Unmanned aerial vehicle signal modulation pattern identification method based on physical priori and dynamic routing

Abstract

The application discloses a kind of unmanned aerial vehicle signal modulation pattern identification methods based on physical priori and dynamic routing, comprising: obtaining original IQ timing data and extracting timing feature vector and physical priori feature vector;Signal lightweight rough judgment is carried out, and the initial probability distribution of the modulation pattern to which the signal belongs is obtained, and confidence evaluation index is calculated;Carry out confidence evaluation, directly output modulation pattern initial probability distribution according to the evaluation result or pass original IQ timing data to next step;IQ data passed is executed double-branch image conversion, generates time-frequency diagram data and constellation diagram data, and features are extracted based on Mamba architecture;Residual gated network is constructed, and the time-frequency texture features and geometric distribution features extracted are fused to generate integrated visual features, which are then input into KAN classifier to output the final modulation pattern identification result.The application greatly reduces the computing energy consumption of the system while maintaining high global recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of wireless communication and signal recognition technology, and more specifically, to a method for identifying UAV signal modulation patterns based on physical priors and dynamic routing. Background Technology

[0002] Drones are increasingly being used in low-altitude logistics, environmental monitoring, and public safety. Automatic Modulation Recognition (AMR) technology has become a key pre-emptive technology for reconnaissance nodes to quickly determine target attributes and take targeted jamming measures. By identifying the modulation pattern of the drone's communication link, surveillance equipment can quickly determine the target's attributes and take targeted jamming or surveillance measures.

[0003] Traditional methods for solving AMR problems mainly employ signal processing techniques, such as likelihood ratio testing based on probabilistic models, higher-order cumulants based on manual statistical feature analysis, and cyclic spectrum features. However, traditional methods have many drawbacks. Probabilistic model-based algorithms have extremely high computational costs, while feature analysis-based methods face difficulties in feature extraction under complex and variable conditions, and both methods have limited accuracy. In recent years, with the rapid development of deep learning, deep learning-based AMR technology (DL-AMR) has become the most advanced technique. By constructing multi-layer neural networks, DL-AMR can automatically extract complex features from modulated signals, thereby significantly improving recognition performance. For time-domain sequence methods, existing techniques often use Convolutional Neural Networks (CNNs) or Long Short-Term Memory (LSTM) networks to directly process IQ orthogonal sequences for recognition. However, when using IQ data, it is difficult to intuitively capture the geometric features of the signal in the frequency and phase domains, and the model convergence speed is slow. Subsequent research fully leveraged the advantages of convolutional neural networks in image classification tasks, preprocessing signals into time-frequency maps or constellation maps and using computer vision models such as ResNet and Visual Geometry Groups for recognition. However, due to severe image noise at low signal-to-noise ratios, recognition was limited. Furthermore, traditional frameworks such as CNNs and Transformers have high computational complexity when processing long sequences or high-resolution images, making it difficult to meet the low-latency and low-power requirements of drone surveillance for edge deployment. During recognition, using only time-frequency maps results in the loss of fine phase information, and the resolution of time-frequency maps generated from short burst signals is low. Using only constellation maps leads to blurred features at low signal-to-noise ratios, making it difficult to distinguish higher-order modulations, such as 16QAM and 64QAM. Therefore, existing single-modal recognition methods cannot achieve high recognition accuracy in complex environments.

[0004] Therefore, researchers have utilized multi-branch convolutional neural networks or attention mechanisms to extract and fuse multimodal features, which has improved recognition accuracy in complex environments to some extent. However, existing multimodal deep learning modulation recognition methods face several extremely severe technical bottlenecks when deployed in actual UAV adversarial scenarios. First, existing feature extraction methods lack awareness of the physical defects of UAV radio frequency hardware. For example, traditional deep learning models (conventional CNNs or feature fusion networks) only fit signals as ordinary data carriers, completely ignoring the aforementioned physical prior information. Second, existing multimodal parallel fusion architectures lead to overloading of computing power at UAV edge reconnaissance nodes. In pursuit of extreme recognition rates, existing multimodal modulation recognition networks generally adopt a "parallel hard computing" architecture, which means that the system operates indiscriminately regardless of the quality of the input signal or the clarity of the features. This parallel computing severely consumes resources. Summary of the Invention

[0005] To address the aforementioned technical problems in existing technologies, this invention proposes a method for identifying UAV signal modulation patterns based on physical priors and dynamic routing. It proposes a transient complex orbit rotation power operator to extract physical features from UAV signals, and integrates a state-space model (Mamba) to extract features from the one-dimensional time-domain sequence, the two-dimensional frequency-domain time-frequency diagram, and the two-dimensional modulation-domain constellation diagram. Furthermore, it combines a computationally adaptive coarse-to-fine two-level cascaded architecture and Kolmogorov-Arnold network (KAN) classification, enabling effective modulation pattern identification in complex environments. The specific technical solution is as follows:

[0006] A method for identifying UAV signal modulation patterns based on physical priors and dynamic routing includes:

[0007] Step 1: Acquire the raw IQ time series data of the UAV communication signal and perform dual-branch lightweight extraction to obtain the time series feature vector and the physical prior feature vector; wherein, a transient complex orbit rotation power operator is applied to the raw IQ time series data, and the transient complex orbit rotation power sequence is obtained by traversing, and the mean, variance, and peak-to-average power ratio of the transient complex orbit rotation power sequence are extracted, and the physical prior feature vector is obtained by concatenating them; the transient complex orbit rotation power operator calculates the product of the transient curvature torque and the transient angular frequency of adjacent sampling points in the complex plane in the time series, and eliminates the denominator term of the angular frequency calculation through energy algebraic coupling to avoid the signal envelope approaching zero and diverging;

[0008] Step 2: Based on the physical prior feature vector and the temporal feature vector, perform a light coarse judgment on the signal to obtain the initial probability distribution of the modulation pattern to which the signal belongs, and calculate the information entropy as the confidence evaluation index.

