Low signal-to-noise ratio modulation signal identification method and system fusing global-local information

Abstract

The invention provides a low-signal-to-noise-ratio modulation signal identification method and system fusing global-local information. I / Q data of signals of different modulation types are input into a ResNet network and a Transform network. The method comprises the following steps: respectively providing local feature information res and global feature information fc, combining the local feature information res and the global feature information fc to obtain a joint feature output of different signal modulation types so as to construct a ResTransformerAtt model, designing a modulation signal identification model for learning global-local context features by utilizing the combination of ResNet and Transformer networks, and identifying the global-local context features in the modulation signal identification model. Features extracted by the two deep learning models are utilized to enhance feature representation, so that signal identification precision and generalization application capability are improved, and stable communication identification under multiple types of signals in a complex electromagnetic environment is completed.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal recognition, and relates to a method and system for recognizing a low signal-to-noise ratio modulation signal by integrating global and local information. Background Art

[0002] The modulation pattern recognition of communication radiation source signals is of great significance both in military and civilian fields. Only by accurately identifying the modulation pattern of the detection signal can the signal modulation parameters be estimated and provide a strong reference for subsequent signal analysis and signal processing. At present, the main methods of traditional modulation pattern recognition can be divided into two categories, namely, modulation recognition methods based on likelihood ratio detection and modulation recognition methods based on feature extraction; Among them, the modulation recognition method based on likelihood ratio detection analyzes the signal characteristics through probability theory, first observes the signal waveform to be identified, and assigns it a candidate modulation type. Then, the true modulation method is determined through the similarity principle. Specifically, this method is based on the likelihood function (Likelihood Function, LF) of the received signal, and makes a decision by comparing the likelihood ratio (Likelihood Ratio, LR) with the corresponding threshold. Its modulation recognition process is a multi-hypothesis testing problem.

[0003] In the 1980s, the average likelihood ratio test theory was first proposed by Kim K et al. and applied to the classification and recognition of modulated signals, overcoming the problem of missing parameters in the maximum likelihood ratio function. [1] . Boiteau D et al. [2] A generalized likelihood ratio test theory is proposed. The generalized likelihood ratio test theory reduces the computational complexity. And when performing non-constant envelope modulation recognition and classification, the performance of the generalized likelihood ratio test theory is better than that of the average likelihood ratio test theory. The hybrid likelihood ratio test theory combines the average likelihood ratio test theory and the generalized likelihood ratio test theory, which can be used to solve the problems of mesh constellations and high-dimensional calculations.

[0004] Modulation recognition methods based on feature extraction rely on expert knowledge in the electromagnetic field and manually set signal feature extraction models. Common signal features include: (1) instantaneous features; (2) statistical features; (3) transform domain features. Feature extraction is a data mapping process, which maps the original signal data to a specific feature space. Instantaneous features describe signal characteristics from the perspectives of electromagnetic signal amplitude, phase, and frequency. Nandi et al. [3]Using 9 different types of signal instantaneous information, the modulation types of 13 different communication signals were identified by selecting the classifier threshold. The statistical features mainly include high-order moments and high-order cumulants. This feature uses appropriate orders, conjugate variables and delay amounts to suppress phase deviation and fading effects for fading channels. Reichert et al. [4] Firstly, the high-order cumulants are used for signal modulation type recognition, which shows the feasibility of high-order cumulants in the field of modulation type recognition and is widely used. The transform domain features calculate the transform coefficients and statistics of the original signal to form a multi-scale feature vector.

[0005] Although the modulation recognition method based on likelihood ratio detection theoretically guarantees the best classification results, it still has the following disadvantages: (1) This method is often only applicable to specific types of modulation signals, not to new or other types of modulation signals, and the recognition range is relatively narrow. (2) The recognition process requires too many known parameters, including carrier frequency, carrier phase, baud rate, signal-to-noise ratio, etc. The determination of parameters often requires a lot of experiments and experience, which is relatively difficult; in addition, due to the existence of unknown parameters, the calculation expression of the likelihood ratio function is very complex, the amount of calculation is large and difficult to handle. If the likelihood ratio function is simplified, it will lead to the loss of classification information and reduce the recognition performance.

[0006] Compared with likelihood ratio detection, artificial signal features require less prior information about the signal and have lower algorithm complexity, but the feature contains a single signal attribute and has limited data characterization capabilities. This method requires pre-selection and extraction of appropriate features. Different modulation methods may require different features to describe their essential characteristics, so signal analysts are required to have rich professional knowledge and experience. In addition, during channel transmission, the signal will be affected by noise and distortion, resulting in the feature extraction algorithm being unable to accurately extract the signal features, thereby reducing the accuracy of modulation recognition. And when a new modulated signal appears, its features will be difficult to extract and process. Summary of the invention

[0007] In view of the problems existing in the prior art, the present invention provides a low signal-to-noise ratio modulated signal recognition method and system that integrates global-local information. A modulated signal recognition model that learns global-local context features is designed by combining a deep residual network (ResNet) and a Transformer network. The features extracted by the two deep learning models are used to strengthen the feature representation, thereby improving the accuracy of signal recognition and the generalization application capability, and completing stable communication recognition of multiple types of signals in complex electromagnetic environments.

