Multi-domain feature fusion communication signal modulation identification method based on deep learning

Abstract

The invention discloses a multi-domain feature fusion communication signal modulation identification method based on deep learning, and belongs to the technical field of communication signal processing. The method comprises the following steps of: extracting a multi-domain feature vector from a complex baseband signal, wherein the multi-domain feature vector covers six categories including a time domain statistical feature, a frequency domain feature, a high-order cumulant feature, a differential phase feature, an I / Q component feature and a PSK (Phase Shift Keying) special feature; constructing a training data set by adopting a logarithmic space uniform sampling strategy, and realizing balanced coverage of a cross-order-of-magnitude symbol rate; and inputting the extracted feature vectors into a layer-by-layer dimensionality reduction neural network adopting a decline Dropout strategy for classification and identification. According to the method, the systematic coverage of 32 modulation types covering nine categories is realized, the natural immune characteristic of the high-order cumulant to additive white Gaussian noise is fully utilized, and the method has the advantages of wide identification range, strong anti-noise capability and good generalization performance.

Description

Technical Field

[0001] This invention belongs to the field of communication signal modulation recognition technology, specifically relating to a multi-domain feature fusion communication signal modulation recognition method based on deep learning. Background Technology

[0002] Automatic modulation identification (AMR) is a key technology in the field of communication signal processing, with significant application value in areas such as spectrum monitoring, electronic reconnaissance, cognitive radio, and signals intelligence. This technology can automatically determine the modulation scheme of a received signal without prior information, laying the foundation for subsequent signal demodulation and information extraction.

[0003] In recent years, the rapid development of deep learning technology has provided new research ideas for modulation recognition. Existing deep learning-based modulation recognition methods mainly fall into three categories: The first category is end-to-end methods, which directly input the original IQ signal or time-domain waveform into deep networks such as convolutional neural networks (CNN) and residual networks (ResNet) for feature learning and classification; the second category is image representation-based methods, which convert the signal into two-dimensional images such as constellation diagrams, time-frequency diagrams, and cyclic spectra before feeding them into image classification networks; and the third category is sequence representation-based methods, which use recurrent neural networks (RNN) or long short-term memory networks (LSTM) to perform time-series modeling of IQ sequences.

[0004] However, existing deep learning-based modulation recognition methods still have the following shortcomings:

[0005] First, end-to-end methods have poor interpretability. While learning features directly from raw signals avoids manual feature design, the features learned by the network lack physical meaning, the model's decision-making process is difficult to interpret, and it is prone to overfitting when training data is insufficient.

[0006] Second, the information utilization of a single representation method is insufficient. Whether using time-domain sequences, frequency-domain features, or image representation, a single data representation format is insufficient to fully characterize the discriminative features of the modulated signal in different transform domains, resulting in insufficient ability to distinguish modulation types with similar characteristics.

[0007] Third, performance degrades significantly under low signal-to-noise ratio (SNR) conditions. Most existing methods rely on networks to automatically learn noise-resistant features, failing to fully utilize proven noise-resistant features in signal processing (such as higher-order cumulants), resulting in a significant reduction in recognition accuracy under low SNR conditions.

[0008] Fourth, the ability to distinguish similar modulation types within the same family is insufficient. Existing methods and publicly available datasets typically do not simultaneously include similar modulation types with the same constellation diagram positions, such as QPSK, OQPSK, and π / 4-QPSK, resulting in a relative lack of research on the identification of such fine-grained differences.

[0009] Fifth, the coverage of modulation types is limited. Existing research experimental verification is mostly based on public datasets such as RadioML, and the modulation types are usually no more than 24. Research on identification methods for large-scale modulation type sets is relatively scarce, which restricts the practical application of modulation identification technology in complex and diverse communication scenarios. Summary of the Invention

[0010] To address the aforementioned problems, this invention proposes a deep learning-based multi-domain feature fusion method for communication signal modulation recognition. This method organically combines expert knowledge in signal processing with deep learning technology. By extracting multi-domain features covering six categories, combining high-order cumulants with natural Gaussian noise suppression capabilities with fine-grained phase analysis features for the PSK family, and utilizing deep neural networks to achieve adaptive feature fusion and classification, it ultimately achieves accurate recognition of 32 modulation types.

[0011] The technical solution of this invention is as follows:

[0012] A deep learning-based multi-domain feature fusion communication signal modulation recognition method includes the following steps:

[0013] S1. From the received communication signal, different signal processing methods are used to extract time-domain statistical features, frequency-domain features, higher-order cumulant features, differential phase features, I / Q component characteristic features, and PSK-specific features, respectively.

