Blind Recognition and Intelligent Classification Prediction Method and System for Signal Modulation Mode

Abstract

This invention provides a method and system for blind identification and intelligent classification prediction of signal modulation patterns, relating to the field of signal processing technology. The method includes obtaining multi-scale time-domain features through wavelet packet decomposition, obtaining instantaneous frequency features through Hilbert transform, and constructing a feature correlation matrix; achieving blind identification of signal modulation patterns using a dynamic routing algorithm and a self-correction mechanism; distinguishing between discrete and continuous modulation based on the feature matrix and signal-to-noise ratio; and predicting signal modulation patterns using hybrid filtering and hierarchical Markov decision. This invention improves the accuracy and stability of signal modulation pattern identification and prediction.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a method and system for blind identification and intelligent classification prediction of signal modulation patterns. Background Technology

[0002] With the rapid development of modern communication technology, signal modulation pattern identification plays an increasingly important role in military reconnaissance, spectrum management, and wireless communication systems. Blind modulation pattern identification refers to the technique of analyzing and processing received signals to determine their modulation pattern without prior knowledge of the transmitted signal. Traditional modulation pattern identification mainly relies on human experience and specific feature extraction methods, such as instantaneous amplitude, instantaneous frequency, and higher-order cumulants. With the rise of deep learning technology, neural network-based modulation pattern identification methods have gradually become a research hotspot. These methods can automatically extract features from the original signal and achieve classification.

[0003] Existing feature extraction methods often focus only on single-dimensional signal features, either analyzing only time-domain features or only frequency-domain features, lacking an effective fusion mechanism for multi-dimensional features. This results in low recognition accuracy in complex signal environments, especially poor performance under low signal-to-noise ratio conditions.

[0004] Traditional modulation scheme recognition algorithms lack in-depth exploration of feature space relationships. Most of them use a combination of static feature extraction and simple classifiers, which cannot adaptively adjust feature representation and classification strategies according to changes in the signal environment, resulting in unstable recognition performance under dynamically changing channel conditions.

[0005] Existing technologies lack the ability to predict future trends in signal modulation patterns. Most methods focus only on signal identification at the current moment, neglecting the study of the regularity of signal modulation patterns evolving over time. This makes it impossible to provide communication systems with forward-looking spectrum resource management and interference avoidance strategies, thus limiting their application value in the field of intelligent communication. Summary of the Invention

[0006] The present invention provides a method and system for blind identification and intelligent classification prediction of signal modulation modes, which can solve the problems in the prior art.

[0007] A first aspect of the present invention provides a method for blind identification and intelligent classification prediction of signal modulation schemes, comprising:

[0008] The signal modulation method is decomposed into multi-scale time-domain features in the time domain by wavelet packet decomposition, and instantaneous frequency features are obtained in the frequency domain by Hilbert transform. The multi-scale time-domain features and the instantaneous frequency features are fused and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimension and combined to construct a feature correlation matrix.

[0009] The spatial relationship of the feature correlation matrix is ​​captured. Based on the spatial relationship, the adaptive clustering and hierarchical representation of the features are realized through a dynamic routing algorithm. At the same time, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix to obtain the blind recognition result of the signal modulation mode.

[0010] Based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation method, the signal modulation method is initially divided into discrete modulation and continuous modulation. Constellation diagram features are constructed by combining the signal modulation parameters. Multi-scale granular features are extracted for discrete modulation and continuous modulation based on the constellation diagram features. Multi-branch parallel convolution is performed on the multi-scale granular features to obtain multi-branch features. The multi-branch features are then fused through adaptive weighting and passed through a fully connected layer to achieve signal modulation method classification.

[0011] The steady-state feature sequence is obtained by filtering the feature correlation matrix using a hybrid filter consisting of Kalman filtering and particle filtering. The evolution trend and dynamic adjustment process of the steady-state feature sequence are captured at different levels using hierarchical Markov decision. The prediction result of the signal modulation mode is obtained by using dual learning to constrain the prediction space.

[0012] The multi-scale time-domain features and the instantaneous frequency features are fused, and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimensionality and combined to construct a feature correlation matrix, including:

[0013] Initial feature matrices are constructed for the multi-scale time-domain features and the instantaneous frequency features, respectively. The initial feature matrices are normalized and projected to obtain reconstructed feature matrices. The Euclidean distance between the reconstructed feature matrices and the initial feature matrices is used as a feature reconstruction term, and the L2 norm of the reconstructed feature matrices is used as a weight regularization term. The fusion weight is obtained by minimizing the weighted sum of the feature reconstruction term and the weight regularization term. The multi-scale time-domain features and the instantaneous frequency features are weighted and fused based on the fusion weight to obtain a fused feature set.

[0014] The fused feature set is divided into multiple feature intervals to obtain a feature histogram. Based on the feature histogram, the joint occurrence frequency of any two features is calculated to obtain a joint probability distribution. The occurrence frequency of a single feature in each interval is calculated to obtain a marginal probability distribution. The ratio of the joint probability distribution to the marginal probability distribution is calculated. The ratio is multiplied by the logarithm of the joint probability distribution and accumulated to obtain the mutual information between features. Features with mutual information greater than a preset relevance threshold are selected and combined to obtain a dimensionality-reduced feature subset.

[0015] The covariance between features is calculated based on the mean and variance of the features in the reduced feature subset. The covariance is then divided by the product of the standard deviations of the corresponding features to obtain the Pearson correlation coefficient. A feature correlation matrix is ​​then constructed based on the Pearson correlation coefficient.

[0016] The spatial relationships of the feature correlation matrix are captured. Based on these spatial relationships, adaptive clustering and hierarchical representation of the features are achieved through a dynamic routing algorithm. Simultaneously, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix, resulting in blind identification of the signal modulation scheme, including:

[0017] The Pearson correlation coefficients in the feature correlation matrix are used to construct a Pearson correlation coefficient matrix. The Pearson correlation coefficient matrix is ​​then divided by the square root of the feature dimension and normalized to obtain the spatial relationship matrix.

[0018] Calculate the similarity score corresponding to the spatial relationship matrix, perform voting iteration based on the similarity score, calculate the consistency degree between feature groups in each iteration, update the voting weight according to the consistency degree, repeat the voting iteration until the voting weight converges, use the voting weight as the clustering weight of the feature group to obtain the adaptive clustering result of the feature group; sort the clustered feature groups according to the voting weight and connect the feature groups in sequence to form a multi-level feature representation structure to obtain the hierarchical representation result;

[0019] The feature channels of the feature correlation matrix are averaged in the spatial dimension to obtain the channel statistical vector. The channel statistical vector is then passed through a fully connected layer to obtain the channel weight vector. The channel weight vectors with values ​​greater than zero are kept unchanged, and the channel weight vectors with values ​​less than zero are set to zero to obtain the updated channel weight vector. The updated channel weight vector is multiplied by the channel statistical vector, and the feature channels are recalibrated based on the product.

[0020] The adaptive clustering results, the hierarchical representation results, and the feature channel recalibration results are fused together, and the blind recognition results of the signal modulation mode are obtained through softmax operation.

[0021] Based on the aforementioned feature correlation matrix and the signal-to-noise ratio of the signal modulation scheme, the signal modulation scheme is initially divided into discrete modulation and continuous modulation. The constellation diagram features constructed by combining the signal modulation parameters include:

[0022] Extract the correlation coefficient from the feature correlation matrix, obtain the curve of the correlation coefficient changing with the signal-to-noise ratio within the signal-to-noise ratio range, if the curve of change has a jump point and the set of function values ​​of the curve of change is countable, then the corresponding signal modulation method is determined to be discrete modulation, otherwise it is continuous modulation;

[0023] For the discrete modulation, the signal modulation method is multiplied by the in-phase component and the quadrature component of the local carrier to obtain a mixed signal. The mixed signal is filtered by a low-pass filter to remove high-frequency components to obtain the in-phase branch signal and the quadrature branch signal. The in-phase branch signal and the quadrature branch signal are symbol-timed synchronized to obtain a complex symbol sequence. The real part of the complex symbol sequence is used as the horizontal axis and the imaginary part is used as the vertical axis to construct the discrete modulation constellation diagram feature.

