Communication signal continuous learning modulation identification method based on signal feature enhancement

Abstract

The invention discloses a communication signal continuous learning modulation identification method based on signal feature enhancement, and belongs to the field of radio communication signal modulation identification. The implementation method comprises the following steps: constructing a modulation identification model and a loss function for continuous learning of a communication signal; preprocessing the signal samples containing various modulation types; taking the three-domain enhanced known class training set as the input of a feature extraction module of the one-dimensional residual hierarchical network, and generating a signal three-domain feature vector u by extracting a three-domain data set; optimizing and adjusting the weights of different dimensions and different areas of the feature vector through an attention weight adjusting module; classifying known modulation signals through a geometric regularization classification module to obtain probability distribution P of classification; completing the training of modulation identification model parameters through back propagation; the instantaneous frequency of an unknown communication signal sample is used as the input of a modulation identification model, the communication signal is identified through the model, and continuous learning identification classification of a satellite downlink signal single-carrier modulation mode is realized.

Description

Technical Field

[0001] The present invention relates to a method for continuously learning communication signal modulation recognition based on signal enhancement, and belongs to the field of radio communication signal modulation recognition. Technical Background

[0002] Wireless communication systems can be categorized as cooperative or non-cooperative, depending on whether the receiver has prior information about the transmitter. In cooperative systems, the receiver is assumed to know various prior information about the transmitter, such as the signal modulation method, maximum frequency, and upconversion frequency band. In non-cooperative communication, however, the receiver is not aware of the transmitter's prior information.

[0003] Communication signal modulation identification is crucial in both non-cooperative civilian and military scenarios. In civilian settings, modulation identification is primarily used for regional radio spectrum management, spectrum monitoring during major events, and interference detection in key areas. In military settings, modulation identification significantly impacts radio communication reconnaissance. However, with the rapid advancement of wireless communication and signal processing technologies, the types of modulation used in communication signals are increasing, and the modulation schemes are becoming increasingly complex. Consequently, radio monitoring and reception of satellite downlink signals face challenges such as a diverse range of identification types, a complex identification environment, and increased identification difficulty.

[0004] Traditional modulation recognition methods are implemented through likelihood-based hypothesis testing or feature extraction for pattern recognition. With the continuous development of artificial intelligence, modulation recognition algorithms combined with deep learning have made significant progress in recognition accuracy and scenario versatility. The deep learning automatic modulation recognition method based on long-short-term neural networks can directly obtain the time domain correlation features from the time-frequency domain data of the signal, and then effectively identify the modulation mode of the unknown signal through calculations of the fully connected layer and the softmax layer. However, for the signal modulation mode in the satellite downlink, there are problems such as environmental attenuation, sidelobe interference, and the difficulty of jointly identifying multiple modulation modes.

[0005] Furthermore, most neural network classification models only consider a subset of known modulation schemes as training and test sets, without considering other modulation schemes outside the closed set. Furthermore, as technology advances and more modulation schemes are being applied, ensuring the recognition of known modulation schemes while avoiding misidentification errors caused by new modulation schemes has become a pressing issue. Summary of the Invention

[0006] To address the current problem of modulation type recognition being unable to identify new samples and being limited to the signal types and orders of the training set when using traditional methods or artificial intelligence neural network methods in satellite downlink signal recognition scenarios, the present invention aims to provide a modulation recognition method for communication signals based on continuous learning of signal feature enhancement. This method performs modulation recognition on satellite downlink signals based on continuous learning of communication signals based on signal feature enhancement, thereby achieving continuous learning, recognition, and classification of single-carrier modulation modes of satellite downlink signals.

[0007] The purpose of the present invention is achieved through the following technical solutions.

[0008] The communication signal continuous learning modulation recognition method based on signal feature enhancement disclosed in the present invention includes the following steps:

[0009] Step 1: Construct a modulation recognition model for continuous learning of communication signals. The modulation recognition model includes a signal three-domain feature enhancement module, a feature extraction module based on a one-dimensional residual hierarchical neural network, a spatial attention weight adjustment module, and a geometric regularization recognition and classification module.

[0010] Step 2: Preprocess the signal samples of various modulation modes. First, use the three-domain feature enhancement module to enhance the features in the frequency domain, power domain, and phase domain to obtain X(n). Then divide X(n) into a known class training set, a known class test set, and an unknown class test set.