[0009] Step 3: Perform confidence assessment based on preset thresholds, and output the initial probability distribution of modulation pattern directly or pass the original IQ time series data to the next step based on the assessment results;

[0010] Step 4: Perform a two-branch image transformation on the IQ data passed in Step 3 to generate time-frequency graph data and constellation graph data with uniform scale and pixel normalization;

[0011] Step 5: Based on the Mamba architecture, feature extraction is performed on the time-frequency map data and constellation map data respectively to obtain time-frequency texture features and geometric distribution features;

[0012] Step 6: Construct a residual gating network to fuse time-frequency texture features and geometric distribution features to generate comprehensive visual features, which are then input into the KAN classifier to output the final modulation pattern recognition result.

[0013] Furthermore, in step 1, the extraction of the temporal feature vector specifically involves: calculating the total energy of the original IQ time series data, and then scaling each sampling point in the time series based on the total energy to complete the global energy normalization of the original IQ time series data; performing adaptive soft thresholding denoising transformation on the energy-normalized time series data: obtaining an adaptive threshold through global average pooling and fully connected layer learning, and using this threshold to perform a shrinking operation on the energy-normalized time series data to achieve noise suppression; inputting the denoised energy-normalized time series data into the Mamba state space model to extract the temporal feature vector.

[0014] Furthermore, step 2 specifically involves: performing heterogeneous feature alignment on the temporal feature vector and the physical prior feature vector, concatenating them along the channel dimension, and then inputting them into a lightweight classifier constructed by a multilayer perceptron. After being mapped by an activation function, the initial probability distribution of the signal belonging to various modulation patterns is output, and the information entropy of the initial probability distribution is calculated as a confidence evaluation index.

[0015] Furthermore, in step 3, the calculated information entropy is compared with a preset gating threshold. If the information entropy is lower than the threshold, it is determined to be an easily identifiable signal, and no further process is performed. The modulation pattern corresponding to the maximum value in the initial probability distribution is directly output. If the information entropy is higher than the threshold, it is determined to be a difficult-to-identify signal, and the original IQ time series data is passed down for deep image analysis.

[0016] Furthermore, in step 4, a smooth pseudo-Wigner-Ville distribution is used to perform time-frequency transformation on the original IQ time series data: First, the discrete instantaneous autocorrelation function of the original IQ time series data is calculated. By introducing a time-domain smoothing window and a frequency-domain smoothing window, a two-dimensional filter is applied to the discrete instantaneous autocorrelation function to obtain the time-frequency image matrix.

[0017] Furthermore, in step 4, the original IQ time series data is mapped into a two-dimensional spatial distribution image: First, coordinate mapping is performed. In defining a two-dimensional network on the complex plane, the I-path data of the original IQ time series data is mapped to the horizontal coordinate and the Q-path data is mapped to the vertical coordinate. Then, the horizontal and vertical coordinates are rasterized to construct a constellation image matrix. The number of samples falling into each pixel grid point is counted, and the constellation image matrix is ​​Gaussian blurred to enhance feature connectivity.

[0018] Furthermore, in step 4, bicubic interpolation is used to enlarge both the time-frequency image matrix and the constellation image matrix to a uniform size, and the image coordinate pixels are mapped to... The interval is then normalized in terms of size and pixels to obtain time-frequency graph data and constellation graph data.

[0019] Furthermore, in step 5, the normalized time-frequency map data and constellation map data are input into the Vision Mamba visual backbone network for bidirectional feature extraction: first, the image is cut into multiple non-overlapping patches and then sequentially flattened. Then, the state space mechanism is used to perform forward and backward bidirectional scanning perception and feature weighted fusion on the flattened sequence to obtain time-frequency texture features and geometric distribution features, respectively.

[0020] Furthermore, in step 6, after aligning the time-frequency texture features and geometric distribution features obtained in step 5, the information is input into a visual gating network constructed from two layers of perceptrons. The weight coefficients of the network are calculated, and the time-frequency texture features and geometric distribution features are dynamically weighted using the weight coefficients. A residual connection mechanism is introduced to generate a comprehensive visual feature vector that contains both frequency evolution information and spatial topological regularity.

[0021] Furthermore, in step 6, the comprehensive visual feature vector is input into a Kolmogorov-Arnold network classifier. The classifier has a learnable nonlinear activation function on the connection edge, which is composed of a weighted combination of basis functions and B-spline functions, and outputs the final modulation pattern recognition result.

[0022] The beneficial effects of the present invention include at least the following:

[0023] (1) A transient complex orbit rotation power operator is proposed. In a physical sense, it is equivalent to extracting the transient rotational mechanical power of UAV signals on the complex plane state manifold, thereby realizing the unique physical characteristics of instantaneously capturing UAV signals.

[0024] (2) A coarse-to-fine two-stage cascaded architecture with adaptive computing power was constructed. For signals with obvious features or at a high signal-to-noise ratio, the system will trigger a mechanism to exit the subsequent process after completing the first stage. This architecture significantly reduces the system's computing power consumption while maintaining extremely high global recognition accuracy.

