Low-complexity modulation signal intelligent identification method and system
By combining a signal feature extractor, BiLSTM, and multi-head attention mechanism, the modulation recognition method solves the problems of time-consuming feature engineering of complex signals and insufficient recognition capability of shallow networks in the existing technology, and achieves efficient and accurate recognition of high-order modulation signals, which is suitable for communication systems in complex electromagnetic environments.
Patent Information
- Application Number
- CN202511615582.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-02-13
AI Technical Summary
Existing technologies rely on time-consuming and difficult-to-automate complex signal feature engineering in high-performance modulation signal recognition, and shallow deep learning architectures lack robustness and accuracy in recognizing high-order complex modulation signals, making it difficult to meet the requirements of engineering applications.
An automatic modulation recognition method is proposed, which employs a joint signal feature extractor, a bidirectional long short-term memory network (BiLSTM), and a multi-head attention mechanism. The method extracts signal features step by step through a three-level REBlock module, a max pooling layer, and a dropout layer. It combines the bidirectional LSTM and the multi-head attention layer for temporal modeling and feature aggregation to achieve the recognition of different modulation types.
It significantly improves robustness and recognition accuracy for complex modulation signals such as high-order QAM, reduces computational complexity, is suitable for non-cooperative communication scenarios, and improves the reliability and efficiency of communication systems.
Smart Images
Figure CN121530799A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of wireless communication technology and artificial intelligence technology, and particularly relates to a low-complexity intelligent signal identification method for complex electromagnetic environment, and especially to an automatic modulation identification method combining a signal feature extractor, a bidirectional long short-term memory network (BiLSTM) and a multi-head attention mechanism. BACKGROUND
[0002] With the rapid development of wireless communication technology, spectrum resources in modern communication systems have become increasingly scarce. Modulation recognition involves automatically identifying the modulation method used by received radio signals, which has wide applications in spectrum monitoring, cognitive radio and other fields. Through accurate identification of signal modulation types, the reliability and efficiency of communication systems can be effectively improved, and information collection and analysis in non-cooperative communication environments can also be facilitated. The research of these two technologies is of great significance to improving the performance, security and intelligent level of modern communication systems. Considering the threat of various non-standard intelligent devices in urban environments, it is necessary to study illegal signal identification technology in complex electromagnetic environments to effectively interfere with navigation and communication frequency bands.
[0003] Domestic scholars have proposed a fast independent component analysis method, which extracts signal feature parameters in time domain, frequency domain and time-frequency hybrid domain, and uses serial and hybrid structures to realize separation and identification of mixed signals of multiple modulation types. The signal feature parameter extraction method based on Welch transform and fractional Fourier transform is studied to realize the identification of multiple signals. The atomic decomposition theory is studied, and the matching degree of time delay and signal frequency offset is used to extract the signal feature parameters, which are then identified through support vector machine (SVM) training, achieving a high recognition probability. The high-order cumulants of the signal are used as feature parameters, and a neural network-based classifier is established, which can also achieve ideal recognition effect after training. Foreign research on signal source identification mainly focuses on the establishment process of the signal feature library. Since the signal recognition result can provide signal information for subsequent applications, signal sorting and identification play a crucial role in electromagnetic warfare. Through analysis of current domestic and foreign research results, it can be concluded that China is still in the theoretical development stage in terms of fast signal recognition method research, and engineering implementation and reliability are severely insufficient.
