Signal modulation recognition method based on multi-modal feature fusion and SNR perception

Abstract

The application discloses a signal modulation recognition method based on multi-modal feature fusion and SNR perception, and belongs to the technical field of wireless communication.The method of the application is as follows: a wireless signal to be recognized is acquired, and complex baseband discretization processing is performed on the wireless signal to obtain a discrete signal; a multi-branch deep neural network structure is constructed by fusing time domain IQ features, frequency domain STFT features and fused features composed of physical statistics, phase difference and SNR; an SNR embedding mechanism is adopted to encode and integrate the signal-to-noise ratio information of the input signal into the feature fusion process, so that the model can perceive the current signal-to-noise ratio condition and affect the model training process through a loss optimization mechanism, adaptive modeling of signal features under different signal-to-noise ratio conditions is realized, and the accuracy and robustness of modulation recognition are improved.The application comprehensively utilizes time domain structure information, frequency domain energy distribution information and statistical features of the signal, so that the model can still maintain high recognition performance under complex electromagnetic environment and low signal-to-noise ratio conditions.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and in particular relates to a signal modulation recognition method based on deep learning, which is applicable to modulation recognition in scenarios such as cognitive radio, spectrum monitoring, and communication reconnaissance. Background Technology

[0002] With the rapid development of wireless communication technology and the increasing scarcity of spectrum resources, cognitive radio technology has become a key means to solve the problem of efficient spectrum resource utilization. In cognitive radio systems, modulation identification, as the core function of spectrum sensing, can identify the modulation mode of the received signal without prior information, providing a fundamental guarantee for subsequent signal demodulation and information recovery.

[0003] Traditional modulation identification methods are mainly divided into two categories: likelihood-based methods and feature-based methods. Likelihood-based methods theoretically have the best identification performance. This method constructs the probability density function of the signal and uses hypothesis testing theory to determine the modulation type. However, this type of method requires precise knowledge of the channel model and noise statistics, resulting in high computational complexity and high sensitivity to model mismatch, making it difficult to meet real-time requirements in practical applications. Feature-based methods distinguish different modulation types by manually designing high-order cumulants, cyclic spectra, and instantaneous features. Although the computational complexity is lower, feature design is highly dependent on expert experience, and identification performance significantly decreases under complex channel conditions such as multipath fading, frequency offset, and phase noise (see Dobre OA, Abdi A, Bar-Ness Y, et al. Survey of automatic modulation classification techniques: classical approaches and new trends[J]. IET communications, 2007,1(2): 137-156). Furthermore, different features have varying abilities to distinguish different modulation types, lacking a unified feature selection standard.

[0004] In recent years, deep learning methods have made significant progress in the field of signal modulation recognition. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have been widely applied to modulation recognition tasks (see O'Shea TJ, Corgan J, Clancy T C. Convolutional radio modulation recognition networks[C] / / International conference on engineering applications of neural networks. Cham: Springer International Publishing, 2016: 213-226). However, most existing deep learning-based modulation recognition methods only use time-domain data or frequency-domain features, failing to fully utilize the joint time-frequency information of the signal, resulting in a significant decrease in recognition accuracy under low signal-to-noise ratio (SNR) conditions. Furthermore, deep learning models, as "black boxes," lack effective utilization of the physical nature of the signal, limiting the model's interpretability and generalization ability. In addition, the SNR dynamically changes in real-world communication environments, and existing methods use the same processing strategy for different SNR conditions, failing to adaptively adjust feature extraction and classification strategies. Therefore, there is an urgent need for a modulation recognition method that can efficiently integrate multimodal features, adapt to changes in SNR, and possess physical interpretability. Summary of the Invention

[0005] To address the problems of insufficient utilization of multimodal information, limited modeling ability of signals to physical characteristics, and unstable recognition performance under different signal-to-noise ratio (SNR) conditions in existing technologies, the purpose of this invention is to provide a signal modulation recognition method based on multimodal feature fusion and SNR perception. By constructing a multi-branch deep neural network structure that integrates time-domain IQ features, frequency-domain STFT features, and fused features composed of physical statistics, phase difference, and SNR, and introducing an SNR embedding mechanism, adaptive modeling of signal features under different SNR conditions is achieved, thereby improving the accuracy and robustness of modulation recognition.