[0011] Step 3: In the known class training phase, the three-domain enhanced known class training set is used as the input of the feature extraction module of the one-dimensional residual hierarchical network. The feature vector u of the three signal domains is generated by extracting the three-domain data set and input into the spatial attention weight adjustment module to obtain a feature vector Sm with a more discrete feature space distribution.

[0012] Step 4: Input Sm into the geometric regularization classification module to classify the modulated signal and obtain the classification probability distribution P; complete the training of model parameters through back propagation to obtain the trained modulation recognition model;

[0013] Step 5: Use the instantaneous frequency of the unknown communication signal sample as the input of the trained modulation recognition model. After model recognition, continuous learning and classification of the single-carrier modulation mode of the satellite downlink signal are achieved.

[0014] Furthermore, the implementation method of step 1 is:

[0015] Step 1.1 builds a signal three-domain feature enhancement module for communication signal modulation recognition, which includes three parts: frequency domain feature enhancement, power domain feature enhancement, and phase domain feature enhancement;

[0016] Step 1.1.1 Feature enhancement in power domain:

[0017]

[0018] Among them, and Represents the real and imaginary parts of the communication signal, and n represents the array index;

[0019] Step 1.1.2 Feature enhancement in phase domain:

[0020]

[0021] Step 1.1.3 Frequency domain feature enhancement:

[0022]

[0023] Among them, f s Indicates the sampling rate;

[0024] The sampled time domain signal is processed by the three-domain feature enhancement module to obtain a vector X(n) of 3*the length of the signal discrete array:

[0025] X(n)=[φ(n),f(n),A(n)]

[0026] The data vector X(n) after the three-domain feature enhancement is divided into a known class training set, a known class test set and an unknown class test set.

[0027] Step 1.2: Construct a feature extraction module based on a one-dimensional residual hierarchical neural network.

[0028] The feature extraction module mainly consists of four residual convolution blocks and a linear transformation layer; each residual convolution block mainly consists of three one-dimensional convolution layers and three batch normalization layers, where two one-dimensional convolution layers and batch normalization layers are connected in sequence to form a convolution block for hierarchical convolution calculation to obtain the output result The remaining one-dimensional convolution layer and batch normalization layer are connected to form a residual block, which is used to convert the output of the residual convolution block into Superimposed on the output of the hierarchical convolution calculation, the output F is obtained (m) ; The formula of the hierarchical one-dimensional residual convolution block is as follows:

[0029]

[0030] ReLu refers to the activation function, BN is the batch normalization operation, It represents a convolution operation calculated in a two-dimensional convolution manner, with a convolution kernel of (1,3) and 3 channels. X is the input data after three-domain feature enhancement, and m represents the number of residual convolution blocks.

[0031] The output result F (4)After the linear transformation layer, the characteristic vector u of the signal is obtained;

[0032] Step 1.3: Construct a spatial attention weight adjustment module, which mainly adjusts the weights of each dimension in the feature vector to enhance the expressiveness of the feature and the discrimination between vectors.

[0033] Step 1.3.1 Through F sq Compress the global information of the multi-dimensional feature vector. The formula is as follows:

[0034]

[0035] Among them F sq Indicates dimension compression, converting W×C input to 1×C output, where C is the number of dimensions; u c represents the output F in step 1.2 (m) A one-dimensional vector in dimension c; W represents the length of dimension c, and z represents the distribution of the one-dimensional feature vector response on the channel;

[0036] Step 1.3.2 by F ex (z, W) calculates the attention vector between dimensions; uses a fully connected linear layer W1 to multiply z, where the dimension of W1 is C / r×C, and r is a scaling parameter; after passing through the ReLU layer δ, the result is multiplied by the fully connected layer W2, and finally passed through the sigmoid function σ to obtain s; the formula is as follows:

[0037] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0038] Step 1.3.3 performs average pooling AvgPool and maximum pooling MaxPool on s to reflect the correlation between different dimensions, and combines it with one-dimensional convolution Conv1D to highlight the information of key dimensions; through the sigmoid function σ, the output optimized and adjusted feature vector S is obtained m As input to the classification layer;

[0039] S m (s)=σ(Conv1D([AvgPool(s);MaxPool(s)]))

[0040] Step 1.4 uses the geometric regularization recognition and classification module to recognize and classify the output of the spatial attention mechanism;

[0041] After global average pooling:

[0042] v=GAP(S m )

[0043] Then map it to the score of the number of categories G through the fully connected layer:

[0044] r=W c ·v+b c

[0045] Among them, W c ∈R G×c is the weight of the fully connected layer, r is the original score, b c ∈R G is the bias of the fully connected layer.