[0025] (3) By using the collaborative mechanism of time-domain features, physical features, time-frequency texture features and geometric distribution features, the blind zone of a single mode under specific channel impairment is effectively overcome, and the recognition robustness is significantly improved. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for identifying UAV signal modulation patterns based on physical priors and dynamic routing according to an embodiment of the present invention;

[0027] Figure 2 This is a diagram of the adaptive soft threshold denoising network structure according to an embodiment of the present invention;

[0028] Figure 3 This is a diagram of the temporal feature extraction network structure according to an embodiment of the present invention;

[0029] Figure 4 This is a diagram of the image feature extraction network structure according to an embodiment of the present invention;

[0030] Figure 5 This is a trend curve of the recognition accuracy of the two-level cascaded and computing power adaptive model using the method of this invention as a function of signal-to-noise ratio;

[0031] Figure 6 This is a graph showing the variation of the first-level network computing power interception rate with the signal-to-noise ratio under different preset gating thresholds in this embodiment of the invention.

[0032] Figure 7 This is a performance report of the two-level cascade and computing power adaptive model proposed by the method of this invention on test cases;

[0033] Figure 8 It is a performance report of a traditional three-modal parallel model under the same test conditions. Detailed Implementation

[0034] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0035] This embodiment discloses a method for identifying UAV signal modulation patterns based on physical priors and dynamic routing. It constructs an asymmetric coarse-to-fine two-level cascaded architecture based on physical priors and adaptive computing power. This architecture uses transient complex orbit rotation power as the underlying physical factor and temporal dynamic evolution characteristics as the first-level lightweight evaluation base, Shannon information entropy to construct gating threshold triggering logic as dynamic routing, and dual-domain visual features of time-frequency texture and constellation topology combined with nonlinear boundaries (Kolmogorov-Arnold Networks KAN) as the second-level deep precision judgment core. Through this cascaded mechanism, the dynamic allocation of computing power under resource-constrained conditions and high-precision identification under extreme channels are achieved.

[0036] like Figure 1 As shown, the UAV signal modulation pattern recognition method includes the following steps:

[0037] Step 1: Acquire the IQ data of the UAV communication signal and perform lightweight dual-branch extraction to obtain the temporal feature vector and the physical prior feature vector. The dual-branch extraction includes a physical prior feature extraction branch and a temporal feature extraction branch.

[0038] Specifically, let the length of the received discrete-time UAV complex baseband communication signal sequence be N, and let the original one-dimensional IQ data at the nth sampling point be represented in complex form:

[0039]

[0040] in, , and These are the in-phase and quadrature components of the signal, respectively. This sequence serves as the input to both the physical prior feature extraction branch and the temporal feature extraction branch.

[0041] In the physical prior feature extraction branch, this invention creatively couples the transient curvature torque and transient angular frequency of the signal, deriving a novel transient complex orbit rotational power operator. Physically, this operator is equivalent to extracting the transient rotational mechanical power of the UAV signal in the complex plane state propagation. Mathematically, the algebraic coupling of the energy factor eliminates the denominator term in the extracted angular frequency, thus eliminating the numerical divergence problem at zero-crossing points under low signal-to-noise ratio conditions.

[0042] Existing technologies for extracting transient angular frequencies In this case, a discrete approximation formula is usually used:

[0043]

[0044] For discrete signals, for , for Because its denominator contains the square of the instantaneous envelope of the signal. When the drone signal experiences deep fading or when the envelope approaches zero due to a low signal-to-noise ratio, a numerical divergence problem due to division-to-zero overflow can occur.

[0045] To address this deficiency from a mathematical perspective, this invention treats the two-dimensional IQ signal as a dynamic particle on a complex plane; different modulation patterns can be considered as different curvatures of the particle's trajectory. First, the first-order transient velocity difference vector between adjacent sampling points is calculated. and the second-order transient acceleration difference vector :

[0046]

[0047]

[0048] In physics, the cross product of velocity and acceleration represents the transient curvature of a particle's motion, and the transient curvature torque characterizing a drone signal. .

[0049] Subsequently, the transient curvature torque Instantaneous envelope energy factor of the signal and transient angular frequency By performing algebraic coupling, a transient complex orbital rotation power operator without numerical divergence risk is derived. :

[0050]

[0051]

[0052]

[0053]

[0054] During the coupling multiplication process, the instantaneous envelope energy factor acts like a dynamic adjustment lever, perfectly offsetting the denominator term in the traditional angular frequency extraction formula. This means that even when the UAV communication link encounters an extremely severe sudden deep fading, causing the signal envelope to drop sharply or even approach zero, the physical operator of this pure multiplication-addition operation will never experience division-to-zero overflow or numerical collapse.

[0055] By using the pure multiplication-addition operators described above to traverse the entire signal sequence, the mean, variance, and peak-to-average ratio of the operator sequence are extracted to construct a lightweight physical prior feature vector. The extraction of the mean, variance, and peak-to-means ratio is as follows:

[0056] First, the mean of the power sequence is calculated to characterize the overall energy level of the signal during the complex plane rotation process. Its expression is:

[0057]

[0058] Secondly, variance is calculated to measure the intensity of dynamic fluctuations caused by the work done by the signal during rotation in space. Since FSK constant envelope modulation has a smooth power manifold, its variance approaches zero; while QAM higher-order quadrature amplitude modulation exhibits drastic sign transitions, resulting in a significantly increased variance. Variance characteristics The calculation formula is:

[0059]

[0060] Furthermore, to characterize the relationship between the peak and average energy in the power sequence, the peak-to-average power ratio (PAPR) is calculated:

[0061]

[0062] By combining the above statistical measures, the dynamic characteristics of UAV signals on the complex plane state manifold can be characterized from different dimensions. Among them, the mean reflects the overall power level, the variance reflects the severity of power fluctuations, and the peak-to-average power ratio can effectively distinguish between high-order modulation signals with sudden sign transitions and smooth, constant-envelope modulation signals.

