Dynamic evolution and periodic structure double-current cross attention fused radio frequency signal classification method and system

Abstract

The invention discloses a radio frequency signal classification method and system based on dynamic evolution and periodic structure double-current cross attention fusion. A radio frequency signal amplitude peak sequence is intercepted through a sliding window, a periodic component is enhanced by adopting adaptive Gaussian filtering, a recurrence plot is generated based on phase-space reconstruction to quantize signal dynamic evolution characteristics, and meanwhile, the sequence is mapped to a polar coordinate space to construct a Gramb angle field to realize periodic topological structure coding. Aiming at complementary characteristics of two types of heterogeneous characteristics, a double-flow cross collaborative fusion architecture is provided. According to the method, the problems of strong signal time-frequency coupling and insufficient classification robustness in a complex electromagnetic environment are solved, and the identification capability on the model, the working mode and the anti-interference characteristics of the unmanned aerial vehicle is remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV radio frequency signal analysis, in particular to a radio frequency signal classification method and system based on the fusion of dynamic evolution and periodic structure dual-stream cross-attention. Background Art

[0002] With the wide application of UAV technology in military, civilian and commercial fields, the monitoring and classification of radio frequency signals have become important technologies to ensure the safe flight and communication of UAVs. Radio frequency signals are the core components of UAV communication, and accurate analysis of radio frequency signals helps to monitor and identify UAV activities. Most of the existing radio frequency signal classification methods rely on traditional signal processing technologies, such as short-time Fourier transform (STFT) and wavelet transform. Although these methods can effectively extract the frequency domain features of signals, they often perform poorly when dealing with complex signal environments, especially when facing problems such as signal interference, noise and multipath effects.

[0003] Due to the trade-off between time and frequency resolution, STFT cannot obtain sufficient accuracy at the same time; although wavelet transform has certain advantages in multi-scale analysis, its adaptability to non-stationary signals is limited, and it is difficult to capture the details of complex signals. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a radio frequency signal classification method based on the fusion of dynamic evolution and periodic structure dual-stream cross-attention. This method is a radio frequency signal classification method based on deep learning. This method combines fast Fourier transform (FFT) sliding window frequency extraction, RP map and Gram angular field transformation technology, and performs feature learning and classification through a multi-scale convolutional neural network (CNN) and Swin Transformer dual-stream model and cross-attention mechanism.

[0005] To achieve the above purpose, the present invention provides the following technical solutions: The radio frequency signal classification method based on the fusion of dynamic evolution and periodic structure dual-stream cross-attention provided by the present invention includes the following steps: S1, number each type of UAV to be predicted; S2, adopt the sliding window technology to extract the frequency point corresponding to the maximum amplitude from each window, and then extract the repeated pattern of the frequency through adaptive Gaussian filtering and downsampling; S3, generate an RP map and a Gram angular field map by transforming the extracted frequency pattern to perform feature representation of the signal; S4. Based on the cross-attention mechanism, obtain the dynamic evolution characteristics of the RF signal by constructing a recursive graph, use the Gram angular field diagram to characterize the periodic structure characteristics and global structure characteristics of the signal, and use a multi-scale convolutional neural network combined with the Swin Transformer framework to achieve cross-scale feature fusion and feature enhancement; S5: Process the fused feature vectors through a fully connected layer, output the predicted UAV RF signal category, and realize the prediction of the classification result.

[0006] Furthermore, in the step S1, number each type of predicted UAV. The numbering includes N types of UAVs to be classified and the background reference signal category, so there are a total of N + 1 types of UAV categories.

[0007] Furthermore, in the step S2, adopt the sliding window technique, extract the frequency point corresponding to the maximum amplitude from each window, and then extract the repetitive pattern of the frequency through adaptive Gaussian filtering and downsampling: Among them, the original signal is , the sliding window signal is , is the index of the window, is the sliding step, then, extract the frequency value corresponding to the maximum amplitude , this frequency value represents the main frequency component of the signal within this window, that is: The filtering process can be expressed as: Then, downsample the filtered frequency values to remove redundant information and extract the repetitive pattern of the frequency.

[0008] Furthermore, in the step S3, transform the extracted frequency pattern to generate an RP graph and a Gram angular field graph for signal feature representation, which specifically includes the following steps: Capture the self-similarity of the signal through the RP graph for the extracted frequency change pattern to describe the dynamic evolution characteristics of the signal; Reveal the periodic structure characteristics of the signal through the polar coordinate transformation of the GASF Gram angular field graph for the extracted frequency change pattern; Convert the one-dimensional time series signal into a multi-dimensional phase space through the delay embedding method of the time series. This process generates a multi-dimensional phase space, and through this phase space, the dynamic behavior of the signal can be described.

[0009] Generally, the Euclidean distance is used to measure the similarity between two time points. Specifically, for two moments and , and its distance calculation formula in the reconstructed space is: Define a threshold , and construct a Heaviside function according to the similarity measure and the threshold to determine whether there is similarity between two time points: When , it means there is a point at the position in the two-dimensional coordinate graph, otherwise there is no point. In this way, a matrix is obtained, representing the similarity between two time points.