[0004] In order to make more efficient use of spectrum resources, automatic modulation recognition (AMR) has become a key task in signal processing and spectrum sensing. AMR mainly identifies the modulation mode of the received signal samples through analysis, providing a basis for subsequent demodulation and data processing. In traditional methods, AMR relies on expert knowledge and hand-designed features, while with the rise of deep learning, data-driven automatic feature extraction methods have gradually shown stronger expressiveness and robustness. However, the current AMR field faces two major challenges. On the one hand, high-performance traditional recognition methods highly depend on time-consuming and difficult-to-automate complex signal feature engineering, such as amplitude and phase distribution-based feature extraction, which limits its application in fast and non-cooperative environments. On the other hand, based on simple convolutional networks and other shallow deep learning architectures, although they can directly process time-domain signals, their feature extraction capabilities are insufficient to deal with high-order complex modulation signals such as 16-quadrature amplitude modulation (QAM) and 64QAM, and their recognition robustness and accuracy performance are poor in these high-difficulty types, making it difficult to meet the requirements of engineering applications. SUMMARY
[0005] The purpose of the present application is to provide a low-complexity modulation signal intelligent recognition algorithm, which can effectively improve the reliability and efficiency of the communication system through accurate recognition of the signal modulation type, and also helps information collection and analysis in non-cooperative communication environments. The present application constructs an automatic modulation recognition method combining signal feature extractor, bidirectional long short-term memory network (BiLSTM) and multi-head attention mechanism modules, which has signal feature extraction and time series modeling capabilities, and can effectively identify signals of different modulation types.
[0006] In one aspect, the present application provides a low-complexity modulation signal intelligent recognition method, a ResAttAMR model designed for accurate recognition of signal modulation type, the method comprising the following steps:
[0007] S0: Collect wireless communication signal data and perform preprocessing, including normalization, denoising and segmentation, to generate an input tensor x with shape [batch, channels1, time_steps], parameters representing data block, channel data and time step, respectively;
[0008] S1: input the preprocessed signal data into a feature extractor composed of three REBlock modules, a MaxPool1d layer, and a Dropout layer, to gradually extract signal features and adjust the number of channels to gradually expand, while reducing the time dimension through max pooling, outputting a feature tensor with a shape of [batch, channels2, time_steps], and the REBlock module effectively solves the defect of insufficient feature extraction depth of shallow convolutional networks;
[0009] S2: adjust the dimensions of the feature tensor, exchange the channel and time step dimensions, and obtain a tensor with a shape of [batch, time_steps, 128] to adapt to the input requirements of BiLSTM;
[0010] S3: input the adjusted signal feature tensor into a bidirectional LSTM network for forward and backward time series modeling, respectively, and output a time series feature with a shape of [batch, time_steps, hidden_size*2], where hidden_size is the number of hidden units of LSTM;
[0011] S4: input the time series feature output by BiLSTM into a Multi-Head Attention layer, learn the feature correlation in different subspaces through parallel calculation of multiple attention heads, and weightedly aggregate key time series information, enhancing the robustness and recognition accuracy of the model for complex modulation signals such as high-order QAM;
[0012] S5: perform Mean Pooling on the output of the Multi-Head Attention layer along the time step dimension to obtain a global feature representation with a shape of [batch, embed_dim];
[0013] S6: input the global feature representation into a fully connected classification layer, and output the probability distribution of each modulation type through a softmax activation function to complete automatic modulation recognition.
[0014] Further, the REBlock module is composed of a convolution layer, a batch normalization (BN) layer, a Squeeze-and-Excitation (SE) module, and a shortcut path, wherein the SE module adaptively adjusts the feature channel weight through a channel attention mechanism to enhance key feature expression; when the input and output channel numbers or steps do not match, a 1x1 convolution is used to adjust the dimensions of the shortcut path; the SE module aims to realize autonomous learning and weighting of discriminative signal features without manual feature engineering.
[0015] Further, the bidirectional LSTM network captures long-term forward and reverse time series dependencies of the signal through forward and backward LSTM layers, respectively, enhancing the model's ability to model global time series characteristics of the modulated signal.
[0016] Further, the multi-head attention mechanism focuses on the most discriminative time in the signal sequence by calculating the attention weight of different time steps, improving the classification robustness in noisy environments.
[0017] Further, the method is suitable for identification of various modulation types, such as Quadrature Phase Shift Keying (QPSK), 16QAM, etc., and has the advantages of high accuracy and low computational complexity in non-cooperative communication scenarios.