[0063] The mean, variance, and peak-to-average ratio are concatenated to form a lightweight physical prior feature vector, which serves as the output of the branch for the first-level lightweight physical feature extraction.

[0064] In the timing feature extraction branch, the original one-dimensional IQ timing data is processed. Since the amplitude fluctuation range of the received signal is large, global energy normalization is performed on the original one-dimensional IQ timing data in order to accelerate network convergence. Here, energy normalization is preferred over amplitude normalization because wireless communication is often affected by impulse noise. Amplitude normalization is easily affected by extreme noise, which can lead to excessive compression of the effective signal. Energy normalization, on the other hand, is based on the overall power statistics of the signal and retains the key peak-to-average power ratio characteristics of different modulation styles, thereby improving the robustness of the model in complex environments.

[0065] Specifically, another expression for the received discrete-time wireless communication signal. for:

[0066]

[0067] in, For in-phase components, These are orthogonal components, where N is the length of the signal sequence. It is a complex unit.

[0068] First, calculate the total energy E of the current sample sequence:

[0069]

[0070] Then, each sampling point in the sequence is scaled to obtain the preprocessed sequence:

[0071]

[0072] Subsequently, based on the Mamba architecture, time-series features were extracted. Specifically:

[0073] The preprocessed sequence Let X be the threshold for adaptive soft thresholding denoising transformation, using the threshold... The input sequence is shrunk to obtain the denoised features. :

[0074]

[0075] This function sets near-zero eigenvalues ​​to zero while retaining negative eigenvalues ​​that contain useful information, thereby achieving noise suppression.

[0076] More specifically, to determine the optimal threshold, an adaptive soft thresholding denoising network is used, and the network structure diagram is as follows: Figure 2 Let the characteristics of the input normalized IQ sequence be... B is the batch size, L is the sequence length, and C is the feature dimension.

[0077] First, the statistical feature vector of the channel dimension is calculated using global average pooling (GAP). :

[0078]

[0079] Subsequently, adaptive scaling parameters are learned through fully connected layers. :

[0080]

[0081] in, , The weights of the two fully connected layers correspond to the following: For the Sigmoid function, This is the activation function.

[0082] The final denoising threshold is calculated by element-wise multiplication. The calculation is as follows:

[0083]

[0084] The obtained denoising features Input a one-dimensional Mamba state-space model and extract long-range temporal dependency features. The core of Mamba is the state equation based on a continuous linear time-invariant system.

[0085]

[0086]

[0087] in, The input is time t; The hidden state vector is N, where N is the state dimension. The state evolution matrix is... For the input projection matrix, This is for outputting the projection matrix.

[0088] To process discretized sequences, the zero-order preservation principle is used to discretize the continuous parameters (A, B). It's not fixed, but rather changes with the current input. A changing function.

[0089] For the input time t The system will dynamically generate the time step, and the specific calculation and generation process is as follows:

[0090]

[0091] Here, Softplus is a smooth activation function to ensure that the generated step size is always positive; Parameter is a learnable bias constant; and Linear is a linear transformation layer that maps the input signal features to the step size space, enabling the model to dynamically generate corresponding weights based on the current signal content. △ t As a time-scale parameter, the model's memory step size can be dynamically adjusted by learning the content of the input signal, thereby enabling the model to adaptively ignore noise or memorize key features.

[0092] Subsequently, the discretized state evolution matrix is ​​calculated using the dynamic time step. With input projection matrix :

[0093]

[0094]

[0095] in , , where I is the identity matrix.

[0096] Subsequent recursive state updates utilize discretization parameters to perform recursive scans:

[0097]

[0098]

[0099] It captures long-range temporal dependencies in IQ time series, and the final output of the Mamba block is combined with a gating mechanism:

[0100]

[0101] Finally, the physical prior feature vector extracted by the transient complex orbit rotation power operator is obtained. Temporal feature vectors extracted with 1D-Mamba .

[0102] Step 2: Perform a light coarse judgment on the physical prior feature vector and temporal feature vector extracted in Step 1, and calculate the confidence evaluation index using information entropy.

[0103] Specifically, for the extracted physical prior feature vectors and time series feature vectors Perform heterogeneous feature alignment.

[0104] The two types of features are mapped to a unified feature dimension through a linear projection layer, and then concatenated along the channel dimension to generate the first-level fused feature vector. :

[0105]

[0106] in, This represents a vector concatenation operation. It is a non-linear activation function. and , , where is the learnable projection matrix and bias.

[0107] Fusion features A lightweight classifier consisting of a fully connected network is input. This classifier maps features to the modulation class space through a linear transformation and uses the Softmax function to calculate the initial probability distribution of the current signal belonging to M modulation patterns. :

[0108]

[0109]

[0110] in, This is the weight matrix of the classification layer. This is the bias vector for the classification layer. The score for the i-th modulation pattern. This represents the predicted probability that the signal belongs to the i-th modulation pattern. It is the Euler number.

[0111] To quantitatively assess the reliability of the first-level coarse-judgment results, the "Shannon entropy" from information theory is introduced as a confidence evaluation index H. Information entropy, in a physical sense, characterizes the determinism of a probability distribution:

[0112]

[0113] The physical logic of information entropy is: when the probability distribution P is steep, that is, for a certain category... When the value is close to 1 and the rest are close to 0, the entropy value H approaches the minimum, indicating that the recognition result has a very high degree of certainty; conversely, when the distribution is flat, the entropy value H increases, indicating that the current signal features are blurred and there is a risk of recognition confusion.

[0114] Step 3: For high confidence, directly output the modulation pattern in the initial probability distribution; for low confidence, pass the original IQ data down.