[0024] For the continuous modulation, after bandpass sampling of the signal modulation method, the real part signal and the imaginary part signal are obtained by orthogonal decomposition. The squares of the real part signal and the imaginary part signal are added together and the square root is obtained to obtain the envelope. The product of the carrier frequency and time is calculated as the carrier phase. The arctangent ratio of the real part signal and the imaginary part signal is subtracted from the carrier phase to obtain the modulation phase. The envelope is used as the amplitude and the modulation phase is used as the angle to construct the continuous modulation constellation diagram features.

[0025] Based on constellation diagram features, multi-scale granular features are extracted for both discrete and continuous modulation. Multi-branch parallel convolution is then performed on these multi-scale granular features to obtain multi-branch features, including:

[0026] For the constellation diagram features of discrete and continuous modulation, multiple sliding windows of different sizes are used to extract multi-scale granular features. Specifically, for the two-dimensional plane of discrete modulation, the Euclidean distance between the current position and the sampling points within the window is calculated, and the local density feature is obtained based on the ratio of the sum of squares of the Euclidean distances to the window size. A second-order matrix is ​​calculated based on the spatial distribution of the sampling points within the window. Eigenvalue decomposition is performed on the second-order matrix to determine eigenvalues, and eigenvectors are calculated to determine the principal direction, resulting in local shape features. The amplitude variance and phase variance of the sampling points within the window are calculated, and polar coordinate correlation features are obtained based on the covariance of the amplitude-phase joint distribution. The local density feature, the local shape feature, and the polar coordinate correlation feature are concatenated along the feature dimension to obtain the multi-scale granular features.

[0027] The multi-scale granular features are input into multiple parallel convolutional branches. Each parallel convolutional branch contains a cascaded convolutional layer with convolutional kernels of different sizes. In the cascaded convolutional layer, each convolutional layer performs convolution operations on the input features to obtain a feature map. The feature map is then processed by pooling and batch normalization and used as the input to the next convolutional layer. The outputs of the last convolutional layer of each parallel convolutional branch are concatenated along the channel dimension to obtain multi-branch features.

[0028] Steady-state feature sequences are obtained by filtering the feature correlation matrix using a hybrid filter consisting of Kalman filtering and particle filtering. Hierarchical Markov decision theory is then used to capture the evolution trend and dynamic adjustment process of the steady-state feature sequences at different levels. Finally, dual learning is used to constrain the prediction space, resulting in predictions of the signal modulation scheme, including:

[0029] The feature correlation matrix is ​​decomposed into feature state vectors and observation vectors;

[0030] The hybrid filter performs linear Kalman prediction on the feature state vector to obtain the predicted state; a sampling interval is determined based on the probability distribution characteristics of the observed vector; uniform sampling is performed within the sampling interval and random perturbations are superimposed to generate particle samples; the likelihood probability of the particle samples is calculated to obtain the particle weights; the predicted state is weighted and resampled based on the particle weights to output a steady-state feature sequence.

[0031] In the hierarchical Markov decision structure, the policy layer receives the steady-state feature sequence and divides it into multiple decision units, calculates the immediate rewards between adjacent decision units to obtain a reward evaluation matrix; the feature layer calculates the feature transition probability matrix based on the temporal change relationship of the decision units; the execution layer calculates the state value of the reward evaluation matrix and the feature transition probability matrix, and selects the action policy corresponding to the maximum state value to obtain the optimal action sequence.

[0032] The optimal action sequence is matched with a preset feature template. Based on the matching result, the optimal action sequence is reverse-mapped to obtain a reconstructed feature sequence. The cyclic consistency loss between the reconstructed feature sequence and the steady-state feature sequence is calculated. The cyclic consistency loss is used as a prediction constraint. The prediction result of the output signal modulation mode is iteratively optimized until the prediction constraint converges.

[0033] A second aspect of the present invention provides a blind identification and intelligent classification prediction system for signal modulation schemes, comprising:

[0034] The first unit is used to obtain multi-scale time-domain features of the signal modulation mode in the time domain through wavelet packet decomposition, and obtain instantaneous frequency features in the frequency domain through Hilbert transform. The multi-scale time-domain features and the instantaneous frequency features are fused and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimension and combined to construct a feature correlation matrix.

[0035] The second unit is used to capture the spatial relationship of the feature correlation matrix. Based on the spatial relationship, the feature adaptive clustering and hierarchical representation are realized through a dynamic routing algorithm. At the same time, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix to obtain the blind recognition result of the signal modulation mode.

[0036] The third unit is used to initially classify signal modulation methods into discrete modulation and continuous modulation based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation method, and construct constellation diagram features by combining the signal modulation parameters; based on the constellation diagram features, multi-scale granular features are extracted for discrete modulation and continuous modulation respectively; multi-branch parallel convolution is performed on the multi-scale granular features to obtain multi-branch features; and the multi-branch features are adaptively weighted and fused and then passed through a fully connected layer to realize the classification of signal modulation methods.

[0037] The fourth unit is used to filter the feature correlation matrix using a hybrid filter composed of Kalman filtering and particle filtering to obtain a steady-state feature sequence. It uses hierarchical Markov decision to capture the evolution trend and dynamic adjustment process of the steady-state feature sequence at different levels, and uses dual learning to constrain the prediction space to obtain the prediction result of the signal modulation mode.

[0038] A third aspect of the embodiments of the present invention,

[0039] An electronic device is provided, comprising:

[0040] processor;

[0041] Memory used to store processor-executable instructions;

[0042] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0043] Fourth aspect of the present invention,

[0044] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0045] The beneficial effects of this application are as follows:

[0046] This invention achieves high-precision feature extraction and dimensionality reduction of signal modulation schemes by using wavelet packet decomposition and Hilbert transform for multi-domain feature fusion, and constructing a feature correlation matrix by combining mutual information calculation. This effectively reduces feature redundancy and improves the accuracy and robustness of signal modulation scheme recognition.

[0047] This invention employs a dynamic routing algorithm and a self-correction mechanism to achieve adaptive clustering and hierarchical representation of features, and combines constellation diagram features for multi-scale granular feature extraction. Through multi-branch parallel convolution and adaptive weighted fusion, it can effectively distinguish between discrete modulation and continuous modulation, and significantly improve the recognition accuracy of signal modulation methods in complex electromagnetic environments.

[0048] This invention introduces a hybrid filter and a hierarchical Markov decision mechanism, which can effectively capture the evolution trend of signal features. By constraining the prediction space through dual learning, it realizes intelligent prediction of signal modulation mode, improves the system's adaptability to unknown modulation mode and prediction accuracy, and provides a reliable basis for the adaptive adjustment of communication system. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the blind identification and intelligent classification prediction method for signal modulation schemes according to an embodiment of the present invention.

[0050] Figure 2 This diagram illustrates the comparison of recognition accuracy under different signal-to-noise ratios.

[0051] Figure 3 This is a schematic diagram of the signal modulation prediction algorithm. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a flowchart illustrating the blind identification and intelligent classification prediction method for signal modulation schemes according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0055] The signal modulation method is decomposed into multi-scale time-domain features in the time domain by wavelet packet decomposition, and instantaneous frequency features are obtained in the frequency domain by Hilbert transform. The multi-scale time-domain features and the instantaneous frequency features are fused and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimension and combined to construct a feature correlation matrix.