[0014] S2. After standardizing the extracted multi-domain feature vectors, input them into a fully connected neural network with progressively reduced dimensionality. After multi-layer feature abstraction, output the prediction scores for each modulation type.

[0015] S3. Determine the modulation type identification result based on the predicted score to classify the modulation method of the communication signal.

[0016] First, a 56-dimensional multi-domain feature system was constructed, covering time-domain statistical features, frequency-domain features, higher-order cumulant features, differential phase features, I / Q component characteristics, and constellation-derived features. This system comprehensively characterizes the modulation signal from multiple perspectives, overcoming the problem that a single feature is insufficient to distinguish modulation types.

[0017] Second, by fully utilizing the natural immunity of higher-order cumulants to additive white Gaussian noise, and combining the sign characteristics and ratio characteristics of cumulants, the recognition performance under low signal-to-noise ratio conditions is significantly improved.

[0018] Third, it achieves systematic coverage of 32 modulation types in nine categories, including phase shift keying, quadrature amplitude, pulse amplitude, amplitude phase shift keying, differential, offset, frequency, amplitude and multicarrier modulation, to meet the diverse identification needs of modern communication systems.

[0019] Fourth, a log-space uniform sampling strategy is used to construct the training dataset, which solves the problem of insufficient sampling in the low-speed segment caused by linear sampling when the symbol rate spans multiple orders of magnitude, and ensures the model's balanced learning ability for signals of different rates.

[0020] Fifth, a strategy of alternating signal-to-noise ratio distribution between training and test sets is adopted to force the model to perform interpolation predictions under conditions where the signal-to-noise ratio is not known, which effectively verifies and improves the model's generalization ability.

[0021] Sixth, a layer-by-layer dimensionality reduction network structure and a decreasing Dropout strategy are adopted. Strong regularization is applied to the front layers of the network to suppress overfitting, while maintaining the expression accuracy of the output layer, thus achieving a balance between classification performance and generalization ability. Attached Figure Description

[0022] Figure 1 This is a flowchart of the modulation recognition method provided in the embodiments of the present invention;

[0023] Figure 2 This is a diagram of the layer-by-layer dimensionality reduction fully connected neural network structure used in the embodiments of the present invention;

[0024] Figure 3 This is a schematic diagram illustrating the staggered signal-to-noise ratio distribution of the training and test sets in an embodiment of the present invention;

[0025] Figure 4 This is a comparison diagram of the symbol rate distribution between logarithmic space sampling and linear sampling in an embodiment of the present invention;

[0026] Figure 5 The following are modulation recognition accuracy curves under different signal-to-noise ratios according to embodiments of the present invention;

[0027] Figure 6 This is the confusion matrix of an embodiment of the present invention under a signal-to-noise ratio of -5 dB;

[0028] Figure 7 This is the confusion matrix of an embodiment of the present invention under a signal-to-noise ratio of 3 dB;

[0029] Figure 8 This is the confusion matrix of an embodiment of the present invention under a signal-to-noise ratio of 11 dB;

[0030] Figure 9 This is the confusion matrix of an embodiment of the present invention under a signal-to-noise ratio of 17 dB;

[0031] Figure 10 The recognition accuracy (%) of each modulation type under different signal-to-noise ratios. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings.

[0033] like Figure 1 As shown, the modulation recognition method of the present invention includes the following steps:

[0034] S1, Multi-domain Feature Extraction

[0035] A 56-dimensional multi-domain feature vector, covering six categories, was extracted from the received communication signal. The extraction process for each type of feature is as follows:

[0036] S1-1, Time-domain statistical characteristics (10 dimensions)

[0037] The 0th dimension feature is defined as the arithmetic mean of the signal amplitude, and its mathematical expression is:

[0038]

[0039] in For signal length, It is a complex signal. This characteristic reflects the average power level of the signal.

[0040] The first dimension of the feature calculates the standard deviation of the signal amplitude, reflecting the degree of dispersion of amplitude variation:

[0041]

[0042] The second feature is the variance of the signal amplitude, which is equal to the square of the standard deviation:

[0043]

[0044] The third-dimensional feature extracts the maximum value in the signal amplitude sequence:

[0045]

[0046] The fourth-dimensional feature identifies the minimum value in the signal amplitude sequence:

[0047]

[0048] The 5th and 6th dimensions of the feature calculate the mean and standard deviation of the real components of the complex signal, respectively, in the following mathematical form:

[0049]

[0050]

[0051] The 7th and 8th dimensions correspond to the mean and standard deviation of the imaginary components of the complex signal, and are expressed as follows:

[0052]

[0053]

[0054] The 9th feature is used to calculate the peak-to-average power ratio (PAPR), which is defined as the ratio of peak power to average power.