[0010] It should be noted that for generating the Gram angle field graph. First, for the time series signal to be processed perform normalization processing so that the value range of the signal is limited between [−1, 1]. The normalization formula is as follows: where and are the maximum and minimum values of the sequence respectively. By this step, the dimension difference is eliminated to ensure the consistency of the data scale.

[0011] Convert the normalized sequence to polar coordinate form: where, represents the phase angle at the th moment, and its value range is , which is used to encode the signal amplitude information; represents the polar radius at the th moment, is the timestamp index, is the time regularization factor (usually taking the total length of the sequence), which is used to linearly map the timestamp to the interval , and retain the time series position characteristics.

[0012] Generate a Gram matrix through the inner product of the sum of phase angles: This matrix captures the global correlation characteristics of the time series signal through the collaborative angle change.

[0013] Radio frequency signals are vulnerable to random noise interference due to their high-frequency characteristics. Traditional Gram Angular Difference Field (GADF) is sensitive to instantaneous phase differences, which may lead to noise amplification. GASF suppresses random interference in high-frequency fluctuations through phase angle sum operation Suppress random interference in high-frequency fluctuations.

[0014] Swin Transformer relies on the attention mechanism to model long-range dependencies. The GASF matrix enhances the model's ability to identify the periodicity and modulation patterns of radio frequency signals by explicitly preserving the amplitude synergy between any two moments, which is consistent with the calculation logic of the self-attention weight matrix.

[0015] Furthermore, in step S4, based on the cross-attention mechanism, joint feature encoding of the input signal is performed through the Gram Angle Sum Field (GASF), which includes the following steps: Perform joint feature encoding on the input signal through the Gram Angle Sum Field (GASF) based on the cross-attention mechanism, including the following steps: Use the dual-stream cross-attention module to achieve feature alignment between the RP map and the GASF map: Calculate the interaction weights through scaled dot-product attention: Input the aligned features into the cascaded multi-scale CNN module to output multi-scale fusion features ; Convert the multi-scale features into a sequence of image patches and input them into the Swin Transformer module to output global semantic features ; Perform element-wise weighted fusion on the aligned multi-modal features , and ; Among them, the splicing process is .

[0016] Furthermore, to achieve semantic alignment of the two types of features between the RP map and the GASF map, the GASF map feature vector is used as the query vector (Query); the RP map feature vector is split into the key vector (Key) and the value vector (Value).

[0017] Among them, , output the aligned fusion features ; Furthermore, in the multi-scale CNN module, the 3×3 convolutional layer group extracts the high-frequency details of the pulse waveform, the 5×5 dilated convolutional layer captures the mid-scale context of the carrier period, and the 7×7 depthwise separable convolutional layer models the global envelope characteristics of the broadband spectrum. After the multi-branch outputs are concatenated by channels, a gated network is used to generate spatially adaptive weights to suppress the noise channels and enhance the discriminative features, and multi-scale fusion features are output. ; Furthermore, in step S5, the fused feature vector is processed through a fully connected layer to output the predicted UAV radio frequency signal category, and the prediction of the classification result is realized, which includes the following steps: A fully connected layer based on softmax normalization is used to construct a category probability mapping space. The mathematical form of this layer is defined as: ; where, is a learnable weight matrix to realize the linear projection from the feature space to the category space; is a bias vector used to adjust the classification decision boundary; here is the feature vector modulated by the global attention mechanism, which is generated by dynamically weighted fusion of multi-scale features and contains cross-modal semantic information.

[0018] In the model training stage, a cross-entropy loss function in the form of negative log-likelihood is used to measure the difference between the predicted distribution and the true label: where, is the number of batch samples, is the total number of categories, represents the indicator function of the true category label of the th sample after one-hot encoding, is the corresponding predicted probability.

[0019] The radio frequency signal classification system with dynamic evolution and periodic structure two-stream cross-attention fusion provided by the present invention includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above method is implemented.

[0020] The beneficial effects of the present invention are as follows: The radio frequency signal classification method based on the fusion of dynamic evolution and periodic structure dual-stream cross-attention provided by the present invention first numbers the UAV devices; then uses the sliding window technique combined with adaptive Gaussian filtering to extract the signal spectrum repetition pattern, and optimizes the feature density through downsampling; uses the recurrence plot (RP) to characterize the dynamic evolution characteristics of the signal, combines with the Gram angle field plot to analyze the periodic structure of the signal, and constructs a dual-stream cross-collaborative fusion architecture; designs a cross-attention mechanism to achieve the dynamic fusion of dynamic evolution and periodic structure features, synchronously uses multi-scale CNN to extract local detail features, combines with Swin Transformer to capture long-range temporal dependence relationships, and forms a multi-scale joint representation feature vector; finally, realizes the radio frequency signal classification decision through a fully connected network.

[0021] This method innovatively integrates dynamic evolution and periodic structure features with a multi-scale deep feature extraction architecture, strengthens feature complementarity through a cross-attention mechanism, and effectively solves the problems of strong time-frequency feature coupling and insufficient classification robustness of UAV signals in complex electromagnetic environments. At the same time, the data preprocessing method significantly improves the reliability and feature recognizability of radio frequency signal analysis through multi-dimensional signal optimization and feature enhancement techniques. The dynamic spectrum interception mechanism based on the sliding window realizes the quasi-stationary processing of non-stationary signals. Combining the cascaded processing of the maximum amplitude frequency point extraction and adaptive Gaussian filtering can intelligently attenuate the in-band noise while retaining the main frequency components of the signal; through the collaborative optimization of downsampling and repetition pattern extraction, while reducing the data dimension, it can still maintain the integrity of key features, providing input features with high signal-to-noise ratio and low redundancy for subsequent deep learning models.

