Low signal-to-noise ratio condition modulation recognition method based on wavelet transform and channel attention

Abstract

The application discloses a low signal-to-noise ratio condition modulation identification method based on wavelet transform and channel attention, comprising: acquiring a received signal; inputting the received signal into a pre-trained neural network identification model to output a classification identification result and a classification accuracy rate. Based on the signal reconstruction and the identification model of the neural network, the wavelet threshold estimation module and the wavelet reconstruction module constitute a signal reconstruction path, the multi-scale feature extraction module and the prediction classification module constitute an identification path, the two paths are mutually enhanced, the advantages of the digital signal processing technology and the neural network are combined, the wavelet threshold estimation module is introduced to predict a denoising threshold through the neural network, and the network parameters of the wavelet threshold estimation module are updated according to the back propagation; the signal reconstruction path and the identification path are combined, the network interpretability is enhanced, the multi-scale modulation features are utilized, and the modulation identification accuracy rate in the low signal-to-noise ratio scene is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of radio signal modulation recognition technology, specifically relating to a low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention. Background Technology

[0002] Automatic Modulation Recognition (AMR) is a technique for identifying the modulation patterns of radio signals, especially non-cooperative signals. It provides basic modulation information for radio signals and has wide applications in modern mobile communications, cognitive radio, spectrum sensing, and signal detection. It plays a crucial role in addressing spectrum scarcity issues in large-scale IoT, autonomous vehicles, and drones. Furthermore, in non-cooperative communication scenarios, AMR is essential for reverse communication, serving as a key intermediate signal processing step between signal detection and demodulation, aiding subsequent demodulation and intelligence analysis. In most communication scenarios, prior information related to the user is not known, especially in non-cooperative communication. Therefore, finding a method that does not rely on prior knowledge of the target signal while maintaining good modulation recognition performance at low signal-to-noise ratios is a pressing problem in signal recognition research.

[0003] Traditional signal recognition techniques are mainly divided into two types: likelihood-based (LB) and feature-based (FB) methods. The LB method achieves optimal performance by considering the probability density function of all unknowns and is often used as a theoretical upper bound in pattern recognition performance comparisons. However, this method has high computational complexity and cannot be applied to situations where there is no matching between the unknown signal and the transceiver. The FB method involves extracting specific features from the received signal and then using a suitable classifier to classify different modulated signals. FB-based methods have relatively low computational complexity and are very reliable in mitigating the negative effects of multipath fading, but the selection of features largely depends on prior information. In recent years, deep learning (DL) has been extensively studied in the AMR field. Deep learning-based AMR captures the relationship between radio signals and modulation categories by training a neural network, automatically extracting easily classifiable signal features. Existing DL-AMR methods include Convolutional Neural Network (CNN) models, Recurrent Neural Network (RNN) models, and hybrid neural network models. CNN models excel at processing spatially relevant data. Their spatial feature extraction capabilities are used to detect signal modulation schemes. Existing CNN-based AMR models can be divided into two categories: CNN models with raw IQ input and CNN models with preprocessed input, such as combining traditional FB-AMR methods with CNN models. Besides spatial correlation characteristics, radio signals also exhibit strong temporal correlation characteristics, which can be learned by RNN models. Therefore, many researchers utilize RNNs for modulation recognition. Benefiting from the respective advantages of RNNs and CNNs, hybrid neural network models have emerged—combining RNNs and CNNs—to further improve AMR performance.

[0004] Many techniques in Digital Signal Processing (DSP) heavily rely on prior information and expert experience, which is impractical in non-cooperative scenarios. Furthermore, existing DL-AMR methods exhibit a sharp drop in accuracy when the signal-to-noise ratio (SNR) is below 0 dB, necessitating improvements and optimizations to deep neural networks to enhance DL-AMR performance under low SNR conditions, such as using networks with stronger feature extraction capabilities. However, AMR performance under low SNR conditions cannot be significantly improved solely by optimizing the neural network, as the signal is almost completely submerged in noise below 0 dB, making it difficult for the network to extract signal features for classification. Moreover, existing methods treat DSP and AMR as independent tasks, preventing them from benefiting from an end-to-end joint framework. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] This invention provides a low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention, the method comprising:

[0007] Acquire received signals;

[0008] The received signal is input into a pre-trained recognition model based on signal reconstruction and neural networks, and the modulation category corresponding to the received signal is output.