[0055]

[0056] PAPR is very effective in distinguishing between constant envelope modulation (PSK, FSK, MSK) and non-constant envelope modulation (QAM, APSK). The PAPR of constant envelope modulation is close to 1, while the PAPR of QAM-type modulation is larger.

[0057] S1-2, Frequency Domain Features (10 Dimensions)

[0058] Perform a Fast Fourier Transform on the signal to obtain its spectral representation, and extract the following features:

[0059] The mean, standard deviation, peak value, and normalized position of the dominant frequency of the power spectral density; the spectrum is divided into four frequency bands, and the energy proportion of each band is calculated; spectral flatness and peak-to-peak ratio.

[0060] Spectral flatness is defined as the ratio of the geometric mean to the arithmetic mean of the power spectrum. A value close to 1 indicates white noise characteristics, while a value much less than 1 indicates the presence of obvious spectral peaks. OFDM has higher spectral flatness, while single-carrier modulation has lower spectral flatness.

[0061] The 10th feature calculates the mean of the power spectral density, defined as:

[0062]

[0063] in This is the discrete Fourier transform of the signal.

[0064] The 11th feature is used to calculate the standard deviation of the power spectral density, quantifying the dispersion of spectral energy.

[0065]

[0066] The 12th dimension identifies the peak value of the power spectral density, corresponding to the strongest frequency component of the signal:

[0067]

[0068] The 13th dimension feature calculates the normalized position of the dominant frequency:

[0069]

[0070] The spectrum is divided into four equal bands, each occupying 25% of the total spectrum width.

[0071] The 14th dimension feature calculates the proportion of energy in the low-frequency band (0-25% of the spectrum) to the total energy:

[0072]

[0073] The energy distribution ratio in the low-frequency band (25%-50% of the spectrum range) in the 15th dimension feature analysis:

[0074]

[0075] The 16th dimension of the feature assessment evaluates the energy concentration in the mid-to-high frequency band (50%-75% of the spectrum):

[0076]

[0077] Energy contribution of the 17th dimension feature quantization high-frequency band (75%-100% spectrum range):

[0078]

[0079] The 18th dimension of the feature calculates spectral flatness, defined as the ratio of the geometric mean to the arithmetic mean of the power spectrum:

[0080]

[0081] A spectral flatness close to 1 indicates white noise characteristics, while a flatness much less than 1 indicates the presence of obvious spectral peaks. OFDM has a higher spectral flatness, while single-carrier modulation has a lower spectral flatness.

[0082] The 19th dimension of the feature calculates the peak-to-peak ratio, reflecting the degree of concentration of the spectrum:

[0083]

[0084] S1-3, Higher-order cumulative features (12 dimensions)

[0085] Higher-order cumulants are naturally immune to Gaussian noise. This property stems from the theoretical basis that the nth-order cumulants (n>2) of a Gaussian random variable are always zero.

[0086] First, the second, fourth, and sixth moments of the signal are calculated. The corresponding cumulative values ​​are then calculated from each moment and power normalization is performed. The extracted features include:

[0087] To eliminate the influence of signal power, the accumulated amount is normalized.

[0088] The 20th dimension feature calculates the normalized fourth-order cumulant. :

[0089]

[0090] This feature is useful for distinguishing BPSK ( ), QPSK ( ) and 8PSK ( It plays a decisive role.

[0091] The 21st dimension feature calculates the normalized fourth-order cumulant. :

[0092]

[0093] The 22nd dimension feature calculates the normalized fourth-order cumulant. :

[0094]

[0095] The 23rd dimension feature calculates the normalized sixth-order cumulant. :

[0096]

[0097] The 24th dimension feature calculates the normalized sixth-order cumulant. :

[0098]

[0099] The 25th feature calculates the ratio of fourth-order cumulants, used to distinguish between PSK and QAM modulation:

[0100]

[0101] in To prevent division by zero of small constants.

[0102] The 26th dimension feature calculates the ratio of the sixth-order to the fourth-order cumulants:

[0103]

[0104] The 27th feature calculates another set of high- and low-order cumulative ratios:

[0105]

[0106] 28th Dimension Feature Extraction The symbol is used to distinguish between BPSK (negative) and QPSK (positive):

[0107]

[0108] 29th Dimension Feature Extraction Symbols:

[0109]

[0110] The 30th dimension feature calculates the fourth-order cumulant of the amplitude sequence, reflecting the non-Gaussianity of the amplitude distribution:

[0111]

[0112] in , .

[0113] The 31st dimension of the feature calculates the amplitude variation coefficient, quantizing the relative stability of the amplitude:

[0114]

[0115] The coefficient of variation of constant envelope modulation is close to 0, while the coefficient of variation of non-constant envelope modulation is larger.

