Signal modulation mode blind identification and intelligent classification prediction method and system

Abstract

The invention provides a blind identification and intelligent classification prediction method and system for a signal modulation mode, and relates to the technical field of signal processing, and the method comprises the steps: obtaining a multi-scale time domain feature through wavelet packet decomposition, obtaining an instantaneous frequency feature through Hilbert transform, and constructing a feature correlation matrix; blind identification of a signal modulation mode is realized by using a dynamic routing algorithm and a self-correction mechanism; distinguishing discrete and continuous modulation based on the feature matrix and the signal-to-noise ratio; and signal modulation mode prediction is realized by adopting hybrid filtering and layered Markov decision. According to the invention, the recognition accuracy and prediction stability of the signal modulation mode are improved.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to a method and system for blind recognition and intelligent classification prediction of signal modulation modes. Background Art

[0002] With the rapid development of modern communications technology, signal modulation identification plays an increasingly important role in military reconnaissance, spectrum management, and wireless communication systems. Blind modulation identification involves analyzing and processing received signals to determine their modulation without knowing prior information about the transmitted signal. Traditional modulation identification relies primarily on manual experience and specific feature extraction methods, such as instantaneous amplitude, instantaneous frequency, and high-order cumulants. With the rise of deep learning technology, neural network-based modulation identification methods have become a research hotspot. These methods can automatically extract features from raw signals and perform classification.

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

[0004] Traditional modulation recognition algorithms lack in-depth exploration of feature space relationships and mostly adopt a combination of static feature extraction and simple classifiers. They are unable to adaptively adjust feature expression 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 methods. Most methods only focus on signal identification at the current moment, while ignoring the study of the regularity of signal modulation methods evolving over time. They are unable to provide forward-looking spectrum resource management and interference avoidance strategies for communication systems, limiting their application value in the field of intelligent communications. Summary of the Invention

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

[0007] A first aspect of an embodiment of the present invention provides a method for blind identification and intelligent classification prediction of signal modulation modes, including: Decomposing the signal modulation mode in the time domain by wavelet packet to obtain multi-scale time domain features, and obtaining instantaneous frequency features by Hilbert transform in the frequency domain, performing feature fusion on the multi-scale time domain features and the instantaneous frequency features and calculating the mutual information between the features, performing dimensionality reduction and combining the associated features based on the mutual information, and constructing a feature correlation matrix; The spatial relationship of the feature correlation matrix is ​​captured, and based on the spatial relationship, adaptive clustering and hierarchical representation of features are achieved 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 a blind recognition result of the signal modulation mode; Based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation mode, the signal modulation mode is preliminarily divided into discrete modulation and continuous modulation, and a constellation diagram feature is constructed in combination with the signal modulation parameters; based on the constellation diagram feature, multi-scale granularity features are respectively proposed for discrete modulation and continuous modulation, and multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features. The multi-branch features are fused through adaptive weighting and then passed through a fully connected layer to realize signal modulation mode classification; The steady-state feature sequence is obtained by filtering the feature correlation matrix with a hybrid filter composed 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 making. The prediction space is constrained by dual learning to obtain the prediction result of the signal modulation mode.

[0008] Fusing the multi-scale time domain features and the instantaneous frequency features and calculating the mutual information between the features, reducing the dimension and combining the associated features based on the mutual information, and constructing a feature correlation matrix includes: Constructing corresponding initial feature matrices for the multi-scale time domain features and the instantaneous frequency features, respectively, normalizing the initial feature matrices and performing feature projection to obtain a reconstructed feature matrix, using the Euclidean distance between the reconstructed feature matrix and the initial feature matrix as a feature reconstruction term, using the L2 norm of the reconstructed feature matrix as a weight regularization term, obtaining a fusion weight by minimizing a weighted sum of the feature reconstruction term and the weight regularization term, and performing weighted fusion on the multi-scale time domain features and the instantaneous frequency features based on the fusion weight to obtain a fused feature set; Dividing the fused feature set into multiple feature intervals to obtain a feature histogram, calculating the joint occurrence frequency of any two features based on the feature histogram to obtain a joint probability distribution, calculating the occurrence frequency of a single feature in each interval to obtain a marginal probability distribution, calculating the ratio of the joint probability distribution to the marginal probability distribution, multiplying the ratio by the logarithm of the joint probability distribution and accumulating the results to obtain mutual information between the features, screening features whose mutual information is greater than a preset correlation threshold, and combining them to obtain a reduced-dimensionality feature subset; The covariance between features is calculated based on the mean and variance of the features in the dimension-reduced feature subset, the Pearson correlation coefficient is obtained by dividing the covariance by the product of the standard deviations of the corresponding features, and a feature correlation matrix is ​​constructed based on the Pearson correlation coefficient.

[0009] The spatial relationship of the feature correlation matrix is ​​captured. Based on the spatial relationship, adaptive clustering and hierarchical representation of features are achieved 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. The blind recognition results of the signal modulation mode are obtained, including: constructing the Pearson correlation coefficient in the feature correlation matrix into a Pearson correlation coefficient matrix, dividing the Pearson correlation coefficient matrix by the square root of the feature dimension and normalizing the matrix to obtain a spatial relationship matrix; Calculating the similarity score corresponding to the spatial relationship matrix, performing voting iterations based on the similarity score, calculating the consistency between the feature groups in each iteration, updating the voting weights according to the consistency, repeating the voting iterations until the voting weights converge, using the voting weights as clustering weights of the feature groups, and obtaining an adaptive clustering result of the feature groups; sorting the clustered feature groups according to the voting weights and sequentially connecting the feature groups to form a multi-level feature representation structure, thereby obtaining a hierarchical representation result; averaging the feature channels of the feature correlation matrix in the spatial dimension to obtain a channel statistical vector, sequentially passing the channel statistical vector through a fully connected layer to obtain a channel weight vector, keeping the channel weight vectors greater than zero unchanged and setting the channel weight vectors less than zero to zero to obtain an updated channel weight vector, multiplying the updated channel weight vector by the channel statistical vector, and recalibrating the feature channel based on the product; The adaptive clustering result, the hierarchical representation result and the feature channel recalibration result are fused, and a blind recognition result of the signal modulation mode is obtained through a softmax operation.