[0046] The softmax function converts r into category probability distribution P:

[0047]

[0048] Among them, y is the vector of the classification label after onehot encoding.

[0049] At this time, the result of model recognition and classification can be output by corresponding the maximum value in P and the category label.

[0050] Furthermore, the model training described in step 1.4 requires calculating the loss function to obtain the loss value, and then updating the model weights through backpropagation and chain derivation.

[0051] The loss function L of the feature extraction module based on the one-dimensional residual hierarchical neural network cl The formula for s is as follows:

[0052]

[0053] The loss function L of the spatial attention weight adjustment module att The formula is as follows:

[0054] L att =||S m (s)||1

[0055] The loss function of the geometric regularization recognition and classification module is constructed as follows:

[0056] According to the extracted feature vector S m (s), construct a potential geometric feature distribution map based on k-NN; the adjacency matrix of the geometric distribution map is A, which is used to represent the connection relationship between nodes, I is the unit matrix, the diagonal matrix is ​​D, and the normalized Laplace matrix L is calculated as follows:

[0057] D ii =∑ j A ij

[0058] L=ID -1 / 2 AD -1 / 2

[0059] The eigenvalues ​​of the Laplacian matrix are λ1≤λ2≤…≤λ n , which can reflect the connectivity of the geometric feature distribution graph;

[0060] The loss function Loss(λ) is obtained according to the feature value:

[0061]

[0062] By L geo (λ) shows that the number of zero eigenvalues ​​of the Laplacian matrix is ​​equal to the number of categories n; the n+1th eigenvalue λ n+1 will be greater than 0, representing the difficulty of decomposing the entire graph into more connected components; by maximizing λ n+1 , making other parts (such as the potential space structure between different classes) more separated, thereby improving classification performance; at the same time, the sum of the first n eigenvalues ​​is minimized, that is, making them as close to 0 as possible, that is, ensuring that each class is indeed connected;

[0063] The loss function L of the entire modulation recognition model is shown below:

[0064] L=L cls +L att +L geo

[0065] After calculating the loss value L, the total loss gradient is calculated by the chain rule. After decomposition, the weights and parameters of the feature extraction module, spatial attention weight adjustment module, and geometric regularization recognition and classification module are updated respectively. The formula is as follows:

[0066]

[0067] After the weights and weights are updated, the training of the modulation recognition model is completed.

[0068] Beneficial effects:

[0069] 1. The communication signal continuous learning modulation recognition method based on signal feature enhancement disclosed in the present invention enhances the three-domain features of the satellite downlink communication signal, expands the characteristic properties of different modulation modes in the modulation domain, and makes the processed three-domain feature vectors have more obvious differences in the frequency domain, phase domain and amplitude domain, which facilitates subsequent inter-class and class-by-class modulation recognition.

[0070] 2. The communication signal continuous learning modulation recognition method based on signal feature enhancement disclosed in the present invention collaboratively improves the one-dimensional residual neural network through the following methods: by modifying the one-dimensional convolution kernel, the data in the three feature domains can be better distinguished; using hierarchical convolution processing, not only deepens the number of neural network layers and extracts deeper feature information, but also combines residual operations to effectively avoid gradient vanishing, slow down the overfitting phenomenon caused by forgetting the original information, and improve the robustness of the model. On this basis, through the spatial attention weight adjustment module, the discreteness of the feature vector in the feature space will be further enhanced, effectively expanding the generalization ability of the feature space in the continuous learning task scenario.

[0071] 3. The present invention discloses a method for continuous learning modulation recognition of communication signals based on signal feature enhancement. By optimizing the geometric regularization recognition and classification module and redesigning the loss function based on the eigenvalues ​​of the Laplace matrix, the model is made more sensitive to new data samples. Furthermore, the geometric regularization recognition and classification module can adapt to new data samples, reducing the potential overlap and confusion between new and existing classes in the feature space, thereby enhancing the model's adaptability to continuous learning in satellite downlink communication environments.