[0006] This invention comprehensively utilizes the temporal structure information, frequency domain spectral amplitude information, and physical statistics, phase difference, and signal-to-noise ratio (SNR) information of signals, enabling the model to maintain high recognition performance even in complex electromagnetic environments and under low SNR conditions. Simultaneously, by introducing physically meaningful feature representations, the interpretability and generalization ability of the model are improved. This invention is applicable to multiple fields such as cognitive radio, spectrum monitoring, electronic countermeasures, communication reconnaissance, wireless network optimization, intelligent spectrum management, and signal analysis.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] This invention discloses a signal modulation recognition method based on multimodal feature fusion and SNR perception. The method establishes a forward propagation model for the signal modulation recognition network; constructs a signal modulation recognition network based on multimodal feature fusion and SNR perception, utilizing time-domain IQ features, frequency-domain STFT features, physical statistical features, phase difference features, and SNR perception information; creates a training dataset; constructs a loss function based on probability distribution; trains the signal modulation recognition network; and uses the trained network for signal modulation recognition.

[0009] The signal modulation recognition method based on multimodal feature fusion and SNR sensing disclosed in this invention includes the following steps:

[0010] Step 1: Acquire the wireless signal to be identified, perform complex baseband representation and discretization processing to obtain a discrete signal.

[0011] Step 2: Input the discrete signal obtained in Step 1 into the tensor As input features, temporal features are extracted to extract temporal structure information.

[0012] Step 3: Perform STFT transformation on the discrete signal obtained in Step 1 to extract its time-frequency domain features. Take the amplitude value to obtain the spectral amplitude, normalize the spectral amplitude, and then input it into the frequency domain feature extraction network for frequency domain feature extraction.

[0013] Step 4: Perform statistical analysis on the discrete signal obtained in Step 1 to extract physical statistical features. Extract the phase of the discrete signal, expand it, and calculate the phase difference between adjacent phases. Input the phase difference sequence into a feature mapping network to obtain phase features. Map the input signal-to-noise ratio (SNR) to discrete intervals to obtain discrete indices, and then map these to embedding vectors to obtain the SNR features. Concatenate these three features to obtain the fused features.

[0014] Step 5: Concatenate the time-domain features, frequency-domain features, and fused features obtained in steps 2, 3, and 4 to obtain... The discriminant score vector is then obtained through nonlinear mapping. .

[0015] Step 6: The dataset consists of discrete signals and their corresponding modulation tags. For each discrete signal sample, after processing in steps 1–5, the final feature representation is obtained. Each modulation category is assigned a corresponding modulation category label. A dataset is then constructed for model training.

[0016] Step 7: Convert the discriminant score vector obtained in Step 5 into a single vector. Transform it into a probability distribution. Construct a loss function based on the probability distribution.

[0017] Step 8: Optimize the trainable parameters in the modulation recognition network using the minimized loss function obtained in Step 7.

[0018] Step 9: Iteratively train the network from step 8 until a trained model is obtained.

[0019] Step 10: Input the signal to be identified into the model trained in step 9 to obtain the corresponding modulation category prediction result.

[0020] Furthermore, the method for obtaining physical characteristics described in step 4 is as follows:

[0021] Input tensor to the discrete signal obtained in step 1 Perform statistical analysis and extract physical statistical features.

[0022]

[0023] in, This represents the physical feature extraction function, which includes the extraction of physical features from amplitude statistics, energy distribution, and higher-order statistics.

[0024] Furthermore, the method for obtaining the phase features described in step 4 is as follows:

[0025] Phase extraction is performed on a discrete signal, and after phase expansion, the instantaneous phase of the discrete signal is expressed as:

[0026]

[0027] in, Indicates the sampling point index. Represents the in-phase component of a discrete signal. Represents the orthogonal components of a discrete signal.

[0028] Calculate the phase difference between adjacent sampling points based on the instantaneous phase:

[0029]

[0030] in, Indicates the length of the sampling sequence.

[0031] This yields the phase difference sequence.

[0032]

[0033] The phase difference sequence is normalized to obtain

[0034]

[0035] Will By inputting a one-dimensional convolutional or fully connected mapping function for feature extraction, a fixed-dimensional phase feature representation is obtained:

[0036]

[0037] in, This represents the phase feature mapping function.

