band, making the communicating party susceptible to interference from other transmitting devices. This problem becomes even more critical when using high-power transmitting equipment; for example, in shipboard communications, communication signals are easily suppressed and interfered with by high-power radar sidelobe signals. Furthermore, in the field of secure communication, normal communication between the sender and receiver is vulnerable to malicious interference from the enemy. Strong suppression jamming is the most common type of interference. This type of interference mainly manifests as the jammer sending interfering signals, preventing the receiver from receiving the original signal normally, thus leading to a decrease in communication quality or communication interruption. (See diagram below.) Figure 1 As shown. To overcome these problems, it is crucial for radio systems to have anti-jamming capabilities in both civilian and military applications.

[0003] Traditional anti-jamming methods include various techniques, such as Direct Sequence Spread Spectrum (DSSS) in the code domain, Frequency Hopping Spread Spectrum (FHSS) and Wiener filtering in the frequency domain, power control techniques in the power domain, and spatial filtering beamforming in the spatial domain. However, DSSS is more suitable for scenarios with high reliability requirements and is not applicable under high-power interference suppression; while FHSS is effective in avoiding interference, it increases the system burden when facing broadband interference suppression and cannot resist the effects of broadband interference; Wiener filtering relies on prior knowledge and assumptions about the characteristics of signals and interference, and cannot achieve anti-jamming when the suppressible interference signal covers the useful signal in both the time and frequency domains. In addition, in most cases, the user's transmit power is limited, and it is impractical for power control techniques to significantly increase the transmit power to mitigate the effects of interference suppression, while spatial filtering methods mainly focus on using multiple antennas to achieve antenna gain and interference suppression, which is not suitable for scenarios with a single antenna. Summary of the Invention

[0004] The purpose of this invention is to provide a communication signal anti-suppression interference method and system based on graph signal processing. It uses deep learning to fully mine the high-dimensional feature representation of the information layer, and can still enhance weak signals under strong interference, thereby improving the anti-interference capability of communication.

[0005] The technical solution to achieve the purpose of this invention is as follows:

[0006] A method for resisting suppression interference in communication signals based on graph signal processing includes a training phase and an application phase, wherein:

[0007] The training phase includes the following steps:

[0008] Step A1: Create a dataset of time-frequency aliasing communication signals, including communication signals and suppressed interference signals, and divide the dataset into a training set and a validation set;

[0009] Step A2: Construct a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolution block, a recurrent convolution network module, and an inverse graph signal processing module;

[0010] Step A3: Create a custom network training objective function;

[0011] Step A4: Based on the network training objective function, use the signal dataset obtained in Step A1 to train and validate the graph signal processing convolutional recurrent network obtained in Step A2, and obtain the trained graph signal processing convolutional recurrent network.

[0012] The application phase includes the following steps:

[0013] Step B1: Acquire the time-frequency aliasing communication signal to be descrambled using a single antenna;

[0014] Step B2: Use the trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal and complete the signal anti-interference.

[0015] A communication signal anti-suppression jamming system based on graph signal processing is provided. This system implements the aforementioned communication signal anti-suppression jamming method based on graph signal processing. The system includes a training module and an application module, wherein:

[0016] The training module includes a dataset creation unit, a graph signal processing convolutional recurrent network construction unit, an objective function creation unit, and a network training unit, wherein:

[0017] The dataset creation unit creates a dataset of time-frequency aliasing communication signals, including communication signals and interference signals, and divides the dataset into a training set and a validation set.

[0018] The network construction unit builds a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolution block, a recurrent convolution network module, and an inverse graph signal processing module.

[0019] The objective function creation unit allows you to create a custom network training objective function.

[0020] The network training unit, based on the network training objective function, uses the signal dataset obtained by the dataset creation unit to train and validate the graph signal processing convolutional recurrent network obtained by the network construction unit, resulting in a trained dual-path complex temporal convolutional network.

[0021] The application module includes a signal acquisition unit and a signal enhancement unit, wherein:

[0022] The signal acquisition unit acquires the time-frequency aliasing communication signal to be descrambled through a single antenna;

[0023] The signal enhancement unit uses a trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal, thus completing the signal anti-interference.

[0024] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned communication signal anti-suppression interference method based on graph signal processing.

[0025] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the graph signal processing-based anti-suppression interference method for communication signals.