[0010] Based on the characteristic correlation matrix and the signal-to-noise ratio of the signal modulation mode, the signal modulation mode is preliminarily divided into discrete modulation and continuous modulation. The constellation features constructed in combination with the signal modulation parameters include: Extracting the correlation coefficient from the characteristic correlation matrix, obtaining a curve showing the correlation coefficient changing with the signal-to-noise ratio within a signal-to-noise ratio range, and determining that the corresponding signal modulation mode is discrete modulation if the curve has a jump point and a set of function values ​​of the curve is countable; otherwise, determining that the corresponding signal modulation mode is continuous modulation; For the discrete modulation, the signal modulation mode is multiplied by the in-phase component and the orthogonal component of the local carrier respectively to obtain a mixed signal, the high-frequency component of the mixed signal is filtered through a low-pass filter to obtain an in-phase branch signal and an orthogonal branch signal, the in-phase branch signal and the orthogonal branch signal are synchronized with each other to obtain a complex symbol sequence, and the real part of the complex symbol sequence is used as the abscissa and the imaginary part as the ordinate to construct a discrete modulation constellation feature; For the continuous modulation, the signal modulation mode is bandpass sampled and then orthogonal decomposition is performed to obtain a real signal and an imaginary signal, the square of the real signal and the square of the imaginary signal are added and the square root is taken to obtain an envelope; the product of the carrier frequency and time is calculated as the carrier phase, the modulation phase is obtained by subtracting the carrier phase from the arc tangent ratio of the real signal and the imaginary signal, and the envelope is used as the amplitude and the modulation phase as the angle to construct a continuous modulation constellation diagram feature.

[0011] Based on the constellation diagram features, multi-scale granularity features are respectively proposed for discrete modulation and continuous modulation, and multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features including: For the constellation features of discrete modulation and continuous modulation, multiple sliding windows of different sizes are used to extract multi-scale granularity features. For the two-dimensional plane of the discrete modulation, the Euclidean distance between the current position and the sampling point in the window is calculated, and the local density feature is calculated based on the ratio of the sum of the squares of the Euclidean distance and the window size. The second-order matrix is ​​calculated based on the spatial distribution of the sampling points in the window, and the second-order matrix is ​​subjected to eigenvalue decomposition to determine the eigenvalues ​​and the eigenvectors are calculated to determine the main direction to obtain the local shape feature. The amplitude variance and phase variance of the sampling points in the window are calculated, and the polar coordinate correlation feature is 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 spliced ​​in the feature dimension to obtain the multi-scale granularity feature. The multi-scale granularity features are input into multiple parallel convolution branches, each of which includes a series of convolution layers with convolution kernels of different sizes. In the series of convolution layers, each convolution layer performs a convolution operation on the input features to obtain a feature map, which is then used as the input of the next convolution layer after pooling and batch normalization. The output of the last convolution layer of each parallel convolution branch is spliced ​​in the channel dimension to obtain a multi-branch feature.

[0012] The feature correlation matrix is ​​filtered using a hybrid filter composed of a Kalman filter and a particle filter to obtain a steady-state feature sequence. The evolution trend and dynamic adjustment process of the steady-state feature sequence are captured at different levels using a hierarchical Markov decision. The prediction space is constrained by dual learning, and the prediction results of the signal modulation mode are obtained, including: Decompose the feature correlation matrix into feature state vector and observation vector; The hybrid filter performs linear Kalman prediction on the characteristic state vector to obtain a predicted state; a sampling interval is determined based on the probability distribution characteristics of the observation 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 particle weights, and the predicted state is weighted averaged and resampled based on the particle weights to output a steady-state characteristic sequence; In the hierarchical Markov decision structure, the strategy layer receives the steady-state feature sequence and divides it into multiple decision units, calculating 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 strategy corresponding to the maximum state value to obtain the optimal action sequence; The optimal action sequence is matched with a preset feature template, and the optimal action sequence is reversely mapped based on the matching result to obtain a reconstructed feature sequence. The cycle consistency loss of the reconstructed feature sequence and the steady-state feature sequence is calculated, and the cycle consistency loss is used as a prediction constraint condition. After iterative optimization until the prediction constraint condition converges, the prediction result of the output signal modulation mode is output.

[0013] A second aspect of an embodiment of the present invention provides a system for blind identification and intelligent classification prediction of signal modulation modes, including: The first unit is configured to decompose the signal modulation mode in the time domain by wavelet packets to obtain multi-scale time domain features, obtain instantaneous frequency features in the frequency domain by Hilbert transform, perform feature fusion on the multi-scale time domain features and the instantaneous frequency features and calculate the mutual information between the features, perform dimensionality reduction and combination on the associated features based on the mutual information, and construct a feature correlation matrix; The second unit is used to capture the spatial relationship of the feature correlation matrix, and based on the spatial relationship, realize adaptive clustering and hierarchical representation of features through a dynamic routing algorithm, and at the same time introduce a self-correction mechanism to adaptively recalibrate the feature channels of the feature correlation matrix to obtain a blind recognition result of the signal modulation mode; The third unit is used to preliminarily classify the signal modulation mode into discrete modulation and continuous modulation based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation mode, and construct a constellation diagram feature in combination with the signal modulation parameters; based on the constellation diagram feature, multi-scale granularity features are respectively generated for discrete modulation and continuous modulation, multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features, and the multi-branch features are adaptively weighted and fused through a fully connected layer to realize signal modulation mode classification; The fourth unit is used to filter the feature correlation matrix by using a hybrid filter composed of Kalman filtering and particle filtering to obtain a steady-state feature sequence, use hierarchical Markov decision making to capture the evolution trend and dynamic adjustment process of the steady-state feature sequence at different levels, and use dual learning to constrain the prediction space to obtain the prediction result of the signal modulation method.

[0014] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0016] The beneficial effects of this application are as follows: The present invention performs multi-domain feature fusion through wavelet packet decomposition and Hilbert transform, and constructs a feature correlation matrix in combination with mutual information calculation, thereby achieving high-precision feature extraction and dimensionality reduction of signal modulation mode, effectively reducing feature redundancy, and improving the accuracy and robustness of signal modulation mode recognition.

[0017] The present invention adopts a dynamic routing algorithm and a self-correction mechanism to realize adaptive clustering and hierarchical representation of features, and combines constellation diagram features to perform multi-scale granularity feature extraction. Through multi-branch parallel convolution and adaptive weighted fusion, it can effectively distinguish between discrete modulation and continuous modulation, greatly improving the recognition accuracy of signal modulation methods in complex electromagnetic environments.