[0009] The identification model includes: a wavelet threshold estimation module, used to estimate the threshold of the I / Q signal components of the received signal based on a neural network to obtain a predicted wavelet denoising threshold; a wavelet reconstruction module, used to denoise and reconstruct the received signal based on the predicted wavelet denoising threshold to obtain a reconstructed signal; a multi-scale feature extraction module, used to extract features from the I-channel signal components, Q-channel signal components, and IQ parallel signal components of the reconstructed signal to obtain I-channel features, Q-channel features, and IQ-channel features; a channel attention network is used to fuse the different channel features of the I-channel features, Q-channel features, and IQ-channel features to obtain fused features, and a long short-term memory network is used to obtain features with temporal relationships of the fused features; and a prediction classification module, used to obtain the modulation category of the received signal based on the features with temporal relationships of the fused features.

[0010] In one embodiment of the present invention, the expression for the received signal is as follows:

[0011]

[0012] Among them, S r (n) represents the received signal. Indicates M modulation methods The modulated signal generated in the process, h(n) represents the impulse response of the wireless transmission channel, * represents the convolution operation, w(n) represents the signal noise, and Δf represents the additional carrier frequency offset during transmission. This represents the random phase shift caused by phase jitter during transmission, where n = 0, 1, ..., N-1, N, and N represents the total length of the received signal.

[0013] In one embodiment of the present invention, the expression for the I / Q signal components of the received signal is as follows:

[0014] The signal component I(n) = real(S)r (n));

[0015] Q-channel signal component Q(n) = imag(S) r (n));

[0016] Among them, real(S) r (n) represents the received signal S r The real part of (n), imag(S) r (n) represents the received signal S r The imaginary part of (n).

[0017] In one embodiment of the present invention, the training process of the recognition model based on signal reconstruction and neural network includes: after several iterations of training, reaching a preset maximum number of iterations, obtaining a trained recognition model based on signal reconstruction and neural network; wherein, one iteration of training includes:

[0018] All received signals are used as a dataset; the dataset is then divided into a training set, a test set, and a validation set.

[0019] For each received signal in the training set, the wavelet threshold estimation module performs threshold estimation on the I / Q signal components of the received signal to obtain the predicted wavelet denoising threshold.

[0020] The wavelet reconstruction module performs denoising and reconstruction on the received signal based on the predicted wavelet denoising threshold to obtain the reconstructed signal.

[0021] Using the multi-scale feature extraction module, features are extracted from the I-channel signal component, Q-channel signal component, and IQ parallel signal component of the reconstructed signal to obtain I-channel features, Q-channel features, and IQ-channel features, respectively. The different channel features of the I-channel features, Q-channel features, and IQ-channel features are then fused using a channel attention network to obtain fused features. Finally, the fused features are processed using a long short-term memory network to obtain features with temporal relationships.

[0022] Using the prediction classification module, a predicted label for the received signal is obtained based on the features with temporal relationships; the loss function value for the current iteration of training is obtained based on the true label and the predicted label; the parameters of the recognition model are updated using the loss function value to obtain the updated recognition model; the updated recognition model is used for the next iteration of training.

[0023] In one embodiment of the present invention, the wavelet threshold estimation module performs threshold estimation on the I / Q signal components of the received signal to obtain a predicted wavelet denoising threshold, including:

[0024] The first fully connected layer and the tanh activation layer in the wavelet threshold estimation module are used to process the I / Q signal components of the received signal in sequence to obtain the feature F. t ;

[0025] The feature F t The expression is as follows:

[0026] Ft=tanh(fc1(reshape(Sr(n))));

[0027] Where tanh() represents the tanh activation function, S r (n) represents the received signal, fc1 represents the processing of the first fully connected layer, and reshape() represents the tensor operation;

[0028] After passing the reconstructed signal through the second fully connected layer, the predicted wavelet denoising threshold is obtained.