[0116] S1-4, Differential Phase Features (8 Dimensions)

[0117] Differential phase characteristics distinguish different phase modulation methods by analyzing the phase transition patterns of the signal, and play an important role in differentiating modulations such as BPSK, QPSK, and 8PSK.

[0118] First, calculate the phase difference sequence:

[0119]

[0120] in This represents the instantaneous phase of the signal. The phase difference has been normalized to... Interval:

[0121]

[0122] The 32nd dimension of the feature calculates the mean absolute value of the phase difference:

[0123]

[0124] The 33rd dimension of the feature calculates the standard deviation of the phase difference, reflecting the degree of dispersion of the phase transition:

[0125]

[0126] The 34th dimension of the feature statistics shows the proportion of phase transitions close to 0, which is used to identify continuous phase modulation (such as MSK, GMSK):

[0127]

[0128] The calculation of the 35th dimension feature is close to... The proportion of phase jumps is a typical characteristic of BPSK modulation:

[0129]

[0130] The 36th dimension feature statistics are close. The proportion of phase transitions reflects the characteristics of QPSK modulation:

[0131]

[0132] The 37th dimension of feature statistics is close. The proportion of phase transitions is an important characteristic of 8PSK and PI / 4-QPSK:

[0133]

[0134] The 38th feature is used to calculate the fourth-order cumulant of the phase difference sequence to determine the sharpness of the phase transition:

[0135]

[0136] The 39th feature is used to calculate the Shannon entropy of the phase jump distribution, quantifying the concentration of phase jumps:

[0137]

[0138] in The phase histogram of the first phase is shown. The normalized probability of each interval. The smaller the entropy value, the more concentrated the phase jumps are at a specific value; the larger the entropy value, the more uniform the distribution of phase jumps.

[0139] S1-5, I / Q component characteristics (6 dimensions)

[0140] I / Q component characteristics describe the relationship between the quadrature components of a complex signal, and the extracted features include:

[0141] The energy ratio and normalized energy difference between the I and Q components; the normalized mean of the Q component; the correlation coefficient between the I and Q components; the shift in the peak position of the I / Q cross-correlation function; and the proportion of samples where the amplitude of the Q component is significantly smaller than that of the I component.

[0142] The I / Q energy ratio is particularly effective for identifying ASK-type modulations because the energy of ASK modulation is mainly concentrated in the I component. The I / Q cross-correlation peak shift characteristic can be used to detect the half-symbol delay characteristics of OQPSK.

[0143] The 40th dimension feature calculates the ratio of the I component energy to the Q component energy:

[0144]

[0145] in To prevent the use of tiny constants for division by zero. For ASK modulation, this ratio is significantly greater than 1.

[0146] The 41st dimension of the feature calculates the normalized difference in I / Q energy:

[0147]

[0148] This characteristic reflects the degree of energy balance of the orthogonal components; the energy difference of ideal orthogonal modulation is close to 0.

[0149] The 42nd feature is used to calculate the normalized mean of the Q component:

[0150]

[0151] For ASK modulation, the normalized mean of the Q component is close to 0.

[0152] The 43rd feature is used to calculate the Pearson correlation coefficient between the I and Q components:

[0153]

[0154] This feature reflects the linear correlation strength of the orthogonal components; the correlation coefficient of ideal orthogonal modulation is close to 0.

[0155] The 44th dimension feature calculates the normalized offset of the peak position of the I / Q cross-correlation function, which is used to detect the half-symbol delay characteristic of OQPSK:

[0156]

[0157] in It is the I / Q cross-correlation function. This represents the maximum search delay. OQPSK modulation exhibits a significant non-zero offset.

[0158] The 45th dimension of feature statistics shows the percentage of samples where the amplitude of the Q component is less than 10% of the mean amplitude of the I component:

[0159]

[0160] This characteristic is a significant marker of ASK modulation, because the Q component of ASK should theoretically be 0.

[0161] S1-6, Cumulative Value Derived Features and PSK-Specific Features (10 Dimensions)

[0162] Cumulative derivative features describe the characteristics of modulated signals from the perspective of constellation diagram structure, including:

[0163] The 46th feature estimates the number of amplitude levels, reflecting the amplitude hierarchy of the constellation through the number of unique quantized values:

[0164]

[0165] in This represents the normalized amplitude. This feature can distinguish the order of M-ASK and M-QAM.

[0166] The 47th feature estimates the number of phase levels, reflecting the phase distribution of the constellation:

[0167]

[0168] This feature can distinguish between different orders of M-PSK modulation.

[0169] The 48th dimension feature calculates the comprehensive magnitude of the fourth-order cumulant:

[0170]

[0171] This feature reflects the degree to which the modulated signal deviates from a Gaussian distribution.