[0022] This method significantly improves the integrity and classification robustness of UAV radio frequency signal representation through cross-modal feature alignment and complementary enhancement techniques for the fusion method of dynamic evolution and periodic structure features. An orthogonal feature space is constructed for the correlation features based on the recurrence plot (RP) and the periodic structure features of the Gram angle field (GAF). The non-linear dynamic characteristics and harmonic structure information of the signal are respectively retained through a two-channel parallel processing mechanism, solving the problem of insufficient representation dimension of complex radio frequency signals by single-modal features.

[0023] In the multi-scale feature pyramid fusion architecture of this method, while retaining the local texture details extracted by the CNN, it deeply fuses the global temporal correlation patterns captured by the Swin Transformer, enhancing the model's collaborative analysis ability between microscopic transient features and macroscopic periodic laws. The cross-attention mechanism significantly improves the effectiveness and robustness of the two-stream feature fusion through dynamic feature interaction and adaptive weight allocation. The cross-modal feature correlation modeling based on the multi-head attention architecture realizes the non-linear mapping learning between the recursive graph dynamic evolution features and the Gram angle field periodic structure features, accurately capturing the harmonic correlation patterns and transient response relationships between features by establishing a cross-channel attention matrix, and solving the problem of information conflict between modalities caused by traditional hard feature splicing. A learnable gating attenuation factor is introduced to dynamically adjust the cross-modal attention weights, suppressing out-of-band noise interference while strengthening the time-domain feature expression of key frequency bands.

[0024] This method intercepts the peak sequence of the radio frequency signal amplitude through a sliding window, enhances the periodic components using adaptive Gaussian filtering, and generates a recurrence plot (RP) based on phase space reconstruction to quantify the dynamic evolution characteristics of the signal. At the same time, the sequence is mapped to the polar coordinate space to construct a Gram angle field (GASF) for periodic topological structure encoding. A two-stream cross-collaborative fusion architecture is proposed for the complementary characteristics of the two types of heterogeneous features: at the local level, a multi-scale convolutional network (MsCNN) is designed to extract the chaotic trajectory texture features of the RP map; at the global level, an improved Swin Transformer is constructed to capture the temporal structure correlation of the GASF field, and cross-level information aggregation is realized through a feature pyramid. A differentiable dynamic gating module is further innovatively introduced to automatically adjust the two-stream contribution weights based on feature importance, and finally, the prediction result is output through a joint classifier. The present invention solves the problems of strong signal time-frequency coupling and insufficient classification robustness in complex electromagnetic environments, and significantly enhances the identification ability of UAV models, working modes, and anti-interference features.

[0025] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following specification. Brief Description of the Drawings

[0026] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration.

[0027] Figure 1 It is a schematic diagram of the simplified process of this method.

[0028] Figure 2 It is a model framework diagram of this method.

[0029] Figure 3 It is a comparison chart of the recognition accuracy rate curve during the training process.

[0030] Figure 4 It is a confusion matrix using only dynamic evolution features.

[0031] Figure 5 It is a confusion matrix of the fusion of dynamic evolution and periodic structure features.

[0032] Figure 6 It is a confusion matrix of the fusion of dynamic evolution, periodic structure features and cross-attention. Specific implementation manners

[0033] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the exemplified embodiments do not limit the present invention.

[0034] Embodiment 1 As Figure 1 shown, the radio frequency signal classification method for the fusion of dynamic evolution and periodic structure two-stream cross-attention provided in this embodiment includes the following steps: S1. Number each type of predicted unmanned aerial vehicle (UAV). S2. Adopt the sliding window technique to extract the frequency point corresponding to the maximum amplitude from each window, and then extract the repeated pattern of the frequency through adaptive Gaussian filtering and downsampling. S3. Generate the RP map and Gram angular field map through transformation of the extracted frequency pattern for signal feature representation. S4: Based on the cross-attention mechanism, obtain the time-domain dynamic evolution features of the radio frequency signal by constructing a recursive graph, use the Gram angular field diagram to characterize the periodic features and global structure features of the signal, and use a multi-scale convolutional neural network combined with the Swin Transformer framework to achieve cross-scale feature fusion and feature enhancement. S5: Process the fused feature vector through a fully connected layer, output the predicted UAV radio frequency signal category, and realize the prediction of the classification result.

[0035] Furthermore, in the step S1, when numbering each type of predicted UAV, if the total number of UAVs to be classified includes N types, then there are a total of N + 1 types of UAV categories, and the supplementary one is the background reference signal category.

