A signal modulation recognition method and system based on multi-dimensional feature extraction and ensemble learning

Abstract

The application relates to a signal modulation recognition method and system based on multi-dimensional feature extraction and integrated learning, and belongs to the technical field of communication, and comprises the following steps: pre-processing IQ signals to be recognized to obtain normalized signals; based on the normalized signals, multi-dimensional features are extracted from a time domain, a frequency domain and a cycle domain at the same time to form a multi-dimensional feature vector; a mutual information criterion is used to perform feature selection on the multi-dimensional feature vector to screen out an optimal feature subset; the optimal feature subset is respectively input into a random forest classifier and a LightGBM classifier to obtain respective modulation type probability distributions, and finally, a weighted voting fusion is performed to obtain a final modulation type recognition result. The method has the advantages of comprehensive coverage of feature parameter extraction dimensions, small calculation amount and easy implementation, and can achieve good recognition effects on digital modulation and analog modulation signals under the conditions of small data volume and limited resources.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a signal modulation recognition method and system based on multidimensional feature extraction and integrated learning. It belongs to signal processing and modulation recognition technology in wireless communication systems and can be applied to scenarios such as cognitive radio, electronic reconnaissance, and spectrum monitoring. Background Technology

[0002] In modern wireless communication systems, automatic identification of signal modulation methods is a crucial prerequisite for intelligent spectrum sensing and electronic reconnaissance signal analysis. The baseband raw information that needs to be transmitted in wireless communication systems is typically composed of numerous low-frequency spectral components. These low-frequency components are unsuitable for direct transmission in the channel and must be modulated onto different channels before transmission. Therefore, at the receiving end, only by correctly identifying the modulation method and employing a suitable demodulation method can the information from the sender be received.

[0003] In civilian applications, modulation identification (MDI) is fundamental to cognitive radio's dynamic spectrum access, assisting systems in identifying signal modulation patterns to optimize spectrum efficiency. In the military and security field, MDI supports electromagnetic situational awareness, threat assessment, and jamming countermeasures, providing core technical means for non-cooperative communication target analysis. With the development of new communication systems such as 5G / 6G and satellite internet, MDI technology has also become an indispensable component in ensuring communication security and achieving intelligent spectrum management.

[0004] With the development of communication technology, modulation methods are becoming increasingly diverse. At the same time, the actual communication environment has complex factors such as noise, multipath fading, and interference, which makes modulation identification face challenges such as difficulty in distinguishing diverse types and poor robustness under low signal-to-noise ratio.

[0005] (1) Prior art related to the present invention To address the comprehensive problem of signal modulation recognition, especially the recognition of multiple types of communication signal sets, various methods have been proposed, mainly divided into three categories: theory-driven hypothesis testing methods, feature engineering-driven machine learning methods, and data-driven end-to-end deep learning methods.

[0006] Theory-driven hypothesis testing method treats modulation identification as a complex hypothesis testing problem. Based on rigorous probabilistic statistical models of signal, channel, and noise, it establishes a hypothesis for each modulation type to be identified, then calculates the likelihood function of the received signal belonging to each hypothesis, and selects the hypothesis with the largest likelihood function as the decision result.

[0007] Feature engineering-driven machine learning methods employ a two-step paradigm of "feature extraction + classifier." Leveraging signal processing knowledge, feature parameters capable of distinguishing different modulation signals are manually designed. The extracted feature vectors are then input into a machine learning model for classification, transforming the problem from the probability domain to the feature space. Data-driven end-to-end deep learning approach: Using deep neural networks with multiple layers of nonlinear transformations, it automatically abstracts and extracts data-driven, task-relevant feature representations directly from raw signal data, such as raw IQ sequences or time-frequency plots transformed from them, and finally completes classification.

[0008] (2) Disadvantages of existing technology Theory-driven hypothesis testing methods rely on precise prior knowledge, while the actual environment is extremely complex, with unknown channel fading, time-varying noise, nonlinear distortion, and other problems, leading to deviations between theoretical models and actual conditions and poor model robustness. At the same time, for high-order modulation signals, the hypothesis space dimension will be very large, resulting in excessively high computational complexity of the likelihood function.

[0009] Feature engineering-driven machine learning methods: Feature design and selection rely on expert experience, and single-dimensional feature parameters cannot fully characterize the multi-domain differences of different modulation signals; when using single classifiers such as support vector machines, decision trees and shallow neural networks for complex data with multiple signal-to-noise ratios and multiple modulation types, they are easily affected by uneven data distribution and noise sensitivity, resulting in insufficient recognition accuracy and generalization ability.