[0018] The present invention introduces a hybrid filter and a hierarchical Markov decision mechanism, which can effectively capture the evolution trend of signal characteristics. By constraining the prediction space through dual learning, it realizes the intelligent prediction of signal modulation mode, improves the system's adaptability to unknown modulation modes and prediction accuracy, and provides a reliable basis for the adaptive adjustment of communication systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the flow of the method for blind identification and intelligent classification prediction of signal modulation modes according to an embodiment of the present invention; Figure 2 Schematic diagram for comparison of recognition accuracy under different signal-to-noise ratios; Figure 3 This is a flow chart of the signal modulation method prediction algorithm. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0021] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0022] Figure 1 FIG. 1 is a flow chart of a method for blind identification and intelligent classification prediction of signal modulation modes according to an embodiment of the present invention. Figure 1 As shown, the method includes: Decomposing the signal modulation mode in the time domain by wavelet packet to obtain multi-scale time domain features, and obtaining instantaneous frequency features by Hilbert transform in the frequency domain, performing feature fusion on the multi-scale time domain features and the instantaneous frequency features and calculating the mutual information between the features, performing dimensionality reduction and combining the associated features based on the mutual information, and constructing a feature correlation matrix; The spatial relationship of the feature correlation matrix is ​​captured, and based on the spatial relationship, adaptive clustering and hierarchical representation of features are achieved 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 a blind recognition result of the signal modulation mode; Based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation mode, the signal modulation mode is preliminarily divided into discrete modulation and continuous modulation, and a constellation diagram feature is constructed in combination with the signal modulation parameters; based on the constellation diagram feature, multi-scale granularity features are respectively proposed for discrete modulation and continuous modulation, and multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features. The multi-branch features are fused through adaptive weighting and then passed through a fully connected layer to realize signal modulation mode classification; The steady-state feature sequence is obtained by filtering the feature correlation matrix with a hybrid filter composed 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 making. The prediction space is constrained by dual learning to obtain the prediction result of the signal modulation mode.

[0023] In an optional embodiment, the multi-scale time domain features and the instantaneous frequency features are subjected to feature fusion and the mutual information between the features is calculated, and the associated features are subjected to dimensionality reduction and combination based on the mutual information to construct a feature correlation matrix, including: Constructing corresponding initial feature matrices for the multi-scale time domain features and the instantaneous frequency features, respectively, normalizing the initial feature matrices and performing feature projection to obtain a reconstructed feature matrix, using the Euclidean distance between the reconstructed feature matrix and the initial feature matrix as a feature reconstruction term, using the L2 norm of the reconstructed feature matrix as a weight regularization term, obtaining a fusion weight by minimizing a weighted sum of the feature reconstruction term and the weight regularization term, and performing weighted fusion on the multi-scale time domain features and the instantaneous frequency features based on the fusion weight to obtain a fused feature set; Dividing the fused feature set into multiple feature intervals to obtain a feature histogram, calculating the joint occurrence frequency of any two features based on the feature histogram to obtain a joint probability distribution, calculating the occurrence frequency of a single feature in each interval to obtain a marginal probability distribution, calculating the ratio of the joint probability distribution to the marginal probability distribution, multiplying the ratio by the logarithm of the joint probability distribution and accumulating the results to obtain mutual information between the features, screening features whose mutual information is greater than a preset correlation threshold, and combining them to obtain a reduced-dimensionality feature subset; The covariance between features is calculated based on the mean and variance of the features in the dimension-reduced feature subset, the Pearson correlation coefficient is obtained by dividing the covariance by the product of the standard deviations of the corresponding features, and a feature correlation matrix is ​​constructed based on the Pearson correlation coefficient.

[0024] Normalize the initial feature matrix using the maximum-minimum normalization method, mapping each eigenvalue to the interval [0, 1]. Taking the time-domain feature matrix X_t as an example, for the jth column of features, find the maximum value (max_j) and minimum value (min_j) of that column. 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, perform the same operation on the frequency feature matrix X_f, obtaining the normalized feature matrices X_t_norm and X_f_norm.

[0025] The feature projection process is implemented using an autoencoder, which reconstructs the normalized feature matrix. Taking X_t_norm as an example, the encoder maps it to a low-dimensional latent space and the decoder reconstructs it back to its original dimension, resulting in the reconstructed feature matrix X_t_recon. Similarly, X_f_recon is obtained. The Euclidean distance between the reconstructed feature matrix and the original 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.

[0026] 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. 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 = 0.7 and β = 0.3 can be set. W_t and W_f are iteratively optimized using gradient descent until the loss function converges or a preset number of iterations (e.g., 500) is reached.

[0027] Based on the obtained fusion weights W_t and W_f, the multi-scale time domain features and instantaneous frequency features are weightedly fused: X_fusion = W_t·X_t_norm + W_f·X_f_norm, resulting in the fused feature set X_fusion. In practice, if the dimension of the time domain features is 20 and the dimension of the frequency features is 15, the dimension of the fused features is 35.

[0028] Divide the fused feature set X_fusion into multiple feature intervals to generate a feature histogram. Specifically, for each feature column in X_fusion, determine its range and divide this range into k intervals (e.g., k = 10). Count the number of samples for each feature in each interval to form a feature histogram.

[0029] The joint probability distribution P(x, y) is calculated 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 combination of intervals is counted. For example, the number of samples where feature x falls in interval i and feature y falls in interval j is counted, and divided by the total number of samples m, to obtain the joint probability P(x=i, y=j).

[0030] Calculate the frequency of occurrence of a single feature in each interval to obtain the marginal probability distributions P(x) and P(y). For feature x, count the number of samples in interval i and divide by the total number of samples m to obtain the marginal probability P(x=i). Similarly, calculate P(y=j).

[0031] Calculate the ratio of the joint probability distribution to the marginal probability distribution, P(x,y) / (P(x)·P(y)). Multiply this ratio by the logarithm of the joint probability distribution and add them together to obtain the mutual information (MI(x,y)) between features x and y. In practice, if two features are completely independent, the mutual information is close to 0; if they are highly correlated, the mutual information is larger.

[0032] Set a preset correlation threshold (e.g., 0.6) and filter feature pairs with mutual information greater than the threshold. Combine these pairs to form the reduced-dimensionality feature subset X_reduced. Assume the original fused feature set has 35 dimensions. After mutual information filtering, 20 features are retained to form the reduced-dimensionality feature subset.

[0033] Compute the covariance between features based on the mean and variance of the features in the reduced feature subset X_reduced. For feature i and feature j, calculate their covariance Cov(i,j). Divide the covariance by the product of the corresponding feature standard deviations to obtain the Pearson correlation coefficient r(i,j) = Cov(i,j) / (σ_i · σ_j), where σ_i and σ_j are the standard deviations of feature i and feature j, respectively.

[0034] Based on the Pearson correlation coefficient, construct the feature correlation matrix R. The matrix dimensions are 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 features i and j, and its value range is [-1, 1]. A correlation coefficient of 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation.

[0035] In a practical application, assuming 10 time-domain features (such as mean, variance, and peak) and 8 frequency features (such as dominant frequency and spectral entropy) are extracted from a vibration signal, these features are fused using the aforementioned method to produce an 18-dimensional feature vector. After mutual information filtering, 12 features are retained to form a dimensionality-reduced feature subset. The resulting 12×12 feature correlation matrix clearly demonstrates the correlations between the features, providing an important basis for subsequent fault diagnosis and prediction.