[0029] The expression for the predicted wavelet denoising threshold is as follows:

[0030]

[0031] Where sigmoid() represents the activation function, and fc2() represents the processing of the second fully connected layer;

[0032] The network parameters of the wavelet threshold estimation module are updated through backpropagation.

[0033] In one embodiment of the present invention, the wavelet reconstruction module performs denoising reconstruction on the received signal based on the predicted wavelet denoising threshold to obtain a reconstructed signal, including:

[0034] The wavelet reconstruction module decomposes the received signal according to a preset wavelet basis and decomposition level using a first formula to obtain low-frequency components and high-frequency components.

[0035] The first formula is as follows:

[0036]

[0037] Among them, S r (n) represents the received signal. S represents r (n) The low-frequency component at scale k obtained after projection into scale space. S represents r (n) The high-frequency component at scale l obtained after wavelet spatial projection;

[0038] The high-frequency components are denoised based on the predicted wavelet denoising threshold to obtain the denoised high-frequency components.

[0039] The reconstructed signal is obtained based on the low-frequency component and the denoised high-frequency component.

[0040] In one embodiment of the present invention, the long short-term memory network includes: two cascaded sub-long short-term memory networks, each with 128 units.

[0041] In one embodiment of the present invention, the prediction classification module is used to obtain a predicted label of the received signal based on the features with temporal relationships, including: the prediction classification module processes the features with temporal relationships to obtain a predicted probability;

[0042] The predicted probability is processed using a prediction formula to obtain the predicted label of the received signal;

[0043] The prediction formula is as follows:

[0044]

[0045] Where p(yx) represents the predicted probability, y represents the true label, x represents the received signal, and argmax() represents the argmax function.

[0046] In one embodiment of the present invention, the expression for the loss function is as follows:

[0047]

[0048] Where Loss represents the loss value, K represents the total number of modulation categories, and y i This represents the actual tag of the i-th received signal. This represents the predicted probability of the i-th received signal.

[0049] The beneficial effects of this invention are:

[0050] In this invention, a recognition model based on signal reconstruction and neural networks comprises a wavelet threshold estimation module and a wavelet reconstruction module forming the signal reconstruction path, and a multi-scale feature extraction module and a prediction classification module forming the recognition path. These two paths reinforce each other. An end-to-end framework integrates signal processing and deep learning, combining the advantages of digital signal processing technology with neural networks to generate the modulation category of the received signal. The wavelet threshold estimation module predicts the wavelet denoising threshold using a neural network and updates its network parameters based on backpropagation to achieve adaptive wavelet transform. Combining the signal reconstruction path with the recognition path enhances the network's interpretability. The adaptive wavelet transform extracts multi-scale modulation features from the reconstructed signal, avoiding reliance on prior information and significantly improving the accuracy of modulation recognition in low signal-to-noise ratio scenarios. Attached Figure Description

[0051] Figure 1 A schematic diagram illustrating the steps of a low signal-to-noise ratio conditional modulation identification method based on wavelet transform and channel attention provided in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the steps of one iteration of training of a recognition model based on signal reconstruction and neural networks provided in an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram of the wavelet threshold estimation module, wavelet reconstruction module, and multi-scale feature extraction module in the recognition model based on signal reconstruction and neural networks provided in this embodiment of the invention.

[0054] Figure 4 The figure shows a comparison of the accuracy of the low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention provided in this embodiment of the invention with existing methods on an open-source dataset.

[0055] Figure 5 This is a graph showing the impact of the selectivity of features in the reconstructed signal provided by the multi-scale feature extraction module in this embodiment of the invention on the algorithm performance. Detailed Implementation

[0056] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0057] To address the issue of a sharp drop in accuracy in existing modulation recognition methods under low signal-to-noise ratio (SNR) scenarios, this invention provides a low SNR conditional modulation recognition method (Wavelet-Based Channel Attention Network, WCAN) based on wavelet transform and channel attention. Figure 1 As shown, it may include:

[0058] S1, acquire the received signal.