[0036] Furthermore, in the step S2, when adopting the sliding window technique to extract the frequency point corresponding to the maximum amplitude from each window, and then extract the repeated pattern of the frequency through adaptive Gaussian filtering and downsampling: Among them, the original signal is , and the sliding window signal is , is the index of the window, is the sliding step size. Then, extract the frequency value corresponding to the maximum amplitude , and this frequency value represents the main frequency component of the signal within this window, that is: The filtering process can be expressed as: Then, downsample the filtered frequency values to remove redundant information and extract the repeated patterns of the frequencies. Through this process, the main changes of the signal in time and frequency can be effectively captured, data redundancy can be reduced, and the efficiency of subsequent processing can be improved; Furthermore, in step S3, the extracted frequency pattern is transformed to generate an RP graph and a Gram angular field graph for signal feature representation, which specifically includes the following steps: Capture the self-similarity of the signal through the RP graph for the extracted frequency change pattern to describe the dynamic evolution characteristics of the signal; Reveal the periodic structure characteristics of the signal through the polar coordinate transformation of the GASF graph for the extracted frequency change pattern; It should be noted that for generating the RP graph. According to the Takens embedding theorem, we can transform a one-dimensional time series signal into a multi-dimensional phase space through the delay embedding method of the time series. This process generates a multi-dimensional phase space, and the dynamic behavior of the signal can be described through this phase space.

[0037] Once the phase space is reconstructed, the similarity between two time points can be calculated. Usually, the Euclidean distance is used to measure the similarity between two time points. Specifically, for two moments and , the distance calculation formula in the reconstructed space is: Next, define a threshold , and construct a Heaviside function according to the similarity measure and the threshold to determine whether there is similarity between two time points: When , it means there is a point at the position in the two-dimensional coordinate graph, otherwise there is no point. In this way, we can obtain a matrix representing the similarity between two time points.

[0038] It should be noted that, in order to generate the Gram angle field diagram, firstly, for the time series signal to be processed The normalization process is performed so that the signal value range is limited to [−1, 1]. The normalization formula is as follows: in, and are the maximum and minimum values ​​of the sequence respectively. This step eliminates the dimension difference and ensures the consistency of data scale.

[0039] The standardized sequence Convert to polar coordinates: in, Indicates The phase angle at the moment has a range of , used to encode signal amplitude information; Indicates The diameter of a moment, is the timestamp index, is the time regularization factor (usually the total length of the sequence), which is used to linearly map timestamps to intervals , retaining the temporal position characteristics.

[0040] Generate the Gram matrix by taking the inner product of the phase angle and: The matrix captures the global correlation characteristics of timing signals through coordinated angle changes.

[0041] RF signals are susceptible to random noise interference due to their high frequency characteristics. Traditional Gram Angle Difference Field (GADF) is sensitive to instantaneous phase differences, which may cause noise amplification. GASF uses phase angle and operation to Suppress random disturbances in high-frequency fluctuations. Experiments show that on the data set used in the experiment, GASF has a 13% higher noise-to-signal ratio (NSR) tolerance than GADF.

[0042] Swin Transformer relies on the attention mechanism to model long-range dependencies. The GASF matrix is ​​obtained by The amplitude synergy between any two moments is explicitly retained, which is consistent with the calculation logic of the self-attention weight matrix, thereby enhancing the model's ability to recognize the periodicity and modulation mode of RF signals; Furthermore, in S4, the input signal is analyzed by Gram Angle Sum Field (GASF) based on the cross attention mechanism. Perform joint feature encoding, which includes the following steps: Using a dual-stream cross-attention module to achieve feature alignment between the RP graph and the GASF graph: The input signal first generates time-domain dynamic features through a Recursive Graph (RP), reflecting the transient behavior and non-stationary characteristics of the signal; meanwhile, through the Gram Angle and Field (GASF), time-domain joint features are generated, encoding the amplitude coupling relationship and periodic pattern of the signal. To achieve semantic alignment of the two types of features, the GASF graph feature vector is used as the query vector (Query); the RP graph feature vector is split into a key vector (Key) and a value vector (Value).

[0043] Calculate the interaction weights through scaled dot-product attention: where, , and the aligned fused feature is output; Multi-scale CNN feature extraction, constructing a cascaded convolutional layer group to extract local-global features of the RF signal: Input the fused feature into the cascaded multi-scale CNN module. After the multi-branch outputs are concatenated along the channel dimension, a gated network is used to generate spatially adaptive weights, suppressing noise channels and enhancing discriminative features, and outputting multi-scale fused features .

[0044] Convert the multi-scale features into a sequence of image patches and input them into the Swin Transformer module. By introducing a cyclic shift window strategy, fixed-window self-attention and shifted-window self-attention are alternately executed in adjacent network layers, and the calculation formula is: where the mask matrix ensures the causal constraint of the shifted window. This mechanism models long-range dependencies while reducing computational complexity, and outputs global semantic features .

[0045] After obtaining the cross-modal aligned features , multi-scale convolutional features and the Swin Transformer global features , adjust the spatial dimensions of all feature tensors to a unified size . Finally, feature tensors , and of a unified size are obtained.

[0046] Concatenate the above-aligned features along the channel dimension to The input three-dimensional gating network generates dynamic fusion weights and outputs the original weight tensor . To ensure weight normalization, for each spatial position apply the Softmax function along the modal dimension to calculate the normalized weight coefficients: Obtain the final weight matrix , satisfying Based on the weight matrix , perform element-wise weighted fusion on the aligned multi-modal features: Input the joint representation into the global average pooling layer to compress it into a feature vector along the spatial dimension, and then map it to the class space through two fully connected layers to output the probability distribution of the RF signal classes.