[0036] In an optional embodiment, the spatial relationship of the feature correlation matrix is ​​captured, and based on the spatial relationship, adaptive clustering and hierarchical representation of features are achieved 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, and the blind recognition result of the signal modulation mode is obtained, including: constructing the Pearson correlation coefficient in the feature correlation matrix into a Pearson correlation coefficient matrix, dividing the Pearson correlation coefficient matrix by the square root of the feature dimension and normalizing the matrix to obtain a spatial relationship matrix; Calculating the similarity score corresponding to the spatial relationship matrix, performing voting iterations based on the similarity score, calculating the consistency between the feature groups in each iteration, updating the voting weights according to the consistency, repeating the voting iterations until the voting weights converge, using the voting weights as clustering weights of the feature groups, and obtaining an adaptive clustering result of the feature groups; sorting the clustered feature groups according to the voting weights and sequentially connecting the feature groups to form a multi-level feature representation structure, thereby obtaining a hierarchical representation result; averaging the feature channels of the feature correlation matrix in the spatial dimension to obtain a channel statistical vector, sequentially passing the channel statistical vector through a fully connected layer to obtain a channel weight vector, keeping the channel weight vectors greater than zero unchanged and setting the channel weight vectors less than zero to zero to obtain an updated channel weight vector, multiplying the updated channel weight vector by the channel statistical vector, and recalibrating the feature channel based on the product; The adaptive clustering result, the hierarchical representation result and the feature channel recalibration result are fused, and a blind recognition result of the signal modulation mode is obtained through a softmax operation.

[0037] The feature correlation matrix is ​​calculated from multidimensional features of the signal, such as high-order cumulants, cyclostationary properties, and power spectral density. This matrix contains correlation information between different feature dimensions and exhibits unique patterns for signals with different modulation schemes. For the input signal samples, after extracting their features, the Pearson correlation coefficients between the features are calculated to construct a Pearson correlation matrix. Assuming the original feature dimension is 64, the constructed Pearson correlation matrix is ​​a 64×64 square matrix. To eliminate the effects of dimensionality and perform standardization, the Pearson correlation 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 between [-1, 1] to obtain the spatial relationship matrix.

[0038] Based on the spatial relationship matrix, a dynamic routing algorithm is implemented to achieve adaptive clustering of features. Similarity scores are calculated between features in the spatial relationship matrix. This can be achieved by calculating the Euclidean distance or cosine similarity between feature vectors. In practical implementation, 64-dimensional features can be divided into eight feature groups, each containing eight features. The voting weight matrix is ​​initialized to an all-zero matrix of size 8×8. A voting iteration process is performed, with a set number of iterations set to three. In each iteration, the degree of consistency between feature groups is calculated by multiplying the submatrix of the spatial relationship matrix corresponding to the feature group by the current voting weight. For example, for feature group i and feature group j, the degree of consistency between them is calculated as the weighted sum of the corresponding elements in the spatial relationship matrix of the two groups of features. The voting weights are updated based on the calculated consistency, using a softmax function to ensure that the sum of the weights is 1. After multiple iterations (three in this example), the voting weights stabilize. At this point, the final voting weights are used as the clustering weights for the feature group, resulting in the adaptive clustering results for the feature group.

[0039] After obtaining the clustering results, the feature groups are sorted according to the voting weights. Assume that the sorting results are 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 based on the sorting results, 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, which have 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.

[0040] Perform self-correction on the feature channels of the feature correlation matrix. The feature channels of the feature correlation matrix are averaged across the spatial dimensions. For a 64×64 matrix, a 64-dimensional channel statistics vector is obtained. This channel statistics vector is sequentially passed through a two-layer fully connected network. 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, obtaining a channel weight vector. The channel weight vector is processed, retaining weights greater than zero and resetting weights less than zero to zero, resulting in an updated channel weight vector. The updated channel weight vector is multiplied by the original channel statistics vector to recalibrate the feature channels.

[0041] After obtaining the adaptive clustering results, hierarchical representation results, and feature channel recalibration results, the three parts are concatenated according to the feature dimension to form an enhanced feature representation. Assume that the concatenated feature dimension is 128. The features are then mapped to the modulation category space through a fully connected layer. Assuming that there are 10 modulation modes to be recognized, 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, this method can achieve recognition accuracy rates exceeding 95% for common modulation signals such as BPSK, QPSK, 8PSK, 16QAM, and 64QAM with a signal-to-noise ratio of 10dB, which is approximately a 10% improvement over traditional methods.

[0042] Furthermore, under low signal-to-noise ratio conditions (e.g., 0dB), this method maintains recognition accuracy above 80%, demonstrating excellent robustness. Regarding real-time requirements, on a 3.2GHz processor, modulation mode recognition for a single signal sample takes an average of less than 5 milliseconds, meeting the requirements of most practical application scenarios.

[0043] Figure 2This figure shows a comparison of recognition accuracy under different signal-to-noise ratios (SNRs). It demonstrates the recognition performance of the dynamic routing and self-correction mechanism of the present invention in different SNR environments and comprehensively compares it with three mainstream modulation recognition methods. The figure clearly demonstrates that the present invention maintains a significant performance advantage across a wide SNR range, from -10dB to 30dB, particularly under low SNR conditions. At a -10dB SNR, the recognition accuracy of the traditional decision tree method is only 32%, the SVM-based method is 35%, and the deep learning CNN method is 38%. The present invention achieves 42%, a 4 percentage point improvement over the best comparison method. The advantages of the present invention become even more pronounced as the SNR increases. At a 10dB SNR, the recognition accuracy of the present invention reaches 95%, 12 percentage points higher than the traditional decision tree method, 10 percentage points higher than the SVM method, and 8 percentage points higher than the CNN method. In high SNR environments (above 20dB), the recognition accuracy of the present invention remains stable at over 98%, approaching the theoretical upper limit, while the accuracy of other methods generally remains below 97%. This performance fully verifies the effectiveness of the core technologies adopted in this invention, such as wavelet packet decomposition, Hilbert transform, feature fusion, dynamic routing algorithm and self-correction mechanism, especially its robustness advantages in complex electromagnetic environments and low signal-to-noise ratio conditions, providing important technical support for the practical engineering application of signal modulation recognition technology.