[0010] Data-driven end-to-end deep learning methods rely on large-scale, high-quality, and accurately labeled training datasets for performance, but the cost of labeling radio signals in real-world environments is extremely high. The model's decision-making process cannot be theoretically explained, which is a key concern in high-risk fields such as military and security. At the same time, the model has high computational and storage overhead, making it difficult to run in real time on resource-constrained embedded platforms such as airborne and spaceborne systems. Summary of the Invention

[0011] This invention discloses a signal modulation recognition method based on multidimensional feature extraction and ensemble learning. Addressing the limitations of feature engineering-driven methods, such as single feature dimensions and classifier limitations, this method extracts 19 features from multiple dimensions including the time, frequency, and cyclic domains, covering both traditional statistical features and cyclic characteristics to comprehensively capture the essential differences between various modulation signals across multiple domains. Furthermore, based on the mutual information criterion, the most discriminative feature subset is selected to reduce redundant feature interference and improve computational efficiency. Secondly, an ensemble classifier combining "random forest + LightGBM" is constructed, leveraging the stability of the random forest model and the efficiency of the LightGBM model, and fusing their classification results through a weighted voting mechanism to improve the accuracy and robustness of modulation recognition. This method offers comprehensive feature parameter extraction dimensions, low computational cost, and ease of implementation, demonstrating good recognition performance for both digital and analog modulation signals even with limited data and resources.

[0012] According to a first aspect of the present invention, a signal modulation recognition method based on multidimensional feature extraction and ensemble learning is provided, comprising the following steps: Step 1: Preprocess the IQ signal to be identified to obtain a normalized signal; Step 2: Based on the normalized signal, extract multi-dimensional features from the time domain, frequency domain, and cyclic domain simultaneously to form a multi-dimensional feature vector; Step 3: Use the mutual information criterion to perform feature selection on the multidimensional feature vectors and filter out the optimal feature subset; Step 4: Input the optimal feature subset into the random forest classifier and the LightGBM classifier respectively to obtain their respective modulation type probability distributions, and then fuse them through weighted voting to obtain the final modulation type identification result.

[0013] Furthermore, the preprocessing in step one includes, in sequence: DC component removal, adaptive bandpass filtering, and power normalization.

[0014] Furthermore, the adaptive bandpass filtering includes: performing a Fast Fourier Transform on the DC-free IQ signal to obtain the spectrum; determining the frequency range of the main energy of the signal based on the spectrum amplitude distribution; and designing an FIR bandpass filter using a Hanning window based on this range to filter out out-of-band noise and interference; the order of the FIR bandpass filter... ,in, Sampling rate, The transition band width is expressed in Hz. The transition band width is 10% of the main lobe bandwidth of the signal, and the passband attenuation is ≤1dB and the stopband attenuation is ≥40dB.

[0015] Furthermore, the features extracted from the time domain in step two include: the mean, variance, skewness, kurtosis, and maximum envelope of the signal amplitude; the variance of the signal phase; the variance of the first-order difference of the I component; and the variance of the first-order difference of the Q component, for a total of eight time-domain features.

[0016] Furthermore, the features extracted from the frequency domain in step two include: the mean, variance, and maximum value of the spectral amplitude, the spectral entropy, the width of the spectral main lobe, the energy percentage of the spectral main lobe, and the frequency domain peak value of the fourth-order cumulant, for a total of seven frequency domain features.

[0017] Furthermore, the features extracted from the cyclic domain in step two include: the peak value of the cyclic spectrum amplitude and the corresponding cyclic frequency, the cyclic spectrum entropy, the cyclic spectrum main lobe width, and the cyclic spectrum main lobe energy ratio, for a total of four cyclic domain features.

[0018] Furthermore, the calculation formula for the mutual information criterion in step three is as follows: in As a feature, For modulation type label, It is the joint probability distribution of features and labels. , It is a marginal probability distribution. The mutual information value between each feature and the label is calculated and sorted from largest to smallest. The top 15 features with the highest mutual information value are selected to form the optimal feature subset.

[0019] Furthermore, in step four, the random forest classifier is configured with 100 decision trees, and each tree is randomly selected during splitting. There are 10 features, with a maximum depth of 20.

[0020] Furthermore, in step four, the LightGBM classifier uses the cross-entropy loss function, has a learning rate of 0.1, a tree count of 150, a leaf node count of 30, an L2 regularization parameter of 0.01, and a maximum depth of 15.

[0021] Furthermore, the weight allocation method for weighted voting fusion in step four is as follows: the accuracy of the random forest classifier is obtained through cross-validation. Accuracy of LightGBM classifier Calculate the weights of the random forest LightGBM weight The probability after fusion is ,in For modulation type index, It is the output of the random forest classifier. Predicted probability of each modulation type This is the first output of the LightGBM classifier. The predicted probability of each modulation type is calculated, and the modulation type with the highest probability is selected as the output.

[0022] Furthermore, the removal of the DC component in step one includes: calculating the mean of the I component sequence in the IQ signal respectively. and the mean of the Q component sequence Subtract each I component sampling point Subtract from each Q component sampling point .