[0038] Furthermore, the method for obtaining the signal-to-noise ratio embedding features in step 4 is as follows:

[0039] A signal-to-noise ratio (SNR) embedding mechanism is introduced to enhance the model's adaptability to different channel conditions. Let the SNR of the discrete signal be... ,Will Mapping to discrete intervals:

[0040]

[0041] in, This represents the minimum signal-to-noise ratio. Indicates the width of the interval.

[0042] Discrete index Mapped to an embedding vector:

[0043]

[0044] in, This indicates an embedded mapping function.

[0045] Furthermore, the SNR sensing fusion method described in step 4 is as follows:

[0046] Physical features, phase features, and signal-to-noise ratio (SNR) are embedded and fused to obtain a fused feature representation. Specifically, after aligning and vectorizing each feature, a fused feature vector is formed through a concatenation operation:

[0047]

[0048] Furthermore, the method for constructing the model training data in step 6 is as follows:

[0049] The dataset consists of discrete signals and their corresponding modulation labels. For each discrete signal sample, after processing in steps 1–5, the final feature representation is obtained. Each sample corresponds to a modulation category label. Let the total number of training samples be... The training dataset can then be represented as:

[0050]

[0051] in, Indicates the first Feature representation of a sample This represents the corresponding real category label.

[0052] Furthermore, the category labels are represented in one-hot encoded form:

[0053]

[0054] in, This indicates the total number of modulation categories.

[0055] To improve the model's generalization ability under different signal-to-noise ratio conditions, a balanced distribution of samples with different signal-to-noise ratios can be ensured when constructing the dataset, thereby avoiding bias during model training.

[0056] Step 7: Construction of the loss function.

[0057] Furthermore, step 7 is implemented as follows:

[0058] For each sample, for the first... For each input sample, the classification model in step 5 outputs a corresponding discriminant score vector:

[0059]

[0060] The scores are converted into a probability distribution using the Softmax function.

[0061]

[0062] in, Indicates the first The sample at the th Class-based discrimination scores Indicates the first The sample at the th Class-based discrimination scores.

[0063] Introducing a label smoothing mechanism, the true labels are represented as:

[0064]

[0065] in, This represents the smoothing coefficient.

[0066] Based on the above definition, construct the cross-entropy loss function:

[0067]

[0068] The loss function optimizes the model parameters by minimizing the difference between the predicted distribution and the true distribution.

[0069] Furthermore, step 8 is implemented as follows:

[0070] After determining the loss function, a gradient-based optimization method is used to optimize and train the trainable parameters in the signal modulation recognition network model. The trainable parameters include the time-domain feature extraction branch, the frequency-domain feature extraction branch, the physical feature and phase feature fusion branch, and the network parameters in the classification model. Let the set of all trainable parameters in the network be denoted as . By minimizing the loss function Continuously update trainable parameters:

[0071]

[0072] in, Indicates the learning rate. This represents the gradient of the loss function with respect to the trainable parameters.

[0073] During training, gradients are calculated using the backpropagation algorithm, and parameters are updated in conjunction with optimization strategies, allowing the model to gradually converge to its optimal state. Furthermore, to improve training stability, regularization methods can be used to constrain the model, such as weight decay to suppress excessively large parameters, thereby improving the model's generalization performance.

[0074] Furthermore, step 9 is implemented as follows:

[0075] Based on the dataset constructed in step 6 and the optimization methods defined in steps 7–8, the model is trained iteratively. In each training iteration, a batch of sample data is input, the corresponding loss is calculated, and the model parameters are updated. As the training process progresses, the loss function gradually decreases, and the model performance gradually improves.

[0076] Let the number of training rounds be... The training process can then be represented as:

[0077]

[0078] During training, performance metrics on the validation set are monitored to determine if the model has converged. Training stops when the model performance stabilizes or reaches preset conditions, and the final model parameters are obtained.

[0079] Furthermore, step 10 is implemented as follows:

[0080] After the model training is completed, the signal to be identified is input into the trained model to obtain the corresponding modulation category prediction result.

[0081] For any input signal, its characteristic representation obtained after processing in steps 1–5 is denoted as: The model outputs the probability distributions for each category in step 7. The final identification result is determined in the following way:

[0082]

[0083] in, This indicates the predicted modulation category.

[0084] The trained model can be used to identify the modulation method of unknown signals.