[0047] Furthermore, in step S5, the fused feature vector is processed through a fully connected layer to output the predicted UAV RF signal class, realizing the prediction of the classification result, specifically carried out in the following manner: For the classification prediction module of the model, a fully connected layer based on softmax normalization is used to construct the class probability mapping space. The mathematical form of this layer is defined as: ; where is the learnable weight matrix to achieve the linear projection from the feature space to the class space; is the bias vector used to adjust the classification decision boundary; the softmax activation function maps the output to a probability distribution through the log-odds transformation , and its -th dimensional element represents the posterior probability that the sample belongs to class . Here is the feature vector modulated by the global attention mechanism, generated through multi-scale feature dynamic weighted fusion, containing cross-modal semantic information.

[0048] In the model training stage, the cross-entropy loss function in the form of negative log-likelihood is used to measure the difference between the predicted distribution and the true label: where is the number of batch samples, is the total number of classes, represents the indicator function of the one-hot encoded true class label of the -th sample, is the corresponding predicted probability. During the backpropagation process, this loss function optimizes the network parameters through the gradient descent algorithm. Its mathematical properties can be decomposed into two terms: the KL divergence measures the information difference between the true distribution and the predicted distribution, and the entropy constraint term suppresses the distribution chaos of the prediction results, thereby enhancing the inter-class separability and improving the model generalization ability simultaneously during the optimization process.

[0049] Example 2 Reference Figure 1 and Figure 2 , Figure 1 is a schematic diagram of the simplified process of the method of the present invention. Figure 2 is the model framework diagram of the present invention. The embodiments of the present invention disclose a radio frequency signal classification method that fuses dynamic evolution and periodic structure two-stream cross-attention, including the following steps: S1. Number the predicted various types of unmanned aerial vehicles. It should be noted that the radio frequency signals to be classified in S1 include radio frequency communication signals from different types of unmanned aerial vehicles, specifically N types of unmanned aerial vehicle signals. In addition to these N types of unmanned aerial vehicle signals, the data set also contains 1 type of background reference signal category for distinguishing background signals other than unmanned aerial vehicles. The background signal category includes environmental noise and other non-unmanned aerial vehicle signals, which are used as negative categories in the classification task to help the model distinguish unmanned aerial vehicle signals from other types of radio signals.

[0050] S2. Adopt the sliding window technique to extract the frequency point corresponding to the maximum amplitude from each window, and then extract the repeated pattern of the frequency through adaptive Gaussian filtering and downsampling. It should be noted that the sliding window technique described in S2 is to extract the frequency point corresponding to the maximum amplitude from each window, and then extract the repeated pattern of the frequency through adaptive Gaussian filtering and downsampling: Among them, the original signal is , the sliding window signal is , the window size is , is the index of the window, is the sliding step. Then, calculate the Fourier transform of each window signal, and convert the time-domain signal into the frequency-domain signal . Through the frequency-domain signal obtained by the Fourier transform, we can extract the frequency value corresponding to the maximum amplitude. This frequency value represents the main frequency component of the signal within this window, that is: After obtaining the frequency After that, first, apply adaptive Gaussian filtering to the extracted frequency sequence. Within each window, determine the appropriate standard deviation of the Gaussian filter by calculating the standard deviation of the local signal: ; Ensure that zero standard deviation does not occur. This standard deviation is used to smooth the frequency signal, ensuring effective smoothing for different noise levels in the signal. Specifically, the filtering process can be expressed as: Then, downsample the filtered frequency values to remove redundant information and extract the repeating patterns of the frequency. Assume the extracted frequency sequence is , and the downsampling process reduces frequency redundancy through sampling at intervals of , that is, the downsampled frequency sequence is obtained: Through this process, the main changes of the signal in time and frequency can be effectively captured, data redundancy can be reduced, and the efficiency of subsequent processing can be improved.

[0051] It should be noted that in this embodiment, the window size is 256, and the window overlap is 0.5. This overlap can ensure the time continuity between adjacent windows and avoid boundary effects and information loss caused by window segmentation. There are 256 sampling points. 1000 samples of 15 types of drones are collected, and the classification categories are N + 1 = 16 categories. The training set and test set are divided according to 8:2. Each sample contains time-series data points of two channels, and each channel contains 140 million data time-series points.

[0052] S3. Generate RP maps and Gram angular field maps through transformation of the extracted frequency patterns for signal feature representation; It should be noted that in S3, generating RP maps and Gram angular field maps through transformation of the extracted frequency patterns for signal feature representation includes the following steps: S301. Capture the self-similarity of the signal through the RP map for the extracted frequency change pattern to describe the dynamic evolution characteristics of the signal S302. Reveal the periodic structure characteristics of the signal through the polar coordinate transformation of the GASF map for the extracted frequency change pattern It should be noted that for generating the RP map. According to the Takens embedding theorem, we can transform the one-dimensional time-series signal into a multi-dimensional phase space through the delay embedding method of the time series. The original time-series signal is , according to the Takens theorem, the embedding dimension is , and the delay time is , then the vector after phase space reconstruction can be expressed as: Among them, is the number of reconstructed vectors, is the time delay is the embedding dimension. This process generates a multi-dimensional phase space through which the dynamic behavior of the signal can be described.