[0059] For S1, the expression for the received signal is as follows:

[0060]

[0061] Among them, S r (n) represents the received signal. Indicates M modulation methods The modulated signal generated in the process, h(n) represents the impulse response of the wireless transmission channel, * represents the convolution operation, w(n) represents the signal noise, and Δf represents the additional carrier frequency offset during transmission. This represents the random phase shift caused by phase jitter during transmission, where n = 0, 1, ..., N-1, N, and N represents the total length of the received signal.

[0062] S2, input the received signal into a pre-trained recognition model based on signal reconstruction and neural network, and output the modulation category corresponding to the received signal;

[0063] Among them, the recognition model, such as Figure 3 As shown, it can include: a wavelet threshold estimation module, used to estimate the threshold of the I / Q signal components of the received signal based on a neural network to obtain a predicted wavelet denoising threshold; a wavelet reconstruction module, used to denoise and reconstruct the received signal based on the predicted wavelet denoising threshold to obtain a reconstructed signal; a multi-scale feature extraction module, used to extract features from the I-channel signal component, Q-channel signal component, and IQ parallel signal component of the reconstructed signal to obtain I-channel features, Q-channel features, and IQ-channel features; a channel attention network is used to fuse the different channel features of the I-channel features, Q-channel features, and IQ-channel features to obtain fused features, and a long short-term memory network is used to obtain features with temporal relationships of the fused features; and a prediction and classification module, used to obtain the modulation category of the received signal based on the features with temporal relationships of the fused features.

[0064] The expressions for the I / Q signal components of the received signal are as follows:

[0065] The signal component I(n) = real(S) r (n));

[0066] Q-channel signal component Q(n) = imag(S) r (n));

[0067] Among them, real(S) r (n) represents the received signal S r The real part of (n), imag(S) r (n) represents the received signal S r The imaginary part of (n).

[0068] Understandably, the I / Q components can also be referred to as in-phase / quadrature components, which are stacked in parallel.

[0069] For S2, the training process of the recognition model based on signal reconstruction and neural networks can include: after several iterations of training, reaching a preset maximum number of iterations, obtaining the trained recognition model based on signal reconstruction and neural networks; wherein, one iteration of training is as follows: Figure 2 As shown, it may include:

[0070] S21, take all received signals as the dataset; divide the dataset into training set, test set and validation set;

[0071] Understandably, the training set in the dataset can be used for the corresponding training process, while the test set and validation set can be used to test and validate the trained recognition model.

[0072] S22, For each received signal in the training set, the wavelet threshold estimation module performs threshold estimation on the I / Q signal components of the received signal to obtain the predicted wavelet denoising threshold;

[0073] For S22, it may include:

[0074] S221, the first fully connected layer and the tanh activation layer in the wavelet threshold estimation module are used to process the I / Q signal components of the received signal in sequence to obtain the feature F. t ;

[0075] Feature F t The expression is as follows:

[0076] Ft=tanh(fc1(reshape(Sr(n))));

[0077] Where tanh() represents the tanh activation function, S r (n) represents the received signal, fc1 represents the processing of the first fully connected layer, and reshape() represents the tensor operation;

[0078] S222, after the reshaped signal is passed through the second fully connected layer, the predicted wavelet denoising threshold is obtained;

[0079] The expression for the predicted wavelet denoising threshold is as follows:

[0080]

[0081] Where sigmoid() represents the activation function, and fc2() represents the processing of the second fully connected layer;

[0082] S223, update the network parameters of the wavelet threshold estimation module through backpropagation.

[0083] Understandably, backpropagation of the neural network is used to update the network parameters in the wavelet threshold estimation module, and these parameters are dynamically adjusted to achieve the best effect in denoising and reconstruction.