[0085] Beneficial effects:

[0086] 1. The signal modulation recognition method based on multimodal feature fusion and SNR sensing disclosed in this invention extracts time-domain and time-frequency domain features through IQ branch and STFT branch respectively, making full use of the complementarity of the two features in terms of information dimensions. IQ features accurately characterize the instantaneous amplitude and phase changes of the signal, while STFT features reflect the frequency energy distribution of the signal. Information fusion is achieved through multimodal feature splicing and nonlinear mapping, which significantly improves the recognition accuracy under low signal-to-noise ratio conditions.

[0087] 2. The signal modulation recognition method based on multimodal feature fusion and SNR perception disclosed in this invention introduces phase feature and physical statistical feature extraction modules, embedding physically interpretable features such as amplitude distribution statistics, phase difference characteristics, and higher-order cumulants of the signal into the deep learning model, thereby enhancing the model's ability to understand the essential laws of the signal, improving the model's generalization ability and interpretability, and helping to distinguish modulation methods with similar time-frequency characteristics but different physical essences.

[0088] 3. The signal modulation recognition method based on multimodal feature fusion and SNR perception disclosed in this invention adopts a signal-to-noise ratio (SNR) embedding mechanism to encode the SNR information of the input signal into an embedding vector and integrate it into the feature fusion process. This enables the model to perceive the current SNR conditions and influence the model training process through SNR embedding and loss weighting mechanisms, effectively solving the problem of large performance fluctuations of traditional methods under different SNR conditions and achieving stable recognition performance over a wide SNR range.

[0089] 4. The signal modulation recognition method based on multimodal feature fusion and SNR perception disclosed in this invention constructs a multimodal feature fusion and SNR perception network using an end-to-end joint training method. By optimizing the design of the multi-task loss function, the feature extraction and classification decision-making processes are constrained simultaneously. Combined with batch normalization and residual connection techniques, the model convergence is accelerated, ensuring recognition accuracy while meeting the real-time requirements of actual communication systems. Attached Figure Description

[0090] Figure 1 This is an overall structural diagram of the signal modulation recognition system based on multimodal feature fusion and SNR perception described in this invention;

[0091] Figure 2 This is a comparison chart of the recognition accuracy of different methods in Embodiment 1 of the present invention on the RadioML2022 dataset;

[0092] Figure 3 This is a graph showing the accuracy of the method of the present invention in identifying various modulation schemes on the RadioML2022 dataset in Embodiment 1 of the present invention.

[0093] Figure 4 This is a confusion matrix diagram of the method of the present invention on the RadioML2022 dataset in Embodiment 1 of the present invention. Detailed Implementation

[0094] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0095] Example 1:

[0096] This embodiment applies the signal modulation recognition method based on multimodal feature fusion and SNR perception to the RadioML2022 dataset. This dataset contains 10 modulation schemes (BPSK, QPSK, 8PSK, QAM16, QAM64, PAM4, CPFSK, GFSK, AM-DSB, WBFM), with a signal length of 128 sampling points and a signal-to-noise ratio range of -20dB to 20dB. In this embodiment, the actual sample range used is -10dB to 20dB.

[0097] like Figure 1As shown in the figure, the signal modulation recognition method based on multimodal feature fusion and SNR sensing disclosed in this embodiment has the following specific implementation steps:

[0098] Step 1: Acquisition and preprocessing of signal data.

[0099] In this embodiment, the publicly available dataset RadioML 2022 is used as the source of experimental data. This dataset contains wireless communication signals with various digital modulation schemes and covers multiple signal-to-noise ratio conditions, making it suitable for modulation recognition task modeling. Each signal sample in the dataset is represented in complex baseband form, and its continuous-time expression is as follows:

[0100]

[0101] in, Indicates the in-phase component of the signal. Represents the orthogonal components of the signal. It represents the imaginary unit.

[0102] After discrete sampling of the input signal, a length of [length missing] is obtained. Discrete sequences:

[0103]

[0104] in, Indicates the sampling point index.

[0105] Furthermore, the discrete signal is reconstructed into a multi-channel real-valued representation for subsequent neural network processing, yielding the input tensor:

[0106]

[0107] in, This indicates the number of channels, with each of the two channels corresponding to an in-phase component. Orthogonal components .

[0108] Step 2: Temporal feature extraction.

[0109] To fully characterize the multidimensional properties of the signal, multimodal features are constructed from three aspects: time domain, frequency domain, and SNR sensing and feature fusion. First, from the time domain perspective, the discrete signal obtained in step 1 is used as the basic input feature to preserve the amplitude and phase information of the original signal.