[0053] Once the phase space is reconstructed, the similarity between two time points can be calculated. Euclidean distance is usually used to measure the similarity between two time points. Specifically, for two moments and , the distance calculation formula in the reconstructed space is: Next, a threshold is defined. According to the similarity measure and the threshold to construct the Heaviside function to determine whether there is similarity between two time points: When , it means there is a point at the position in the two-dimensional coordinate graph, otherwise there is no point. In this way, we can obtain a matrix representing the similarity between two time points.

[0054] It should be noted that for generating the Gram angle field diagram. First, for the time series signal to be processed is normalized so that the value range of the signal is limited between [−1,1]. The normalization formula is as follows: Among them, and are the maximum and minimum values of the sequence respectively. By this step, the dimension difference is eliminated to ensure the consistency of the data scale.

[0055] The normalized sequence is converted to polar coordinate form: Among them, represents the phase angle at the th moment, and its value range is , which is used to encode the signal amplitude information; represents the polar radius at the th moment, is the timestamp index, is the time regularization factor (usually taking the total length of the sequence) and is used to linearly map the timestamps to the interval , and retain the temporal position features.

[0056] Construct two types of Gram matrices through phase angle operations. The specific method is as follows: Generate the Gram matrix through the inner product of the sum of phase angles: This matrix captures the global correlation features of the time series signal through the collaborative angle change.

[0057] Generate the Gram matrix through the inner product of the phase angle difference: This matrix characterizes the dynamic fluctuation features of the time series signal through local angle differences.

[0058] Due to the high-frequency characteristics, radio frequency signals are vulnerable to random noise interference. The traditional Gram Angle Difference Field (GADF) is sensitive to instantaneous phase differences, which may lead to noise amplification. GASF suppresses the random perturbations in the high-frequency fluctuations through the sum of phase angle operations The experiment shows that on the dataset used in the experiment, the noise-to-signal ratio (NSR) tolerance of GASF is improved by 13% compared with GADF.

[0059] Swin Transformer relies on the attention mechanism to model long-range dependencies. The GASF matrix explicitly retains the amplitude collaboration between any two moments, which is consistent with the calculation logic of the self-attention weight matrix, thereby enhancing the model's ability to identify the periodicity and modulation mode of radio frequency signals; S4: Based on the cross-attention mechanism, obtain the dynamic evolution features of radio frequency signals by constructing a recursive graph, use the Gram angle field diagram to characterize the periodic features and global structure features of the signal, and use a multi-scale convolutional neural network combined with the Swin Transformer framework to achieve cross-scale feature fusion and feature enhancement: It should be noted that in step S4, the input signal is jointly feature-encoded through the Gram Angle Sum Field (GASF), and its core process is as follows: S401. Design a two-stream cross-attention module to achieve feature alignment between the RP graph and the GASF graph: The input signal first generates dynamic evolution features through the recursive graph (RP), reflecting the transient behavior and non-stationary characteristics of the signal; at the same time, generate periodic structure joint features through the Gram Angle Sum Field (GASF) to encode the amplitude coupling relationship and periodic pattern of the signal. Achieve semantic alignment of the two types of features, and use the GASF graph feature vector as the query vector (Query); use the RP graph feature vector as the query vector (Query); use the RP graph feature vector Split into a key vector (Key) and a value vector (Value).

[0060] Calculate the interaction weight through scaled dot-product attention: Among them, , output the aligned fused features ; S402, Multi-scale CNN feature extraction, construct a cascaded convolutional layer group to extract local-global features of the RF signal: Input the fused features into the cascaded multi-scale CNN module, where the 3×3 convolutional layer group extracts the high-frequency details of the pulse waveform, the 5×5 dilated convolutional layer captures the mid-scale context of the carrier period, and the 7×7 depthwise separable convolutional layer models the global envelope characteristics of the broadband spectrum. After the multi-branch outputs are concatenated along the channel dimension, a gated network is used to generate spatially adaptive weights to suppress noise channels and enhance discriminative features, and output multi-scale fused features ; S403, Swin Transformer global dependency modeling: Convert the multi-scale features into a sequence of image patches and input them into the Swin Transformer module. By introducing the cyclic shift window strategy, fixed-window self-attention and shifted-window self-attention are alternately executed in adjacent network layers, and the calculation formula is: Among them, the mask matrix ensures the causal constraint of the shifted window. This mechanism models long-range dependencies while reducing the computational complexity and outputs global semantic features ; S404, Joint representation generation for adaptive weighted fusion of multi-modal features: After obtaining the cross-modal aligned features , multi-scale convolutional features and the Swin Transformer global features , multi-modal feature fusion and classification are realized through the following process. First, for the heterogeneous feature tensors output by different branches, bilinear interpolation and 1×1 convolution are used to align the spatial resolution and the number of channels. Adjust the spatial dimensions of all feature tensors to the same size . Finally, feature tensors of a unified size are obtained , and .

[0061] Concatenate the above aligned features along the channel dimension into The input three-dimensional gating network generates dynamic fusion weights and outputs the original weight tensor . To ensure weight normalization, for each spatial position apply the Softmax function along the modal dimension to calculate the normalized weight coefficients: Obtain the final weight matrix , satisfying Based on the weight matrix , perform element-wise weighted fusion on the aligned multi-modal features: Input the joint representation into the global average pooling layer to compress it into a feature vector , and then map it to the class space through two fully connected layers to output the class probability distribution of the RF signal.