[0056] The spatial relationship of the feature correlation matrix is ​​captured. Based on the spatial relationship, the adaptive clustering and hierarchical representation of the features are realized through a dynamic routing algorithm. At the same time, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix to obtain the blind recognition result of the signal modulation mode.

[0057] Based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation method, the signal modulation method is initially divided into discrete modulation and continuous modulation. Constellation diagram features are constructed by combining the signal modulation parameters. Multi-scale granular features are extracted for discrete modulation and continuous modulation based on the constellation diagram features. Multi-branch parallel convolution is performed on the multi-scale granular features to obtain multi-branch features. The multi-branch features are then fused through adaptive weighting and passed through a fully connected layer to achieve signal modulation method classification.

[0058] The steady-state feature sequence is obtained by filtering the feature correlation matrix using a hybrid filter consisting of Kalman filtering and particle filtering. The evolution trend and dynamic adjustment process of the steady-state feature sequence are captured at different levels using hierarchical Markov decision. The prediction result of the signal modulation mode is obtained by using dual learning to constrain the prediction space.

[0059] In one optional implementation, the multi-scale time-domain features and the instantaneous frequency features are fused, and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimensionality and combined to construct a feature correlation matrix, including:

[0060] Initial feature matrices are constructed for the multi-scale time-domain features and the instantaneous frequency features, respectively. The initial feature matrices are normalized and projected to obtain reconstructed feature matrices. The Euclidean distance between the reconstructed feature matrices and the initial feature matrices is used as a feature reconstruction term, and the L2 norm of the reconstructed feature matrices is used as a weight regularization term. The fusion weight is obtained by minimizing the weighted sum of the feature reconstruction term and the weight regularization term. The multi-scale time-domain features and the instantaneous frequency features are weighted and fused based on the fusion weight to obtain a fused feature set.

[0061] The fused feature set is divided into multiple feature intervals to obtain a feature histogram. Based on the feature histogram, the joint occurrence frequency of any two features is calculated to obtain a joint probability distribution. The occurrence frequency of a single feature in each interval is calculated to obtain a marginal probability distribution. The ratio of the joint probability distribution to the marginal probability distribution is calculated. The ratio is multiplied by the logarithm of the joint probability distribution and accumulated to obtain the mutual information between features. Features with mutual information greater than a preset relevance threshold are selected and combined to obtain a dimensionality-reduced feature subset.

[0062] The covariance between features is calculated based on the mean and variance of the features in the reduced feature subset. The covariance is then divided by the product of the standard deviations of the corresponding features to obtain the Pearson correlation coefficient. A feature correlation matrix is ​​then constructed based on the Pearson correlation coefficient.

[0063] The initial feature matrix is ​​normalized using a maximum-minimum normalization method, mapping each feature value to the interval [0,1]. Taking the time-domain feature matrix X_t as an example, for the j-th column feature, the maximum value max_j and minimum value min_j of that column are found. For the element x_ij in the i-th row and j-th column of the matrix, the normalized value is (x_ij-min_j) / (max_j-min_j). Similarly, the same operation is performed on the frequency feature matrix X_f to obtain the normalized feature matrices X_t_norm and X_f_norm.

[0064] The feature projection process is implemented through an autoencoder to reconstruct the normalized feature matrix. Taking X_t_norm as an example, the encoder maps it to a low-dimensional latent space, and then the decoder reconstructs it back to the original dimension, obtaining the reconstructed feature matrix X_t_recon. Similarly, X_f_recon is obtained. The Euclidean distance between the reconstructed feature matrix and the initial feature matrix is ​​calculated as the feature reconstruction term L_recon. Specifically, for time-domain features, L_recon_t = ||X_t_norm - X_t_recon||_2; for frequency features, L_recon_f = ||X_f_norm - X_f_recon||_2.

[0065] The L2 norm of the reconstructed feature matrix is ​​calculated as the weight regularization term L_reg. For time-domain features, L_reg_t = ||X_t_recon||_2; for frequency features, L_reg_f = ||X_f_recon||_2. The fusion weights W_t and W_f are obtained by minimizing the weighted sum of the feature reconstruction term and the weight regularization term. The total loss function is L = α(L_recon_t + L_recon_f) + β(L_reg_t + L_reg_f), where α and β are balancing factors. In practical applications, α can be set to 0.7 and β to 0.3. W_t and W_f are iteratively optimized using gradient descent until the loss function converges or the preset number of iterations (e.g., 500) is reached.

[0066] Based on the obtained fusion weights W_t and W_f, the multi-scale time-domain features and instantaneous frequency features are weighted and fused: X_fusion = W_t·X_t_norm + W_f·X_f_norm, resulting in the fused feature set X_fusion. In practical applications, if the time-domain feature dimension is 20 and the frequency feature dimension is 15, then the fused feature dimension is 35.

[0067] The fused feature set X_fusion is divided into multiple feature intervals to obtain a feature histogram. Specifically, for each column of features in X_fusion, its value range is determined, and this range is divided into k equal intervals (e.g., k=10). The number of samples for each feature in each interval is counted to form a feature histogram.

[0068] The joint probability distribution P(x,y) is obtained by calculating the joint frequency of any two features based on the feature histogram. For features x and y, the frequency of their co-occurrence in each interval combination is counted. For example, the number of samples where feature x falls in the i-th interval and feature y falls in the j-th interval is counted, and divided by the total number of samples m, to obtain the joint probability P(x=i,y=j).

[0069] The marginal probability distributions P(x) and P(y) are obtained by calculating the frequency of occurrence of a single feature in each interval. For feature x, the number of its samples in the i-th interval is counted and divided by the total number of samples m to obtain the marginal probability P(x=i). Similarly, P(y=j) is calculated.

[0070] Calculate the ratio P(x,y) / (P(x)·P(y)) between the joint probability distribution and the marginal probability distribution. Multiply this ratio by the logarithm of the joint probability distribution and sum them to obtain the mutual information MI(x,y) between features x and y. In practical applications, if the two features are completely independent, the mutual information value is close to 0; if they are highly correlated, the mutual information value is large.

[0071] Set a preset relevance threshold (e.g., 0.6), filter feature pairs with mutual information greater than this threshold, and combine them to obtain a dimensionality-reduced feature subset X_reduced. Assuming the original fused feature set has a dimension of 35, after mutual information filtering, 20 features are retained, forming a dimensionality-reduced feature subset.

[0072] The covariance between features is calculated based on the mean and variance of the features in the dimensionality-reduced feature subset X_reduced. For features i and j, their covariance Cov(i,j) is calculated. The Pearson correlation coefficient r(i,j) = Cov(i,j) / (σ_i·σ_j) is obtained by dividing the covariance by the product of the standard deviations of the corresponding features.

[0073] A feature correlation matrix R is constructed based on the Pearson correlation coefficient. The matrix dimension is the number of reduced features × the number of reduced features (e.g., 20×20). The element in the i-th row and j-th column of the matrix is ​​the Pearson correlation coefficient r(i,j) between feature i and feature j, with a value range of [-1,1]. A correlation coefficient of 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation.

[0074] In a practical application, assuming 10 time-domain features (such as mean, variance, and peak value) and 8 frequency features (such as dominant frequency and spectral entropy) are extracted from the vibration signal, they are fused using the method described above to obtain an 18-dimensional feature vector. After mutual information filtering, 12 features are retained to form a dimensionality-reduced feature subset. The final 12×12 feature correlation matrix clearly shows the correlation between the features, providing an important basis for subsequent fault diagnosis and prediction.