[0018] In another aspect of the present application, a low-complexity modulation signal intelligent identification system is also provided, comprising the following modules:
[0019] A data preprocessing module is used to collect wireless communication signal data and perform preprocessing;
[0020] A feature extraction module is used to input the preprocessed signal data into a feature extractor composed of three REBlock modules, a max-pooling layer and a dropout layer, to gradually extract signal features and adjust the number of channels;
[0021] A time series information module is used to input the signal feature tensor after adjusting the dimension into a BiLSTM network for forward and backward time series modeling, respectively, and input the time series features output by the BiLSTM into a multi-head attention layer to learn the feature correlation in different subspaces and weightedly aggregate key time series information;
[0022] A prediction output module is used to average pool the output of the multi-head attention layer along the time step dimension to obtain a global feature representation, which is input into a fully connected classification layer and output through a softmax activation function to obtain the probability distribution of each modulation type, completing automatic modulation identification.
[0023] Compared with the prior art, the present application has the following advantages and positive effects:
[0024] Compared with the traditional single-stage convolutional network technology, the hierarchical feature extractor composed of three REBlock modules is innovatively adopted, the progressive channel expansion (2→32→64→128) is cooperated with the maximum pooling dimension reduction, the calculation efficiency is maintained, and the feature expression capability is significantly improved. And through the REBlock structure integrated with the SE module, the channel level adaptive feature re-labeling is realized. Secondly, the ResAttAMR training model of the application innovatively fuses the combination mechanism of "multi-head attention + average pooling", realizes multi-scale feature attention through parallel attention heads, cooperates with the timing dimension pooling, makes the model maintain the timing sensitivity while realizing the feature compression, and greatly reduces the model training time compared with the prior art.
[0025] Furthermore, the application combines the feature extraction strategy of three-level progressive channel expansion, the bidirectional long short-term memory network and the multi-head attention mechanism to build a highly cooperative recognition architecture. The design aims to overcome the inherent defects of the simple deep learning network feature extraction depth, and significantly enhances the model recognition ability of 16QAM, 64QAM and other high-order complex modulation signals. Especially, the long-range dependence captured by the bidirectional LSTM and the focus of the multi-head attention mechanism on the discriminative timing information ensure that the model has excellent robustness and accuracy to complex electromagnetic environment and high-order modulation types under the premise of maintaining low computational complexity. BRIEF DESCRIPTION OF DRAWINGS
[0026] The technical solutions of the application will be described in detail below with reference to the drawings. It should be noted that the drawings are only used to exemplarily describe the embodiments of the application, so that those skilled in the art can better understand the technical solutions of the application. Those skilled in the art can obtain other implementation manners according to the drawings without departing from the core idea of the application.
[0027] Figure 1 The ResAttAMR model structure diagram disclosed by the application is shown in Figure 1;
[0028] Figure 2 The feature extractor module structure diagram is shown in Figure 2;
[0029] Figure 3 The REBlock module structure diagram is shown in Figure 3;
[0030] Figure 4 The SE module structure diagram is shown in Figure 4;
[0031] Figure 5 The model training loss curve is shown in Figure 5;
[0032] Figure 6 The modulation type prediction rate bar chart is shown in Figure 6;
[0033] Figure 7 The system overall confusion matrix diagram is shown in Figure 7;
[0034] Figure 8 Confusion matrix graphs under different signal-to-noise ratios. DETAILED DESCRIPTION
[0035] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.
[0036] The overall structure of the ResAttAMR model method designed by the present application is shown in Figure 1 The structure mainly includes four modules, which are a feature extractor module, a bidirectional long short-term memory network module, a multi-head attention mechanism module and a full connection layer module. The specific steps of the ResAttAMR model of the present application are as follows:
[0037] S0: Collect wireless communication signal data and pre-process, including normalization, denoising and segmentation, to generate an input tensor x with a shape of [batch, channels, time_steps]; wherein batch is the number of samples input into the model at a time, channels is the dimension or feature channel number of the input signal, and time_steps is the time series length of a single sample.