[0062] S5: Process the fused feature vector through a fully connected layer to output the predicted UAV RF signal class, realizing the prediction of the classification result.

[0063] For the classification prediction module of the model, a fully connected layer based on softmax normalization is used to construct the class probability mapping space. The mathematical form of this layer is defined as: where is the learnable weight matrix to realize the linear projection from the feature space to the class space; is the bias vector used to adjust the classification decision boundary; the softmax activation function maps the output to a probability distribution through the logit transformation , and its -dimensional element represents the posterior probability that the sample belongs to class . Here is the feature vector modulated by the attention mechanism, generated by multi-scale feature dynamic weighted fusion, and contains cross-modal semantic information.

[0064] In the model training stage, the cross-entropy loss function in the form of negative log-likelihood is used to measure the difference between the predicted distribution and the true label: where is the number of batch samples, is the total number of classes, represents the indicator function after one-hot encoding of the true class label of the -th sample, is the corresponding predicted probability. During the backpropagation process, this loss function optimizes the network parameters through the gradient descent algorithm. Its mathematical properties can be decomposed into two terms: the KL divergence measures the information difference between the true distribution and the predicted distribution, and the entropy constraint term suppresses the distribution chaos of the prediction results, thereby enhancing the inter-class separability and improving the model generalization ability simultaneously during the optimization process. To improve numerical stability, logarithmic summation techniques are used in actual calculations to avoid probability underflow, and label smoothing techniques are introduced to mitigate the overfitting risk.

[0065] Reference Figure 3 , Figure 4 , Figure 5 and Figure 6 , Figure 3 is the comparison graph of the recognition accuracy curves during the training process. It can be seen from the graph that the recognition accuracy using the fusion of dynamic evolution and periodic structure features and cross-attention is better than that using only dynamic evolution features and the joint representation without cross-attention. Figure 4 , Figure 5 and Figure 6 are the confusion matrices using only dynamic evolution features, the confusion matrix of the fusion of dynamic evolution and periodic structure features, and the confusion matrix of the fusion of dynamic evolution and periodic structure features and cross-attention respectively. It can be seen from the graph that the fusion of dynamic evolution and periodic structure features and cross-attention has a better classification effect on the UAV category.

[0066] The above-described embodiments are only preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are within the protection scope of the present invention. The protection scope of the present invention is subject to the claims.

Claims

1. A radio frequency signal classification method based on dynamic evolution and periodic structure dual-stream cross-attention fusion, characterized by: The following steps are involved: S1, numbers the predicted types of drones; S2, using sliding window technology, extracts the frequency point corresponding to the maximum amplitude from each window, and then extracts the repetitive pattern of frequency through adaptive Gaussian filtering and downsampling; S3, transforming the extracted frequency pattern to generate RP diagram and Gram angle field diagram to represent the characteristics of the signal; S4, based on the cross-attention mechanism, obtains the dynamic evolution characteristics of the RF signal by constructing a recursive graph, uses the Gram angular field diagram to represent the periodic structure characteristics and global structure characteristics of the signal, and uses a multi-scale convolutional neural network combined with the SwinTransformer framework to achieve cross-scale feature fusion and feature enhancement; S5: Process the fused feature vector through a fully connected layer and output the predicted drone RF signal category to achieve the prediction of the classification result.

2. The radio frequency signal classification method of dynamic evolution and periodic structure dual-stream cross-attention fusion according to claim 1 is characterized by: In the step S1, the predicted types of drones are numbered, and the numbering includes the total number N of drones to be classified and the background reference signal category, which includes a total of N+1 drone categories.

3. The radio frequency signal classification method of dynamic evolution and periodic structure dual-stream cross-attention fusion according to claim 1 is characterized by: In step S2, a sliding window technique is used to extract the frequency point corresponding to the maximum amplitude from each window, and then the repetitive pattern of the frequency is extracted by adaptive Gaussian filtering and downsampling: The original signal is , the sliding window signal is , is the index of the window, is the sliding step length; Next, extract the frequency value corresponding to the maximum amplitude , the frequency value represents the main frequency component of the signal in the window, that is: The filtering process can be expressed as: in, is the smoothed output value, Indicates the original letter is at position The value at is the total number of traversed points, is the local standard deviation, is the index of the traversed neighborhood data point; is the time point of the smoothed result; Then, the filtered frequency values ​​are downsampled to remove redundant information and extract the repetitive pattern of frequencies.