[0075] In one optional implementation, the spatial relationship of the feature correlation matrix is ​​captured. Based on the spatial relationship, adaptive clustering and hierarchical representation of the features are achieved through a dynamic routing algorithm. Simultaneously, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix, resulting in blind identification of the signal modulation scheme, including:

[0076] The Pearson correlation coefficients in the feature correlation matrix are used to construct a Pearson correlation coefficient matrix. The Pearson correlation coefficient matrix is ​​then divided by the square root of the feature dimension and normalized to obtain the spatial relationship matrix.

[0077] Calculate the similarity score corresponding to the spatial relationship matrix, perform voting iteration based on the similarity score, calculate the consistency degree between feature groups in each iteration, update the voting weight according to the consistency degree, repeat the voting iteration until the voting weight converges, use the voting weight as the clustering weight of the feature group to obtain the adaptive clustering result of the feature group; sort the clustered feature groups according to the voting weight and connect the feature groups in sequence to form a multi-level feature representation structure to obtain the hierarchical representation result;

[0078] The feature channels of the feature correlation matrix are averaged in the spatial dimension to obtain the channel statistical vector. The channel statistical vector is then passed through a fully connected layer to obtain the channel weight vector. The channel weight vectors with values ​​greater than zero are kept unchanged, and the channel weight vectors with values ​​less than zero are set to zero to obtain the updated channel weight vector. The updated channel weight vector is multiplied by the channel statistical vector, and the feature channels are recalibrated based on the product.

[0079] The adaptive clustering results, the hierarchical representation results, and the feature channel recalibration results are fused together, and the blind recognition results of the signal modulation mode are obtained through softmax operation.

[0080] The feature correlation matrix is ​​calculated from multi-dimensional features of the signal, such as higher-order cumulants, cyclostationary characteristics, and power spectral density. This matrix contains correlation information between different feature dimensions and exhibits unique patterns for signals with different modulation schemes. For an input signal sample, after extracting its features, the Pearson correlation coefficients between the features are calculated to construct the Pearson correlation coefficient matrix. Assuming the original feature dimension is 64, the constructed Pearson correlation coefficient matrix is ​​a 64×64 square matrix. To eliminate the influence of dimension and perform standardization, the Pearson correlation coefficient matrix is ​​divided by the square root of the feature dimension (8 in this example), and then normalized so that the matrix element values ​​fall within the range of [-1, 1], resulting in the spatial relationship matrix.

[0081] Based on the spatial relation matrix, a dynamic routing algorithm is executed to achieve adaptive clustering of features. The similarity score between features in the spatial relation matrix is ​​calculated, which can be achieved by calculating the Euclidean distance or cosine similarity between feature vectors. In the actual implementation, for 64-dimensional features, they can be divided into 8 feature groups, each containing 8 features. The voting weight matrix is ​​initialized as an all-zero matrix with a size of 8×8. A voting iteration process is executed, with the number of iterations set to 3. In each iteration, the consistency degree between feature groups is calculated, which is achieved by multiplying the submatrix of the corresponding feature group in the spatial relation matrix with the current voting weight. For example, for feature group i and feature group j, the consistency degree between them is calculated as the weighted sum of the corresponding elements of the two feature groups in the spatial relation matrix. The voting weights are updated according to the calculated consistency degree, and a softmax function is used during the update to ensure that the weight sum is 1. After multiple iterations (3 times in this example), the voting weights tend to stabilize. At this point, the final voting weights are used as the clustering weights for the feature groups, obtaining the adaptive clustering results for the feature groups.

[0082] After obtaining the clustering results, the feature groups are sorted according to their voting weights. Assume the sorting result is feature group 5, feature group 2, feature group 7, feature group 1, feature group 4, feature group 3, feature group 8, and feature group 6. Connections are established between the feature groups sequentially based on the sorting result, forming a multi-level feature representation structure. In this embodiment, a three-layer structure can be constructed: the first layer includes feature group 5 and feature group 2 with the highest weights; the second layer includes feature group 7, feature group 1, and feature group 4; and the third layer includes the remaining feature groups. Different layers are connected by the connection strength determined by the voting weights, forming a hierarchical representation result.

[0083] Self-calibration is performed on the feature channels of the feature correlation matrix. The feature channels of the feature correlation matrix are averaged across the spatial dimension, resulting in a 64×64 matrix with a 64-dimensional channel statistical vector. This channel statistical vector is then passed through two fully connected network layers. The first layer reduces the 64-dimensionality to 16-dimensionality using the ReLU activation function; the second layer restores the 16-dimensionality to 64-dimensionality, yielding the channel weight vector. The channel weight vector is then processed, retaining weights greater than zero and setting weights less than zero to zero, resulting in an updated channel weight vector. Finally, the updated channel weight vector is multiplied by the original channel statistical vector to recalibrate the feature channels.

[0084] After obtaining the adaptive clustering results, hierarchical representation results, and feature channel recalibration results, the three results are concatenated according to the feature dimensions to form an enhanced feature representation, assuming the concatenated feature dimension is 128-dimensional. Then, a fully connected layer maps the features to the modulation scheme category space; assuming there are 10 modulation schemes to be identified, the output layer dimension is 10. Finally, a softmax operation is used to convert the output into a probability distribution, and the category with the highest probability is the recognition result. In actual tests, for common modulation signals such as BPSK, QPSK, 8PSK, 16QAM, and 64QAM with a signal-to-noise ratio of 10dB, this method can achieve a recognition accuracy of over 95%, improving performance by approximately 10% compared to traditional methods.

[0085] Furthermore, even under low signal-to-noise ratio conditions (such as 0dB), the recognition accuracy of this method remains above 80%, demonstrating good robustness. Regarding real-time requirements, on a processor with a clock speed of 3.2GHz, the average time to complete the modulation scheme recognition of a single signal sample is no more than 5 milliseconds, meeting the needs of most practical application scenarios.

[0086] Figure 2This diagram illustrates the comparison of recognition accuracy under different signal-to-noise ratios (SNRs). It showcases the recognition performance of the dynamic routing and self-correction mechanism of this invention under varying SNR environments and provides a comprehensive comparative analysis with three mainstream modulation recognition methods. The diagram clearly demonstrates that this invention maintains a significant performance advantage across a wide SNR range from -10dB to 30dB, particularly under low SNR conditions. At -10dB, the recognition accuracy of traditional decision tree methods is only 32%, SVM-based methods 35%, and deep learning CNN methods 38%, while this invention achieves 42%, a 4 percentage point improvement over the best comparison method. As the SNR increases, the advantages of this invention become more pronounced. At a 10dB SNR, the recognition accuracy reaches 95%, 12 percentage points higher than traditional decision tree methods, 10 percentage points higher than SVM methods, and 8 percentage points higher than CNN methods. In high SNR environments (above 20dB), the recognition accuracy of this invention remains stable above 98%, approaching the theoretical upper limit, while the accuracy of other methods is generally below 97%. This performance fully verifies the effectiveness of the core technologies employed in this invention, such as wavelet packet decomposition, Hilbert transform, feature fusion, dynamic routing algorithm, and self-correction mechanism. In particular, its robustness under complex electromagnetic environments and low signal-to-noise ratio conditions provides important technical support for the practical engineering application of signal modulation recognition technology.