[0026] Compared with the prior art, the significant advantages of this invention are:

[0027] (1) Deep learning technology is introduced to enhance weak target signals. Complex enhancement networks are used to process communication signals. The convolutional layer is extended from real domain operation to complex domain operation. Short-time graph signal processing is used to replace the traditional short-time Fourier transform as the signal processing front end, which reduces complexity and further improves signal enhancement performance.

[0028] (2) Establish a neural network that can satisfy the enhancement of complex communication signals. This network can make full use of the correlation of signal I / Q data, realize feature extraction of communication signals in the graph domain through graph signal transformation, and use convolutional recurrent network to achieve anti-suppression interference.

[0029] (3) The convolutional recurrent network for signal processing can achieve good anti-interference ability while maintaining low computational complexity. It can effectively remove interference suppression signals in the same frequency band, improve communication quality, and resist malicious suppression and interference from the enemy, thus achieving anti-interference communication. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of a communication scenario facing strong suppression and interference.

[0031] Figure 2 This is a schematic diagram of a graph signal processing convolutional recurrent network architecture used for signal enhancement.

[0032] Figure 3 It is a segmented diagram representing the structure.

[0033] Figure 4 This is a schematic diagram of using a graph signal processing convolutional recurrent network to remove strong suppression interference spectrum when the interference-to-signal ratio is 30dB. Detailed Implementation

[0034] To overcome the drawbacks of traditional anti-interference techniques, this invention focuses on using neural networks to achieve anti-suppression interference. Specifically, it introduces deep learning technology to enhance weak target signals. Generally, deep learning enhancement methods focus on speech enhancement, primarily used to improve the accuracy of Automatic Speech Recognition (ASR) modules. Traditional speech enhancement frameworks are mainly performed in the time or time-frequency domain. Since communication signals are typically represented in I / Q complex form, complex enhancement networks are needed to process these signals, extending convolutional layers from real-domain operations to complex-domain operations.

[0035] To reduce complexity and further improve signal enhancement performance, this invention employs Short-Time Graph Signal Processing (SGSP) to replace the traditional Short-Time Fourier Transform (STFT) as the signal processing front-end. While the Fourier Transform serves as a bridge between the time and frequency domains, it has several limitations, primarily: it lacks time and frequency localization capabilities; it has limitations with non-stationary signals; and it has resolution limitations. Many methods overcome these limitations, such as Gabor expansion, Cohen distribution, and wavelet transform. This invention uses graph signal processing instead. In Graph Signal Processing (GSP), data dependencies are represented by a graph, where nodes label data and edges capture the dependencies between nodes. Generally, GSP focuses on data originating from irregular domains, which are often characterized by complex and irregular structures that traditional signal processing cannot directly handle. However, treating the communication signal as a regular directed graph structure allows for capturing changes in signal characteristics from a new spatial transformation perspective. A communication signal is treated as a graph signal represented by a weighted adjacency matrix. Long-term signals are processed into segmented graph signals by dividing them into frames. Then, graph signal processing is performed according to graph theory. By combining graph signal processing with a modified convolutional recurrent network, a graph signal processing convolutional recurrent network is established for the enhancement of single-channel communication signals.

[0036] This invention discloses a method for resisting suppression interference in communication signals based on graph signal processing, comprising a training phase and an application phase, wherein:

[0037] The training phase includes the following steps:

[0038] Step A1: Create a dataset of time-frequency aliasing communication signals, including communication signals and suppressed interference signals, and divide the dataset into a training set and a validation set;

[0039] Step A2: Construct a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolution block, a recurrent convolution network module, and an inverse graph signal processing module;

[0040] Step A3: Create a custom network training objective function;

[0041] Step A4: Based on the network training objective function, use the signal dataset obtained in Step A1 to train and validate the graph signal processing convolutional recurrent network obtained in Step A2, and obtain the trained graph signal processing convolutional recurrent network.

[0042] The application phase includes the following steps:

[0043] Step B1: Acquire the time-frequency aliasing communication signal to be descrambled using a single antenna;

[0044] Step B2: Use the trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal and complete the signal anti-interference.

[0045] Furthermore, the time-frequency aliasing communication signal in step A1 can be represented as:

[0046]

[0047] Where x represents the aliased signal, and the target signal is a complex signal. Suppressing interference signals Let T represent the complex field, and let T represent the signal length.

[0048] As a specific example, the dataset in step A1 can be constructed using either of the following two methods:

[0049] The first method involves using computer software to generate communication signals with various modulation methods and to suppress interference signals. The modulation methods can be digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband blocking interference, and frequency sweeping interference.