[0072] 4. The communication signal continuous learning modulation recognition method based on signal feature enhancement disclosed in the present invention realizes continuous learning and analysis recognition of modulated signal modulation through a three-domain feature enhancement module, a feature extraction module based on a one-dimensional residual hierarchical network, a spatial attention weight adjustment layer and a geometric regularization recognition and classification layer, thereby improving the accuracy and efficiency of signal modulation recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 Schematic diagram of a communication signal open set modulation recognition method based on signal enhancement according to the present invention;

[0074] Figure 2 It is a structural diagram of the open set recognition model of modulation mode based on signal enhancement of the present invention. DETAILED DESCRIPTION

[0075] In order to further explain the implementation ideas of the present invention, the following will be combined with the accompanying drawings in the embodiments of the present invention to carefully and clearly describe the technical solutions in the embodiments of the present invention. The aforementioned and other technical contents, features and effects of the present invention can be clearly presented in the following detailed description of the specific embodiments with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and specific understanding of the technical means and effects adopted by the present invention to achieve the intended purpose can be obtained.

[0076] Example:

[0077] The communication signal debugging and identification continuous learning method of this embodiment uses the data in the RML2018.01A dataset to train, test and verify the open set modulation identification network model of the communication signal. Figure 1 As shown, the communication signal continuous learning modulation recognition method based on signal feature enhancement disclosed in this embodiment is specifically implemented in the following steps:

[0078] Step 1: Construct a modulation recognition model for continuous learning of communication signals. The model includes a signal three-domain feature enhancement module, a feature extraction module based on a one-dimensional residual hierarchical neural network, a spatial attention weight adjustment module, and a geometric regularization recognition and classification module.

[0079] The implementation method of step one is:

[0080] Step 1.1 builds a signal three-domain feature enhancement module for communication signal modulation recognition, which includes three parts: frequency domain feature enhancement, power domain feature enhancement, and phase domain feature enhancement;

[0081] Step 1.1.1 Feature enhancement in power domain:

[0082]

[0083] Among them, and Represents the real and imaginary parts of the communication signal, and n represents the array index;

[0084] Step 1.1.2 Feature enhancement in phase domain:

[0085]

[0086] Step 1.1.3 Frequency domain feature enhancement:

[0087]

[0088] Among them, f s Indicates the sampling rate;

[0089] The sampled time domain signal is processed by the three-domain feature enhancement module to obtain a vector X(n) of 3*the length of the signal discrete array:

[0090] X(n)=[φ(n),f(n),A(n)]

[0091] The data vector X(n) after the three-domain feature enhancement is divided into a known class training set, a known class test set and an unknown class test set.

[0092] Step 1.2: Construct a feature extraction module based on a one-dimensional residual hierarchical neural network.

[0093] The feature extraction module mainly consists of four residual convolution blocks and a linear transformation layer; each residual convolution block mainly consists of three one-dimensional convolution layers and three batch normalization layers, where two one-dimensional convolution layers and batch normalization layers are connected in sequence to form a convolution block for hierarchical convolution calculation to obtain the output result The remaining one-dimensional convolution layer and batch normalization layer are connected to form a residual block, which is used to convert the output of the residual convolution block into Superimposed on the output of the hierarchical convolution calculation, the output F is obtained (m) ; The formula of the hierarchical one-dimensional residual convolution block is as follows:

[0094]

[0095] ReLu refers to the activation function, BN is the batch normalization operation, It represents a convolution operation calculated in a two-dimensional convolution manner, with a convolution kernel of (1,3) and 3 channels. X is the input data after three-domain feature enhancement, and m represents the number of residual convolution blocks.

[0096] Finally, the output result F (4) After the linear transformation layer, the characteristic vector u of the signal is obtained;

[0097] Step 1.3: Construct a spatial attention weight adjustment module, which mainly adjusts the weights of each dimension in the feature vector to enhance the expressiveness of the feature and the discrimination between vectors.