[0087] In one optional implementation, based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation scheme, the signal modulation scheme is initially divided into discrete modulation and continuous modulation. Constructing constellation diagram features by combining the signal modulation parameters includes:

[0088] Extract the correlation coefficient from the feature correlation matrix, obtain the curve of the correlation coefficient changing with the signal-to-noise ratio within the signal-to-noise ratio range, if the curve of change has a jump point and the set of function values ​​of the curve of change is countable, then the corresponding signal modulation method is determined to be discrete modulation, otherwise it is continuous modulation;

[0089] For the discrete modulation, the signal modulation method is multiplied by the in-phase component and the quadrature component of the local carrier to obtain a mixed signal. The mixed signal is filtered by a low-pass filter to remove high-frequency components to obtain the in-phase branch signal and the quadrature branch signal. The in-phase branch signal and the quadrature branch signal are symbol-timed synchronized to obtain a complex symbol sequence. The real part of the complex symbol sequence is used as the horizontal axis and the imaginary part is used as the vertical axis to construct the discrete modulation constellation diagram feature.

[0090] For the continuous modulation, after bandpass sampling of the signal modulation method, the real part signal and the imaginary part signal are obtained by orthogonal decomposition. The squares of the real part signal and the imaginary part signal are added together and the square root is obtained to obtain the envelope. The product of the carrier frequency and time is calculated as the carrier phase. The arctangent ratio of the real part signal and the imaginary part signal is subtracted from the carrier phase to obtain the modulation phase. The envelope is used as the amplitude and the modulation phase is used as the angle to construct the continuous modulation constellation diagram features.

[0091] For the correlation coefficient in the feature correlation matrix, a correlation coefficient value can be obtained every 1dB within a signal-to-noise ratio (SNR) range, such as 0dB to 20dB, and a curve showing the correlation coefficient changing with SNR can be plotted. For example, in a binary phase-shift keying (BPSK) signal, when the SNR changes from 5dB to 6dB, the correlation coefficient abruptly changes from 0.65 to 0.85, and the set of correlation coefficient values ​​within the entire SNR range can be represented as {0.45, 0.55, 0.65, 0.85, 0.91, 0.95...}, which is a countable set. Therefore, BPSK is determined to be discrete modulation. In contrast, in a frequency-modulated (FM) signal, the correlation coefficient changes smoothly with SNR, such as gradually increasing from 0.4 to 0.9 without a significant jump. The correlation coefficient value can take any value within the range [0.4, 0.9], which is an uncountable set, and is determined to be continuous modulation.

[0092] For discrete modulation methods, such as Quadrature Phase Shift Keying (QPSK), the received signal needs to be mixed with the local carrier. Assuming the received signal frequency is 10MHz, a local carrier signal with the same frequency is generated, and its in-phase component is cos(2π×10⁻⁶). 7 t), the orthogonal component is sin(2π×10). 7 The QPSK signal is multiplied by these two components respectively to obtain two mixed signals. A low-pass filter with a cutoff frequency of 1.5 times the signal bandwidth (e.g., 2MHz) is applied to these two mixed signals to filter out high-frequency components, resulting in the in-phase branch signal I(t) and the quadrature branch signal Q(t).

[0093] For symbol timing synchronization, the Gardner algorithm can be used, with the sampling rate being an integer multiple of the symbol rate (e.g., 4 times). For the synchronized I(t) and Q(t) signals, the optimal sampling point is taken for each symbol period, forming a complex symbol sequence z(n) = I(n) + jQ(n). Taking QPSK as an example, with a signal-to-noise ratio of 10dB, the complex symbol sequence obtained is as follows: {0.95 + 0.98j, -0.97 + 0.94j, -0.96 - 0.97j, 0.98 - 0.95j}. By using the real parts of these complex numbers as the abscissa and the imaginary parts as the ordinate, the QPSK modulation constellation characteristics, where the constellation points are concentrated around (±1, ±1), can be plotted.

[0094] For continuous modulation methods, such as frequency modulation (FM) signals, bandpass sampling is performed first. Assuming the FM signal has a center frequency of 100MHz and a bandwidth of 200kHz, according to the sampling theorem, the sampling rate must be at least 400kHz. After sampling, orthogonal decomposition is performed using Hilbert transform to obtain the real part signal s_r(t) and the imaginary part signal s_i(t).

[0095] To calculate the envelope of a signal, add the squares of the real part s_r(t) and the squares of the imaginary part s_i(t), then take the square root to obtain the envelope r(t). For example, at a certain time t=0.001s, s_r(0.001)=0.5, s_i(0.001)=0.866, then r(0.001)=sqrt(0.5) 2 +0.866 2 =1.

[0096] Simultaneously, the carrier phase is calculated. Assuming the carrier frequency f_c = 100MHz, the carrier phase at time t is 2π × f_c × t. At t = 0.001s, the carrier phase is 2π × 10⁻⁶. 8 ×0.001=2π×10 5 (rad)

[0097] To calculate the modulation phase, subtract the carrier phase from the arctangent ratio of the real and imaginary signals, arctan(s_i(t) / s_r(t)). At t=0.001s, arctan(0.866 / 0.5)=π / 3 (rad). Subtracting an integer multiple of the carrier phase, we obtain the modulation phase Φ(0.001)=π / 3-2π×10. 5 (mod 2π).

[0098] Using the envelope r(t) as the radius in polar coordinates and the modulation phase Φ(t) as the angle, or using r(t)×cos(Φ(t)) as the abscissa and r(t)×sin(Φ(t)) as the ordinate, a constellation diagram feature of continuous modulation can be constructed. For FM signals, due to their characteristics, the envelope r(t) remains basically constant (approximately 1), while the phase Φ(t) varies with the modulation signal, exhibiting a circular constellation diagram feature.

[0099] Using the above method, discrete and continuous modulation can be automatically distinguished based on signal characteristics, and corresponding constellation diagram features can be constructed, providing a valid basis for subsequent modulation identification and signal processing. This method is applicable to various wireless communication scenarios, especially applications requiring modulation analysis of unknown signals.

[0100] In one optional implementation, multi-scale granular features are extracted based on constellation diagram features for both discrete and continuous modulation. Multi-branch parallel convolution is then performed on these multi-scale granular features to obtain multi-branch features, including:

[0101] For the constellation diagram features of discrete and continuous modulation, multiple sliding windows of different sizes are used to extract multi-scale granular features. Specifically, for the two-dimensional plane of discrete modulation, the Euclidean distance between the current position and the sampling points within the window is calculated, and the local density feature is obtained based on the ratio of the sum of squares of the Euclidean distances to the window size. A second-order matrix is ​​calculated based on the spatial distribution of the sampling points within the window. Eigenvalue decomposition is performed on the second-order matrix to determine eigenvalues, and eigenvectors are calculated to determine the principal direction, resulting in local shape features. The amplitude variance and phase variance of the sampling points within the window are calculated, and polar coordinate correlation features are obtained based on the covariance of the amplitude-phase joint distribution. The local density feature, the local shape feature, and the polar coordinate correlation feature are concatenated along the feature dimension to obtain the multi-scale granular features.

[0102] The multi-scale granular features are input into multiple parallel convolutional branches. Each parallel convolutional branch contains a cascaded convolutional layer with convolutional kernels of different sizes. In the cascaded convolutional layer, each convolutional layer performs convolution operations on the input features to obtain a feature map. The feature map is then processed by pooling and batch normalization and used as the input to the next convolutional layer. The outputs of the last convolutional layer of each parallel convolutional branch are concatenated along the channel dimension to obtain multi-branch features.

[0103] For discrete modulation signals, such as QPSK and 8PSK, the constellation points are distributed at fixed positions; while for continuous modulation signals, such as FM and AM, the constellation points are continuously distributed. In this embodiment, the sliding window sizes used include 3×3, 5×5, and 7×7 pixels. The window slides across the constellation map with a step size of 1, covering the entire constellation map plane.