[0172] The 49th feature calculates the combined ratio of higher-order to lower-order cumulants:

[0173]

[0174] This feature reflects the relative strength of higher-order statistical properties.

[0175] PSK-specific features are specifically designed to distinguish easily confused modulation types within the PSK family, including 8PSK, QPSK, OQPSK, and PI / 4-QPSK. These modulations have similar cumulant characteristics and require more refined phase analysis for effective differentiation. These include:

[0176] Calculation of the 50th dimension feature The proportion of nearby phase transitions is judged using a more stringent threshold.

[0177]

[0178] 8PSK and PI / 4-QPSK have a large number of The proportion of QPSK and OQPSK is lower.

[0179] Calculation of the 51st dimension feature Percentage of phase jumps in the vicinity:

[0180]

[0181] QPSK and OQPSK The main focus is on transitions.

[0182] Calculation of the 52nd dimension feature Percentage of phase jumps in the vicinity:

[0183]

[0184] PI / 4-QPSK simultaneously possesses and The jump is a key feature that distinguishes it from 8PSK.

[0185] Calculation of the 53rd dimension feature Percentage of phase jumps in the vicinity:

[0186]

[0187] The 54th feature is used to calculate the uniformity of the phase transition, defined as the coefficient of variation of the absolute value of the phase difference:

[0188]

[0189] 8PSK has eight equidistant phase points, resulting in a more uniform distribution of phase transitions and a higher eigenvalue. QPSK, on ​​the other hand, has only four phase points, with phase transitions concentrated at specific values, resulting in a lower eigenvalue.

[0190] The 55th feature dimension is specifically used to detect the half-symbol delay characteristic of the I / Q components in OQPSK. The Q component of OQPSK has a delay of half a symbol period relative to the I component, which causes a shift in the peak position of the I / Q cross-correlation.

[0191]

[0192] in This represents the offset of the peak value of the I / Q cross-correlation relative to the center. This defines the search scope.

[0193] Specific calculation steps: 1. Calculate the I / Q cross-correlation function 2. Search for peak locations near the center. 3. Calculate and normalize the peak offset.

[0194] The eigenvalue of OQPSK is significantly greater than 0, while the eigenvalue of QPSK and 8PSK is close to 0.

[0195] S2, Neural Network Classification

[0196] like Figure 2 As shown, a fully connected neural network with layer-by-layer dimensionality reduction is used for classification.

[0197] The extracted 56-dimensional feature vectors are z-score normalized before being input into the network. The network uses multiple stacked fully connected layers, each followed by a batch normalization layer, a ReLU activation function, and a Dropout regularization layer.

[0198] The Dropout regularization layer employs a decreasing strategy, ensuring that the Dropout rate near the input layer is higher than that near the output layer. Stronger regularization is applied to the front layers of the network to prevent overfitting while maintaining the expressive power of the back layers, thus ensuring the accuracy of the final classification.

[0199] The output layer is mapped to the prediction scores of 32 modulation types through the Softmax function.

[0200] S3, Modulation Type Identification

[0201] The modulation type identification result is determined based on the predicted score, and the category with the highest score is taken as the final identification result.

[0202] Dataset construction strategy

[0203] The training dataset construction of this invention follows the core concept of "the training set emphasizes systematic coverage, and the test set emphasizes the ability to generalize to real-world scenarios".

[0204] For modulation types with symbol rates spanning multiple orders of magnitude, a log-space uniform sampling strategy is employed. For example... Figure 4 As shown, the symbol rate ranges from 200 SPS to 80 MSPS, spanning 5 to 6 orders of magnitude. Logarithmic sampling ensures uniform sampling point density within each order of magnitude, avoiding insufficient sampling in the low-speed range caused by linear sampling, and maintaining a balance in the model's learning weights for low-speed, medium-speed, and high-speed signals.

[0205] like Figure 3 As shown, the signal-to-noise ratio (SNR) values ​​in the test set and the training set are distributed using an interleaved strategy. The training set SNR values ​​are selected from -4 dB to 18 dB (12 points in total), spaced 2 dB apart; the test set SNR values ​​are selected from -5 dB to 17 dB (12 points in total), spaced 2 dB apart, completely interleaved with the training set. This strategy forces the model to interpolate and predict between the training set SNR values, rigorously examining the model's generalization ability.