[0098] Step 1.3.1 Through F sq Compress the global information of the multi-dimensional feature vector. The formula is as follows:

[0099]

[0100] Among them F sq Indicates dimension compression, converting W×C input to 1×C output, where C is the number of dimensions; u c represents the output F in step 1.2 (m) A one-dimensional vector in dimension c; W represents the length of dimension c, and z represents the distribution of the one-dimensional feature vector response on the channel;

[0101] Step 1.3.2 by F ex (z, W) calculates the attention vector between dimensions; uses a fully connected linear layer W1 to multiply z, where the dimension of W1 is C / r×C, and r is a scaling parameter; after passing through the ReLU layer δ, the result is multiplied by the fully connected layer W2, and finally passed through the sigmoid function σ to obtain s; the formula is as follows:

[0102] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0103] Step 1.3.3 performs average pooling AvgPool and maximum pooling MaxPool on s to reflect the correlation between different dimensions, and combines it with one-dimensional convolution Conv1D to highlight the information of key dimensions; through the sigmoid function σ, the output optimized and adjusted feature vector S is obtained m As input to the classification layer;

[0104] S m (s)=σ(Conv1D([AvgPool(s);MaxPool(s)]))

[0105] Step 1.4 uses the geometric regularization recognition and classification module to process the output of the spatial attention mechanism;

[0106] First, global average pooling:

[0107] v=GAP(S m )

[0108] Then map it to the score of the number of categories G through the fully connected layer:

[0109] r=W c ·v+b c

[0110] Among them, W c ∈R G×C is the weight of the fully connected layer, r is the original score, b c ∈R G is the bias of the fully connected layer.

[0111] The softmax function converts r into category probability distribution P:

[0112]

[0113] Among them, y is the vector of the classification label after onehot encoding.

[0114] At this time, the result of model recognition and classification can be output by corresponding the maximum value in P and the category label t.

[0115] The model training described in step 1.4 requires calculating the loss function to obtain the loss value, and then updating the model weights through backpropagation and chain derivation.

[0116] The loss function L of the feature extraction module based on the one-dimensional residual hierarchical neural network cls The formula is as follows:

[0117]

[0118] The loss function L of the spatial attention weight adjustment module att The formula is as follows:

[0119] L att =||S m (s)||1

[0120] The loss function of the geometric regularization recognition and classification module is constructed as follows:

[0121] First, according to the extracted feature vector S m (s), construct a potential geometric feature distribution map based on k-NN; the adjacency matrix of the geometric distribution map is A, which is used to represent the connection relationship between nodes, I is the unit matrix, the diagonal matrix is ​​D, and the normalized Laplace matrix L is calculated as follows:

[0122] D ii =∑ j A ij

[0123] L=ID -1 / 2 AD -1 / 2

[0124] The eigenvalues ​​of the Laplacian matrix are λ1≤λ2≤…≤λ n , which can reflect the connectivity of the geometric feature distribution graph;

[0125] The loss function Loss(λ) is obtained according to the feature value:

[0126]

[0127] By L geo (λ) It can be seen that the number of zero eigenvalues ​​of the Laplacian matrix is ​​equal to the number of categories n; the n+1th eigenvalue λ n+1 will be greater than 0, representing the difficulty of decomposing the entire graph into more connected components; by maximizing λ n+1 , which may make other parts (such as the potential space structure between different classes) more separated, thereby improving classification performance; at the same time, the sum of the first n eigenvalues ​​is minimized, that is, we hope that they are as close to 0 as possible, that is, to ensure that each class is indeed connected;

[0128] The loss function L of the entire model is shown below

[0129] L=L cls +L att +L geo

[0130] After calculating the loss value L, the total loss gradient is calculated using the chain rule. After decomposition, the weights and parameters of the feature extraction module, spatial attention weight adjustment module, and geometric regularization recognition and classification module are updated through back propagation. The formula is as follows:

[0131]

[0132] After the weights and weights are updated, the training of the modulation recognition model is completed.

[0133] Step 2: The known class sample set for the model training phase is selected from 14 types of modulated communication signals, namely 4ASK, 8ASK, BPSK, 8PSK, 16PSK, 16APSK, 32APSK, 16QAM, 32QAM, 2FSK, 4FSK, 8FSK, MSK, and OFDM. The signal samples containing various modulation types are preprocessed. First, the three-domain feature enhancement module is used to enhance the features in the frequency domain, power domain, and phase domain to obtain X(n). The data of X(n) is divided into a known class training set and a known class test set in an 8:2 ratio.

[0134] Step 3: In the known class training phase, the three-domain enhanced known class training set is used as the input of the feature extraction module of the one-dimensional residual hierarchical network, and the feature vectors of the three domains of the signal are generated by extracting the three-domain data set;

[0135] Step 4: Input the feature vector into the spatial attention weight adjustment module to optimize and adjust the weights of different dimensions and different regions of the feature vector to obtain a feature vector Sm with a more discrete feature space distribution;

[0136] Step 5: Input the feature vector into the geometric regularization classification module to obtain the classification probability distribution P and classify the modulated signal. At the same time, the model parameters are trained through chain rule back propagation to the trained modulation recognition model.