[0104] For two-dimensional plane processing of discrete modulation, local density features are extracted by calculating the Euclidean distance between the current position and the sampling points within the window. Assume the current window center point coordinates are (x0, y0), and the window contains n sampling points, with the coordinates of the i-th sampling point being (x0, y0). i , y i Then calculate the Euclidean distance d between each sampling point and the center point. iFor a 5×5 window with 25 points, calculate the Euclidean distances between these 25 points and the center point. Divide the sum of the squares of these distances by the window size of 25 to obtain the local density value. For example, if the sum of the squares of the Euclidean distances between the sampling points within the window is 125, then the local density feature value is 125 / 25 = 5. This feature reflects the degree of clustering of constellation points. For discrete modulation such as QPSK, the local density feature is higher at constellation points and lower in non-constellation point regions.

[0105] For n sampling points within a window, calculate their covariance matrix relative to the window center. This matrix reflects the distribution and correlation of the sampling points along the x and y axes. Perform eigenvalue decomposition on this second-order matrix to obtain two eigenvalues, λ1 and λ2 (assuming λ1 ≥ λ2). The magnitude of the eigenvalues ​​reflects the dispersion of data along the corresponding eigenvector direction. Simultaneously, calculate the corresponding eigenvectors to determine the principal direction. For example, for BPSK modulation, its constellation points are mainly distributed on the real axis, so the eigenvector direction corresponding to a larger eigenvalue is close to the horizontal direction; while for QPSK, its constellation points are squarely distributed, and the two eigenvalues ​​are similar. The ratio of the eigenvalues, λ2 / λ1, and the direction angle of the eigenvectors constitute the local shape feature.

[0106] Each sampling point (x) within the window i , y i Convert to polar coordinates (r) i , θ i ), where r i Indicates amplitude, θ i The phase is represented. The mean and variance of the amplitude, as well as the mean and variance of the phase, are calculated for all points within the window. For amplitude-stable modulation schemes such as PSK, the amplitude variance is small; while for amplitude-varying modulation schemes such as QAM, the amplitude variance is large. Simultaneously, the covariance between amplitude and phase is calculated to reflect their correlation. For example, for 16QAM modulation, where the constellation points are distributed on a 4×4 grid, the amplitude and phase exhibit a certain correlation.

[0107] The extracted local density features, local shape features, and polar coordinate correlation features are concatenated along the feature dimension to form a multi-scale granular feature. A corresponding feature map is generated for each sliding window size. For example, for a constellation diagram with an input size of 64×64, after extracting features using a 3×3 window, a 62×62×d feature map is obtained, where d is the feature dimension, containing a combination of density, shape, and polar coordinate correlation features.

[0108] Multi-scale granular features are input into multiple parallel convolutional branches for processing. This embodiment uses three parallel branches, each containing cascaded convolutional layers with kernels of different sizes. The first branch uses a 3×3 kernel, the second branch uses a 5×5 kernel, and the third branch uses a 7×7 kernel. Each branch contains three convolutional layers with 64, 128, and 256 channels, respectively.

[0109] In the cascaded convolutional layers, each layer performs convolution operations on the input features to obtain a feature map. Taking the first branch as an example, the first layer uses 64 3×3 convolutional kernels to convolve the input features with a stride of 1 and padding of 1, maintaining the spatial size of the feature map. The convolved feature map then undergoes max pooling with a 2×2 pooling window and a stride of 2, halving the feature map size. Batch normalization is then performed to standardize the feature values ​​to a distribution with a mean of 0 and a variance of 1, improving training stability. The processed feature map serves as the input to the second convolutional layer, which uses 128 3×3 convolutional kernels. After pooling and normalization, the feature map is passed to the third layer. The third layer uses 256 3×3 convolutional kernels for final feature extraction.

[0110] The processing flow for the other two branches is similar, differing only in the size of the convolutional kernels. Larger kernels capture a wider range of spatial dependencies, while smaller kernels focus on local details. Finally, the outputs of the last convolutional layer from each parallel convolutional branch are concatenated along the channel dimension to obtain multi-branch features. Assuming each branch's final output feature map is 8×8×256, the concatenated map yields an 8×8×768 feature map, containing modulation signal features from different receptive fields, providing rich feature representations for subsequent modulation recognition tasks.

[0111] In one optional implementation, a steady-state feature sequence is obtained by filtering the feature correlation matrix using a hybrid filter consisting of Kalman filtering and particle filtering. Hierarchical Markov decision-making is then used to capture the evolution trend and dynamic adjustment process of the steady-state feature sequence at different levels. Finally, dual learning is used to constrain the prediction space, resulting in a prediction of the signal modulation scheme, including:

[0112] The feature correlation matrix is ​​decomposed into feature state vectors and observation vectors;

[0113] The hybrid filter performs linear Kalman prediction on the feature state vector to obtain the predicted state; a sampling interval is determined based on the probability distribution characteristics of the observed vector; uniform sampling is performed within the sampling interval and random perturbations are superimposed to generate particle samples; the likelihood probability of the particle samples is calculated to obtain the particle weights; the predicted state is weighted and resampled based on the particle weights to output a steady-state feature sequence.

[0114] In the hierarchical Markov decision structure, the policy layer receives the steady-state feature sequence and divides it into multiple decision units, calculates the immediate rewards between adjacent decision units to obtain a reward evaluation matrix; the feature layer calculates the feature transition probability matrix based on the temporal change relationship of the decision units; the execution layer calculates the state value of the reward evaluation matrix and the feature transition probability matrix, and selects the action policy corresponding to the maximum state value to obtain the optimal action sequence.

[0115] The optimal action sequence is matched with a preset feature template. Based on the matching result, the optimal action sequence is reverse-mapped to obtain a reconstructed feature sequence. The cyclic consistency loss between the reconstructed feature sequence and the steady-state feature sequence is calculated. The cyclic consistency loss is used as a prediction constraint. The prediction result of the output signal modulation mode is iteratively optimized until the prediction constraint converges.

[0116] like Figure 3 As shown, the method includes:

[0117] The input feature correlation matrix is ​​received and decomposed into feature state vectors and observation vectors using matrix factorization. The feature state vectors represent the intrinsic states of the signal characteristics, while the observation vectors represent the observed values ​​obtained from actual measurements. In a specific case, the feature correlation matrix can be decomposed into a 100×1 dimensional state vector and a 20×1 dimensional observation vector.

[0118] The hybrid filter processing section consists of two core components: a linear Kalman filter and a particle filter. The linear Kalman filter first predicts the characteristic state vector, taking into account the state transition matrix and system noise. For example, for sampling time t, the system state transition matrix can be set as a diagonal matrix, where the diagonal elements are 0.95, indicating that the state has a certain temporal continuity. The system noise can be set as Gaussian white noise with a mean of 0 and a standard deviation of 0.01. The predicted state is obtained by multiplying the state transition matrix by the state at the previous time step, and then adding the system noise.

[0119] The particle filtering section determines the sampling interval based on the probability distribution characteristics of the observation vector. In practical applications, the sampling interval, centered at the mean and with a radius three times the variance, can be determined by calculating the mean and variance of the observation vector. Uniform sampling is performed within this interval, for example, generating 200 particle samples, and a random Gaussian perturbation with a standard deviation of 0.05 is superimposed on each sample to increase sample diversity. For each particle sample, its corresponding observed value is calculated using the observation equation, and then compared with the actual observed value to calculate the likelihood probability as the particle weight. In the specific implementation, the likelihood probability can be calculated using a Gaussian probability density function, where the mean is the observed value predicted by the particle, and the variance is the measurement noise variance (e.g., 0.01).

[0120] After the particle weights are calculated, a weighted average is applied to the Kalman predicted states. The weighted average is calculated by multiplying the state value of each particle by the sum of its corresponding normalized weight. To prevent particle degradation, when the effective number of particles falls below a threshold (e.g., 50% of the total number of particles), system resampling is performed, retaining high-weight particles and replacing low-weight particles. The resampled particle set represents the posterior distribution of the system, and its mean is output as the steady-state feature sequence.