[0115] Specifically, the system executes adaptive computing power allocation and gating routing logic. The calculated information entropy H is compared with a preset gating threshold. Perform dynamic comparison:

[0116] when When the system determines that the current input is a regular, easily identifiable signal with significant features and minimal interference, it immediately triggers an Early-Exit instruction, refusing to initiate subsequent time-consuming time-frequency diagram / constellation diagram generation and high-dimensional visual network inference, and directly outputs the modulation pattern corresponding to the maximum value in the current initial probability distribution.

[0117] when When the system determines that the current signal features are ambiguous, such as being at an extremely low signal-to-noise ratio or encountering severe co-channel interference, there is a risk of recognition confusion. In this case, the gating network marks it as a problematic signal, activates and wakes up the second-level heavy computing resources as needed, passes the original IQ data down, and initiates the subsequent dual-branch deep image analysis process to salvage the features. This logic ensures that good signals take a shortcut and poor signals go to deeper layers, significantly reducing average inference power consumption.

[0118] Step 4: Perform dual-branch image transformation on the IQ data signal passed in Step 3 to generate time-frequency map data and constellation map data with uniform scale and pixel normalization.

[0119] To overcome the cross-term interference problem inherent in the traditional Wigner-Ville distribution (WVD) while maintaining better time-frequency clustering than the Short-Time Fourier Transform (STFT), a smoothed pseudo-Wigner-Ville distribution (SPWVD) is used for time-frequency transformation. Compared to STFT, SPWVD is not limited by the uncertainty principle and has extremely high time-frequency resolution; compared to ordinary WVD, the dual-window smoothing mechanism effectively suppresses cross-term interference between multi-component signals, resulting in a clearer time-frequency texture. First, the discrete instantaneous autocorrelation function of the signal is calculated. For complex baseband signals The formula is shown in expression (1.11):

[0120]

[0121] in, Indicates a time index. represents the time delay variable, and * denotes the complex conjugate operation. The SPWVD formula introduces a time-domain smoothing window. and frequency domain smoothing window right For two-dimensional filtering, its discrete expression is:

[0122]

[0123] Where k is the frequency index and K is the total number of frequency points (corresponding to the height of the generated time-frequency map). It is a time-domain smoothing window function with a window length of . It is used in the time domain for smoothing and to suppress cross terms; It is a frequency domain smoothing window function with a window length of . This is used for smoothing in the frequency domain. Subsequent calculations... The modulus value is obtained and the output time-frequency image matrix is ​​mapped to reflect the information of frequency change over time.

[0124] Another approach maps the one-dimensional IQ time-series signal into a two-dimensional spatial distribution image. First, coordinate mapping is performed; a two-dimensional network is defined on the complex plane, mapping the I-channel data into a coordinate system. Q-path data is mapped to the ordinate. For signals Its discrete coordinates in the image matrix The calculation is as follows:

[0125]

[0126]

[0127] Where W and H are the width and height of the target image, This indicates rounding down to the nearest integer.

[0128] Then, to adapt to Vision Mamba, the discrete coordinates were rasterized to construct a constellation image matrix. Count the number of samples falling into each pixel grid point:

[0129]

[0130] in Let be the impulse function.

[0131] To enhance the connectivity of features, Apply Gaussian blur:

[0132]

[0133] Among them Gaussian kernel. The processed image serves as the input to the third branch, reflecting the signal amplitude and phase distribution.

[0134] Bicubic interpolation was used on both the time-frequency image matrix and the constellation image matrix to reduce the original size. Enlarged to a uniform size The pixels are normalized to obtain time-frequency graph data and constellation graph data, which are beneficial for the subsequent learning features of the network.

[0135] First, find the floating-point coordinates in the original image. Take the surrounding area A total of 16 neighboring points in the integer grid. ,in And the weight function Calculated 16 point pairs of target pixel coordinates The contribution weights are based on the neighbor and neighbor relationships. distance :

[0136]

[0137] in Take -0.5 or -0.75. Target coordinate pixel value. Calculated using the weighted convolution formula:

[0138]

[0139] in, This represents the pixel value of the original image at the specified coordinates. These are the weighting coefficients in the horizontal direction. These are the weighting coefficients in the vertical direction. Thus, the target image was obtained. Subsequently, to accelerate the convergence speed of Vision Mamba, pixel-level normalization was performed on the image, mapping its values ​​to... Interval:

[0140]

[0141] The final result The standard image for the Vision Mamba module is input to the two branches.

[0142] Step 5: Based on the Mamba architecture, extract features from the time-frequency map data and constellation map data obtained in Step 4 to obtain time-frequency texture features and geometric distribution features.

[0143] Specifically, for the two branches, bidirectional feature extraction is performed using the Vision Mamba visual backbone network. The input is a standardized image. The image is divided into N patches and flattened into a sequence. :

[0144]

[0145] in Let represent the vector of the i-th flattened image patch; E is a learnable linear projection matrix used to map image patches in pixel space to a high-dimensional feature space. This is a sequence concatenation operation; The location embedding vector is used to supplement the feature sequence with the absolute spatial location information of the image patch in the original image, so that the subsequent bidirectional scanning mechanism can perceive the topological structure of the image.

[0146] A bidirectional state-space scan is then performed, with the forward flow following the sequence. Scanning, simulating texture perception from the top left to the bottom right of the image:

[0147]

[0148]

[0149] in, Forward hidden state, represented as the state variable at position i, storing context information from the beginning of the sequence (top left of the image) to the current position. Let be the forward state transition matrix, representing the hidden state of the control at the previous time step. The proportion of updates passed to the current time step is the core of the State space midel (SSM), which determines the model's memory retention capability. This is the previous time step's forward hidden state. Let be the forward input projection matrix, representing the current input features. Mapped to the state space. Let be the feature of the input sequence at the i-th position. The forward output feature represents the result of projecting the hidden state onto the output, and includes texture information from the forward perception. The forward output projection matrix represents the projection of the hidden state. Projecting the features onto the output feature space yields the forward features at the current position.