[0137] Step 6: Test the model with a known class test set to verify the model classification and recognition effect;

[0138] Step 7: Use the unknown class test set as the input of the model to obtain the model classification and recognition results;

[0139] Step 7.1: In this embodiment, the unknown communication signal is collected using Ku / Ka-band low-orbit satellite signal acquisition equipment. The satellite downlink communication signal is filtered and output as a baseband signal. The digital AD is processed on the board to generate an IQ signal. The downlink communication signal collected by the Ku / Ka-band low-orbit satellite signal acquisition equipment is filtered and output as two-way IQ data. This is then enhanced in the frequency, power, and phase domains by the three-domain feature enhancement module to create an unknown test set.

[0140] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A communication signal continuous learning modulation recognition method based on signal feature enhancement, characterized by: The following steps are included: Step 1: Construct a modulation recognition model for continuous learning of communication signals. The modulation recognition model includes a signal three-domain feature enhancement module, a feature extraction module based on a one-dimensional residual hierarchical neural network, a spatial attention weight adjustment module, and a geometric regularization recognition and classification module. Step 2: Preprocess the signal samples of various modulation modes. First, use the three-domain feature enhancement module to enhance the features in the frequency domain, power domain, and phase domain to obtain X(n). Then divide X(n) into a known class training set, a known class test set, and an unknown class test set. Step 3: In the known class training phase, the three-domain enhanced known class training set is used as the input of the feature extraction module of the one-dimensional residual hierarchical network. The feature vector u of the three signal domains is generated by extracting the three-domain data set and input into the spatial attention weight adjustment module to obtain a feature vector Sm with a more discrete feature space distribution. Step 4: Input Sm into the geometric regularization classification module to classify the modulated signal and obtain the classification probability distribution P; complete the training of model parameters through back propagation to obtain the trained modulation recognition model; Step 5: Use the instantaneous frequency of the unknown communication signal sample as the input of the trained modulation recognition model. After model recognition, continuous learning and classification of the single-carrier modulation mode of the satellite downlink signal are achieved.

2. The method for continuous learning modulation recognition of communication signals based on signal enhancement according to claim 1, wherein: The implementation method of step one is: Step 1.1 builds a signal three-domain feature enhancement module for communication signal modulation recognition, which includes three parts: frequency domain feature enhancement, power domain feature enhancement, and phase domain feature enhancement; Step 1.2, construct a feature extraction module based on one-dimensional residual hierarchical neural network; Step 1.3: Construct a spatial attention weight adjustment module, which mainly adjusts the weights of each dimension in the feature vector to enhance the expressiveness of the feature and the discrimination between vectors. Step 1.4 uses the geometric regularization recognition and classification module to recognize and classify the output of the spatial attention mechanism.

3. The method for continuous learning modulation recognition of communication signals based on signal enhancement according to claim 2, characterized in that: Step 1.1 is implemented as follows: Step 1.1.1 Feature enhancement in power domain: Among them, and Represents the real and imaginary parts of the communication signal, and n represents the array index; Step 1.1.2 Feature enhancement in phase domain: Step 1.1.3 Frequency domain feature enhancement: Among them, f s Indicates the sampling rate; The sampled time domain signal is processed by the three-domain feature enhancement module to obtain a vector X(n) of 3*the length of the signal discrete array: X(n)=[φ(n),f(n),A(n)] The data vector X(n) after the three-domain feature enhancement is divided into a known class training set, a known class test set and an unknown class test set.

4. The method for continuous learning modulation recognition of communication signals based on signal enhancement according to claim 2, wherein: In step 1.2, The feature extraction module mainly consists of four residual convolution blocks and a linear transformation layer; each residual convolution block mainly consists of three one-dimensional convolution layers and three batch normalization layers, where two one-dimensional convolution layers and batch normalization layers are connected in sequence to form a convolution block for hierarchical convolution calculation to obtain the output result The remaining one-dimensional convolution layer and batch normalization layer are connected to form a residual block, which is used to convert the output of the residual convolution block into Superimposed on the output of the hierarchical convolution calculation, the output F is obtained (m) ; The formula of the hierarchical one-dimensional residual convolution block is as follows: ReLU refers to the activation function, BN is the batch normalization operation, It represents a convolution operation calculated in a two-dimensional convolution manner, with a convolution kernel of (1,3) and 3 channels. X is the input data after three-domain feature enhancement, and m represents the number of residual convolution blocks. The output result F (4) The signal’s feature vector u is obtained through the linear transformation layer.