[0084] S23, Wavelet reconstruction module, performs denoising and reconstruction on the received signal based on the predicted wavelet denoising threshold to obtain the reconstructed signal, which may include:

[0085] S231, the wavelet reconstruction module decomposes the received signal according to the preset wavelet basis and decomposition level using the first formula to obtain low-frequency components and high-frequency components.

[0086] The first formula is as follows:

[0087]

[0088] Where Sr(n) represents the received signal, S CAk (n) represents the low-frequency component of Sr(n) at scale k after projection into scale space, S CDl (n) represents the high-frequency component of Sr(n) at scale l after wavelet spatial projection;

[0089] φ represents the scaling factor, k represents the scale, t represents time, and φ represents the time factor. k,t (n) represents the scaling function. Denotes wavelet coefficients, ψ l,t (n) represents the wavelet function.

[0090] The received signal is decomposed into wavelet coefficients of different scales by selecting an appropriate wavelet basis and a number of decomposition levels using wavelet transform. In each decomposition level, the received signal is decomposed into two parts: a low-frequency component and a high-frequency component. Therefore, the received signal can be represented by the sum of a low-frequency signal and a finite number of high-frequency signals.

[0091] S232, Denoise the high-frequency components based on the predicted wavelet denoising threshold to obtain the denoised high-frequency components;

[0092] It can contain the main characteristics of the received signal, while It can include details of the received signal, such as sharp variations in noise. Noise reduction can be achieved by discarding some high-frequency wavelet coefficients.

[0093] S233, based on the low-frequency component and the denoised high-frequency component, obtains the reconstructed signal.

[0094] S24. Using a multi-scale feature extraction module, features are extracted from the I-channel signal component, Q-channel signal component, and IQ parallel signal component of the reconstructed signal to obtain I-channel features, Q-channel features, and IQ-channel features respectively. Based on a channel attention network, the different channel features of the I-channel features, Q-channel features, and IQ-channel features are fused to obtain fused features. Based on a long short-term memory network, the fused features are processed to obtain features with temporal relationships.

[0095] Optionally, the long short-term memory network may include: two cascaded sub-long short-term memory networks with 128 units each.

[0096] The fused features are processed using a long short-term memory network to obtain features with temporal relationships.

[0097] Understandably, features with temporal relationships that are fused can represent temporal features.

[0098] Specifically, the independent information obtained from the I, Q, and IQ channels is fused along the channel dimension. An attention network (Squeeze and Excitation, SE) suitable for different channel features of the signal is used to obtain features with channel attention, which are then input into a two-dimensional convolution to explore spatial correlations. Considering the temporal characteristics of the modulated signal, the fused features are passed through two cascaded Long Short-Term Memory (LSTM) layers with 128 units each to effectively process sequential data and extract temporal correlations. Then, two fully connected layers with 128 neurons and a scaling factor are added to obtain features with temporal relationships in the fused features.

[0099] S25. Using the prediction classification module, based on features with temporal relationships, the predicted label of the received signal is obtained; the loss function value of the current iteration training is obtained according to the real label and the predicted label; the parameters of the recognition model are updated using the loss function value to obtain the updated recognition model; the updated recognition model is used for the next iteration training.

[0100] In S25, the predicted label of the received signal is obtained using the prediction classification module based on temporal features, which may include:

[0101] The prediction and classification module processes features with temporal relationships to obtain prediction probabilities;

[0102] The predicted probability is processed using the prediction formula to obtain the predicted label of the received signal;

[0103] The prediction formula is as follows:

[0104]

[0105] Where p(yx) represents the predicted probability, y represents the true label, x represents the received signal, and argmax() represents the argmax function.

[0106] For S25, the expression for the loss function is as follows:

[0107]

[0108] Where Loss represents the loss value, K represents the total number of modulation categories, and y i This represents the actual tag of the i-th received signal. This represents the predicted probability of the i-th received signal.

[0109] Understandably, the predicted label serves as the classification and recognition result; argmax() represents the function for predicting the label, which obtains the final predicted label based on the predicted probability; the number of iterations required to train the recognition model can be determined by setting the maximum number of iterations Epoch and the number of categories N.