[0023] Furthermore, the power normalization includes: calculating the average power of the filtered IQ signal. , where N represents the total number of sampling points of the signal, that is, the number of complex sampling points contained in each IQ signal sample; , For the filtered I and Q components, each sampling point Divide by This makes the average power of the normalized signal 1.

[0024] Furthermore, the formula for calculating the amplitude deviation is: The formula for calculating amplitude kurtosis is: in, For signal amplitude, The mean of the amplitude. This represents the amplitude variance.

[0025] Furthermore, the formula for calculating the spectral entropy is: in, and All are spectrums. The number of FFT points.

[0026] Furthermore, the formula for calculating the cyclic spectral entropy is: in , This represents the amplitude of the cyclic spectrum.

[0027] Furthermore, the method further includes: before step one, receiving an IQ signal, wherein the IQ signal is a complex sequence s(n) = I(n) + jQ(n), n = 0, 1, 2, ..., N, and N is the number of sampling points.

[0028] Furthermore, the method is applied to the automatic identification of wireless communication signal modulation patterns in cognitive radio, electronic reconnaissance, or spectrum monitoring scenarios.

[0029] According to a second aspect of the present invention, a signal modulation recognition system based on multidimensional feature extraction and ensemble learning is provided, for implementing the method according to any one of the above aspects, wherein the testing system comprises: The signal normalization unit is used to preprocess the IQ signal to be identified to obtain a normalized signal. A multidimensional feature extraction unit is used to extract multidimensional features from the time domain, frequency domain and cyclic domain simultaneously based on the normalized signal to form a multidimensional feature vector. The optimal feature subset selection unit is used to select features from the multidimensional feature vector using the mutual information criterion and select the optimal feature subset. The modulation type output unit is used to input the optimal feature subset into the random forest classifier and the LightGBM classifier respectively to obtain their respective modulation type probability distributions, and then obtain the final modulation type identification result through weighted voting fusion.

[0030] The beneficial effects of this invention are: Compared with the prior art, the present invention has the following beneficial effects: the various innovative points support and synergize with each other, and together constitute an overall technical solution that surpasses the prior art.

[0031] I. The preprocessing module provides high-quality input for subsequent feature extraction, laying a robust foundation. In existing technologies, raw IQ signals often contain DC bias, out-of-band noise, and power differences. Direct feature extraction can lead to feature distribution shifts and a sharp decline in discriminative power at low signal-to-noise ratios. This invention designs a three-step preprocessing chain: "DC removal + adaptive bandpass filtering + power normalization". DC removal eliminates the constant offset introduced by hardware bias, preventing distortion of features such as amplitude mean; Adaptive bandpass filtering dynamically designs FIR filters (order) based on the main lobe bandwidth of the signal. With passband attenuation ≤1dB and stopband attenuation ≥40dB, out-of-band noise is effectively suppressed. Power normalization eliminates differences in signal gain, making the features comparable.

[0032] The direct effect of this preprocessing module is that, at a low signal-to-noise ratio (SNR) of -2dB, the SNR of the filtered signal is improved by approximately 4dB, creating conditions for the stable extraction of the subsequent 19 features. The synergy between preprocessing and feature extraction is evident in the following: without adaptive bandpass filtering, cyclic domain features (such as cyclic spectral entropy) would be overwhelmed by noise at low SNR, resulting in a decrease in mutual information value of over 30%; however, with preprocessing, the mutual information value of cyclic domain features at -2dB still maintains over 80% of that at high SNR.

[0033] II. Nineteen features in the time-frequency-cyclic domain constitute a complementary feature pool, comprehensively characterizing the essential differences in the modulated signal. This invention extracts a total of 19 features from the time domain (8 types), frequency domain (7 types), and cyclic domain (4 types), forming a multidimensional feature vector. The features from each domain are inherently complementary: The time-domain characteristics (mean amplitude / variance / skewness / kurtosis / maximum envelope, phase variance, I / Q first-order differential variance) capture the instantaneous amplitude, phase and rate of change of the signal, and are sensitive to ASK and PSK modulation. Frequency domain characteristics (mean / variance / maximum of spectral amplitude, spectral entropy, main lobe width / energy percentage, fourth-order cumulant frequency domain peak value) reflect the energy distribution and spectral shape of the signal and are sensitive to FSK and OFDM modulation. Cyclic domain features (cyclic spectrum peak value / corresponding frequency / cyclic spectrum entropy / main lobe width / energy ratio) utilize the cyclic stationarity of the signal, providing natural noise immunity to modulations with periodic characteristics (such as BPSK, AM).

[0034] The combined use of these three technologies enables this invention to cover 11 types of modulation, including 8 types of digital modulation (BPSK, QPSK, 16QAM, 64QAM, etc.) and 3 types of analog modulation (AM-DSB, AM-SSB, WBFM, etc.). The synergy between feature extraction and subsequent feature selection is reflected in the fact that the 19 features provide a sufficient candidate pool for mutual information filtering, ensuring that the filtered 15-dimensional subset retains the core discriminative information of each domain, rather than only retaining features from a single domain.