5. The method for continuous learning modulation recognition of communication signals based on signal enhancement according to claim 4, characterized in that: Step 1.3 is implemented as follows: Step 1.3.1 Through F sq Compress the global information of the multi-dimensional feature vector. The formula is as follows: Among them F sq Indicates dimension compression, converting W×C input to 1×C output, where C is the number of dimensions; u c represents the output F in step 1.2 (m) A one-dimensional vector in dimension c; W represents the length of dimension c, and z represents the distribution of the one-dimensional feature vector response on the channel; Step 1.3.2 by F ex (z, W) calculates the attention vector between dimensions; uses a fully connected linear layer W1 to multiply z, where the dimension of W1 is C / r×C, and r is a scaling parameter; after passing through the ReLU layer δ, the result is multiplied by the fully connected layer W2, and finally passed through the sigmoid function σ to obtain s; the formula is as follows: s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) Step 1.3.3 performs average pooling AvgPool and maximum pooling MaxPool on s to reflect the correlation between different dimensions, and combines it with one-dimensional convolution Conv1D to highlight the information of key dimensions; through the sigmoid function σ, the output optimized and adjusted feature vector S is obtained m As input to the classification layer; S m (s)=σ(Conv1D([AvgPool(s);MaxPool(s)])) 6. The method for continuous learning modulation recognition of communication signals based on signal enhancement according to claim 5, characterized in that: In step 1.4, After global average pooling: v=GAP(S m ) Then map it to the score of the number of categories G through the fully connected layer: r=W c ·v+b c Among them, W c ∈R G×C is the weight of the fully connected layer, r is the original score, b c ∈R G is the bias of the fully connected layer; The softmax function converts r into category probability distribution P: Among them, y is the vector of the classification label of the modulation mode after onehot encoding; The correspondence between the maximum value in P and the category label is used to output the classification result of the model.

7. The method for continuous learning modulation recognition of communication signals based on signal enhancement according to claim 6, characterized in that: The model training described in step 1.4 requires calculating the loss function to obtain the loss value. The calculation method is as follows: The loss function L of the feature extraction module based on the one-dimensional residual hierarchical neural network cls The formula is as follows: The loss function L of the spatial attention weight adjustment module att The formula is as follows: L att =||S m (s)||1 The loss function of the geometric regularization recognition and classification module is constructed as follows: According to the extracted feature vector S m (s), construct a potential geometric feature distribution map based on k-NN; the adjacency matrix of the geometric distribution map is A, which is used to represent the connection relationship between nodes, I is the unit matrix, the diagonal matrix is ​​D, and the normalized Laplace matrix L is calculated as follows: D ii =∑ j A ij L=I-D -1 / 2 AD -1 / 2 The eigenvalues ​​of the Laplacian matrix are λ1≤λ2≤…≤λ n , which can reflect the connectivity of the geometric feature distribution graph; The loss function Loss(λ) is obtained according to the feature value: By L geo (λ) shows that the number of zero eigenvalues ​​of the Laplacian matrix is ​​equal to the number of categories n; the n+1th eigenvalue λ n+1 will be greater than 0, representing the difficulty of decomposing the entire graph into more connected components; by maximizing λ n+1 , making the other parts more separated, thereby improving the classification performance; at the same time, the sum of the first n eigenvalues ​​is minimized, that is, they are as close to 0 as possible, that is, to ensure that each class is indeed connected; The loss function L of the entire model is shown below L=L cls +L att +L geo 8. The method for continuous learning modulation recognition of communication signals based on signal enhancement according to claim 7, wherein: The model training described in step 1.4 requires the loss value to complete the model parameter and weight adjustment, the method is as follows: After calculating the loss value L, the total loss gradient is calculated by the chain rule. After decomposition, the weights and parameters of the feature extraction module, spatial attention weight adjustment module, and geometric regularization recognition and classification module are updated respectively. The formula is as follows: At this point, the model training is completed.

Images