[0121] The hierarchical Markov decision structure consists of a policy layer, a feature layer, and an execution layer. The policy layer receives the steady-state feature sequence and divides it into decision units according to time windows (e.g., every 10 time points). Adjacent decision units are evaluated using the difference in feature vectors (e.g., Euclidean distance or cosine similarity) as an immediate reward, forming a reward evaluation matrix. In practical cases, if the feature difference between two decision units is 0.2, it can be mapped to a reward value of 0.8, indicating a high similarity between them.

[0122] The feature layer, based on the temporal changes of decision units, statistically analyzes the transition frequencies between different feature states to generate a feature transition probability matrix. For example, the probability of feature state i transitioning to state j can be calculated as the number of times state i transitions to state j divided by the total number of times state i occurs. In practical applications, if the state space is large, state clustering can be performed to merge similar states into representative states, thereby reducing computational complexity.

[0123] The execution layer comprehensively utilizes the reward evaluation matrix and the feature transition probability matrix to calculate the value of each state. The calculation of the state value considers the sum of the immediate reward and the discounted reward to be obtained in the future. In a specific implementation, the discount factor can be set to 0.9, and the calculation is iterative until the state value converges (e.g., the value change between two consecutive iterations is less than 0.001). Based on the state value, the action strategy that maximizes the long-term cumulative reward is selected to generate the optimal action sequence.

[0124] The dual learning constrained prediction space part matches the optimal action sequence with pre-defined feature templates. These templates contain typical feature patterns from different modulation schemes, such as QPSK, 16QAM, and 64QAM. The matching process employs a dynamic time warping algorithm to calculate the similarity between the optimal action sequence and each template. Based on the matching results, the optimal action sequence is reverse-mapped to generate a reconstructed feature sequence.

[0125] Cyclic consistency constraints require that the reconstructed feature sequence should be consistent with the original steady-state feature sequence. The mean squared error between the two is calculated as the cyclic consistency loss, which is used as a prediction constraint. The prediction process ends when the loss value is lower than a preset threshold (e.g., 0.01) or the number of iterations reaches an upper limit (e.g., 100 times). Finally, based on the optimally matched feature template, the prediction result of the signal modulation scheme, such as QPSK or 16QAM, is output.

[0126] Experimental results show that the hybrid filtering and hierarchical Markov decision method can achieve an accuracy of 92.5% for signals with a signal-to-noise ratio of 5dB in signal modulation pattern recognition tasks, which is about 15 percentage points higher than the traditional method; the processing speed is increased by 40%, and about 2,000 signal samples can be processed per second, which meets the needs of real-time communication systems.

[0127] This invention relates to a signal modulation scheme blind identification and intelligent classification prediction system, the system comprising:

[0128] The first unit is used to obtain multi-scale time-domain features of the signal modulation mode in the time domain through wavelet packet decomposition, and obtain instantaneous frequency features in the frequency domain through Hilbert transform. The multi-scale time-domain features and the instantaneous frequency features are fused and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimension and combined to construct a feature correlation matrix.

[0129] The second unit is used to capture the spatial relationship of the feature correlation matrix. Based on the spatial relationship, the feature adaptive clustering and hierarchical representation are realized through a dynamic routing algorithm. At the same time, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix to obtain the blind recognition result of the signal modulation mode.

[0130] The third unit is used to initially classify signal modulation methods into discrete modulation and continuous modulation based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation method, and construct constellation diagram features by combining the signal modulation parameters; based on the constellation diagram features, multi-scale granular features are extracted for discrete modulation and continuous modulation respectively; multi-branch parallel convolution is performed on the multi-scale granular features to obtain multi-branch features; and the multi-branch features are adaptively weighted and fused and then passed through a fully connected layer to realize the classification of signal modulation methods.

[0131] The fourth unit is used to filter the feature correlation matrix using a hybrid filter composed of Kalman filtering and particle filtering to obtain a steady-state feature sequence. It uses hierarchical Markov decision to capture the evolution trend and dynamic adjustment process of the steady-state feature sequence at different levels, and uses dual learning to constrain the prediction space to obtain the prediction result of the signal modulation mode.

[0132] A third aspect of the present invention provides an electronic device, comprising:

[0133] processor;

[0134] Memory used to store processor-executable instructions;

[0135] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0136] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0137] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for blind identification and intelligent classification prediction based on signal modulation, characterized in that, include: The signal modulation method is decomposed into multi-scale time-domain features in the time domain by wavelet packet decomposition, and instantaneous frequency features are obtained in the frequency domain by Hilbert transform. The multi-scale time-domain features and the instantaneous frequency features are fused and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimension and combined to construct a feature correlation matrix. The spatial relationship of the feature correlation matrix is ​​captured. Based on the spatial relationship, the adaptive clustering and hierarchical representation of the features are realized through a dynamic routing algorithm. At the same time, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix to obtain the blind recognition result of the signal modulation mode. Based on the aforementioned feature correlation matrix and the signal-to-noise ratio of the signal modulation method, the signal modulation method is initially divided into discrete modulation and continuous modulation, and constellation diagram features are constructed by combining the signal modulation parameters; Based on constellation diagram features, multi-scale granular features are extracted for discrete modulation and continuous modulation respectively. Multi-branch parallel convolution is performed on the multi-scale granular features to obtain multi-branch features. The multi-branch features are then fused through adaptive weighting and passed through a fully connected layer to classify the signal modulation mode. The steady-state feature sequence is obtained by filtering the feature correlation matrix using a hybrid filter consisting of Kalman filtering and particle filtering. The evolution trend and dynamic adjustment process of the steady-state feature sequence are captured at different levels using hierarchical Markov decision. The prediction result of the signal modulation mode is obtained by using dual learning to constrain the prediction space.

2. The method according to claim 1, characterized in that, The multi-scale time-domain features and the instantaneous frequency features are fused, and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimensionality and combined to construct a feature correlation matrix, including: Initial feature matrices are constructed for the multi-scale time-domain features and the instantaneous frequency features, respectively. The initial feature matrices are normalized and projected to obtain reconstructed feature matrices. The Euclidean distance between the reconstructed feature matrices and the initial feature matrices is used as a feature reconstruction term, and the L2 norm of the reconstructed feature matrices is used as a weight regularization term. The fusion weight is obtained by minimizing the weighted sum of the feature reconstruction term and the weight regularization term. The multi-scale time-domain features and the instantaneous frequency features are weighted and fused based on the fusion weight to obtain a fused feature set. The fused feature set is divided into multiple feature intervals to obtain a feature histogram. Based on the feature histogram, the joint occurrence frequency of any two features is calculated to obtain a joint probability distribution. The occurrence frequency of a single feature in each interval is calculated to obtain a marginal probability distribution. The ratio of the joint probability distribution to the marginal probability distribution is calculated. The ratio is multiplied by the logarithm of the joint probability distribution and accumulated to obtain the mutual information between features. Features with mutual information greater than a preset relevance threshold are selected and combined to obtain a dimensionality-reduced feature subset. The covariance between features is calculated based on the mean and variance of the features in the reduced feature subset. The covariance is then divided by the product of the standard deviations of the corresponding features to obtain the Pearson correlation coefficient. A feature correlation matrix is ​​then constructed based on the Pearson correlation coefficient.