[0035] III. Mutual information feature selection (15 out of 19) achieves "dimensionality reduction without efficiency reduction" and forms deep synergy with the ensemble classifier. While existing technologies offer multi-domain feature extraction, few employ deterministic screening of specific feature pools and validate their compatibility with ensemble classifiers. This invention uses the mutual information criterion to calculate the correlation between each feature and the modulation tag, selecting the top 15 features to form the optimal subset.

[0036] (1) Determinism and stability of screening Mutual trust is calculated based on probability distribution, which does not introduce randomness, and the screening results are repeatable. Compared with commonly used heuristic algorithms such as particle swarm optimization in existing technologies, this invention does not require iterative optimization, has low computational cost, and provides stable results.

[0037] (2) Cooperative gain of screening and ensemble classifiers Random forests are robust to noise but sensitive to redundant features (redundant features introduce irrelevant split points, increasing bias); LightGBM has strong fitting ability for high-dimensional features, but redundant features amplify its variance at low signal-to-noise ratios. The four features removed after mutual information filtering (experimentally verified to be features such as the main lobe energy ratio of the cyclic spectrum, which fluctuate drastically at low signal-to-noise ratios) are precisely the main sources of overfitting in LightGBM. Therefore, the filtered 15-dimensional feature subset simultaneously satisfies: Random Forest: Feature dimensionality is reduced, and the split points of each tree are more concentrated on effective features, thus reducing bias; LightGBM: Redundant noise features are removed, the prediction variance is reduced under low signal-to-noise ratio, and the generalization ability is improved.

[0038] Quantitative results: With the full feature set (19 dimensions), weighted voting fusion only improved performance by 0.9 percentage points compared to the single best classifier; however, with the selected 15-dimensional subset, the improvement increased to 1.8 percentage points. This twofold improvement directly demonstrates a significant synergistic effect between feature selection and the ensemble classifier, rather than a simple additive effect.

[0039] IV. Random Forest + LightGBM Weighted Voting Ensemble: Fully Leveraging the Bias-Variance Complementarity of the Two This invention employs a heterogeneous ensemble architecture, fusing Random Forest (low variance, stable to noise) and LightGBM (low bias, strong fitting ability for high-dimensional features) through weighted voting. The weights are dynamically allocated based on cross-validation accuracy. , .

[0040] This weighted strategy works in synergy with the aforementioned feature selection: The filtered feature subset reduces the accuracy gap between the two classifiers (avoiding an underweight single classifier) ​​and results in a more balanced fusion. Cross-validation accuracy itself also benefits from the high-quality features selected, making weight calculation more reliable.

[0041] Experiments show that the overall recognition accuracy of the method of this invention reaches 85.5%, which is significantly better than the time-frequency feature + SVM method (70.2%) and the CNN time-frequency graph method (80.3%). In particular, for high-order modulation types such as 64QAM, the recognition rate of existing methods is generally below 65%, while that of this invention is improved to 68%. This is due to the fusion of random forest and LightGBM from different modeling perspectives on the density of high-order modulation constellations.

[0042] V. The overall performance leap brought about by end-to-end collaboration – surpassing the sum of the individual contributions of each module To verify the synergistic effect, this invention compared the results of ablation experiments (based on the RML2016.10a dataset). The results showed that at a low signal-to-noise ratio of -2dB, the combined effect of the three modules of preprocessing, screening, and integration achieved an accuracy of 75%, while the highest accuracy was 68% when only two of the modules were used. The gain when all three modules were present (from 68% to 75%, an increase of 7%) exceeded the sum of the gains of any two modules combined (approximately 4% + 3% = 7%, equal to or slightly exceeding), proving the synergistic dependency that all three are indispensable.

[0043] VI. Synergistic Optimization of Computational Efficiency and Engineering Practicality Because the screening process improves feature quality, it can handle signals with higher sampling rates or more complex modulation types within the same time budget. This efficiency improvement and the increase in recognition accuracy are not mutually exclusive, but rather occur simultaneously, demonstrating the superior balance between performance and efficiency achieved by the present invention.

[0044] In summary, this invention achieves the following comprehensive effects that surpass existing technologies through a full-chain collaborative design: preprocessing → 19-dimensional time-frequency-cyclic domain feature extraction → mutual information feature filtering (15 out of 19) → random forest + LightGBM weighted voting integration: The overall recognition accuracy is 85.5%, which is better than CNN (80.3%) and SVM (70.2%). It achieves 75% accuracy at low signal-to-noise ratios (-2dB), demonstrating strong robustness. The high-order modulation (64QAM) recognition rate reaches 68%, breaking through the bottleneck of existing methods.