[0038] S1: input the pre-processed signal data into the feature extractor module, the specific structure of which is shown in Figure 2 The feature extractor is composed of three REBlock modules, a maximum pooling layer (MaxPool1d) and a dropout layer (Dropout), which gradually extracts signal features and adjusts the number of channels (from 2 to 32, 64, and finally to 128), while reducing the time dimension through maximum pooling, and outputs a feature tensor with a shape of [batch, 128, time_steps];
[0039] As shown in Figure 3 The REBlock module described in the present application is composed of a convolution layer, a BN layer, an SE block and possibly a shortcut, and its specific working principle is as follows: when the step is not 1 or the number of input and output channels is different, a shortcut is needed to adapt the dimension, that is, if the shapes of the input and output are different (i.e. the step is not 1 or the number of input and output channels is different), a 1x1 convolution is used to adjust the size of the input tensor to match the output. It combines the advantages of residual network and SE module, which can alleviate the gradient vanishing problem and enhance the feature extraction ability.
[0040] As shown in Figure 4As shown, the SE block in the REBlock module of the present application realizes adaptive re-labeling of feature channels through a channel attention mechanism, which allows the network to adaptively re-label the importance of feature channels, which is a mechanism for enhancing feature representation in a convolutional neural network. The SE module re-labels the importance of each channel by learning the interdependence between different channels, so that feature information can be more effectively utilized. Its workflow consists of four steps: (1) Squeeze: the input tensor x (whose shape size is [batch, channels, _], _ is any size) is squeezed into a two-dimensional tensor [batch, channels] by spatial dimension mean pooling three-dimensional tensor x; (2) Channel compression: when the tensor passes through the ReLU activation function of the fully connected layer fc1, the channel number is proportionally compressed according to the preset hyperparameter reduction; (3) Channel recovery: the original channel number is restored through linear transformation by fc2 and sigmoid activation function, and training weights between 0 and 1 are generated; (4) Excitation: after reshaping the weights into a [batch, channels, 1] three-dimensional vector, they are multiplied with the input x channel by channel, and the re-labeled feature x*se is output. This process significantly improves the representation ability of key features by adaptively learning the interdependence between channels. Among them, the ReLU activation function is ReLU(x) = max(0, x), and the sigmoid activation function is
[0041] The present application introduces an SE module in the REBlock to realize a channel attention mechanism, and the core innovation lies in that it enables the model to adaptively learn and weight those endogenous features that are proven to be crucial for classification in traditional recognition methods, such as amplitude or phase distribution characteristics reflecting the difference between modulation types. This mechanism effectively replaces the cumbersome, expert knowledge-dependent manual feature calculation process, realizes the automatic and dynamic re-labeling of key features, and thus greatly improves the generality and efficiency of the model.
[0042] In addition, the present application adopts a three-stage progressive channel expansion feature extraction strategy (with a maximum channel number of 128), which solves the problem of insufficient depth of shallow convolutional networks in extracting complex high-dimensional signal features. This deep and wide feature extraction design ensures that the model has obtained sufficient rich and high expression power of signal representation before subsequent time series modeling.
[0043] S2: Adjust the dimension of the feature tensor, exchange the channel and time step dimensions, and obtain a tensor with a shape of [batch, time_steps, 128] to adapt to the input requirements of BiLSTM (BiLSTM);
[0044] S3: The adjusted feature tensor is input into a bidirectional LSTM network for forward and backward time series modeling, respectively, outputting time series features with a shape of [batch, time_steps, hidden_size*2], where hidden_size is the number of hidden units of the LSTM;
[0045] The LSTM network effectively alleviates the gradient vanishing / explosion problem by introducing gating mechanisms (input gate, forget gate, output gate) and memory cells (CellState), and can learn the long-term dependencies between modulation signal data. Its output expression is as follows:
[0046] f t =sigmoid(W f [h t-1 ,x t ]+b f )
[0047] In the formula, W f is the weight matrix, b f is the bias term, h t-1 is the memory cell processing value at the previous time, and the output range is [0, 1], 0 represents complete forgetting, and 1 represents complete retention
[0048] In the S3 process, the modulation signal features are input into a bidirectional LSTM network. Bidirectional LSTM can extract information from both forward and backward time series of signals to capture historical and future information features of modulation signals, which is crucial for wireless signals with strong time series dependence. This allows the model to capture the global time dependence of different modulation signals, thereby enhancing classification performance.