4. The radio frequency signal classification method of dynamic evolution and periodic structure dual-stream cross-attention fusion according to claim 1 is characterized by: In step S3, the extracted frequency mode is transformed to generate an RP diagram and a Gram angle field diagram to represent the characteristics of the signal, which specifically includes the following steps: The extracted frequency variation pattern is used to capture the self-similarity of the signal through an RP diagram to describe the dynamic evolution characteristics of the signal; The extracted frequency variation pattern is transformed into polar coordinates of the GASF Gram angle field diagram to reveal the periodic structure characteristics of the signal; The one-dimensional time series signal is transformed into a multi-dimensional phase space through the delayed embedding method of the time series. This process generates a multi-dimensional phase space, through which the dynamic behavior of the signal can be described; Euclidean distance is usually used to measure the similarity between two time points. and , and the distance calculation formula in the reconstruction space is: in, Represents the Euclidean distance between high-dimensional state vectors; , A state vector representing the delayed embedding construction; , express; represents the offset value; the observation value of the original time series; , Indicates offset The observed value after Based on the similarity measure Construct the Heaviside function to determine whether there is similarity between two time points: in, represents a similarity measure; represents the normalized distance threshold; when When , it means that in the two-dimensional coordinate graph There is a point on the position, otherwise there is no point. In this way, we get a The matrix represents the similarity between two time points; To generate the Gram angle field diagram, firstly, the time series signal to be processed The normalization process is performed so that the signal value range is limited to [−1, 1]. The normalization formula is as follows: in, Represents normalized time series data; and are the maximum and minimum values ​​of the sequence respectively; The standardized sequence Convert to polar coordinates: in, Indicates The phase angle at the moment has a range of , used to encode signal amplitude information; Indicates The diameter of a moment, is the timestamp index, is the time regularization factor, used to linearly map timestamps to intervals , retaining the temporal position characteristics; Generate the Gram matrix by taking the inner product of the phase angle and: in, represents the Gram sum field matrix; , Represents the angle in the polar coordinate system, which is obtained by mapping the normalized value of the original time series to the polar angle; The Gram sum field matrix captures the global correlation characteristics of the time series signal through coordinated angle changes; The GASF is calculated by phase angle and Suppress random disturbances in high-frequency fluctuations; The Swin Transformer relies on the attention mechanism to model long-range dependencies, and the Gram sum field matrix is ​​passed through Explicitly preserve the amplitude coherence between any two moments.

5. The radio frequency signal classification method of dynamic evolution and periodic structure dual-stream cross-attention fusion according to claim 1 is characterized by: In step S4, the input signal is analyzed by Gram angle and field GASF based on the cross attention mechanism. The joint feature encoding includes the following steps: Using the two-stream cross attention module, the feature alignment of the RP graph and the GASF graph is achieved: The interaction weights are calculated by scaled dot product attention: Align Features Input cascade multi-scale CNN module and output multi-scale fusion features ; Multi-scale features Converted into image block sequence input Swin Transformer module output global semantic features ; The aligned multimodal features , and Perform element-by-element weighted fusion; In obtaining cross-modal alignment features , multi-scale convolutional features and Swin Transformer global features Then adjust the spatial dimensions of all feature tensors to a uniform size , and finally obtain a feature tensor of uniform size , and ; The splicing process is ; in, Represents the multimodal features after alignment of the two-stream cross-attention module; Represents the fusion weights of different features at corresponding spatial positions; represents the joint feature matrix obtained by element-by-element weighted fusion; Represents the aligned features of the cascaded multi-scale CNN module output, including multi-scale information from local to mid-level; Represents the global semantic features output by the Swin Transformer module, capturing long-range dependencies and non-local relationships.

6. The radio frequency signal classification method of dynamic evolution and periodic structure dual-stream cross-attention fusion according to claim 5 is characterized by: The following steps are used to align the features of the RP graph and the GASF graph: The GASF graph feature vector As the query vector Query; the RP graph feature vector Split into key vector Key and value vector Value; in, , output aligned fusion features .

7. The radio frequency signal classification method of dynamic evolution and periodic structure dual-stream cross-attention fusion according to claim 5 is characterized by: In the multi-scale CNN module, the following steps are followed: a 3×3 convolutional layer group is used to extract high-frequency details of the pulse waveform, a 5×5 dilated convolutional layer captures the mid-scale context of the carrier period, a 7×7 depth-separable convolutional layer models the global envelope characteristics of the broadband spectrum, and after channel splicing of multi-branch outputs, a gating network is used to generate spatial adaptive weights, suppress noisy channels and enhance discriminative features, and output multi-scale fusion features. .

8. The radio frequency signal classification method of dynamic evolution and periodic structure dual-stream cross attention fusion according to claim 1 is characterized by: In step S5, the fused feature vector is processed through a fully connected layer, and the predicted drone radio frequency signal category is output to achieve the prediction of the classification result, including the following steps: The category probability mapping space is constructed using a fully connected layer based on softmax normalization, and the mathematical form is defined as: ； in, Represents the final probability output of the model, indicating the confidence distribution of the input data in different categories; It is a learnable weight matrix that realizes the linear projection from feature space to category space; is the bias vector used to adjust the classification decision boundary; It is the feature vector modulated by the global attention mechanism, generated by dynamic weighted fusion of multi-scale features, and contains cross-modal semantic information; During the model training phase, the cross entropy loss function in the form of negative log-likelihood is used to measure the difference between the predicted distribution and the true label: in, Represents the difference between the model's predicted probability distribution and the true label distribution; is the number of batch samples, is the total number of categories, Indicates The indicative function of the true category label of each sample after one-hot encoding, is the corresponding predicted probability.

9. A radio frequency signal classification system with dynamic evolution and periodic structure dual-stream cross-attention fusion, 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, the method according to any one of claims 1 to 8 is implemented.

Images