[0206] Simulation Experiment

[0207] This embodiment constructs an identification system containing 32 modulation types, covering nine major categories:

[0208] Phase shift keying modulation types: BPSK, QPSK, 8PSK;

[0209] Quadrature amplitude modulation types: 16QAM, 32QAM, 64QAM, 160QAM;

[0210] Pulse amplitude modulation types: 2PAM, 4PAM;

[0211] Amplitude phase shift keying modulation types: 16APSK, 32APSK, 64APSK;

[0212] Differential modulation types: DBPSK, SDBPSK, DQPSK, SDQPSK;

[0213] Offset modulation types: OQPSK, PI / 4-QPSK;

[0214] Frequency modulation types: MSK, GMSK, 2FSK, 4FSK, FM;

[0215] Amplitude modulation types: AM, 2ASK, 4ASK, DSB, SSB;

[0216] Constellation mapping modulation types: 8ARY, 16ARY, 32ARY;

[0217] Multicarrier modulation type: OFDM.

[0218] The training set contains 768,000 samples, with 2,000 samples per modulation type per SNR. The test set contains 76,800 samples, with 200 samples per modulation type per SNR. The SNR of the test set is completely staggered with that of the training set to verify the model's generalization ability under unseen SNR conditions.

[0219] Experimental results are as follows Figures 5 to 10 As shown.

[0220] Figure 5 The overall recognition accuracy curves of the method of the present invention under different signal-to-noise ratio conditions are shown. Figure 6 As shown, the method of this invention achieves an overall recognition accuracy of 90.02% at 3 dB, 94.38% at 5 dB, 96.84% at 7 dB, and remains stable above 98% at 11 dB and above. Even under extremely low signal-to-noise ratio conditions of -5 dB, it still maintains a recognition accuracy of 58.95%, far exceeding the 3.125% (1 / 32) of random guessing.

[0221] Figures 6 to 9 The confusion matrices are shown for signal-to-noise ratios of -5 dB, 3 dB, 11 dB, and 17 dB, respectively. Figure 10 The display shows the recognition accuracy (%) of each modulation type under different signal-to-noise ratios. The experimental results fully verify the technical advantages of the multi-domain feature fusion method of this invention, specifically reflected in the following aspects:

[0222] First, perfect or near-perfect recognition of multiple modulation types is achieved. 4FSK and OFDM achieve 100% recognition accuracy across the entire tested signal-to-noise ratio range from -5 dB to 17 dB; 2FSK achieves 100% accuracy at -3 dB and above; SDBPSK, DSB, and SSB achieve over 99.5% accuracy at -3 dB and above; and 4PAM and BPSK achieve 94.5% and 99.5% accuracy, respectively, at -5 dB. The reliable recognition capability of these modulation types even at extremely low signal-to-noise ratios fully demonstrates the effective characterization capability of the multi-domain feature system for signals with different modulation mechanisms.

[0223] Second, higher-order cumulant characteristics significantly improve noise immunity. Thanks to the inherent immunity of higher-order cumulants to Gaussian noise, BPSK achieves 99.5% recognition accuracy at -5 dB, 4PAM reaches 94.5% at -5 dB, 2PAM reaches 98.5% at -3 dB, and GMSK reaches 94.5% at -5 dB. The excellent performance of these modulation types at extremely low signal-to-noise ratios validates the noise immunity advantages of higher-order cumulant characteristics.

[0224] Third, the frequency modulation category demonstrates excellent overall performance. 4FSK maintains 100% recognition accuracy across all signal-to-noise ratios, 2FSK achieves 100% accuracy at -3 dB and above, GMSK reaches 94.5% at -5 dB, and MSK reaches 93.5% at -1 dB. The frequency domain characteristics and differential phase characteristics effectively capture the essential characteristics of frequency modulation.

[0225] Fourth, the differential modulation recognition performance is stable and excellent. SDBPSK achieves 100% accuracy at -3 dB and above, SDQPSK at -1 dB and above, DQPSK at -1 dB, and DBPSK maintains above 90% accuracy across all tested signal-to-noise ratios. The differential phase characteristics effectively support accurate recognition of differential coded modulation.

[0226] Fifth, PSK-specific features effectively address the issue of intra-family confusion. OQPSK achieves 99.5% accuracy at 1 dB and 100% accuracy at 3 dB and above; 8PSK achieves 97% accuracy at 5 dB and 100% accuracy at 7 dB and above. The I / Q cross-correlation peak shift feature successfully captures the half-symbol delay characteristics of OQPSK, and the fine phase analysis feature effectively distinguishes similar modulation types within the PSK family.

[0227] Sixth, QAM family modulation enables accurate identification at high signal-to-noise ratios. 32QAM achieves 95% at 5 dB, 64QAM achieves 92.5% at 5 dB, and 16QAM achieves 93.5% at 7 dB. All three achieve over 99% at 11 dB and above. The high-order cumulant ratio characteristic effectively distinguishes the constellation structure differences of QAMs of different orders.