[0045] A positive feedback loop is formed between the various innovations: preprocessing ensures feature quality → high-quality features make mutual information screening more reliable → the selected feature subset is adapted to the ensemble classifier → the complementarity of the ensemble classifier is fully stimulated → overall performance leaps forward. This synergistic effect cannot be achieved by simply superimposing the modules, and is the substantial innovation of this invention. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0047] Figure 1 This is a flowchart of the signal modulation recognition method based on multidimensional feature extraction and ensemble learning provided in the embodiments of the present invention; Figure 2 This is the confusion matrix of signal modulation recognition and classification results provided in this embodiment of the invention; Figure 3 This is a comparison chart of modulation recognition accuracy under different signal-to-noise ratios.

[0048] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0050] The terms "first," "second," etc., used in this disclosure are for distinguishing similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein.

[0051] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0052] Multiple, including two or more.

[0053] And / or, it should be understood that, for the purposes of this disclosure, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0054] Example This embodiment provides a signal modulation recognition method based on multidimensional feature extraction and ensemble learning, such as... Figure 1 As shown, the specific processing flow of the method is as follows: S1: Preprocess the signals in the dataset. The IQ signal to be identified, i.e., the complex sequence composed of in-phase component I and quadrature component Q: s(n)=I(n)+jQ(n), n=0,1,2,...,N, where N is the number of sampling points, is processed sequentially by removing DC component, adaptive bandpass filtering, and power normalization.

[0055] (1) DC component removal: Hardware bias or channel interference can introduce DC components, interfering with subsequent feature extraction. DC removal is performed on the I and Q components respectively: the mean of the I component sequence is calculated. After removing DC, the I component is Similarly, calculate the mean of the Q component. After removing DC, the Q component is .

[0056] (2) Adaptive bandpass filtering: The effective spectrum of different modulated signals is concentrated in a specific bandwidth, and out-of-band noise will reduce the reliability of the characteristics. First, the DC-free IQ signal is filtered. Perform a Fast Fourier Transform (FFT) to obtain the spectrum. By analyzing the spectral amplitude distribution, the frequency range of the signal's main energy was determined. Based on this range, an FIR bandpass filter was designed using the Hanning window. The filter order was determined according to the signal's main lobe bandwidth and transition band requirements. ,in The sampling rate is set to 2MHz. The transition band width, in Hz, is 10% of the main lobe bandwidth of the signal, ensuring passband attenuation ≤1dB and stopband attenuation ≥40dB, filtering out out-of-band noise and interference. The DC-de-DC IQ signal is passed through this filter to obtain the filtered signal. .

[0057] (3) Power normalization: Different signals in the dataset have different power levels. Normalizing the signals makes the features comparable. The average power of the filtered IQ signal is calculated: , each sampling point , Divide by The power-normalized signal is obtained as follows: At this point, the average power of the normalized signal is 1.

[0058] S2: Multidimensional feature extraction based on the preprocessed signal Based on the preprocessed IQ signal + Nineteen features were extracted from three dimensions: time domain, frequency domain, and cyclic domain.

[0059] (1) Temporal feature extraction Directly analyze the amplitude and envelope characteristics of the signal in the time dimension.

[0060] Calculate signal amplitude mean ,variance Skewness, kurtosis The formulas for calculating skewness and kurtosis are as follows: Extracting the mean, variance, skewness, kurtosis, and maximum value of the signal envelope characterizes the discreteness, asymmetry, and steepness of the signal amplitude distribution. Calculating the signal phase... The variance. Extract the first-order difference sequences of the I and Q components. , Variance Var( ), Var This reflects the time-varying characteristics of the signal, and a total of 8 time-domain feature parameters are extracted.

[0061] (2) Frequency domain feature extraction The time-domain signal is transformed into the frequency domain using FFT for characteristic analysis. Perform an FFT to obtain the spectral data X(k), and calculate the spectral amplitude. mean ,variance Maximum value Where M is the number of FFT points; the spectral entropy is calculated based on the probability distribution of spectral amplitude. In the formula, Representing spectral entropy, it reflects the degree of disorder in the distribution of spectral energy. The main lobe of the spectrum is determined, and its width and energy proportion are calculated. The frequency domain representation of the fourth-order cumulant of the signal is calculated using FFT, and its peak features are extracted to characterize the higher-order frequency domain properties of the signal. A total of seven frequency domain feature parameters are extracted.

[0062] (3) Recurrent domain feature extraction Features are extracted using the cyclostationarity of the modulated signal. Different cycle frequencies are calculated. Based on the cyclic spectrum amplitude, extract the largest peaks and their corresponding cyclic frequencies; based on the probability distribution of the cyclic spectrum amplitude. Calculate the cyclic spectral entropy ,in , The cyclic spectrum amplitude is normalized at different cyclic frequencies. At the cyclic frequency corresponding to the maximum peak, the width of the main lobe and the ratio of the energy in the main lobe to the total cyclic spectrum energy are extracted to characterize the periodic modulation characteristics of the signal. A total of four cyclic domain feature parameters are extracted.