[0050] Based on the generated communication signal and interference signal, the communication signal and interference signal are mixed according to the interference-to-signal ratio (J / S) to form an aliased signal. According to the network training requirements, the dataset of time-frequency aliased communication signal is divided into training set and validation set according to the ratio of 80% and 20% respectively.

[0051] The second approach involves sampling actual communication devices and jamming devices to obtain a dataset of time-frequency aliasing communication signals. Based on network training requirements, the dataset of time-frequency aliasing communication signals is divided into a training set and a validation set at a ratio of 80% and 20%, respectively.

[0052] As a specific example, in step A2, a graph signal processing convolutional recurrent network is constructed, including a graph signal processing module, a complex convolutional block, a recurrent convolutional network module, and an inverse graph signal processing module, as detailed below:

[0053] The graph signal processing module treats a communication signal as a graph signal represented by a weighted adjacency matrix. The graph signal is represented by G = (V, E, A), where V = {v1, v2, ..., v...} N} represents N vertices, E = {i,j,w} ij Let} denote a set of edges i,j∈V, A denote the adjacency matrix, and w ij The weight value representing the connectivity between vertex i and vertex j;

[0054] For a long sequence of complex signals, it is represented as Define a graphics shift window of size m, and then process the signal s into a segmented graph signal by dividing it into frames. Where k = n / m represents the frame number. Let l represent the l-th frame of the segmented signal, where l∈1,...,k.

[0055] Define a degree matrix D to describe the number of connections each node has with other nodes; the Laplacian matrix is ​​then represented as L = DA. Perform singular value decomposition (SVD) on the Laplacian matrix L, denoted as SVD(L) = P × Λ × U, where P, Λ, and U represent the left singular matrix, eigenvalues, and right singular matrix, respectively. For each frame of signal... Short-time graph signal processing will be performed, denoted as G. sp The converted segmented signal is represented as follows: k transformed It will be pieced together in a new dimension;

[0056] The short-time graph signal processing is represented as: S Gsp :R n →R m×k For a signal s, the transformed graphical signal is defined as S. g =S Gsp (s);

[0057] Each sampling point in the graph is considered a vertex and connected to other vertices. Considering the directional connection of each vertex, a new adjacency matrix A is defined for the segmented signal. m ;

[0058] Define different weights for each direction, with the adjacency matrix weights a of size m×m. ij Represented as

[0059] a ij = (i-1)*m+j-1,i,j∈1,...,m

[0060] The degree matrix D is obtained. m =diag{(i-1 / 2)m 2 -m / 2} m×m The values ​​of i = 1, ..., m are given by the Laplace matrix L. m =D m -A m After SVD processing, it becomes SVD(L) m ) = P m ×Λ m ×U m ;

[0061] The converted framed signal is represented as follows:

[0062]

[0063] After short-time graph signal processing, the signal is converted into... Still in complex signal form, the real and imaginary parts are stacked as two channel inputs, and S... g In a new dimension reconstruction;

[0064] The complex convolution module is implemented using complex two-dimensional convolution to increase the number of channels in the input signal, which is set to 256; the complex-valued convolution filter W is defined as W = W r +jW i W r and W i Let and represent the real and imaginary parts of the complex one-dimensional convolution kernel, respectively; therefore, through complex convolution, the new real and imaginary parts are defined as y. r =x r *W r -x i *W i and y i =x r *W i -x i *W r ;

[0065] The recurrent convolutional network module consists of four complex encoders (corresponding to [32, 64, 128, 256] channels), one complex LSTM layer, and four complex decoders (corresponding to [256, 128, 64, 32] channels). The complex encoders and decoders each consist of complex convolutional layers, complex regularization layers, and real PreLU activation layers, differing only in the number of convolutional channels. The operation of the complex LSTM layer is similar to complex convolution, with the new real and imaginary parts represented as follows:

[0066] L r =LSTM r (Y r )-LSTM i (Y i )

[0067] L i =LSTM r (Y i )+LSTM i (Y r )

[0068] Among them, L r and L i This represents the new real and imaginary parts of the LSTM output. r and LSTM i Refers to the LSTM operation involving the real and imaginary parts;

[0069] After the signal is processed by the recurrent convolutional network module, the output is represented as a graph domain mask. This is called a complex graph ratio mask (CGRM). CGRM is defined as follows:

[0070]

[0071] Among them, S g and X g The SGSP diagrams represent the mixed signals of the target and interference, respectively. S represents g The real and imaginary parts, and They represent X respectively g The real and imaginary parts;

[0072] Reshape the mask as Represented in polar coordinates:

[0073]

[0074]

[0075] The output of the graph signal processing module is processed with CGRM as follows to obtain the graph domain estimate of the network output:

[0076]

[0077] After passing through the inverse graph signal processing module, it is represented as:

[0078]

[0079] Therefore, the entire signal s is recovered in the following way.