[0049] S4: The time series features output by the BiLSTM are input into a multi-head attention (Multi-Head Attention) layer, which learns the feature correlation in different subspaces through parallel computation of multiple attention heads and weighted aggregation of key time series information. The Attention mechanism assigns weights to each time step, allowing the model to focus on the most representative moments in the signal sequence. This not only improves the model's interpretability, but also enhances its classification ability in complex signal and noise backgrounds. Its specific expression is as follows
[0050]
[0051] where Q is the query matrix, representing the target position for which attention needs to be calculated. K is the key matrix, used for reference positions matching the query. V is the value matrix, containing actual feature information, which is finally weighted and aggregated. The dot product of the query and key is calculated through matrix multiplication to obtain a similarity score matrix QK T , the final weighted aggregation of the value matrix V is calculated as follows: Scaling is performed to prevent the gradient from vanishing due to the large dot product value. Finally, the softmax function is normalized, and the weighted aggregation is used to calculate the attention
[0052] S5: The output of the multi-head attention layer is averaged pooled along the time step dimension to obtain a global feature representation with a shape of [batch, embed_dim];
[0053] S6: The global feature representation is input into a fully connected classification layer, and the probability distribution of each modulation type is output through the softmax activation function, completing automatic modulation recognition. The end of the fully connected layer network, after feature extraction by the convolutional layer and dimension reduction by the pooling layer, converts the feature map into a one-dimensional vector and inputs it into the fully connected layer for classification or regression operation.
[0054] The present application uses a multi-head attention mechanism to process time sequence features in a parallel subspace modeling manner. This is equivalent to the model adopting multiple independent and differentiated feature perspectives in the decision-making process, thereby avoiding excessive dependence on a single signal feature. This multi-perspective feature aggregation mechanism significantly improves the model's ability to finely distinguish various modulation signals. Secondly, the long-range temporal dependence captured by the bidirectional LSTM and the weighted aggregation of key temporal information by the multi-head attention mechanism enable the model to effectively smooth the influence of transient noise and outliers in the signal. This makes the ResAttAMR model exhibit higher robustness and stable generalization ability than simple deep learning architectures when facing unknown data and noise interference, helping to avoid the overfitting problem that traditional shallow networks are prone to.
[0055] In another aspect, the present application also provides a low-complexity modulation signal intelligent recognition system, comprising the following modules:
[0056] A data preprocessing module is used to collect wireless communication signal data and perform preprocessing;
[0057] A feature extraction module is used to input the preprocessed signal data into a feature extractor composed of three REBlock modules, a max pooling layer and a dropout layer, to gradually extract signal features and adjust the number of channels;
[0058] A time sequence information module is used to input the signal feature tensor after adjusting the dimension into a BiLSTM network for forward and backward time sequence modeling, respectively, and input the time sequence features output by the BiLSTM into a multi-head attention layer to learn the feature correlation of different subspaces and weighted aggregate key time sequence information;
[0059] The prediction output module is used for average pooling of the output of the multi-head attention layer along the time step dimension to obtain a global feature representation, and the global feature representation is output through a softmax activation function after a fully connected classification layer to complete automatic modulation recognition.
[0060] Embodiments
[0061] The model training convergence test is performed through the S0-S6 steps, and the Loss is used as an evaluation signal to predict the model training performance, the training specific parameter settings are shown in Table 1, and 11 types of modulation signals are classified. Figure 5 As shown in the training loss curve obtained by learning and optimizing the algorithm to effectively reduce the loss function through training of the input sample data, the training loss gradually decreases over time, proving that the present application can be well trained in modulation signal intelligent identification.
[0062] Table 1 Network parameter configuration table
[0063]
[0064]
[0065] The trained model parameters are used for modulation mode prediction of the test set data and compared with the true value, and it can be seen from Figure 6 It can be seen that the intelligent recognition accuracy of the model for most modulation types is almost close to 100%, and the overall accuracy of the test set reaches 90%, meeting the performance requirements.