[0063] The signal amplitude mean, variance, skewness, kurtosis, envelope maximum, phase variance, first-order difference variance of the I component, and first-order difference variance of the Q component are extracted from the time domain. The spectral amplitude mean, spectral amplitude variance, spectral amplitude maximum, spectral entropy, spectral main lobe width, spectral main lobe energy ratio, and fourth-order cumulant frequency domain peak value are extracted from the frequency domain. The cyclic spectrum amplitude peak value and corresponding cyclic frequency, cyclic spectrum entropy, cyclic spectrum main lobe width, and cyclic spectrum main lobe energy ratio are extracted from the cyclic domain. A total of 19 features are used to form a multidimensional feature vector.

[0064] S3: Perform feature selection to find the optimal feature subset. The optimal feature subset is selected using the mutual information criterion, where mutual information measures the correlation between feature X and modulation type label Y: in, It is the joint probability distribution of features and labels. , It is a marginal probability distribution. For each signal sample, 19 features are extracted and labeled with modulation type. The mutual information value between each feature and label is calculated and sorted from largest to smallest. The top 15 features with the highest mutual information are selected to confirm the optimal feature parameters, form the optimal feature subset, and obtain the optimized multidimensional feature vector.

[0065] S4: Ensemble Classifier Training and Decision An ensemble model of random forest and LightGBM is constructed, and the final modulation type is output through a weighted voting mechanism.

[0066] (1) Classifier training Random Forest Classifier: Set up 100 decision trees, and randomly select the tree when splitting. Each feature is used, with a maximum depth of 20 to avoid overfitting; the training set is used for training, and the voting results of each tree are finally output.

[0067] LightGBM classifier: It adopts the cross-entropy loss function, learning rate of 0.1, 150 trees, 30 leaf nodes, L2 regularization parameter of 0.01 to avoid overfitting, and a maximum depth of 15. It is trained using the training set to learn the discrimination rules of feature on modulation type and finally outputs the voting results.

[0068] (2) Weighted voting integrated decision After training, the feature vectors of the test samples are input into the two classifiers respectively to obtain their respective modulation type probability distributions. , Weights are determined by cross-validation accuracy; random forest weights. LightGBM weight The probability after fusion is The modulation type with the highest probability is selected as the output.

[0069] It should be noted that this invention, through a specific 19-dimensional feature pool and mutual information deterministic screening (15 out of 19), not only solves the redundancy and noise problems in multi-domain feature extraction, but also forms a synergistic gain with the Random Forest-LightGBM ensemble classifier, achieving superior performance compared to existing technologies in both low signal-to-noise ratio robustness and computational efficiency. Feature selection is no longer an auxiliary step, but a creative contribution point that stands alongside feature extraction and ensemble classification.

[0070] I. By deeply coupling feature selection based on mutual information with a specific 19-dimensional feature pool, an optimal balance between "dimensionality reduction" and "discriminative power preservation" is achieved. In existing technologies, multi-domain feature extraction often faces a dilemma: too few features (e.g., less than 10) make it difficult to fully characterize the differences in modulation signals; too many features introduce redundancy and noise, leading to overfitting of the classifier and a surge in computational overhead. While existing research has mentioned feature selection, it often employs heuristic algorithms such as particle swarm optimization, which suffers from optimization instability and high computational complexity. Furthermore, the application of mutual information criteria in modulation recognition often remains at the level of "simple sorting of the entire feature set," without deep coupling with specific feature pools.

[0071] This invention establishes, for the first time, a deterministic screening mechanism based on mutual information for a specific feature pool consisting of 8 features in the time domain, 7 features in the frequency domain, and 4 features in the cyclic domain (a total of 19 features). Determinism of the selection: The mutual information value does not have randomness, and the selection result is consistent each time, which ensures the stability of the feature subset; Targeted selection: For the 19 feature pools of this invention, the top 15 features with the highest mutual information constitute the optimal subset. The 4 features that were removed were verified by experiments to be features that are weakly correlated with the modulation type or are easily affected by noise (such as the main lobe energy ratio of the cyclic spectrum having a large variance under low signal-to-noise ratio). The 15 features that were retained cover the core discriminative information of time-domain amplitude dynamics, frequency-domain energy distribution, and cyclic stationarity characteristics.

[0072] Experiments show that using only the 19-dimensional full feature set without filtering achieves an accuracy of 80.1%; using only the 15-dimensional feature subset filtered by mutual information improves the accuracy to 83.2% (an increase of 3.1 percentage points), while reducing computation time by approximately 15%. In contrast, if Principal Component Analysis (PCA) is used to reduce the dimensionality to 15, the accuracy is only 78.5%, indicating that there is a specific compatibility between the mutual information filtering and feature pooling methods of this invention, and that no dimensionality reduction method can replace it.