[0228] Seventh, the model's generalization ability was fully validated. The test set uses a signal-to-noise ratio distribution that is completely interleaved with the training set, requiring the model to perform interpolation predictions without knowing the signal-to-noise ratio. Figure 5 As shown, the recognition accuracy curve rises smoothly without significant fluctuations, proving that the method of the present invention has good generalization ability and can adapt to continuously changing channel conditions in practical applications.

[0229] Boundary signal-to-noise ratio analysis for each modulation type achieving 90% recognition accuracy

[0230] To quantify the reliable identification boundary of each modulation type, a 90% identification accuracy rate was used as the threshold to analyze the boundary signal-to-noise ratio of each modulation type:

[0231] The modulation types that can achieve a recognition accuracy of over 90% at -5 dB include: 4FSK, OFDM, 2FSK, BPSK, SDBPSK, 4PAM, GMSK, and DBPSK, a total of 8 types;

[0232] The modulation types that can achieve a recognition accuracy of over 90% at -3 dB include: 2PAM, SDQPSK, 8ARY, DSB, and SSB, a total of 5 types;

[0233] Modulation types that can achieve a recognition accuracy of over 90% above -1 dB include: QPSK, DQPSK, and MSK, a total of 3 types;

[0234] Modulation types that can achieve a recognition accuracy of over 90% at a depth of 1 dB include: OQPSK and 16ARY, a total of 2 types;

[0235] The modulation types that can achieve a recognition accuracy of over 90% at 3 dB or higher include: 2ASK and 16OQAM, a total of 2 types;

[0236] The modulation types that can achieve a recognition accuracy of over 90% above 7 dB include: 16QAM, 16APSK, 32APSK, PI / 4QPSK, 8PSK, 32QAM, 64QAM, and 4ASK, totaling 8 types.

[0237] The modulation types that can achieve a recognition accuracy of over 90% at 9 dB include: 64APSK and AM, a total of 2 types.

[0238] In summary, 28 out of the 32 modulation types can achieve a recognition accuracy of over 90% at 9 dB or below, accounting for 87.5%. Among them, 18 types can achieve a recognition accuracy of over 90% at -1 dB or below, accounting for 56.25%, demonstrating the excellent recognition capability of the method of this invention under low signal-to-noise ratio conditions.

[0239] In summary, the communication signal modulation identification method based on multi-domain feature fusion proposed in this invention achieved excellent identification performance under test conditions of 32 modulation types. The overall identification accuracy reached 94.38% at a signal-to-noise ratio (SNR) of 5 dB, 96.84% at 7 dB, and remained stable above 98% at 11 dB and above. 4FSK and OFDM achieved 100% identification accuracy across the entire tested SNR range, and the identification accuracy for the 18 modulation types reached above 90% at -1 dB and below. The experimental results fully verify the effectiveness of the 56-dimensional multi-domain feature system of this invention, the noise resistance advantage of high-order cumulant features, and the ability of PSK-specific features to distinguish easily confused modulation types within a family, providing a reliable technical solution for communication signal modulation identification in complex electromagnetic environments.

Claims

1. A method for identifying modulation of communication signals based on multi-domain feature fusion using deep learning, characterized in that, Includes the following steps: S1. From the received communication signal, different signal processing methods are used to extract time-domain statistical features, frequency-domain features, higher-order cumulant features, differential phase features, I / Q component characteristic features, and PSK-specific features to obtain multi-domain feature vectors. Based on the multi-domain feature vectors and their corresponding modulation type labels, training datasets and test datasets are constructed. For modulation types with symbol rates spanning multiple orders of magnitude, a logarithmic space uniform sampling strategy is used to generate symbol rate distributions, ensuring balanced sampling point density across orders of magnitude. The signal-to-noise ratio (SNR) values ​​in the test dataset and the training dataset are obtained using an interleaved distribution strategy, forcing the model to interpolate and predict the SNR values ​​in the training dataset to rigorously verify the model's generalization ability. The time-domain statistical features are represented by statistics of signal amplitude and complex components, used to characterize the basic energy distribution characteristics of the signal and the differences between constant envelope and non-constant envelope modulation. The frequency domain features are represented by statistics of signal power spectral density and segmented energy distribution, which are used to reflect the differences in spectral structure of different modulation methods in the frequency domain; The higher-order cumulants feature is represented by normalized fourth-order and sixth-order cumulants and their derived features, and is used to distinguish different modulation types by utilizing the natural immunity of higher-order cumulants to Gaussian noise. The differential phase characteristics are represented by the statistics of the phase difference sequence and the proportion of specific phase jumps, and are used to distinguish modulation methods with different phase jump modes. The I / Q component characteristics are represented by the energy ratio, correlation, and cross-correlation shift characteristics of the orthogonal components, and are used to identify ASK-type modulation and detect the half-symbol delay characteristics of OQPSK. The PSK-specific features are represented by fine phase transition ratio, phase uniformity, and I / Q delay detection features, which are used to distinguish easily confused modulation types within the PSK family; S2. After standardizing the extracted multi-domain feature vectors, input them into a layer-by-layer dimensionality-reducing fully connected neural network. After multi-layer feature abstraction, the prediction scores for each modulation type are output. The layer-by-layer dimensionality-reducing fully connected neural network includes a feature abstraction module composed of multiple stacked fully connected layers. Each fully connected layer is followed by a batch normalization layer, an activation function layer, and a Dropout regularization layer. The Dropout regularization layer adopts a decreasing strategy, making the Dropout rate closer to the input layer higher than the Dropout rate closer to the output layer, so as to maintain the expression accuracy of the output layer while suppressing overfitting. S3. Determine the modulation type identification result based on the predicted score to classify the modulation method of the communication signal.