[0080]

[0081] in, Indicates S gsp The reverse processing, concat represents the processing of k segmented signals. Then, the parts are assembled.

[0082] As a specific example, in graph signal processing convolutional recurrent networks, the scale-invariant source-noise ratio loss SI-SNR in the time domain and the mean square error MSE in the graph domain are used as loss functions;

[0083] SI-SNR is an effective loss function, defined as:

[0084]

[0085]

[0086]

[0087] Among them, s and Represent the source signal and the estimated signal;

[0088] Will adopt and M g MSE as a constraint:

[0089]

[0090] Therefore, the loss function of the graph signal processing convolutional recurrent network is designed as follows:

[0091] Loss = SI - SNR + MSE

[0092] As a specific example, step A4 describes using the signal dataset obtained in step A1 to train and validate the graph signal processing convolutional recurrent network obtained in step A2, as follows:

[0093] The training and validation sets are grouped and packaged into mini-batches, which are then fed into a dual-path complex temporal convolutional network. The network parameters are set, and the Adam optimizer is used with an initial learning rate of 0.001. The learning rate is halved when the validation set loss no longer decreases in three consecutive training rounds. The total training takes 100 rounds, and training stops when the validation set loss no longer decreases in six consecutive rounds. After the model has been trained on the training and validation sets, the optimal parameter model is saved.

[0094] The present invention also provides a communication signal anti-suppression interference system based on graph signal processing, the system comprising a training module and an application module, wherein:

[0095] The training module includes a dataset creation unit, a graph signal processing convolutional recurrent network construction unit, an objective function creation unit, and a network training unit, wherein:

[0096] The dataset creation unit creates a dataset of time-frequency aliasing communication signals, including communication signals and interference signals, and divides the dataset into a training set and a validation set.

[0097] The network construction unit builds a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolution block, a recurrent convolution network module, and an inverse graph signal processing module.

[0098] The objective function creation unit allows you to create a custom network training objective function.

[0099] The network training unit, based on the network training objective function, uses the signal dataset obtained by the dataset creation unit to train and validate the graph signal processing convolutional recurrent network obtained by the network construction unit, resulting in a trained dual-path complex temporal convolutional network.

[0100] The application module includes a signal acquisition unit and a signal enhancement unit, wherein:

[0101] The signal acquisition unit acquires the time-frequency aliasing communication signal to be descrambled through a single antenna;

[0102] The signal enhancement unit uses a trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal, thus completing the signal anti-interference.

[0103] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned communication signal anti-suppression interference method based on graph signal processing.

[0104] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the method for resisting interference of communication signals based on graph signal processing.

[0105] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0106] Example

[0107] Traditional anti-interference methods suffer from various prior knowledge assumptions and unsatisfactory anti-interference effects. This invention establishes a neural network that can enhance complex communication signals. This network can make full use of the correlation of signal I / Q data, realize feature extraction of communication signals in the graph domain through graph signal transformation, and use convolutional recurrent networks to achieve anti-interference.

[0108] Combination Figure 1 This invention demonstrates a communication scenario where the transmitting end faces suppression from high power at the same frequency;

[0109] Combination Figure 2 This invention provides a method for resisting suppression interference of communication signals based on graph signal processing, including a training phase and an application phase. The training phase includes the following steps:

[0110] Step A1: Create a dataset of time-frequency aliasing communication signals, including communication signals and suppressed interference signals, and divide the dataset into a training set and a validation set;

[0111] Step A2: Construct a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolution block, a recurrent convolution network module, and an inverse graph signal processing module;

[0112] Step A3: Create a custom network training objective function;

[0113] Step A4: Based on the network training objective function, use the signal dataset obtained in Step A1 to train and validate the graph signal processing convolutional recurrent network obtained in Step A2, and obtain the trained graph signal processing convolutional recurrent network.

[0114] The application phase includes the following steps:

[0115] Step B1: Acquire the time-frequency aliasing communication signal to be descrambled using a single antenna;

[0116] Step B2: Use the trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal and complete the signal anti-interference.