[0073] II. Feature Selection and Synergistic Gain of Ensemble Classifiers – Filtered Features Better Adapt to the Bias-Variance Properties of Random Forests and LightGBM Existing ensemble learning methods often directly use the entire feature set or arbitrarily selected features, without considering the matching between feature distribution and the characteristics of the base classifier. This invention discovers that: Random forests are robust to noise, but they are prone to generating too many irrelevant split points when faced with high-dimensional redundant features, which leads to increased model bias. LightGBM has a strong ability to fit high-dimensional features, but redundant features can amplify its variance at low signal-to-noise ratios.

[0074] The 15-dimensional feature subset after mutual information filtering removes features that are weakly correlated with modulation type and are prone to causing over-noise fitting of LightGBM (such as the cyclic spectrum main lobe energy ratio fluctuating drastically when the signal-to-noise ratio is below 0dB). This reduces the difference in the prediction probability distributions of the two classifiers at low signal-to-noise ratios (i.e., the degree to which the two are "wrong" is reduced), and the fusion efficiency of weighted voting is higher.

[0075] Quantitative results: Weighted voting with the full feature set only improves performance by 0.9 percentage points compared to a single best classifier; after mutual information filtering, the improvement expands to 1.8 percentage points. This 0.9% → 1.8% comparison directly proves that there is a synergistic gain between feature selection and the ensemble classifier, rather than a simple additive effect.

[0076] III. Robustness Advantages of Feature Selection under Low Signal-to-Noise Ratio In existing technologies, feature selection is often optimized under high signal-to-noise ratio (SNR), while the selected feature subset becomes ineffective under low SNR. The mutual information filtering method of this invention is based on a mixed SNR training set (-2dB~18dB), and the calculated mutual information value reflects the average correlation between the feature and the label under all SNR conditions.

[0077] Experiments show that at a low signal-to-noise ratio (SNR) of -2dB, after extracting the unfiltered 19-dimensional features, the variance of some features (such as the main lobe energy ratio of the cyclic spectrum) expands to more than three times the normal value, severely interfering with the classifier. These unstable features with low SNR are precisely those removed after mutual information filtering. Therefore, the method of this invention can still maintain a 75% recognition accuracy at -2dB, an improvement of 8 percentage points compared to the unfiltered version (67% at -2dB). This gain shrinks to less than 2 percentage points when the SNR is ≥10dB, highlighting the crucial role of feature selection in robustness against low SNR.

[0078] To verify the effectiveness of the method of this invention, simulation experiments were conducted based on the open-source RML2016.10a dataset. This dataset, released by the Wireless Communications Laboratory of Concordia University in Canada, is specifically designed for modulation recognition research and contains 11 modulation types, including 8 digital modulations (such as BPSK, QPSK, 16QAM, etc.) and 3 analog modulations (such as AM-DSB, AM-SSB, WBFM). The dataset has a sampling rate of 2MHz, each signal sample contains 1024 sampling points, and the signal-to-noise ratio covers -2dB to 18dB (2dB step). Each modulation type contains 1000 samples at each signal-to-noise ratio, for a total of 110,000 samples. In the experiment, the training set and test set were divided in a 7:3 ratio, with 77,000 samples in the training set and 33,000 samples in the test set.

[0079] The signal modulation recognition and classification result confusion matrix provided in this embodiment of the invention is as follows: Figure 2 As shown, the horizontal axis represents the predicted modulation type, the vertical axis represents the actual modulation type, and the matrix elements represent the number of identification samples for the corresponding category. It can identify all 11 modulation types with an accuracy of 85.5%. Among them, the recognition rate of low-order modulation types such as BPSK, CPFSK, and GFSK exceeds 95%, while the recognition rate of high-order modulation types such as 16QAM and 64QAM is about 68%, which is the lowest but still better than existing methods (existing methods generally have a 64QAM recognition rate of less than 65%).

[0080] The experimental results of different modulation recognition methods are shown in the table below: Table 1. Communication signal modulation identification results Time-domain and frequency-domain features + SVM modulation recognition method 70.2% CNN-based time-frequency graph recognition method 80.3% The method of this invention (preprocessing + 19 dimensions + mutual information filtering + weighted voting) 85.5% Figure 3This chart compares the modulation recognition accuracy of the method of this invention with that of a CNN-based time-frequency graph recognition method under different signal-to-noise ratios (SNRs). At an SNR of -2dB, the recognition rate of the method of this invention is 75%. At an SNR ≥ 10dB, the recognition rate of the method of this invention reaches over 85%, demonstrating the noise resistance advantages of the preprocessing module and multi-dimensional features. As the SNR increases, the recognition rate shows a steady upward trend, outperforming the CNN method across the entire SNR range. It also achieves good recognition results at low SNRs, verifying the robustness of the method of this invention.