[0110] Specifically, the discrete I / Q signal obtained in step 1 is used as the input tensor. As input features, the tensor is fed into a temporal feature extraction module composed of a multi-layer convolutional neural network. This module includes convolutional layers, batch normalization layers, and non-linear activation functions, and incorporates residual structures to enhance feature representation. During feature extraction, pooling operations are used to progressively compress the feature dimension and expand the receptive field, thereby extracting local and global temporal structure information. After the above processing, the temporal feature representation is obtained:

[0111]

[0112] in, This represents the time-domain feature extraction function.

[0113] Step 3: Frequency domain feature extraction.

[0114] In order to characterize the energy distribution of the input signal in the frequency domain, a frequency domain feature extraction branch is constructed to perform time-frequency analysis on the discrete signal obtained in step 1.

[0115] Specifically, a Short-Time Fourier Transform (STFT) is performed on the discrete signal to obtain its time-frequency characteristics. In this embodiment, a length of [length missing] is selected. The Fourier transform window, and set the frame shift to... Its expression is:

[0116]

[0117] in, Represents the window function. Indicates frequency index, Indicates the time frame index.

[0118] The amplitude of the frequency representation is obtained by taking the amplitude value:

[0119]

[0120] The normalized data is then input into a frequency domain feature extraction network. This network extracts spectral structure features through convolution and pooling operations, and can incorporate a spectral enhancement mechanism to highlight key frequency components. The final frequency domain feature representation is obtained as follows:

[0121]

[0122] in, This represents the frequency domain feature extraction function.

[0123] Step 4: Physical features, phase features, and SNR sensing fusion.

[0124] In addition to time and frequency domain features, to further introduce the physical meaning information of the signal to be identified, physical feature and phase feature extraction branches are constructed, and perceptual modeling is performed in combination with signal-to-noise ratio information.

[0125] First, perform statistical analysis on the discrete signal obtained in step 1 to extract physical statistical features.

[0126]

[0127] in, This represents the physical feature extraction function, which includes the extraction of physical features from amplitude statistics, energy distribution, and higher-order statistics.

[0128] Furthermore, to characterize the phase change pattern of the signal to be identified, phase extraction is performed on the discrete signal. To avoid phase jumps, phase unwrap processing is used, and the instantaneous phase of the discrete signal is defined as:

[0129]

[0130] Based on this, the phase is expanded and the phase difference between adjacent sampling points is calculated:

[0131]

[0132] This yields the phase difference sequence.

[0133]

[0134] Furthermore, the phase difference sequence is normalized to obtain...

[0135]

[0136] Will By inputting a one-dimensional convolutional or fully connected mapping function for feature extraction, a fixed-dimensional phase feature representation is obtained:

[0137]

[0138] in, This represents the phase feature mapping function.

[0139] Considering the significant impact of signal-to-noise ratio (SNR) on signal characteristic distribution, an SNR embedding mechanism is introduced to enhance the model's adaptability to different channel conditions. Let the SNR of the discrete signal be... Map it to a discrete interval:

[0140]

[0141] in, This represents the minimum signal-to-noise ratio. Indicates the width of the interval.

[0142] Discrete index Mapped to an embedding vector:

[0143]

[0144] in, This indicates an embedded mapping function.

[0145] Physical features, phase features, and signal-to-noise ratio (SNR) are embedded and fused to obtain a fused feature representation. Specifically, after aligning and vectorizing each feature, a fused feature vector is formed through a concatenation operation:

[0146]

[0147] Through the above fusion process, a unified expression of the multidimensional features of discrete signals is achieved, thereby improving the model's ability to distinguish different modulation methods.

[0148] Step 5: Multimodal feature fusion and classification decision.

[0149] After obtaining the time-domain features, frequency-domain features, and fused features, a unified model is performed on the multimodal features. First, the features from each branch are concatenated:

[0150]

[0151] Based on the fusion, feature dimensionality reduction and information reconstruction are achieved through nonlinear mapping to obtain the discrimination score vector:

[0152]

[0153] in, and They represent the learnable parameters, This represents a non-linear activation function.

[0154] Through the above steps, effective fusion of multimodal features and determination of signal modulation type are achieved.

[0155] Step 6: Model training data construction.