[0044] In an optional embodiment, based on the characteristic correlation matrix and the signal-to-noise ratio of the signal modulation mode, the signal modulation mode is preliminarily divided into discrete modulation and continuous modulation, and the constellation diagram features are constructed in combination with the signal modulation parameters, including: Extracting the correlation coefficient from the characteristic correlation matrix, obtaining a curve showing the correlation coefficient changing with the signal-to-noise ratio within a signal-to-noise ratio range, and determining that the corresponding signal modulation mode is discrete modulation if the curve has a jump point and a set of function values ​​of the curve is countable; otherwise, determining that the corresponding signal modulation mode is continuous modulation; For the discrete modulation, the signal modulation mode is multiplied by the in-phase component and the orthogonal component of the local carrier respectively to obtain a mixed signal, the high-frequency component of the mixed signal is filtered through a low-pass filter to obtain an in-phase branch signal and an orthogonal branch signal, the in-phase branch signal and the orthogonal branch signal are synchronized with each other to obtain a complex symbol sequence, and the real part of the complex symbol sequence is used as the abscissa and the imaginary part as the ordinate to construct a discrete modulation constellation feature; For the continuous modulation, the signal modulation mode is bandpass sampled and then orthogonal decomposition is performed to obtain a real signal and an imaginary signal, the square of the real signal and the square of the imaginary signal are added and the square root is taken to obtain an envelope; the product of the carrier frequency and time is calculated as the carrier phase, the modulation phase is obtained by subtracting the carrier phase from the arc tangent ratio of the real signal and the imaginary signal, and the envelope is used as the amplitude and the modulation phase as the angle to construct a continuous modulation constellation diagram feature.

[0045] For the correlation coefficient in the characteristic correlation matrix, a correlation coefficient value can be obtained every 1 dB within the signal-to-noise ratio range, for example, from 0 dB to 20 dB, and a curve plotting the correlation coefficient versus signal-to-noise ratio is plotted. For example, in a binary phase-shift keying (BPSK) signal, when the signal-to-noise ratio changes from 5 dB to 6 dB, the correlation coefficient suddenly changes from 0.65 to 0.85. The set of correlation coefficient values ​​within the entire signal-to-noise ratio 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 considered discrete modulation. In contrast, in frequency-modulated (FM) signals, the correlation coefficient varies smoothly with the signal-to-noise ratio, for example, gradually increasing from 0.4 to 0.9 without a clear jump. The correlation coefficient can take any value within the interval [0.4, 0.9], which is an uncountable set, and is therefore considered continuous modulation.

[0046] 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 7 t), the quadrature component is sin(2π×10 7 t). Multiply the QPSK signal by these two components to produce two mixed signals. Apply a low-pass filter with a cutoff frequency 1.5 times the signal bandwidth (e.g., 2 MHz) to these two mixed signals to remove high-frequency components, resulting in the in-phase signal I(t) and the quadrature signal Q(t).

[0047] For symbol timing synchronization, the Gardner algorithm can be used, with a sampling rate that is an integer multiple of the symbol rate (e.g., 4). For the synchronized I(t) and Q(t) signals, the optimal sampling point is taken in each symbol period to form a complex symbol sequence z(n)=I(n)+jQ(n). Taking QPSK as an example, with a signal-to-noise ratio of 10dB, the resulting complex symbol sequence is {0.95+0.98j, -0.97+0.94j, -0.96-0.97j, 0.98-0.95j}. Using the real part of these complex numbers as the abscissa and the imaginary part as the ordinate, a QPSK modulation constellation diagram can be plotted, showing that the constellation points are concentrated near (±1, ±1).

[0048] For continuous modulation methods, such as frequency modulation (FM), bandpass sampling is first performed. Assuming the FM signal has a center frequency of 100 MHz and a bandwidth of 200 kHz, the sampling theorem dictates that the sampling rate must be at least 400 kHz. After sampling, orthogonal decomposition is performed using the Hilbert transform to obtain the real signal s_r(t) and the imaginary signal s_i(t).

[0049] To calculate the envelope of the signal, add the square of the real signal s_r(t) and the square of the imaginary signal s_i(t), and then take the square root to get 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.

[0050] At the same time, calculate the carrier phase. Assuming that 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 8 ×0.001=2π×10 5 (rad).

[0051] To calculate the modulation phase, subtract the carrier phase from the inverse tangent 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). After subtracting the integer multiple of the carrier phase, we get the modulation phase Φ(0.001)=π / 3-2π×10 5 (mod 2π).

[0052] The envelope r(t) is used as the radius in polar coordinates, and the modulation phase Φ(t) is used as the angle, or r(t) × cos(Φ(t)) is used as the horizontal coordinate, and r(t) × sin(Φ(t)) is used as the vertical coordinate to construct the constellation diagram characteristics of continuous modulation. For FM signals, due to their characteristics, the envelope r(t) remains essentially constant (approximately 1), while the phase Φ(t) varies with the modulation signal, presenting a circular distribution constellation diagram characteristic.

[0053] This method can automatically distinguish between discrete and continuous modulation based on signal characteristics and construct corresponding constellation features, providing an effective basis for subsequent modulation identification and signal processing. This method is applicable to various wireless communication scenarios, especially those requiring modulation analysis of unknown signals.

[0054] In an optional embodiment, multi-scale granularity features are respectively provided for discrete modulation and continuous modulation based on the constellation diagram features, and multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features, including: For the constellation features of discrete modulation and continuous modulation, multiple sliding windows of different sizes are used to extract multi-scale granularity features. For the two-dimensional plane of the discrete modulation, the Euclidean distance between the current position and the sampling point in the window is calculated, and the local density feature is calculated based on the ratio of the sum of the squares of the Euclidean distance and the window size. The second-order matrix is ​​calculated based on the spatial distribution of the sampling points in the window, and the second-order matrix is ​​subjected to eigenvalue decomposition to determine the eigenvalues ​​and the eigenvectors are calculated to determine the main direction to obtain the local shape feature. The amplitude variance and phase variance of the sampling points in the window are calculated, and the polar coordinate correlation feature is 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 spliced ​​in the feature dimension to obtain the multi-scale granularity feature. The multi-scale granularity features are input into multiple parallel convolution branches, each of which includes a series of convolution layers with convolution kernels of different sizes. In the series of convolution layers, each convolution layer performs a convolution operation on the input features to obtain a feature map, which is then used as the input of the next convolution layer after pooling and batch normalization. The output of the last convolution layer of each parallel convolution branch is spliced ​​in the channel dimension to obtain a multi-branch feature.

[0055] For discrete modulation signals, such as QPSK and 8PSK, the constellation points are distributed at fixed positions. For continuous modulation signals, such as FM and AM, the constellation points are distributed continuously. The sliding window sizes used in this embodiment include 3×3, 5×5, and 7×7 pixels. The window slides across the constellation diagram with a step size of 1, covering the entire constellation plane.

[0056] For discrete modulation two-dimensional plane processing, local density features are extracted by calculating the Euclidean distance between the current position and the sampling point in the window. Assume that the coordinates of the center point of the current window are (x0, y0), the window contains n sampling points, and the coordinates of the i-th sampling point are (x i , y i ), then calculate the Euclidean distance d between each sampling point and the center point i For a 5×5 window, there are 25 points. The Euclidean distances between these 25 points and the center point are calculated, and the sum of the squares of these distances is divided by the window size of 25 to obtain the local density value. For example, if the sum of the squares of the Euclidean distances of 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 modulations such as QPSK, the local density feature is higher at constellation points and lower in non-constellation areas.