[0117] Furthermore, the time-frequency aliasing communication signal in step A1 can be represented as:

[0118]

[0119] Where x represents the aliased signal, and the target signal is the complex signal s = s r +js i ,s∈C T×1 Suppressing interference signals is T represents the signal length.

[0120] Furthermore, the dataset in step A1 is constructed using either of the following two methods:

[0121] The first method involves using computer software to generate communication signals with various modulation methods and to suppress interference signals. The modulation methods can be digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband blocking interference, and frequency sweeping interference.

[0122] Based on the generated communication signal and interference signal, the communication signal and interference signal are mixed according to the interference-to-signal ratio (J / S) to form an aliased signal. According to the network training requirements, the dataset of time-frequency aliased communication signal is divided into training set and validation set according to the ratio of 80% and 20% respectively.

[0123] The second approach involves sampling actual communication devices and jamming devices to obtain a dataset of time-frequency aliasing communication signals. Based on network training requirements, the dataset of time-frequency aliasing communication signals is divided into a training set and a validation set at a ratio of 80% and 20%, respectively.

[0124] Further, in step A2, a graph signal processing convolutional recurrent network is constructed, including a graph signal processing module, a complex convolutional block, a recurrent convolutional network module, and an inverse graph signal processing module, as detailed below:

[0125] The graph signal processing module treats a communication signal as a graph signal represented by a weighted adjacency matrix. The graph signal is represented by G = (V, E, A), where V = {v1, v2, ..., v...} N} represents N vertices, E = {i,j,w} ij Let} denote a set of edges i,j∈V, A denote the adjacency matrix, and w ij The weight value representing the connectivity between vertex i and vertex j;

[0126] For a long sequence of complex signals, it is represented as Define a graphics shift window of size m, and then process the signal s into a segmented graph signal by dividing it into frames. Where k = n / m represents the frame number. Let L represent the l-th frame of the segmented signal, where l ∈ 1, ..., k. Define a degree matrix D to describe the number of connections each node has with other nodes; the Laplace matrix is ​​then represented as L = DA. Perform singular value decomposition (SVD) on the Laplace matrix L, denoted as SVD(L) = P × Λ × U, where P, Λ, and U represent the left singular matrix, eigenvalues, and right singular matrix, respectively. For each frame of the signal... Short-time graph signal processing will be performed, denoted as G. sp The converted segmented signal is represented as follows: k transformed It will be pieced together in a new dimension;

[0127] The short-time graph signal processing is represented as: S Gsp :R n →R m×k For a signal s, the transformed graphical signal is defined as S. g =S Gsp (s);

[0128] Each sampling point in the graph is considered a vertex on the graph, such as Figure 3 As shown, and connected to other vertices, considering the directional connections of each vertex, a new adjacency matrix A is defined for the segmented signal. m ;

[0129] Define different weights for each direction, with the adjacency matrix weights a of size m×m. ij Represented as

[0130] a ij = (i-1)*m+j-1,i,j∈1,...,m

[0131] The degree matrix D is obtained. m =diag{(i-1 / 2)m 2 -m / 2} m×m The values ​​of i = 1, ..., m are given by the Laplace matrix L. m =D m -A m After SVD processing, it becomes SVD(L) m ) = P m ×Λ m ×U m ;

[0132] The converted framed signal is represented as follows:

[0133]

[0134] After short-time graph signal processing, the signal is converted into... Still in complex signal form, the real and imaginary parts are stacked as two channel inputs, and S... g In a new dimension reconstruction;

[0135] The complex convolution module is implemented using complex two-dimensional convolution to increase the number of channels in the input signal, which is set to 256; the complex-valued convolution filter W is defined as W = W r +jW i W r and W i Let and represent the real and imaginary parts of the complex one-dimensional convolution kernel, respectively; therefore, through complex convolution, the new real and imaginary parts are defined as y. r =x r *W r -x i *W i and y i =x r *W i -x i *W r ;

[0136] The recurrent convolutional network module consists of four complex encoders (corresponding to [32, 64, 128, 256] channels), one complex LSTM layer, and four complex decoders (corresponding to [256, 128, 64, 32] channels). The complex encoders and decoders each consist of complex convolutional layers, complex regularization layers, and real PreLU activation layers, differing only in the number of convolutional channels. The operation of the complex LSTM layer is similar to complex convolution, with the new real and imaginary parts represented as follows:

[0137] L r =LSTM r (Y r )-LSTM i (Y i )