[0156] Based on the feature representations and label information obtained in steps 1–5, a dataset for model training is constructed. In this embodiment, the data comes from the RadioML 2022 dataset. For each signal sample in the dataset, after processing in steps 1–5, the final feature representation is obtained. Each sample corresponds to a modulation category label. Let the total number of training samples be... The training dataset can then be represented as:

[0157]

[0158] in, Indicates the first Feature representation of a sample This represents the corresponding real category label.

[0159] Furthermore, the category labels are represented in one-hot encoded form:

[0160]

[0161] in, This indicates the total number of modulation categories.

[0162] To improve the model's generalization ability under different signal-to-noise ratio conditions, a balanced distribution of samples with different signal-to-noise ratios can be ensured when constructing the dataset, thereby avoiding bias during model training.

[0163] Step 7: Construction of the loss function.

[0164] To achieve effective classification of modulation categories, a loss function based on probability distribution is constructed.

[0165] First, regarding the first For each input sample, the classification model in step 5 outputs a corresponding discriminant score vector:

[0166]

[0167] The scores are converted into a probability distribution using the Softmax function.

[0168]

[0169] in, Indicates the first The sample at the th Class-based discrimination scores Indicates the first The sample at the th Class-based discrimination scores.

[0170] To alleviate model overfitting and improve generalization ability, a label smoothing mechanism is introduced, representing the true labels as follows:

[0171]

[0172] in, This represents the smoothing coefficient.

[0173] Based on the above definition, construct the cross-entropy loss function:

[0174]

[0175] By minimizing this loss function, the model's predicted probability distribution gradually approximates the true distribution, thereby achieving effective differentiation of modulation categories.

[0176] Step 8: Model parameter optimization.

[0177] After determining the loss function, a gradient-based optimization method is used to optimize and train the trainable parameters in the signal modulation recognition network model. The trainable parameters include the time-domain feature extraction branch, the frequency-domain feature extraction branch, the physical feature and phase feature fusion branch, and the network parameters in the classification model. Let the set of all trainable parameters in the network be denoted as . By minimizing the loss function Continuously update trainable parameters:

[0178]

[0179] in, Indicates the learning rate. This represents the gradient of the loss function with respect to the trainable parameters.

[0180] During training, gradients are calculated using the backpropagation algorithm, and parameters are updated in conjunction with optimization strategies, allowing the model to gradually converge to its optimal state. Furthermore, to improve training stability, regularization methods can be used to constrain the model, such as weight decay to suppress excessively large parameters, thereby improving the model's generalization performance.

[0181] Step 9: Model training and convergence determination.

[0182] Based on the dataset constructed in step 6 and the optimization methods defined in steps 7–8, the model is trained iteratively. In each training iteration, a batch of sample data is input, the corresponding loss is calculated, and the model parameters are updated. As the training process progresses, the loss function gradually decreases, and the model performance gradually improves.

[0183] Let the number of training rounds be... The training process can then be represented as:

[0184]

[0185] During training, performance metrics on the validation set are monitored to determine if the model has converged. Training stops when the model performance stabilizes or reaches preset conditions, and the final model parameters are obtained.

[0186] Step 10: Signal modulation identification.

[0187] After the model training is completed, the signal to be identified is input into the trained model to obtain the corresponding modulation category prediction result.

[0188] For any input signal, its characteristic representation obtained after processing in steps 1–5 is denoted as: The model outputs the probability distributions for each category in step 7. The final identification result is determined in the following way:

[0189]

[0190] in, This indicates the predicted modulation category.

[0191] Through the above process, the modulation method of unknown signals can be identified.

[0192] This embodiment will be tested on the RadioML2022 dataset and compared with existing methods to illustrate the effectiveness of the present invention.

[0193] 1. Experimental conditions

[0194] The hardware testing conditions for this experiment were: Intel i9-13980HX, 16GB RAM, and an RTX 4060 GPU with 4GB VRAM. The signal samples used in the test were from the RadioML2022 dataset, and the actual sample range used was -10dB to 20dB.