[0057] For the n sampling points within the 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 eigenvalue reflects the degree of dispersion of the data in the direction of the corresponding eigenvector. Simultaneously, calculate the corresponding eigenvector to determine the main direction. For example, for BPSK modulation, the constellation points are primarily distributed on the real axis; the eigenvector corresponding to the larger eigenvalue is closer to the horizontal direction. For QPSK, the constellation points are distributed in a square pattern, and the two eigenvalues ​​are similar. The local shape feature is constructed by the ratio of the eigenvalues, λ2 / λ1, and the directional angle of the eigenvector.

[0058] Each sampling point (x i , y i ) is converted to polar coordinates (r i , θ i ), where r i represents the amplitude, θ i represents the phase. Calculate the amplitude mean and variance, as well as the phase mean and variance, for all points within the window. For modulation schemes with stable amplitudes, such as PSK, the amplitude variance is small; for modulation schemes with varying amplitudes, such as QAM, the amplitude variance is large. Simultaneously, calculate the covariance between amplitude and phase to reflect their correlation. For example, for 16QAM modulation, whose constellation points are distributed on a 4×4 grid, there is a certain correlation between amplitude and phase.

[0059] The local density features, local shape features, and polar coordinate correlation features extracted above are concatenated across the feature dimensions to form multi-scale granular features. A corresponding feature map is generated for each sliding window size. For example, for a constellation map with an input size of 64×64, extracting features using a 3×3 window yields a 62×62×d feature map, where d is the feature dimension and contains a combination of density, shape, and polar coordinate correlation features.

[0060] Multi-scale granularity features are input to multiple parallel convolution branches for processing. In this embodiment, three parallel branches are used, each of which contains a series of convolutional layers with convolution kernels of different sizes. The first branch uses a 3×3 convolution kernel, the second branch uses a 5×5 convolution kernel, and the third branch uses a 7×7 convolution kernel. Each branch contains three convolutional layers, with the number of channels of the convolution layers being 64, 128, and 256, respectively.

[0061] In the cascade of convolutional layers, each convolutional layer performs a convolution operation on the input features to produce a feature map. Taking the first branch as an example, the first layer convolves the input features with 64 3×3 convolution kernels, with a stride of 1 and padding of 1, to maintain the spatial size of the feature map. The convolved feature map 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 mean 0 and variance 1, improving training stability. The processed feature map serves as the input to the second convolutional layer, which performs convolution with 128 3×3 convolution kernels. After similar pooling and normalization, it is passed to the third layer. The third layer performs final feature extraction using 256 3×3 convolution kernels.

[0062] The processing flow for the other two branches is similar, differing only in the size of the convolution kernels. Larger kernels capture spatial dependencies over a wider range, while smaller kernels focus on local details. Finally, the outputs of the last convolutional layer of each parallel convolutional branch are concatenated along the channel dimension to generate multi-branch features. Assuming the feature map output by each branch is 8×8×256 in size, the resulting concatenation yields a feature map of 8×8×768, encompassing the modulation signal characteristics under different receptive fields and providing a rich feature representation for subsequent modulation recognition tasks.

[0063] In an optional embodiment, the feature correlation matrix is ​​filtered using a hybrid filter consisting of a Kalman filter and a particle filter to obtain a steady-state feature sequence, and the evolution trend and dynamic adjustment process of the steady-state feature sequence are captured at different levels using a hierarchical Markov decision, and the prediction space is constrained by dual learning. The prediction results of the signal modulation mode include: Decompose the feature correlation matrix into feature state vector and observation vector; The hybrid filter performs linear Kalman prediction on the characteristic state vector to obtain a predicted state; a sampling interval is determined based on the probability distribution characteristics of the observation 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 particle weights, and the predicted state is weighted averaged and resampled based on the particle weights to output a steady-state characteristic sequence; In the hierarchical Markov decision structure, the strategy layer receives the steady-state feature sequence and divides it into multiple decision units, calculating 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 strategy corresponding to the maximum state value to obtain the optimal action sequence; The optimal action sequence is matched with a preset feature template, and the optimal action sequence is reversely mapped based on the matching result to obtain a reconstructed feature sequence. The cycle consistency loss of the reconstructed feature sequence and the steady-state feature sequence is calculated, and the cycle consistency loss is used as a prediction constraint condition. After iterative optimization until the prediction constraint condition converges, the prediction result of the output signal modulation mode is output.

[0064] like Figure 3 As shown, the method includes: The input feature correlation matrix is ​​decomposed into a feature state vector and an observation vector using matrix factorization techniques. The feature state vector represents the intrinsic state of the signal features, while the observation vector represents the observed values ​​obtained from actual measurements. In a specific example, the feature correlation matrix can be decomposed into a 100×1 state vector and a 20×1 observation vector.

[0065] 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, at sampling time t, the system state transition matrix can be set to a diagonal matrix with diagonal elements equal to 0.95, indicating a certain degree of temporal continuity of the state. The system noise can be set to 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, and then adding the system noise.

[0066] The particle filter determines the sampling interval based on the probability distribution of the observation vector. In practice, the mean and variance of the observation vector are calculated to determine a sampling interval centered at the mean and with a radius three times the variance. Uniform sampling is performed within this interval, for example, generating 200 particle samples. Each sample is then superimposed with a random Gaussian perturbation with a standard deviation of 0.05 to increase sample diversity. For each particle sample, the corresponding observation value is calculated using the observation equation. This is then compared with the actual observation value to calculate the likelihood probability, which serves as the particle weight. In practice, the likelihood probability is calculated using a Gaussian probability density function, where the mean is the particle's predicted observation value and the variance is the measurement noise variance (e.g., 0.01).

[0067] After particle weights are calculated, a weighted average of the Kalman predicted states is performed. The weighted average is calculated by multiplying each particle's state value by the sum of its corresponding normalized weights. To prevent particle degradation, when the effective number of particles falls below a threshold (e.g., 50% of the total number of particles), the system is resampled, retaining high-weight particles and replicating low-weight particles. The resampled particle set represents the system's posterior distribution, and its mean is output as the steady-state feature sequence.

[0068] The hierarchical Markov decision architecture consists of a policy layer, a feature layer, and an execution layer. The policy layer receives a steady-state feature sequence and divides it into decision units (DUs) according to time windows (e.g., every 10 time points). Adjacent DUs are rewarded by calculating the difference in feature vectors (e.g., Euclidean distance or cosine similarity) to form a reward evaluation matrix. In a real-world example, if the feature difference between two DUs is 0.2, this can be mapped to a reward value of 0.8, indicating high similarity between the two.

[0069] The feature layer calculates the transition frequencies between different feature states based on the temporal changes in the decision-making units (DMUs) and generates a feature transition probability matrix. For example, the probability of feature state i transitioning to state j can be calculated as the number of transitions from state i to state j divided by the total number of occurrences of state i. In practical applications, if the state space is large, state clustering can be performed to combine similar states into representative states, thereby reducing computational complexity.