[0150] Backflow is along the sequence Scanning, simulating texture perception from the bottom right to the top left of the image:

[0151]

[0152]

[0153] in, For backward hiding, it is represented as the state variable at the i-th position, storing context information from the end of the sequence (bottom right of the image) to the current position. This is the backward state transition matrix, representing the hidden state that controls the next time step. Updated to the current time. This is the backward hidden state for the next time step. The backward input projection matrix represents the current input features. Mapped to the state space. Let be the feature of the input sequence at the i-th position. The backward output feature represents the result of the hidden state after being projected onto the output, and includes the texture information perceived backward. The backward output projection matrix represents the projection of the hidden state. Projecting the features onto the output feature space yields the forward features at the current position.

[0154] The forward and backward features are then weighted and fused to obtain a feature sequence that includes the global context.

[0155]

[0156] in The sigmoid activation function is used. Subsequent average pooling of the sequence z yields the time-frequency texture features. and geometric distribution characteristics .

[0157] Step 6: Construct a residual gating network to fuse time-frequency texture features and geometric distribution features to generate comprehensive visual features, which are then input into the KAN classifier to output the final modulation pattern recognition result.

[0158] Specifically, the time-frequency texture features obtained in step 5 and geometric distribution characteristics First, it is mapped to a unified fusion space dimension through linear projection. After alignment, the time-frequency texture feature vector and constellation geometric feature vector are obtained, expressed as:

[0159]

[0160]

[0161] in, Indicates the aligned features. and These are the learnable weight matrix and the bias, respectively. This is a combination of activation function and normalization.

[0162] Then, visual feature fusion is performed, utilizing a complementary gating mechanism to fuse the aligned time-frequency texture feature vectors. Geometric eigenvectors of constellations Generate comprehensive visual features .

[0163] More specifically, a two-layer perceptron is first constructed as a visual gating network. The two visual features are concatenated and then input into the network to calculate the weight coefficients. :

[0164]

[0165]

[0166] in, This is a vector concatenation operation. This is the weight matrix. For activation functions; This is the Sigmoid function.

[0167] Subsequent use The two sets of features are dynamically weighted and then fused complementaryly to obtain a comprehensive visual feature vector.

[0168]

[0169] This complementary fusion mechanism can automatically prioritize time-frequency texture features when the constellation diagram is scattered, i.e., the frequency bias is severe, and automatically prioritize topological geometric features when the time-frequency cross terms increase, i.e. the amplitude and phase distortion is severe, thus realizing dynamic feature self-rescue under extremely poor channel conditions.

[0170] Input visual features using the KAN classifier , For the feature dimension, the output is , Let be the total number of modulation pattern categories. A linear layer constructed using a Kolmogorov-Arnold Network (KAN) maps features to the modulation category space and outputs the decision result. Unlike traditional Multilayer Perceptrons (MLPs) that use fixed activation functions (such as ReLU) and linear weight matrices, the core of the KAN layer lies in placing learnable nonlinear activation functions on the connection edges. For the j-th component of the output vector... Its calculation formula follows the Kolmogorov-Arnold formula:

[0171]

[0172] This uses a single-layer mapping formula. Let i be the i-th component of the input feature vector. For the bias term of the j-th output node, Let be a one-dimensional activation function connecting the i-th input node and the j-th output node. To achieve efficient fitting of complex decision boundaries, A weighted combination of basis functions and B-spline functions:

[0173]

[0174] The SiLU function serves as the basis function to capture the overall linear trend; a linear combination of k-order B-spline basis functions is used to fit local nonlinear fluctuations.

[0175]

[0176]

[0177] Where G is the number of grid points, which determines the fitting precision of the spline function; To define the k-th B-spline basis function on the mesh; Learnable spline control coefficients.

[0178] During training, the network updates the weights. , It will also dynamically adjust the control coefficient. This changes the function. The curve shape is then determined. The output vector z is subsequently input into the Softmax function to calculate the probability that the signal belongs to the j-th modulation category. :

[0179]

[0180] The final output recognition result is Take the category index with the highest probability.

[0181] In one specific embodiment, validation was performed using the publicly available dataset RML2016.10a. The experimental environment consisted of a Windows 10 operating system, the PyTorch deep learning framework, and a single NVIDIA RTX 5060ti GPU.

[0182] Dataset Description: This dataset contains 11 modulation styles: 8 digital modulations (BPSK, QPSK, 8PSK, 16QAM, 64QAM, GFSK, CPFSK, PAM4) and 3 analog modulations (WB-FM, AM-SSB, AM-DSB). The signal-to-noise ratio (SNR) ranges from -20dB to +18dB with a step size of 2dB. Each sample is of length... The IQ orthogonal sequence has a data dimension of (2, 128).

[0183] The specific steps of the UAV signal modulation pattern recognition method are as follows:

[0184] The dataset was divided into a 50% training set and a 50% test set, with a batch size of 32, 50 training rounds, and a gating threshold of 0.4.

[0185] Step 1: Feed the raw IQ data into two branches of the first-level lightweight network. In the physical prior feature extraction branch, first-order and second-order difference operations are performed on the input IQ sequence to extract the cross-difference term representing the signal curvature torque, and simultaneously extract the cross-difference term of the transient angular frequency molecule. By multiplying the absolute values ​​of the two and eliminating the risk of division-by-zero divergence through energy coupling, the transient complex orbit rotation power is obtained. The mean, variance, and peak-to-average power ratio of the power sequence are calculated, and a one-dimensional physical prior feature vector is output. .