[0195] 2. Experimental Results

[0196] To verify the modulation recognition performance, the results of this invention were compared with those of ULCNN (see Guo L, Wang Y, Liu Y, et al. Ultralight convolutional neural network for automatic modulation classification in internet of unmanned aerial vehicles[J]. IEEE Internet of Things Journal, 2024, 11(11): 20831-20839.), DAE (see Ke Z, Vikalo H. Real-time radio technology and modulation classification via anLSTM auto-encoder[J]. IEEE Transactions on Wireless Communications, 2021, 21(1): 370-382.), CDSCNN (see Xiao C, Yang S, Feng Z. Complex-valued depthwiseseparable convolutional neural network for automatic modulation classification[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 1-10.), and MCNET (see Huynh-The T, Hua CH, Pham QV, et al.) on the RadioML2022 dataset. MCNet: An efficient CNN architecture for robust automatic modulation classification[J].IEEE Communications Letters, 2020, 24(4): 811-815.) Comparison of the results of four existing deep learning modulation recognition algorithms.To quantitatively measure the quality of the reconstruction results, recognition accuracy is used as the main evaluation metric to measure the model's overall recognition ability for different modulation types. It is defined as the ratio of the number of correctly predicted samples to the total number of samples, thus reflecting the model's classification performance in the global scope. Meanwhile, to further analyze the robustness of the model under different channel conditions, a segmented accuracy metric based on signal-to-noise ratio (SNR) is introduced. That is, the model's classification accuracy is statistically calculated separately in different SNR intervals to measure the difference in the model's recognition ability under low and high SNR conditions.

[0197] The experimental results are shown in Table 1:

[0198] Table 1. Comparison of recognition accuracy (%) of different methods on the RadioML2022 dataset

[0199]

[0200] As shown in Table 3, the method of this invention outperforms existing methods under all signal-to-noise ratio conditions on the RadioML2022 dataset, especially under low signal-to-noise ratio conditions. Compared with the overall suboptimal method MCNET, the accuracy under low signal-to-noise ratio conditions is improved by 23.8%, and the average accuracy is improved by 5.9%.

[0201] Figure 2 The figure shows the recognition accuracy curves of different methods under various signal-to-noise ratio (SNR) conditions on the RadioML2022 dataset. As can be seen from the figure, the method of this invention maintains high recognition accuracy across the entire SNR range, with a particularly significant advantage in the low SNR region.

[0202] The signal modulation recognition method based on multimodal feature fusion and SNR perception introduced a multimodal feature fusion mechanism and signal-to-noise ratio embedding information, enabling the model to adaptively adjust feature expression under different signal-to-noise ratio conditions, thereby maintaining high recognition accuracy and stability in complex electromagnetic environments.

[0203] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A signal modulation recognition method based on multimodal feature fusion and SNR sensing, characterized in that: Includes the following steps, Step 1: Acquire the wireless signal to be identified, perform complex baseband representation and discretization processing to obtain a discrete signal; Step 2: Input the discrete signal tensor obtained in Step 1 As input features, temporal feature extraction is performed to extract temporal structure information; Step 3: Perform STFT transformation on the discrete signal obtained in Step 1 to extract its time-frequency domain features; The amplitude is obtained by taking its value, and the normalized amplitude is then input into the frequency domain feature extraction network for frequency domain feature extraction. Step 4: Perform statistical analysis on the discrete signal obtained in Step 1 to extract physical statistical features; extract the phase of the discrete signal, expand it, calculate the adjacent phase difference, and input the phase difference sequence into the feature mapping network to obtain the phase features; The input signal-to-noise ratio is mapped to a discrete interval to obtain a discrete index, which is then mapped to an embedding vector to obtain the SNR feature. The above three features are concatenated to obtain the fused feature; Step 5: Concatenate the time-domain features, frequency-domain features, and fused features obtained in steps 2, 3, and 4 to obtain... ; Then, the discriminant score vector is obtained through nonlinear mapping. ; Step 6: The dataset consists of discrete signals and their corresponding modulation tags; for each discrete signal sample, after processing in steps 1–5, the final feature representation is obtained. Each modulation category is assigned a corresponding modulation category label; a dataset for model training is constructed. Step 7: Calculate the discriminant score vector obtained in Step 5. Transform it into a probability distribution; construct a loss function based on the probability distribution; Step 8: Optimize the trainable parameters in the modulation recognition network using the minimized loss function obtained in Step 7; Step 9: Iteratively train the network from step 8 until a trained model is obtained; Step 10: Input the signal to be identified into the model trained in step 9 to obtain the corresponding modulation category prediction result.

2. The method as described in claim 1, characterized in that: The method for obtaining physical characteristics in step 4 is as follows: Input tensor to the discrete signal obtained in step 1 Perform statistical analysis and extract physical statistical features. in, This represents the physical feature extraction function, which includes the extraction of physical features from amplitude statistics, energy distribution, and higher-order statistics.