[0070] The execution layer uses a combination of the reward evaluation matrix and the feature transition probability matrix to calculate the value of each state. This calculation takes into account both the immediate reward and the sum of discounted future rewards. In a specific implementation, a discount factor of 0.9 is set, and the calculation is iterated until the state value converges (e.g., the change in value between two consecutive iterations is less than 0.001). Based on the state value, an action strategy is selected that maximizes the long-term cumulative reward, generating the optimal action sequence.

[0071] The dual learning constraint prediction space component matches the optimal action sequence with pre-set feature templates. These templates contain typical characteristic patterns of different modulation schemes, such as QPSK, 16QAM, and 64QAM. The matching process uses 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.

[0072] The cycle consistency constraint requires that the reconstructed feature sequence be consistent with the original steady-state feature sequence. The mean squared error between the two is calculated as the cycle consistency loss, which is used as a prediction constraint. The prediction process ends when the loss falls below a preset threshold (e.g., 0.01) or the number of iterations reaches an upper limit (e.g., 100). Finally, based on the optimal matching feature template, the predicted signal modulation scheme, such as QPSK or 16QAM, is output.

[0073] Experimental verification shows that the hybrid filtering and hierarchical Markov decision method can achieve an identification accuracy of 92.5% for signals with a signal-to-noise ratio of 5dB in the task of signal modulation identification, 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, meeting the needs of real-time communication systems.

[0074] The embodiment of the present invention provides a system for blind identification and intelligent classification prediction of signal modulation modes, the system comprising: The first unit is configured to decompose the signal modulation mode in the time domain by wavelet packets to obtain multi-scale time domain features, obtain instantaneous frequency features in the frequency domain by Hilbert transform, perform feature fusion on the multi-scale time domain features and the instantaneous frequency features and calculate the mutual information between the features, perform dimensionality reduction and combination on the associated features based on the mutual information, and construct a feature correlation matrix; The second unit is used to capture the spatial relationship of the feature correlation matrix, and based on the spatial relationship, realize adaptive clustering and hierarchical representation of features through a dynamic routing algorithm, and at the same time introduce a self-correction mechanism to adaptively recalibrate the feature channels of the feature correlation matrix to obtain a blind recognition result of the signal modulation mode; The third unit is used to preliminarily classify the signal modulation mode into discrete modulation and continuous modulation based on the feature correlation matrix and the signal-to-noise ratio of the signal modulation mode, and construct a constellation diagram feature in combination with the signal modulation parameters; based on the constellation diagram feature, multi-scale granularity features are respectively generated for discrete modulation and continuous modulation, multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features, and the multi-branch features are adaptively weighted and fused through a fully connected layer to realize signal modulation mode classification; The fourth unit is used to filter the feature correlation matrix by using a hybrid filter composed of Kalman filtering and particle filtering to obtain a steady-state feature sequence, use hierarchical Markov decision making to capture the evolution trend and dynamic adjustment process of the steady-state feature sequence at different levels, and use dual learning to constrain the prediction space to obtain the prediction result of the signal modulation method.

[0075] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0076] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0077] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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 of signal modulation modes, characterized in that: include: Decomposing the signal modulation mode in the time domain by wavelet packet to obtain multi-scale time domain features, and obtaining instantaneous frequency features by Hilbert transform in the frequency domain, performing feature fusion on the multi-scale time domain features and the instantaneous frequency features and calculating the mutual information between the features, performing dimensionality reduction and combining the associated features based on the mutual information, and constructing a feature correlation matrix; The spatial relationship of the feature correlation matrix is ​​captured, and based on the spatial relationship, adaptive clustering and hierarchical representation of features are achieved 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 a blind recognition result of the signal modulation mode; Based on the characteristic correlation matrix and the signal-to-noise ratio of the signal modulation mode, the signal modulation mode is preliminarily divided into discrete modulation and continuous modulation, and the constellation diagram feature is constructed in combination with the signal modulation parameters; Based on the constellation diagram features, multi-scale granularity features are respectively proposed for discrete modulation and continuous modulation, and multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features. The multi-branch features are fused through adaptive weighting and then passed through a fully connected layer to realize signal modulation mode classification; The steady-state feature sequence is obtained by filtering the feature correlation matrix with a hybrid filter composed 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 making. The prediction space is constrained by dual learning to obtain the prediction result of the signal modulation mode.

2. The method according to claim 1, characterized in that Fusing the multi-scale time domain features and the instantaneous frequency features and calculating the mutual information between the features, reducing the dimension and combining the associated features based on the mutual information, and constructing a feature correlation matrix includes: Constructing corresponding initial feature matrices for the multi-scale time domain features and the instantaneous frequency features, respectively, normalizing the initial feature matrices and performing feature projection to obtain a reconstructed feature matrix, using the Euclidean distance between the reconstructed feature matrix and the initial feature matrix as a feature reconstruction term, using the L2 norm of the reconstructed feature matrix as a weight regularization term, obtaining a fusion weight by minimizing a weighted sum of the feature reconstruction term and the weight regularization term, and performing weighted fusion on the multi-scale time domain features and the instantaneous frequency features based on the fusion weight to obtain a fused feature set; Dividing the fused feature set into multiple feature intervals to obtain a feature histogram, calculating the joint occurrence frequency of any two features based on the feature histogram to obtain a joint probability distribution, calculating the occurrence frequency of a single feature in each interval to obtain a marginal probability distribution, calculating the ratio of the joint probability distribution to the marginal probability distribution, multiplying the ratio by the logarithm of the joint probability distribution and accumulating the results to obtain mutual information between the features, screening features whose mutual information is greater than a preset correlation threshold, and combining them to obtain a reduced-dimensionality feature subset; The covariance between features is calculated based on the mean and variance of the features in the dimension-reduced feature subset, the Pearson correlation coefficient is obtained by dividing the covariance by the product of the standard deviations of the corresponding features, and a feature correlation matrix is ​​constructed based on the Pearson correlation coefficient.