[0110] The dataset can include a training set, a test set, and a validation set. The recognition model is trained epochs using the training set. The trained recognition model can then be used to perform modulation and recognition on all samples in the test set.

[0111] Specifically, all samples in the test set are sequentially retrieved and input into the trained recognition model for classification prediction, and the predicted labels for each sample are output. The predicted labels of the samples are compared with the true labels of the samples to obtain the classification results. The classification accuracy of the model can be obtained by statistically analyzing the recognition results of all samples.

[0112] The low signal-to-noise ratio (SNR) conditional modulation recognition method based on wavelet transform and channel attention provided in this invention, in its recognition model based on signal reconstruction and neural networks, comprises a wavelet threshold estimation module and a wavelet reconstruction module forming the signal reconstruction path, and a multi-scale feature extraction module and a prediction classification module forming the recognition path, with the two paths mutually reinforcing each other. By integrating signal processing and deep learning through an end-to-end framework, it combines the advantages of digital signal processing technology and neural networks to generate the modulation category of the received signal. The wavelet threshold estimation module predicts the wavelet denoising threshold through a neural network and updates its network parameters based on backpropagation to achieve adaptive wavelet transform. Combining the signal reconstruction path with the recognition path enhances the interpretability of the network. The adaptive wavelet transform extracts multi-scale modulation features from the reconstructed signal, avoiding dependence on prior information and significantly improving the accuracy of modulation recognition in low SNR scenarios.

[0113] To demonstrate the beneficial effects of the low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention proposed in this embodiment of the invention, the beneficial effects of this embodiment of the invention will be further explained through simulation experiments below.

[0114] Open source dataset experiments

[0115] The simulation experiments used the standard open-source dataset RML2016.10a, which includes 220,000 modulated signals with signal-to-noise ratios (SNR) ranging from -20dB to +18dB, and considers 11 commonly used modulations: WBFM, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, QPSK, and 8PSK. Each signal in the dataset has 128 time I / Q samples in complex form and was generated in a harsh simulated propagation environment, which is affected by AWGN, multipath fading, sampling rate offset, and center frequency offset to simulate real-world conditions.

[0116] The experiment divided the dataset into three parts in a 6:3:1 ratio: training set, test set, and validation set. Specifically, for each modulation type and signal-to-noise ratio, 1000 signals were randomly divided into 600 signals for training and 300 signals for testing. The Adam optimizer was used in model training, and the learning rate was set to change with the number of training epochs, with a learning rate λ = 1 / (0.1*Epoch+1). The batch size of gradient updates during training could be 256.

[0117] To evaluate the effectiveness of the low signal-to-noise ratio (SNR) conditional modulation (LMD) recognition method based on wavelet transform and channel attention under low SNR conditions, this invention selected 10 baseline methods for comparison. Specifically, these included CNN1, CNN2, ResNet, LSTM, GRU, CLDNN, DenseNet, CGDNet, PET-CGDNN, and MAC models. Experiments were conducted using the above datasets on the proposed wavelet transform and channel attention-based LNR conditional modulation recognition method and baseline methods. For accuracy comparisons with existing methods on open-source datasets, please refer to the graphs. Figure 4 .from Figure 4 As can be seen, this method achieves the best recognition performance, especially when the signal-to-noise ratio (SNR) is below 0 dB. The low SNR conditional modulation recognition method based on wavelet transform and channel attention proposed in this embodiment of the invention outperforms other existing methods in terms of recognition accuracy. This result proves that combining DSP with deep neural networks is effective. The WCAN proposed in this embodiment of the invention can perform adaptive threshold denoising and reconstruction for signals with different SNRs. When the SNR is -4 dB, it also achieves an 8.21% accuracy improvement on the RML2016.10A dataset compared to the current highest-accuracy method.

[0118] The test accuracy and the number of learning parameters of different methods under all signal-to-noise ratios are compared, as shown in Table 1, which compares the complexity and accuracy of the proposed method with other models.

[0119] Table 1. Comparison of the complexity and accuracy of the method proposed in this invention with other models.