[0138] L i =LSTM r (Y i )+LSTM i (Y r )

[0139] Among them, L r and L i This represents the new real and imaginary parts of the LSTM output. r and LSTM i Refers to the LSTM operation involving the real and imaginary parts;

[0140] After the signal is processed by the recurrent convolutional network module, the output is represented as a graph domain mask. This is called a complex graph ratio mask (CGRM). CGRM is defined as follows:

[0141]

[0142] Among them, S g and X g The SGSP diagrams represent the mixed signals of the target and interference, respectively. S represents g The real and imaginary parts, and They represent X respectively g The real and imaginary parts;

[0143] Reshape the mask as Represented in polar coordinates:

[0144]

[0145]

[0146] The output of the graph signal processing module is processed with CGRM as follows to obtain the graph domain estimate of the network output:

[0147]

[0148] After passing through the inverse graph signal processing module, it is represented as:

[0149]

[0150] Therefore, the entire signal s is recovered in the following way.

[0151]

[0152] in, Indicates S gsp The reverse processing, concat represents the processing of k segmented signals. Then, the parts are assembled.

[0153] Furthermore, in the graph signal processing convolutional recurrent network, the scale-invariant source-noise ratio loss SI-SNR in the time domain and the mean square error MSE in the graph domain are used as loss functions.

[0154] SI-SNR is an effective loss function, defined as:

[0155]

[0156]

[0157]

[0158] Among them, s and Represent the source signal and the estimated signal;

[0159] Will adopt and M g MSE as a constraint:

[0160]

[0161] Therefore, the loss function of the graph signal processing convolutional recurrent network is designed as follows:

[0162] Loss = SI - SNR + MSE

[0163] Furthermore, step A4 involves using the signal dataset obtained in step A1 to train and validate the graph signal processing convolutional recurrent network obtained in step A2, as detailed below:

[0164] The training and validation sets are grouped and packaged into mini-batches, which are then fed into a dual-path complex temporal convolutional network. The network parameters are set, and the Adam optimizer is used with an initial learning rate of 0.001. The learning rate is halved when the validation set loss no longer decreases in three consecutive training rounds. The total training takes 100 rounds, and training stops when the validation set loss no longer decreases in six consecutive rounds. After the model has been trained on the training and validation sets, the optimal parameter model is saved.

[0165] This embodiment also provides a communication signal anti-suppression interference system based on graph signal processing. The system includes a training module and an application module, wherein:

[0166] The training module includes a dataset creation unit, a graph signal processing convolutional recurrent network construction unit, an objective function creation unit, and a network training unit, wherein:

[0167] The dataset creation unit creates a dataset of time-frequency aliasing communication signals, including communication signals and interference signals, and divides the dataset into a training set and a validation set.

[0168] The network construction unit builds a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolution block, a recurrent convolution network module, and an inverse graph signal processing module.

[0169] The objective function creation unit allows you to create a custom network training objective function.

[0170] The network training unit, based on the network training objective function, uses the signal dataset obtained by the dataset creation unit to train and validate the graph signal processing convolutional recurrent network obtained by the network construction unit, resulting in a trained dual-path complex temporal convolutional network.

[0171] The application module includes a signal acquisition unit and a signal enhancement unit, wherein:

[0172] The signal acquisition unit acquires the time-frequency aliasing communication signal to be descrambled through a single antenna;

[0173] The signal enhancement unit uses a trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal, thus completing the signal anti-interference.

[0174] Figure 4 The diagram shows the removal of the strong suppression interference spectrum using a convolutional recurrent network for signal processing when the interference-to-signal ratio (J / S) is 27dB. Green represents the mixed signal spectrum after interference, orange represents the source signal spectrum, and blue represents the recovered signal spectrum. Figure 4 The spectrum of the target signal and the effect of the jamming suppression are shown. It can be seen that the proposed method can enhance the capabilities of deep learning networks and achieve enhanced performance under ultra-high jamming suppression.

[0175] Simulation results show that the graph signal processing convolutional recurrent network proposed in this invention achieves excellent anti-interference capabilities while maintaining low computational complexity. This invention can effectively remove interference signals in the same frequency band, improve communication quality, and counteract malicious suppression and interference from adversaries, achieving anti-interference communication.