[0186] In the temporal feature extraction branch, the L2 norm energy of the original IQ data is calculated, which is done by calculating the square root of the sum of squares of the IQ data and adding 0.000001 to prevent division by zero overflow. Each sampling point of the IQ sequence is divided by this energy factor to complete global energy normalization, preserving the peak-to-average ratio (PAR) feature. The normalized IQ data is then fed into an adaptive soft-threshold denoising module that combines global average pooling (GAP) and a two-layer perceptron to suppress channel noise. Figure 2 As shown. Subsequently, a linear layer is used to map to 64 dimensions, which is then fed into a stacked two-layer Mamba Block to extract long-range temporal dependency features. The network structure diagram for temporal feature extraction is as follows: Figure 3 As shown.

[0187] Step 2: The one-dimensional physical prior feature vector and temporal feature vector extracted in Step 1 are mapped to a unified dimension through a linear projection layer and then concatenated along the channel dimension. The first fused feature after concatenation is input into a lightweight classifier composed of a multilayer perceptron. After mapping by the Softmax activation function, the output is the initial probability distribution of the current signal belonging to the 11 modulation patterns in the RML2016.10a dataset. Subsequently, the Shannon entropy H of this initial probability distribution is calculated as the confidence evaluation index for the first-level preliminary judgment.

[0188] Step 3: Compare the calculated Shannon entropy H with the preset gating threshold ( =0.4) is compared to allocate computing power. When When the system determines that the current signal is a easily identifiable signal with significant features, it triggers a computational power cutoff command, rejects subsequent image extraction networks, and directly outputs the recognition result. When the system determines that the current signal is a difficult signal, it initiates a subsequent process for fine-tuning and passes the original IQ data to the subsequent process.

[0189] Step 4: Transform the 128-point sequence using a smoothed pseudo-Wigner-Ville distribution (SPWVD). Set the time-domain window to a Hamming window of length 31 and the frequency-domain window to a Hamming window of length 63. The generated time-frequency plot reflects the energy distribution of the signal in the time-frequency plane.

[0190] Map the I-channel data to the x-axis and the Q-channel data to the y-axis, normalizing the mapping range to [-1, 1]. Construct a 224*224 raster matrix, count the number of sampling points falling into each pixel, and apply... A Gaussian kernel is used for heatmap rendering to enhance the topological connectivity of sparse constellation points.

[0191] The generated time-frequency graph and constellation graph are uniformly adjusted to 224*224 pixels using a bicubic interpolation algorithm, and the pixel values ​​are normalized to the [0, 1] interval to adapt to the input requirements of Vision Mamba.

[0192] Step 5: Input the (B, 1, 224, 224) image. Segment the image into 16*16 patches, with a patch sequence length L=196, and linearly map it to a 192-dimensional feature space. Input the Vision Mamba encoder and extract time-frequency texture features using a bidirectional SSM scanning mechanism. and spatial topological features The image feature extraction network structure diagram is as follows: Figure 4 As shown.

[0193] Step 6: Construct a visual residual gating network to fuse two pure visual features. The time-frequency texture features and geometric distribution features are concatenated and fed into a two-layer perceptron to calculate dynamic weight coefficients. These weights are then used to dynamically weight the two visual features, generating weights via an MLP. Generate a comprehensive visual feature vector.

[0194] A KAN linear layer is constructed as the classification head, with an input dimension of 256 fusion feature dimensions and an output dimension of 11, corresponding to the 11 modulation categories of RML2016.10a. SiLU basis functions and B-spline functions are used as edge activation functions, with the number of basis functions G=5 and the order k=3. These are precisely mapped to the 11-dimensional modulation category space, and the final modulation pattern probability distribution is output by the Softmax function. The category corresponding to the highest probability is taken as the final recognition result.

[0195] Finally, the test results are evaluated. For example... Figure 5 As shown in the figure, this illustrates the trend of recognition accuracy of the two-stage cascaded and computationally adaptive models applying the method of this invention as a function of signal-to-noise ratio (SNR). According to the curve trend and the evaluation in the text, when the SNR is in the lower range of approximately -20dB to -10dB, the recognition accuracy is low and increases slowly. After the SNR crosses -10dB, the accuracy begins to rise sharply. Above -4dB, the model's recognition accuracy remains high, and then stabilizes in the subsequent high SNR range, reaching over 80%.

[0196] Figure 6 The graph presents the relationship between the first-level network's interception rate and the signal-to-noise ratio (SNR), and compares three different preset gate thresholds: 0.4, 0.6, and 0.8. As can be observed from the graph, with the increase of SNR, signal quality improves, and the interception rate rises rapidly before stabilizing in the high SNR region. Furthermore, the higher the gate threshold γ is set, such as γ=0.8, the higher the interception rate is at each SNR level.

[0197] Figure 7 This is a performance report on the test cases of the two-level cascaded and computationally adaptive model proposed using the method of this invention. The core data indicator is: with a total of 110,000 test samples, the total inference time of the model is 9.5204 seconds.

[0198] Figure 8 This is a performance report of a traditional trimodal parallel model under the same test conditions. The core data metric is: when processing the same total of 110,000 test samples, the total inference time of this parallel model increased to 12.9900 seconds.