2. The method according to claim 1, characterized in that, In step S1, the extraction process of time-domain statistical features is specifically as follows: Calculate the mean, standard deviation, variance, maximum and minimum values ​​of the signal amplitude; calculate the mean and standard deviation of the real and imaginary parts of the complex signal; calculate the peak-to-average power ratio, which is defined as the ratio of peak power to average power and is used to distinguish between constant envelope modulation and non-constant envelope modulation.

3. The method according to claim 1, characterized in that, In step S1, the extraction process of frequency domain features is specifically as follows: The signal is subjected to Fast Fourier Transform to obtain a spectral representation; the mean, standard deviation, peak value and normalized position of the dominant frequency of the power spectral density are calculated; the spectrum is divided into multiple frequency bands and the energy proportion of each frequency band is calculated; the spectral flatness and peak value ratio are calculated.

4. The method according to claim 1, characterized in that, In step S1, the extraction process of higher-order cumulant features is specifically as follows: Calculate the second, fourth, and sixth moments of the signal; calculate the corresponding cumulative quantities from each moment and perform power normalization; calculate the ratio characteristics between different order cumulative quantities. Extract the sign characteristics of the cumulative amount; Calculate the higher-order statistics and coefficient of variation of the signal amplitude sequence.

5. The method according to claim 1, characterized in that, In step S1, the extraction process of differential phase features is specifically as follows: Calculate the instantaneous phase of the signal and obtain the phase difference sequence; normalize the phase difference sequence; calculate the mean and standard deviation of the phase difference; and calculate the percentage of phase jumps close to zero, π, ±π / 2, ±π / 4, etc., respectively. Calculate the fourth-order cumulant and the entropy of the phase jump distribution of the phase difference sequence.

6. The method according to claim 1, characterized in that, In step S1, the extraction process of I / Q component characteristic features is specifically as follows: Calculate the energy ratio of the I component to the Q component and the normalized energy difference; Calculate the normalized mean of the Q component; calculate the correlation coefficient between the I and Q components; Calculate the peak position offset of the I / Q cross-correlation function; The percentage of samples in which the amplitude of the Q component is significantly smaller than that of the I component is counted.

7. The method according to claim 1, characterized in that, In step S1, the extraction process of PSK-specific features is as follows: The proportion of phase jumps near π / 4, π / 2, 3π / 4, and π was statistically analyzed using fine thresholds; the uniformity of the phase jump distribution was calculated. The normalized offset of the I / Q cross-correlation peak relative to the center is calculated to detect the half-symbol delay characteristics of the I / Q components in OQPSK modulation.

8. The method according to claim 1, characterized in that, The modulation types cover nine major categories, including phase shift keying modulation, quadrature amplitude modulation, pulse amplitude modulation, amplitude phase shift keying modulation, differential modulation, offset modulation, frequency modulation, amplitude modulation, and multicarrier modulation, achieving systematic coverage of the mainstream modulation methods of modern communication systems.

9. The method according to claim 8, characterized in that, The modulation types specifically include the following 32 types: Phase shift keying modulation types: BPSK, QPSK, 8PSK; Quadrature amplitude modulation types: 16QAM, 32QAM, 64QAM, 160QAM; Pulse amplitude modulation types: 2PAM, 4PAM; Amplitude phase shift keying modulation types: 16APSK, 32APSK, 64APSK; Differential modulation types: DBPSK, SDBPSK, DQPSK, SDQPSK; Offset modulation types: OQPSK, PI / 4-QPSK; Frequency modulation types: MSK, GMSK, 2FSK, 4FSK, FM; Amplitude modulation types: AM, 2ASK, 4ASK, DSB, SSB; Constellation mapping modulation types: 8ARY, 16ARY, 32ARY; Multicarrier modulation type: OFDM.

Images