Claims

1. A method for resisting suppression interference in communication signals based on graph signal processing, characterized in that, It includes a training phase and an application phase, in which: The training phase includes the following steps: Step A1: Create a dataset of time-frequency aliasing communication signals, including communication signals and suppressed interference signals, and divide the dataset into a training set and a validation set; Step A2: Construct a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolutional block, a recurrent convolutional network module, and an inverse graph signal processing module, as detailed below: The graph signal processing module treats a communication signal as a graph signal represented by a weighted adjacency matrix. For a long sequence of complex signals... Define size as The graphic shift window, then the signal Framing processing results in segmented image signals; For each frame of the segmented signal, perform the Short Time Graph Processing (SGSP) operation. The resulting segmented signal is represented as follows: , A converted It will be spliced ​​together in a new dimension and transformed. After processing by the complex convolution module, the number of channels of the input signal is increased, and then further processed by the recurrent convolution network module; the recurrent convolution network module consists of 4 complex encoders, 1 complex LSTM layer, and 4 complex decoders; After the signal is processed by the recurrent convolutional network module, the output is represented as a graph domain mask. This is called the Complex Graph Ratio Mask (CGRM). CGRM is defined as follows: in, and The SGSP diagrams represent the mixed signals of the target and interference, respectively. , They represent The real and imaginary parts, and They represent The real and imaginary parts; The output of the graph signal processing module is used to perform a Hadamard operation with the CGRM to obtain the graph domain estimate of the network output; then the signal is recovered by the inverse graph signal processing module. ; Step A3: Create a custom network training objective function; Step A4: Based on the network training objective function, use the signal dataset obtained in Step A1 to train and validate the graph signal processing convolutional recurrent network obtained in Step A2, and obtain the trained graph signal processing convolutional recurrent network. The application phase includes the following steps: Step B1: Acquire the time-frequency aliasing communication signal to be descrambled using a single antenna; Step B2: Use the trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal and complete the signal anti-interference.

2. The communication signal anti-suppression interference method based on graph signal processing according to claim 1, characterized in that, The time-frequency aliasing communication signal in step A1 is represented as follows: in, This indicates an aliased signal, and the target signal is a complex signal. Suppressing interference signals is , Represents the field of complex numbers. Indicates the signal length.

3. The communication signal anti-suppression interference method based on graph signal processing according to claim 1, characterized in that, The dataset in step A1 is constructed in the following way: Computer software is used to generate communication signals with different modulation methods and to suppress interference signals. The modulation methods are digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband blocking interference, and frequency sweeping interference. Based on the generated communication signal and interference signal, the communication signal and interference signal are mixed according to the interference-to-signal ratio (J / S) to form an aliased signal. According to the network training requirements, the dataset of time-frequency aliased communication signal is divided into training set and validation set according to the ratio of 80% and 20%.

4. The communication signal anti-suppression interference method based on graph signal processing according to claim 1, characterized in that, The dataset in step A1 is constructed in the following way: The second approach involves sampling actual communication devices and jamming devices to obtain a dataset of time-frequency aliasing communication signals. Based on network training requirements, the dataset of time-frequency aliasing communication signals is divided into a training set and a validation set, with each set comprising 80% and 20% of the data, respectively.