[0199] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying UAV signal modulation patterns based on physical priors and dynamic routing, characterized in that, include: Step 1: Acquire the raw IQ time series data of the UAV communication signal and extract the time series feature vector and physical prior feature vector; wherein, a transient complex orbit rotation power operator is applied to the raw IQ time series data, and the transient complex orbit rotation power sequence is obtained by traversing, and the mean, variance, and peak-to-average power ratio of the transient complex orbit rotation power sequence are extracted, and the physical prior feature vector is obtained by splicing them; the transient complex orbit rotation power operator calculates the product of the transient curvature torque and transient angular frequency of adjacent sampling points in the complex plane in the time series, and eliminates the denominator term of the angular frequency calculation through energy algebraic coupling to avoid the signal envelope approaching zero and diverging; Step 2: Based on the physical prior feature vector and the temporal feature vector, perform a light coarse judgment on the signal to obtain the initial probability distribution of the modulation pattern to which the signal belongs, and calculate the information entropy as the confidence evaluation index. Step 3: Perform confidence assessment based on preset thresholds, and output the initial probability distribution of modulation pattern directly or pass the original IQ time series data to the next step based on the assessment results; Step 4: Perform a two-branch image transformation on the IQ data passed in Step 3 to generate time-frequency graph data and constellation graph data with uniform scale and pixel normalization; Step 5: Based on the Mamba architecture, feature extraction is performed on the time-frequency map data and constellation map data respectively to obtain time-frequency texture features and geometric distribution features; Step 6: Construct a residual gating network to fuse time-frequency texture features and geometric distribution features to generate comprehensive visual features, which are then input into the KAN classifier to output the final modulation pattern recognition result.

2. The UAV signal modulation pattern recognition method as described in claim 1, characterized in that, In step 1, the extraction of the time series feature vector specifically involves: calculating the total energy of the original IQ time series data, and then scaling each sampling point in the time series based on the total energy to complete the global energy normalization of the original IQ time series data; An adaptive soft thresholding denoising transformation is performed on the energy-normalized time-series data: an adaptive threshold is learned through global average pooling and fully connected layers, and this threshold is used to shrink the energy-normalized time-series data to achieve noise suppression; the denoised energy-normalized time-series data is then input into the Mamba state-space model to extract the time-series feature vector.

3. The UAV signal modulation pattern recognition method as described in claim 1, characterized in that, Step 2 specifically involves: performing heterogeneous feature alignment between the temporal feature vector and the physical prior feature vector, concatenating them along the channel dimension, and then inputting them into a lightweight classifier constructed by a multilayer perceptron. After being mapped by an activation function, the initial probability distribution of the signal belonging to various modulation patterns is output, and the information entropy of the initial probability distribution is calculated as a confidence evaluation index.

4. The UAV signal modulation pattern recognition method as described in claim 1, characterized in that, In step 3, the calculated information entropy is compared with a preset gating threshold. If the information entropy is lower than the threshold, it is determined to be an easily identifiable signal, and no further process is performed. The modulation pattern corresponding to the maximum value in the initial probability distribution is directly output. If the information entropy is higher than the threshold, it is determined to be a difficult-to-identify signal, and the original IQ time series data is passed down for deep image analysis.

5. The UAV signal modulation pattern recognition method as described in claim 1, characterized in that, In step 4, a smooth pseudo-Wigner-Ville distribution is used to perform time-frequency transformation on the original IQ time series data: First, the discrete instantaneous autocorrelation function of the original IQ time series data is calculated. By introducing a time-domain smoothing window and a frequency-domain smoothing window, a two-dimensional filter is applied to the discrete instantaneous autocorrelation function to obtain the time-frequency image matrix.

6. The UAV signal modulation pattern recognition method as described in claim 5, characterized in that, In step 4, the original IQ time series data is mapped into a two-dimensional spatial distribution image: First, coordinate mapping is performed. In the two-dimensional network defined on the complex plane, the I-path data of the original IQ time series data is mapped to the horizontal coordinate and the Q-path data is mapped to the vertical coordinate. Then, the horizontal and vertical coordinates are rasterized to construct a constellation image matrix. The number of samples falling into each pixel grid point is counted, and the constellation image matrix is ​​Gaussian blurred to enhance feature connectivity.

7. The UAV signal modulation pattern recognition method as described in claim 6, characterized in that, In step 4, bicubic interpolation is used to enlarge both the time-frequency image matrix and the constellation image matrix to a uniform size, and the image coordinate pixels are mapped to... The interval is then normalized in terms of size and pixels to obtain time-frequency graph data and constellation graph data.

8. The UAV signal modulation pattern recognition method as described in claim 1, characterized in that, In step 5, the normalized time-frequency map data and constellation map data are input into the Vision Mamba visual backbone network for bidirectional feature extraction: first, the image is cut into multiple non-overlapping patches and then sequentially flattened. Then, the state space mechanism is used to perform forward and backward bidirectional scanning perception and feature weighted fusion on the flattened sequence to obtain time-frequency texture features and geometric distribution features, respectively.

9. The UAV signal modulation pattern recognition method as described in claim 1, characterized in that, In step 6, after aligning the time-frequency texture features and geometric distribution features obtained in step 5, the information is input into a visual gating network constructed from two layers of perceptrons. The weight coefficients of the network are calculated, and the time-frequency texture features and geometric distribution features are dynamically weighted using the weight coefficients. A residual connection mechanism is introduced to generate a comprehensive visual feature vector that contains both frequency evolution information and spatial topological regularity.

10. The UAV signal modulation pattern recognition method as described in claim 1, characterized in that, In step 6, the comprehensive visual feature vector is input into the Kolmogorov-Arnold network classifier. The classifier has a learnable nonlinear activation function on the connection edge, which is composed of a weighted combination of basis functions and B-spline functions, and outputs the final modulation pattern recognition result.

Images