3. The method as described in claim 1, characterized in that: The method for obtaining phase features in step 4 is as follows: Phase extraction is performed on a discrete signal, and after phase expansion, the instantaneous phase of the discrete signal is expressed as: in, Indicates the sampling point index. Represents the in-phase component of a discrete signal. Represents the orthogonal components of a discrete signal; Calculate the phase difference between adjacent sampling points based on the instantaneous phase: in, Indicates the length of the sampling sequence; This yields the phase difference sequence. The phase difference sequence is normalized to obtain Will By inputting a one-dimensional convolutional or fully connected mapping function for feature extraction, a fixed-dimensional phase feature representation is obtained: in, This represents the phase feature mapping function.

4. The method as described in claim 1, characterized in that: The method for obtaining the signal-to-noise ratio embedding features in step 4 is as follows: A signal-to-noise ratio (SNR) embedding mechanism is introduced to enhance the model's adaptability to different channel conditions; let the SNR of the discrete signal be... ,Will Mapping to discrete intervals: in, This represents the minimum signal-to-noise ratio. Indicates the interval width; Discrete index Mapped to an embedding vector: in, This indicates an embedded mapping function.

5. The method as described in claim 1, characterized in that: The SNR sensing fusion method described in step 4 is as follows: The physical features, phase features, and SNR embeddings are fused to obtain a fused feature representation; specifically, after aligning and vectorizing each feature, a fused feature vector is formed through a concatenation operation. 。 6. The method as described in claim 1, characterized in that: The method for constructing the model training data in step 6 is as follows: The dataset consists of discrete signals and their corresponding modulation labels; for each discrete signal sample, after processing in steps 1–5, the final feature representation is obtained. Each training sample corresponds to a modulation category label; let the total number of training samples be... The training dataset can then be represented as: in, Indicates the first Feature representation of a sample This represents the corresponding real category label; Representing category labels in one-hot encoded form: in, This indicates the total number of modulation categories.

7. The method as described in claim 1, characterized in that: Step 7 is implemented as follows: For each sample, for the first... For each input sample, the classification model in step 5 outputs a corresponding discriminant score vector: The scores are converted into a probability distribution using the Softmax function. in, Indicates the first The sample at the th Class-based discrimination scores Indicates the first The sample at the th Class-based discrimination scores; Introducing a label smoothing mechanism, the true labels are represented as: in, Indicates the smoothing coefficient; Based on the above definition, construct the cross-entropy loss function: The loss function optimizes the model parameters by minimizing the difference between the predicted distribution and the true distribution.

8. The method as described in claim 1, characterized in that: Step 8 is implemented as follows: After determining the loss function, a gradient-based optimization method is used to optimize and train the trainable parameters in the signal modulation recognition network. These trainable parameters include the time-domain feature extraction branch, the frequency-domain feature extraction branch, the physical feature and phase feature fusion branch, and the network parameters in the classification model. Let the set of all trainable parameters in the network be denoted as . By minimizing the loss function Continuously update trainable parameters: in, Indicates the learning rate. This represents the gradient of the loss function with respect to the trainable parameters; During training, gradients are calculated using the backpropagation algorithm and combined with optimization strategies to update parameters, thereby gradually converging the model to the optimal state. To improve training stability, regularization methods can be used to constrain the model.

9. The method as described in claim 1, characterized in that: Step 9 is implemented as follows: Based on the dataset constructed in step 6 and the optimization methods defined in steps 7–8, the model is trained iteratively. In each training iteration, a batch of sample data is input, the corresponding loss is calculated, and the model parameters are updated. As the training process progresses, the loss function gradually decreases, and the model performance gradually improves. Let the number of training rounds be... The training process can then be represented as: During training, the model convergence is determined by monitoring performance metrics on the validation set. When the model performance stabilizes or reaches the preset conditions, training is stopped, and the trained model parameters are obtained.

10. The method as described in claim 1, characterized in that: Step 10 is implemented as follows: After the model training is completed, the signal to be identified is input into the trained model to obtain the corresponding modulation category prediction result; For any input signal, its characteristic representation obtained after processing in steps 1–5 is denoted as: The model outputs the probability distributions for each category in step 7. The final identification result is determined in the following way: in, Indicates the predicted modulation category; The trained model enables the identification of unknown signal modulation methods.

Images