[0120] Methods Learning parameters Accuracy △(↑％) CNN1 1.58M 53.14％ 9.38％ CNN2 0.86M 58.35％ 4.17％ ResNet 3.11M 59.11％ 3.41％ LSTM 0.23M 60.21％ 2.31％ GRU 0.15M 58.20％ 4.32％ PET-CGDNN 0.10M 59.95％ 2.57％ DenseNet 4.23M 60.28％ 2.24％ CLDNN 0.51M 58.02％ 4.50％ MAC 2.74M 60.15％ 2.37％ CGDNet 0.12M 58.95％ 3.57％ WCAN 0.72M 62.52％ -

[0121] In Table 1, Methods represents the method, Learning parameters represent the learning parameters, Accuracy represents the overall accuracy, and Δ represents the difference between the baseline method and the WCAN model. As can be seen from Table 1, the WCAN proposed in this embodiment of the invention still achieves satisfactory performance with a total accuracy of 62.52% with a relatively small number of model parameters, demonstrating the effectiveness of the proposed WCAN method in combining DSP and deep neural networks.

[0122] Model performance analysis experiment

[0123] To verify the feature selectivity of the multi-scale feature extraction module for the reconstructed signal, the provided model and the results after removing the attention network SE were compared with current mainstream attention modules, including spatial attention (SA), convolutional block attention (CBAM), and triplet attention (TA). For a graph showing the impact of the multi-scale feature extraction module's feature selectivity on the reconstructed signal on the algorithm's performance, please refer to [link to graph]. Figure 5 .from Figure 5 As can be seen, the attention network SE can selectively handle multi-scale features along the channel dimension and achieve the best recognition results. The selectable I-channel, Q-channel, and I / Q multi-channel maximize the use of complementary information after wavelet reconstruction.

[0124] Simulation experiments further demonstrate that the wavelet threshold estimation module and wavelet reconstruction module constitute the signal reconstruction path, while the multi-scale feature extraction module and prediction classification module constitute the recognition path in the recognition model based on signal reconstruction and neural networks proposed in this embodiment of the invention. The two paths reinforce each other. By integrating signal processing and deep learning through an end-to-end framework, the advantages of digital signal processing technology and neural networks are combined to generate the modulation category of the received signal. The wavelet threshold estimation module is introduced to predict the wavelet denoising threshold through a neural network and update its own network parameters according to backpropagation to achieve adaptive wavelet transform. Combining the signal reconstruction path and the recognition path enhances the interpretability of the network. By using adaptive wavelet transform to extract multi-scale modulation features from the reconstructed signal, the dependence on prior information is avoided, and the accuracy of modulation recognition in low signal-to-noise ratio scenarios is significantly improved.

[0125] It should be noted that, in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0126] The above description is merely a preferred 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 are included within the scope of protection of the present invention.

Claims

1. A low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention, characterized in that, include: Acquire received signals; The received signal is input into a pre-trained recognition model based on signal reconstruction and neural networks, and the modulation category corresponding to the received signal is output. The identification model includes: a wavelet threshold estimation module, used to perform threshold estimation on the I / Q signal components of the received signal based on a neural network to obtain a predicted wavelet denoising threshold; a wavelet reconstruction module, used to perform denoising reconstruction on the received signal based on the predicted wavelet denoising threshold to obtain a reconstructed signal; a multi-scale feature extraction module, used to extract features from the I-channel signal component, Q-channel signal component, and IQ parallel signal component of the reconstructed signal respectively to obtain I-channel features, Q-channel features, and IQ-channel features; fusing different channel features of the I-channel features, Q-channel features, and IQ-channel features based on a channel attention network to obtain fused features, and obtaining temporally related features of the fused features through a long short-term memory network; a prediction classification module, used to obtain the modulation category of the received signal based on the temporally related features of the fused features; and a wavelet threshold estimation module to perform threshold estimation on the I / Q signal components of the received signal to obtain a predicted wavelet denoising threshold, including: The first fully connected layer and the tanh activation layer in the wavelet threshold estimation module are used to process the I / Q signal components of the received signal sequentially to obtain the features. ; The features The expression is as follows: ； in, express Activation function Indicates receiving signal, This indicates the processing of the first fully connected layer. Represents tensor operations; The feature After passing through the second fully connected layer, the predicted wavelet denoising threshold is obtained; The expression for the predicted wavelet denoising threshold is as follows: ； in, This represents the activation function. This indicates processing by the second fully connected layer; The network parameters of the wavelet threshold estimation module are updated through backpropagation; The wavelet reconstruction module performs denoising and reconstruction on the received signal based on the predicted wavelet denoising threshold to obtain a reconstructed signal, including: The wavelet reconstruction module decomposes the received signal according to a preset wavelet basis and decomposition level using a first formula to obtain low-frequency components and high-frequency components. The first formula is as follows: ； in, Indicates receiving signal, express The result after scale-space projection The low-frequency component at this scale, express The result obtained after wavelet spatial projection High-frequency components at the scale; The high-frequency components are denoised based on the predicted wavelet denoising threshold to obtain the denoised high-frequency components. The reconstructed signal is obtained based on the low-frequency component and the denoised high-frequency component.