[0066] The application analyzes the recognition accuracy performance of different modulation modes under different signal-to-noise ratios (SNR). For digital modulation signals, high-order modulation such as 64QAM has an accuracy rate of more than 90% under high SNR conditions, such as SNR≥10dB, which benefits from the clear and distinguishable constellation distribution characteristics under high SNR conditions. However, under low SNR conditions, such as SNR≤0dB, 64QAM is easily confused with 16QAM, mainly because the phase characteristics of the two are similar under low SNR, making it difficult for the model to accurately distinguish the mapping positions of each symbol. Binary phase shift keying (BPSK) and QPSK rely on obvious phase jump characteristics, and the accuracy rate is 95% in the whole SNR range. Among them, the accuracy rate of BPSK reaches 100% when SNR=12dB, and the simple two-phase modulation characteristics make the signal features in the time-frequency domain easy to capture and identify; for analog modulation signals, amplitude modulation-double side band modulation (AM-DSB) and amplitude modulation-single side band modulation (AM-SSB) have high mutual error rates when SNR≤5dB due to similar frequency spectrum structures, at which time the noise interference seriously affects the subtle differences of the amplitude envelopes of the two, making it difficult for the model to distinguish. However, as the SNR improves, the model can effectively distinguish between the two by extracting the change trend and bandwidth difference characteristics of the amplitude envelope, and the accuracy rates are 97.5% and 85.5% respectively when SNR=18dB. Wide band frequency modulation (WBFM) has a significant anti-interference characteristic, and uses frequency modulation to effectively suppress the influence of noise on the signal, so that the accuracy rate is stable at more than 98% in the whole SNR range, and it shows good reliability in actual communication scenarios.
[0067] Furthermore, as shown in Figure 7 , the application gives a system overall confusion matrix diagram, which can accurately locate high-frequency misjudgment, such as the recognition of signals such as AM-DSB and AM-SSB, 16QAM and 64QAM. The application gives part of the confusion matrix diagram under different SNR, as shown in Figure 8 , and the specific results are as follows:
[0068] SNR=0dB, the model shows performance degradation, the overall accuracy of the present application is 82%. The classification effect of binary phase shift keying (BPSK) is poor, a large number of them are misclassified as QPSK. The classification accuracy of Gauss frequency shift keying (GFSK) is still high, and there is classification confusion between 64QAM and 16QAM. QPSK performs relatively well, but the overall results show that there is a significant classification problem;
[0069] SNR=4dB, the accuracy of the model is 85%, and the classification of BPSK, GFSK and QPSK performs well.
[0070] 64QAM and 16QAM are often confused, especially the case of misclassifying QAM16 as QAM64 occurs 80 times. GFSK achieves perfect classification, while 64QAM has poor recognition performance;
[0071] SNR=8dB, the accuracy of the model reaches 85%, and the classification of BPSK, GFSK and QPSK performs extremely well.
[0072] There is still classification confusion between 64QAM and 16QAM. GFSK achieves perfect classification, and among high-order modulation modes, QPSK has the highest prediction accuracy;
[0073] SNR=12dB, the overall accuracy of the model is 85%, and the prediction of BPSK and GFSK is completely correct. There is obvious classification confusion between 64QAM and 16QAM. QPSK has excellent classification performance, indicating that its classification result is reliable;
[0074] SNR=16dB, the accuracy of the model reaches 84%, and the classification results of BPSK and GFSK are excellent. There is still classification confusion between 64QAM and 16QAM, which leads to the decrease of their precision and recall rate. However, the overall performance of the model of the present application remains balanced and performs well.
[0075] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or blocks. These computer program instructions can also be stored in a computer readable storage medium that can guide the computer or other programmable data processing devices to work. Figure 1 one flow or multiple flows and / or blocks Figure 1 code of the functions specified in one block or multiple blocks.
[0076] The above descriptions are only the preferred embodiments of the present application, not intended to limit the present application in other forms. Any person skilled in the art can make changes or modifications to the equivalent embodiments with the disclosed technical contents. However, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the present application and according to the technical essence of the present application still belong to the protection scope of the present application.