3. The method according to claim 1, characterized in that, The spatial relationships of the feature correlation matrix are captured. Based on these spatial relationships, adaptive clustering and hierarchical representation of the features are achieved through a dynamic routing algorithm. Simultaneously, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix, resulting in blind identification of the signal modulation scheme, including: The Pearson correlation coefficients in the feature correlation matrix are used to construct a Pearson correlation coefficient matrix. The Pearson correlation coefficient matrix is ​​then divided by the square root of the feature dimension and normalized to obtain the spatial relationship matrix. Calculate the similarity score corresponding to the spatial relationship matrix, perform voting iteration based on the similarity score, calculate the consistency degree between feature groups in each iteration, update the voting weight according to the consistency degree, repeat the voting iteration until the voting weight converges, use the voting weight as the clustering weight of the feature group to obtain the adaptive clustering result of the feature group; sort the clustered feature groups according to the voting weight and connect the feature groups in sequence to form a multi-level feature representation structure to obtain the hierarchical representation result; The feature channels of the feature correlation matrix are averaged in the spatial dimension to obtain the channel statistical vector. The channel statistical vector is then passed through a fully connected layer to obtain the channel weight vector. The channel weight vectors with values ​​greater than zero are kept unchanged, and the channel weight vectors with values ​​less than zero are set to zero to obtain the updated channel weight vector. The updated channel weight vector is multiplied by the channel statistical vector, and the feature channels are recalibrated based on the product. The adaptive clustering results, the hierarchical representation results, and the feature channel recalibration results are fused together, and the blind recognition results of the signal modulation mode are obtained through softmax operation.

4. The method according to claim 1, characterized in that, Based on the aforementioned feature correlation matrix and the signal-to-noise ratio of the signal modulation scheme, the signal modulation scheme is initially divided into discrete modulation and continuous modulation. The constellation diagram features constructed by combining the signal modulation parameters include: Extract the correlation coefficient from the feature correlation matrix, obtain the curve of the correlation coefficient changing with the signal-to-noise ratio within the signal-to-noise ratio range, if the curve of change has a jump point and the set of function values ​​of the curve of change is countable, then the corresponding signal modulation method is determined to be discrete modulation, otherwise it is continuous modulation; For the discrete modulation, the signal modulation method is multiplied by the in-phase component and the quadrature component of the local carrier to obtain a mixed signal. The mixed signal is filtered by a low-pass filter to remove high-frequency components to obtain the in-phase branch signal and the quadrature branch signal. The in-phase branch signal and the quadrature branch signal are symbol-timed synchronized to obtain a complex symbol sequence. The real part of the complex symbol sequence is used as the horizontal axis and the imaginary part is used as the vertical axis to construct the discrete modulation constellation diagram feature. For the continuous modulation, after bandpass sampling of the signal modulation method, the real part signal and the imaginary part signal are obtained by orthogonal decomposition. The squares of the real part signal and the imaginary part signal are added together and the square root is obtained to obtain the envelope. The product of the carrier frequency and time is calculated as the carrier phase. The arctangent ratio of the real part signal and the imaginary part signal is subtracted from the carrier phase to obtain the modulation phase. The envelope is used as the amplitude and the modulation phase is used as the angle to construct the continuous modulation constellation diagram features.

5. The method according to claim 1, characterized in that, Based on constellation diagram features, multi-scale granular features are extracted for both discrete and continuous modulation. Multi-branch parallel convolution is then performed on these multi-scale granular features to obtain multi-branch features, including: For the constellation diagram features of discrete and continuous modulation, multiple sliding windows of different sizes are used to extract multi-scale granular features. Specifically, for the two-dimensional plane of discrete modulation, the Euclidean distance between the current position and the sampling points within the window is calculated, and the local density feature is obtained based on the ratio of the sum of squares of the Euclidean distances to the window size. A second-order matrix is ​​calculated based on the spatial distribution of the sampling points within the window. Eigenvalue decomposition is performed on the second-order matrix to determine eigenvalues, and eigenvectors are calculated to determine the principal direction, resulting in local shape features. The amplitude variance and phase variance of the sampling points within the window are calculated, and polar coordinate correlation features are obtained based on the covariance of the amplitude-phase joint distribution. The local density feature, the local shape feature, and the polar coordinate correlation feature are concatenated along the feature dimension to obtain the multi-scale granular features. The multi-scale granular features are input into multiple parallel convolutional branches. Each parallel convolutional branch contains a cascaded convolutional layer with convolutional kernels of different sizes. In the cascaded convolutional layer, each convolutional layer performs convolution operations on the input features to obtain a feature map. The feature map is then processed by pooling and batch normalization and used as the input to the next convolutional layer. The outputs of the last convolutional layer of each parallel convolutional branch are concatenated along the channel dimension to obtain multi-branch features.

6. The method according to claim 1, characterized in that, Steady-state feature sequences are obtained by filtering the feature correlation matrix using a hybrid filter consisting of Kalman filtering and particle filtering. Hierarchical Markov decision theory is then used to capture the evolution trend and dynamic adjustment process of the steady-state feature sequences at different levels. Finally, dual learning is used to constrain the prediction space, resulting in predictions of the signal modulation scheme, including: The feature correlation matrix is ​​decomposed into feature state vectors and observation vectors; The hybrid filter performs linear Kalman prediction on the feature state vector to obtain the predicted state; a sampling interval is determined based on the probability distribution characteristics of the observed vector; uniform sampling is performed within the sampling interval and random perturbations are superimposed to generate particle samples; the likelihood probability of the particle samples is calculated to obtain the particle weights; the predicted state is weighted and resampled based on the particle weights to output a steady-state feature sequence. In the hierarchical Markov decision structure, the policy layer receives the steady-state feature sequence and divides it into multiple decision units, calculates the immediate rewards between adjacent decision units to obtain a reward evaluation matrix; the feature layer calculates the feature transition probability matrix based on the temporal change relationship of the decision units; the execution layer calculates the state value of the reward evaluation matrix and the feature transition probability matrix, and selects the action policy corresponding to the maximum state value to obtain the optimal action sequence. The optimal action sequence is matched with a preset feature template. Based on the matching result, the optimal action sequence is reverse-mapped to obtain a reconstructed feature sequence. The cyclic consistency loss between the reconstructed feature sequence and the steady-state feature sequence is calculated. The cyclic consistency loss is used as a prediction constraint. The prediction result of the output signal modulation mode is iteratively optimized until the prediction constraint converges.

7. A signal modulation mode blind identification and intelligent classification prediction system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to obtain multi-scale time-domain features of the signal modulation mode in the time domain through wavelet packet decomposition, and obtain instantaneous frequency features in the frequency domain through Hilbert transform. The multi-scale time-domain features and the instantaneous frequency features are fused and the mutual information between the features is calculated. Based on the mutual information, the associated features are reduced in dimension and combined to construct a feature correlation matrix. The second unit is used to capture the spatial relationship of the feature correlation matrix. Based on the spatial relationship, the feature adaptive clustering and hierarchical representation are realized through a dynamic routing algorithm. At the same time, a self-correction mechanism is introduced to adaptively recalibrate the feature channels of the feature correlation matrix to obtain the blind recognition result of the signal modulation mode. The third unit is used to initially classify the signal modulation method into discrete modulation and continuous modulation based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation method, and to construct constellation diagram features by combining the signal modulation parameters. Based on constellation diagram features, multi-scale granular features are extracted for discrete modulation and continuous modulation respectively. Multi-branch parallel convolution is performed on the multi-scale granular features to obtain multi-branch features. The multi-branch features are then fused through adaptive weighting and passed through a fully connected layer to classify the signal modulation mode. The fourth unit is used to filter the feature correlation matrix using a hybrid filter composed of Kalman filtering and particle filtering to obtain a steady-state feature sequence. It uses hierarchical Markov decision to capture the evolution trend and dynamic adjustment process of the steady-state feature sequence at different levels, and uses dual learning to constrain the prediction space to obtain the prediction result of the signal modulation mode.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images