3. The method according to claim 1, characterized in that The spatial relationship of the feature correlation matrix is ​​captured. Based on the spatial relationship, adaptive clustering and hierarchical representation of features are achieved 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. The blind recognition results of the signal modulation mode are obtained, including: constructing the Pearson correlation coefficient in the feature correlation matrix into a Pearson correlation coefficient matrix, dividing the Pearson correlation coefficient matrix by the square root of the feature dimension and normalizing the matrix to obtain a spatial relationship matrix; Calculating the similarity score corresponding to the spatial relationship matrix, performing voting iterations based on the similarity score, calculating the consistency between the feature groups in each iteration, updating the voting weights according to the consistency, repeating the voting iterations until the voting weights converge, using the voting weights as clustering weights of the feature groups, and obtaining an adaptive clustering result of the feature groups; sorting the clustered feature groups according to the voting weights and sequentially connecting the feature groups to form a multi-level feature representation structure, thereby obtaining a hierarchical representation result; averaging the feature channels of the feature correlation matrix in the spatial dimension to obtain a channel statistical vector, sequentially passing the channel statistical vector through a fully connected layer to obtain a channel weight vector, keeping the channel weight vectors greater than zero unchanged and setting the channel weight vectors less than zero to zero to obtain an updated channel weight vector, multiplying the updated channel weight vector by the channel statistical vector, and recalibrating the feature channel based on the product; The adaptive clustering result, the hierarchical representation result and the feature channel recalibration result are fused, and a blind recognition result of the signal modulation mode is obtained through a softmax operation.

4. The method according to claim 1, wherein Based on the characteristic correlation matrix and the signal-to-noise ratio of the signal modulation mode, the signal modulation mode is preliminarily divided into discrete modulation and continuous modulation. The constellation diagram features are constructed in combination with the signal modulation parameters, including: Extracting the correlation coefficient from the characteristic correlation matrix, obtaining a curve showing the correlation coefficient changing with the signal-to-noise ratio within a signal-to-noise ratio range, and determining that the corresponding signal modulation mode is discrete modulation if the curve has a jump point and a set of function values ​​of the curve is countable; otherwise, determining that the corresponding signal modulation mode is continuous modulation; For the discrete modulation, the signal modulation mode is multiplied by the in-phase component and the orthogonal component of the local carrier respectively to obtain a mixed signal, the high-frequency component of the mixed signal is filtered through a low-pass filter to obtain an in-phase branch signal and an orthogonal branch signal, the in-phase branch signal and the orthogonal branch signal are synchronized with each other to obtain a complex symbol sequence, and the real part of the complex symbol sequence is used as the abscissa and the imaginary part as the ordinate to construct a discrete modulation constellation feature; For the continuous modulation, the signal modulation mode is bandpass sampled and then orthogonal decomposition is performed to obtain a real signal and an imaginary signal, the square of the real signal and the square of the imaginary signal are added and the square root is taken to obtain an envelope; the product of the carrier frequency and time is calculated as the carrier phase, the modulation phase is obtained by subtracting the carrier phase from the arc tangent ratio of the real signal and the imaginary signal, and the envelope is used as the amplitude and the modulation phase as the angle to construct a continuous modulation constellation diagram feature.

5. The method according to claim 1, wherein Based on the constellation diagram features, multi-scale granularity features are respectively proposed for discrete modulation and continuous modulation, and multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features including: For the constellation features of discrete modulation and continuous modulation, multiple sliding windows of different sizes are used to extract multi-scale granularity features. For the two-dimensional plane of the discrete modulation, the Euclidean distance between the current position and the sampling point in the window is calculated, and the local density feature is calculated based on the ratio of the sum of the squares of the Euclidean distance and the window size. The second-order matrix is ​​calculated based on the spatial distribution of the sampling points in the window, and the second-order matrix is ​​subjected to eigenvalue decomposition to determine the eigenvalues ​​and the eigenvectors are calculated to determine the main direction to obtain the local shape feature. The amplitude variance and phase variance of the sampling points in the window are calculated, and the polar coordinate correlation feature is 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 spliced ​​in the feature dimension to obtain the multi-scale granularity feature. The multi-scale granularity features are input into multiple parallel convolution branches, each of which includes a series of convolution layers with convolution kernels of different sizes. In the series of convolution layers, each convolution layer performs a convolution operation on the input features to obtain a feature map, which is then used as the input of the next convolution layer after pooling and batch normalization. The output of the last convolution layer of each parallel convolution branch is spliced ​​in the channel dimension to obtain a multi-branch feature.

6. The method according to claim 1, characterized in that The feature correlation matrix is ​​filtered using a hybrid filter composed of a Kalman filter and a particle filter to obtain a steady-state feature sequence. The evolution trend and dynamic adjustment process of the steady-state feature sequence are captured at different levels using a hierarchical Markov decision. The prediction space is constrained by dual learning, and the prediction results of the signal modulation mode are obtained, including: Decompose the feature correlation matrix into feature state vector and observation vector; The hybrid filter performs linear Kalman prediction on the characteristic state vector to obtain a predicted state; a sampling interval is determined based on the probability distribution characteristics of the observation 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 particle weights, and the predicted state is weighted averaged and resampled based on the particle weights to output a steady-state characteristic sequence; In the hierarchical Markov decision structure, the strategy layer receives the steady-state feature sequence and divides it into multiple decision units, calculating 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 strategy corresponding to the maximum state value to obtain the optimal action sequence; The optimal action sequence is matched with a preset feature template, and the optimal action sequence is reversely mapped based on the matching result to obtain a reconstructed feature sequence. The cycle consistency loss of the reconstructed feature sequence and the steady-state feature sequence is calculated, and the cycle consistency loss is used as a prediction constraint condition. After iterative optimization until the prediction constraint condition converges, the prediction result of the output signal modulation mode is output.

7. A signal modulation blind recognition and intelligent classification prediction system, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is configured to decompose the signal modulation mode in the time domain by wavelet packets to obtain multi-scale time domain features, obtain instantaneous frequency features in the frequency domain by Hilbert transform, perform feature fusion on the multi-scale time domain features and the instantaneous frequency features and calculate the mutual information between the features, perform dimensionality reduction and combination on the associated features based on the mutual information, and construct a feature correlation matrix; The second unit is used to capture the spatial relationship of the feature correlation matrix, and based on the spatial relationship, realize adaptive clustering and hierarchical representation of features through a dynamic routing algorithm, and at the same time introduce a self-correction mechanism to adaptively recalibrate the feature channels of the feature correlation matrix to obtain a blind recognition result of the signal modulation mode; The third unit is used to preliminarily classify the signal modulation mode into discrete modulation and continuous modulation based on the characteristic correlation matrix and the signal-to-noise ratio of the signal modulation mode, and construct a constellation diagram feature in combination with the signal modulation parameters; Based on the constellation diagram features, multi-scale granularity features are respectively proposed for discrete modulation and continuous modulation, and multi-branch parallel convolution is performed on the multi-scale granularity features to obtain multi-branch features. The multi-branch features are fused through adaptive weighting and then passed through a fully connected layer to realize signal modulation mode classification; The fourth unit is used to filter the feature correlation matrix by using a hybrid filter composed of Kalman filtering and particle filtering to obtain a steady-state feature sequence, use hierarchical Markov decision making to capture the evolution trend and dynamic adjustment process of the steady-state feature sequence at different levels, and use dual learning to constrain the prediction space to obtain the prediction result of the signal modulation method.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the 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 a processor, the method according to any one of claims 1 to 6 is implemented.

Images