Claims
1. A low-complexity intelligent recognition method for modulation signals, characterized in that, Includes the following steps: S1. Collect wireless communication signal data and perform preprocessing; S2. Input the preprocessed signal data into the feature extractor, which consists of three REBlock modules, a max pooling layer and a dropout layer, to gradually extract signal features and adjust the number of channels. S3. After adjusting the dimension of the signal feature tensor, input it into the BiLSTM network to perform forward and backward time-series modeling respectively. S4. Input the temporal features output by BiLSTM into the multi-head attention layer. By computing multiple attention heads in parallel, learn the feature correlations of different subspaces and aggregate key temporal information in a weighted manner. S5. Perform average pooling on the output of the multi-head attention layer along the time step dimension to obtain the global feature representation; S6. Input the global feature representation into the fully connected classification layer, and output the probability distribution of each modulation type through the softmax activation function to complete the automatic modulation recognition.
2. The low-complexity intelligent identification method for modulation signals according to claim 1, characterized in that, The preprocessing includes normalization, denoising, and segmentation, generating an input tensor x with shape [batch, channels1, time_steps], where the parameters represent data blocks, channel data, and time steps, respectively.
3. The method for intelligent recognition of low-complexity modulation signals according to claim 2, characterized in that, The feature extractor consists of three REBlock modules, a max pooling layer, and a dropout layer. It gradually extracts signal features and adjusts the number of channels, gradually increasing the number of channels. At the same time, it reduces the time dimension through max pooling and outputs a feature tensor with the shape [batch, channels2, time_steps].
4. The low-complexity intelligent identification method for modulation signals according to claim 3, characterized in that, The REBlock module consists of convolutional layers, batch normalization layers, squeeze-excitation network (SE) module, and shortcut path. The SE module adaptively adjusts the feature channel weights through a channel attention mechanism to enhance the expression of key features. When the number of input / output channels or stride does not match, use 1×1 convolution to adjust the dimensions of the shortcut path.
5. The low-complexity intelligent identification method for modulation signals according to claim 4, characterized in that, The SE module recalibrates the importance of each channel by learning the interdependencies between different channels. Its workflow consists of four steps: (1) Squeezing: The input tensor x is compressed into a two-dimensional tensor [batch, channels] by mean pooling of the spatial dimension; (2) Channel compression: When the tensor passes through the ReLU activation function of the fully connected layer fc1, the number of channels is compressed proportionally according to the preset hyperparameters; (3) Channel recovery: The original number of channels is recovered by linear transformation of fc2 and sigmoid activation function, and training weights between 0 and 1 are generated; (4) Activation: The weights are reshaped into a three-dimensional vector [batch, channels, 1] and multiplied with the input x channel by channel to output the recalibrated features.
6. The low-complexity intelligent identification method for modulation signals according to claim 5, characterized in that, Step S3 is specifically implemented as follows: swapping the channel and time step dimensions to obtain a tensor of shape [batch, time_steps, channels2]. The adjusted feature tensor is input into a bidirectional LSTM network to perform forward and backward temporal modeling, and outputs temporal features of shape [batch, time_steps, hidden_size*2], where hidden_size is the number of hidden units in the LSTM.
7. A low-complexity intelligent modulation signal recognition system, used to implement the intelligent modulation signal recognition method according to any one of claims 1 to 6, characterized in that, Includes the following modules: The data preprocessing module is used to acquire wireless communication signal data and perform preprocessing. The feature extraction module is used to input the preprocessed signal data into the feature extractor, which consists of three REBlock modules, a max pooling layer, and a dropout layer, to gradually extract signal features and adjust the number of channels. The temporal information module is used to adjust the dimension of the signal feature tensor and input it into the BiLSTM network to perform forward and backward temporal modeling respectively. The temporal features output by BiLSTM are then input into the multi-head attention layer to learn the feature correlations of different subspaces and to aggregate key temporal information in a weighted manner. The prediction output module is used to perform average pooling on the output of the multi-head attention layer along the time step dimension to obtain a global feature representation. The global feature representation is then passed through a fully connected classification layer, and the probability distribution of each modulation type is output through the softmax activation function to complete automatic modulation recognition.