2. The low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention according to claim 1, characterized in that, The expression for the received signal is as follows: ； in, Indicates receiving signal, Indicates from modulation method The modulated signal generated in the middle, The impulse response of a wireless transmission channel. This represents the convolution operation. Indicates signal noise. This indicates the additional carrier frequency offset during transmission. This indicates the random phase shift caused by phase jitter during transmission. , This indicates the total length of the received signal.

3. The low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention according to claim 1, characterized in that, The expressions for the I / Q signal components of the received signal are as follows: I-channel signal component ; Q-channel signal quantity ; in, Indicates received signal The real part, Indicates received signal The imaginary part.

4. The low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention according to claim 3, characterized in that, The training process of the recognition model based on signal reconstruction and neural networks includes: after several iterations of training, reaching a preset maximum number of iterations, obtaining the trained recognition model based on signal reconstruction and neural networks; wherein, one iteration of training includes: All received signals are used as a dataset; the dataset is then divided into a training set, a test set, and a validation set. For each received signal in the training set, the wavelet threshold estimation module performs threshold estimation on the I / Q signal components of the received signal to obtain the predicted wavelet denoising threshold. The wavelet reconstruction module performs denoising and reconstruction on the received signal based on the predicted wavelet denoising threshold to obtain the reconstructed signal. Using the multi-scale feature extraction module, features are extracted from the I-channel signal component, Q-channel signal component, and IQ parallel signal component of the reconstructed signal to obtain I-channel features, Q-channel features, and IQ-channel features, respectively. The different channel features of the I-channel features, Q-channel features, and IQ-channel features are then fused using a channel attention network to obtain fused features. Finally, the fused features are processed using a long short-term memory network to obtain features with temporal relationships. Using the prediction classification module, a predicted label for the received signal is obtained based on the features with temporal relationships; the loss function value for the current iteration of training is obtained based on the true label and the predicted label; the parameters of the recognition model are updated using the loss function value to obtain the updated recognition model; the updated recognition model is used for the next iteration of training.

5. The low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network includes two cascaded sub-LSTM networks, each with 128 units.

6. The low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention according to claim 4, characterized in that, Using the prediction classification module, based on the features with temporal relationships, the predicted label of the received signal is obtained, including: the prediction classification module processes the features with temporal relationships to obtain the predicted probability; The predicted probability is processed using a prediction formula to obtain the predicted label of the received signal; The prediction formula is as follows: ； in, Indicates the predicted probability. Indicates the true label, Indicates receiving signal, This refers to the argmax function.

7. The low signal-to-noise ratio conditional modulation recognition method based on wavelet transform and channel attention according to claim 4, characterized in that, The expression for the loss function is as follows: ； in, Indicates the loss value. This indicates the total number of modulation categories. Indicates the first A real tag that receives signals. Indicates the first The predicted probability of a received signal.

Images