5. The communication signal anti-suppression interference method based on graph signal processing according to claim 1, characterized in that, In step A2, a graph signal processing convolutional recurrent network is constructed, including a graph signal processing module, a complex convolutional block, a recurrent convolutional network module, and an inverse graph signal processing module, as detailed below: The graph signal processing module treats a communication signal as a graph signal represented by a weighted adjacency matrix. The graph signal is... It means that among them express One vertex, express A set of edges, Represents the adjacency matrix. Represents vertices and vertex The weighted values ​​of the connectivity between them; For a long sequence of complex signals, it is represented as Define size as The graphic shift window, then the signal Framing and segmenting the signal ,in Indicates the frame number. The segmented signal represents the first segment. frame, ; Define degree matrix The Laplace matrix is ​​used to describe the number of connections each node has with other nodes; in this case, it is represented as: ; Laplace matrix Perform singular value decomposition (SVD), denoted as: ,in , and Represent the left singular matrix, eigenvalues, and right singular matrix, respectively; for each frame of signal Short-time graph signal processing will be performed, represented as The converted segmented signal is represented as follows: , A converted It will be pieced together in a new dimension; The short-time graph signal processing can be represented as: For signals The transformed graph structure signal is defined as ; Each sampling point in the graph is treated as a vertex and connected to other vertices. Considering the directional connection of each vertex, a new adjacency matrix is ​​defined for the segmented signal. ; Define different weights for each direction, with a size of [value missing]. Adjacency matrix weights Represented as Determine the degree matrix And the Laplace matrix After SVD processing, it becomes ; The converted framed signal is represented as follows: After short-time graph signal processing, the signal is converted into... It remains a complex signal, with the real and imaginary parts stacked as two channel inputs. In a new dimension reconstruction; The complex convolution module is implemented using complex two-dimensional convolution to increase the number of channels in the input signal, set to 256; the complex-valued convolution filter... Defined as , and Let represent the real and imaginary parts of the complex one-dimensional convolution kernel, respectively; therefore, through complex convolution, the new real and imaginary parts are defined as... and ; The recurrent convolutional network module consists of four complex encoders, one complex LSTM layer, and four complex decoders. The four complex encoders correspond to [32, 64, 128, 256] channels, and the four complex decoders correspond to [256, 128, 64, 32] channels. Both the complex encoders and decoders consist of complex convolutional layers, complex regularization layers, and real PreLU activation layers, differing only in the number of convolutional channels. The operation of the complex LSTM layer is similar to complex convolution, with the new real and imaginary parts represented as follows: in, and This represents the new real and imaginary parts of the complex LSTM output. and Refers to the real and imaginary parts of LSTM operation; After the signal is processed by the recurrent convolutional network module, the output is represented as a graph domain mask. This is called the Complex Graph Ratio Mask (CGRM). CGRM is defined as follows: in, and The SGSP diagrams represent the mixed signals of the target and interference, respectively. , They represent The real and imaginary parts, and They represent The real and imaginary parts; Reshape the mask as Represented in polar coordinates as: The output of the graph signal processing module is processed with CGRM as follows to obtain the graph domain estimate of the network output: After passing through the inverse graph signal processing module, it is represented as: Therefore, for the entire signal Restore using the following methods in, Indicates to Reverse processing, Indicates to Segmented signal Then, the parts are assembled.

6. The communication signal anti-suppression interference method based on graph signal processing according to claim 5, characterized in that, In graph signal processing convolutional recurrent networks, the scale-invariant source-noise ratio loss SI-SNR in the time domain and the mean square error MSE in the graph domain are used as loss functions. SI-SNR is an effective loss function, defined as: in, and Represent the source signal and the estimated signal; Will adopt and MSE as a constraint: Therefore, the loss function of the graph signal processing convolutional recurrent network is designed as follows: 。 7. The communication signal anti-suppression interference method based on graph signal processing according to claim 1, characterized in that, In step A4, the signal dataset obtained in step A1 is used to train and validate the graph signal processing convolutional recurrent network obtained in step A2, as detailed below: The training and validation sets are grouped and packaged into mini-batches, which are then fed into a dual-path complex temporal convolutional network. The network parameters are set, and the Adam optimizer is used with an initial learning rate of 0.

001. The learning rate is halved when the validation set loss no longer decreases in three consecutive training rounds. The total training takes 100 rounds, and training stops when the validation set loss no longer decreases in six consecutive rounds. After the model has been trained on the training and validation sets, the optimal parameter model is saved.

8. A communication signal anti-suppression interference system based on graph signal processing, characterized in that, This system is used to implement the communication signal anti-suppression interference method based on graph signal processing as described in any one of claims 1 to 7, the system comprising a training module and an application module, wherein: The training module includes a dataset creation unit, a graph signal processing convolutional recurrent network construction unit, an objective function creation unit, and a network training unit, wherein: The dataset creation unit creates a dataset of time-frequency aliasing communication signals, including communication signals and interference signals, and divides the dataset into a training set and a validation set. The network construction unit builds a graph signal processing convolutional recurrent network, including a graph signal processing module, a complex convolution block, a recurrent convolution network module, and an inverse graph signal processing module. The objective function creation unit allows you to create a custom network training objective function. The network training unit, based on the network training objective function, uses the signal dataset obtained by the dataset creation unit to train and validate the graph signal processing convolutional recurrent network obtained by the network construction unit, thus obtaining a trained graph signal processing convolutional recurrent network. The application module includes a signal acquisition unit and a signal enhancement unit, wherein: The signal acquisition unit acquires the time-frequency aliasing communication signal to be descrambled through a single antenna; The signal enhancement unit uses a trained graph signal processing convolutional recurrent network to recover the useful signal from the aliased signal, thus completing the signal anti-interference.

9. A mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the communication signal anti-suppression interference method based on graph signal processing as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the communication signal anti-suppression interference method based on graph signal processing as described in any one of claims 1 to 7.

Images