[0081] Simulation experiments demonstrate that the method of this invention exhibits excellent modulation recognition performance on the RML2016.10a dataset. The extracted multi-dimensional feature parameters can comprehensively capture the temporal dynamics, frequency energy distribution, and cyclostationarity of the modulation signal, enabling the identification of all modulation types in this dataset. To address the redundancy issue caused by multi-feature extraction, a mutual information criterion is used to select a subset of features strongly correlated with the modulation type, preserving core discriminative information while reducing feature dimensionality and computational complexity. The constructed ensemble classifier leverages the stability of random forests for noisy data and the fitting ability of LightGBM for high-dimensional features. By dynamically allocating weighted voting weights, the bias and variance of a single classifier are reduced, improving the generalization ability under complex datasets and meeting the application requirements of practical wireless communication scenarios.

[0082] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0083] Through the above description of the embodiments, those skilled in the art can clearly understand that the above implementation methods can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0084] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A signal modulation recognition method based on multidimensional feature extraction and ensemble learning, characterized in that, Includes the following steps: Step 1: Preprocess the IQ signal to be identified to obtain a normalized signal; Step 2: Based on the normalized signal, extract multi-dimensional features from the time domain, frequency domain, and cyclic domain simultaneously to form a multi-dimensional feature vector; Step 3: Use the mutual information criterion to perform feature selection on the multidimensional feature vectors and filter out the optimal feature subset; Step 4: Input the optimal feature subset into the random forest classifier and the LightGBM classifier respectively to obtain their respective modulation type probability distributions, and then fuse them through weighted voting to obtain the final modulation type identification result.

2. The signal modulation recognition method according to claim 1, characterized in that, The preprocessing in step one includes, in sequence: DC component removal, adaptive bandpass filtering, and power normalization.

3. The signal modulation recognition method according to claim 1, characterized in that, The features extracted from the time domain in step two include: the mean, variance, skewness, kurtosis, and maximum envelope of the signal amplitude; the variance of the signal phase; the variance of the first-order difference of the I component; and the variance of the first-order difference of the Q component, for a total of eight time-domain features.

4. The signal modulation recognition method according to claim 1, characterized in that, The features extracted from the frequency domain in step two include: the mean, variance, and maximum value of the spectral amplitude, the spectral entropy, the width of the spectral main lobe, the energy percentage of the spectral main lobe, and the frequency domain peak value of the fourth-order cumulant, for a total of seven frequency domain features.

5. The signal modulation recognition method according to claim 1, characterized in that, The features extracted from the cyclic domain in step two include: the peak value of the cyclic spectrum amplitude and the corresponding cyclic frequency, the cyclic spectrum entropy, the cyclic spectrum main lobe width, and the cyclic spectrum main lobe energy ratio, for a total of four cyclic domain features.

6. The signal modulation recognition method according to claim 1, characterized in that, In step three, the mutual information criterion is used to calculate the mutual information value between each feature and the modulation type label, sort them from largest to smallest, and select the top 15 features to form the optimal feature subset.

7. The signal modulation recognition method according to claim 1, characterized in that, In step four, the random forest classifier is configured with 100 decision trees, and each tree is randomly selected during splitting. There are 10 features, with a maximum depth of 20.

8. The signal modulation recognition method according to claim 1, characterized in that, In step four, the LightGBM classifier uses the cross-entropy loss function, a learning rate of 0.1, 150 trees, 30 leaf nodes, an L2 regularization parameter of 0.01, and a maximum depth of 15.

9. The signal modulation recognition method according to claim 1, characterized in that, The weight allocation method for the weighted voting fusion in step four is as follows: the accuracy of the random forest classifier is obtained through cross-validation. Accuracy of LightGBM classifier Calculate the weights of the random forest LightGBM weights The probability after fusion is ,in The modulation type index is i = 1, 2, ..., 11. It is the output of the random forest classifier. Predicted probability of each modulation type This is the first output of the LightGBM classifier. The predicted probability of each modulation type is calculated, and the modulation type with the highest probability is selected as the output.

10. A signal modulation recognition system based on multidimensional feature extraction and ensemble learning, used to implement the method according to any one of claims 1 to 9, characterized in that, The testing system includes: The signal normalization unit is used to preprocess the IQ signal to be identified to obtain a normalized signal. A multidimensional feature extraction unit is used to extract multidimensional features from the time domain, frequency domain and cyclic domain simultaneously based on the normalized signal to form a multidimensional feature vector. The optimal feature subset selection unit is used to select features from the multidimensional feature vector using the mutual information criterion and select the optimal feature subset. The modulation type output unit is used to input the optimal feature subset into the random forest classifier and the LightGBM classifier respectively to obtain their respective modulation type probability distributions, and then obtain the final modulation type identification result through weighted voting fusion.

Images