[0008] The present invention is achieved through the following technical solutions: A low signal-to-noise ratio modulation signal recognition method integrating global-local information, comprising: The in-phase and quadrature (IQ) data of different modulation types are input into the ResNet network, firstly passed through the Convolutional Block Attention Module (CBAM) to obtain the weighted feature x, and then extracted through the ResNet network to obtain the local feature information res; The IQ data of signals of different modulation types are input into the Transformer network. First, the long short-term memory model (LSTM) is used to extract the temporal features. Then, the encoder module of the Transformer network extracts the global features of the IQ data of signals of different modulation types. After normalization, the global feature information fc is obtained. The local feature information res and the global feature information fc are combined to obtain the joint feature output of different signal modulation types, thereby constructing a modulated signal recognition (ResTransformerAtt) model. Based on the ResTransformerAtt model, the recognition of low signal-to-noise ratio modulated signals that fuse global-local information is realized.

[0009] Preferably, the IQ data of signals of different modulation types are input into the ResNet network to extract local feature information res, specifically: On the ResNet network, the IQ data of signals of different modulation types first pass through the CBAM module to obtain the weighted feature x of the comprehensive attention weight, and then pass through ResNet extraction to obtain the local feature information res; the local feature information res is input into the Dropout layer to alleviate the overfitting phenomenon and obtain the simplified local feature information res.

[0010] Preferably, the CBAM module includes a channel attention module and a spatial attention module.

[0011] The comprehensive attention weight is obtained by multiplying the attention weights of the channel attention module and the spatial attention module.

[0012] Preferably, the process of acquiring the attention weight of the channel attention module is: The elements in each channel of the IQ data features of the input signals of different modulation types are averaged to obtain the average value of each channel, thereby obtaining a global average pooling vector reflecting each channel; Extract the global maximum value in each channel of the IQ data features of signals of different modulation types, so as to obtain the most significant maximum pooling feature vector reflecting each channel; The maximum pooling vector and the global average pooling vector are nonlinearly transformed through a shared multi-layer perceptron to generate the attention weights of each channel of the maximum pooling vector and the global average pooling vector respectively; Add the attention weights of each channel of the maximum pooling vector and the global average pooling vector, fuse the two statistical information, and get the comprehensive weight of each channel; The comprehensive weight is input into the Sigmoid activation function for normalization to generate the attention weight of the channel attention module.

[0013] Preferably, the process of acquiring the attention weight of the spatial attention module is: Perform maximum pooling and average pooling operations on the input features to obtain two different spatial information vectors; Two different spatial information vectors are stacked to obtain a vector containing different pooling information for each spatial position; Use a convolutional layer to fuse the vectors containing different pooled information for each spatial position, learn the relationship between spatial features, and generate an attention weight for each position; The attention weight generated at each position is input into the Sigmoid activation function for normalization to generate the attention weight of the spatial attention module.

[0014] Preferably, the IQ data of signals of different modulation types are input into the Transformer network to extract the global feature information fc, specifically: On the Transformer network, the IQ data of signals of different modulation types are first subjected to time series feature extraction by the LSTM module, and then subjected to the Transformer network coding module coder to extract the global features of the IQ data of signals of different modulation types to obtain global feature information codex, and the global feature information codex is subjected to a linear layer Fc to obtain normalized global feature information fc; the linear layer Fc includes a normalization layer and a linear layer.

[0015] Preferably, the Long Short-Term Memory (LSTM) network is a special type of Recurrent Neural Network (RNN) that aims to solve the gradient vanishing or gradient exploding problems encountered by traditional RNN when processing long sequence data.

[0016] The process of extracting time series features is as follows: The IQ data of signals of different modulation types are used as the input sequence of the LSTM module. The input sequence is controlled by the forget gate, input gate and output gate in the LSTM module. The features of some sequences are discarded from the cell state, the features of some new sequences are written into the cell state, and the features of some sequences are output to the hidden state. The cell state in the LSTM module will gradually contain the local and global feature information of the sequence. After the LSTM module processes the entire input sequence of IQ data of signals of different modulation types, the extracted hidden state is used as the representation of the timing feature.

[0017] Furthermore, the input sequence includes: external input at the current moment, hidden state that conveys the temporal features of the previous time step, long-term temporal features that preserve information, and the cell state at the previous moment that is updated at each time step.

[0018] The forget gate is calculated to determine which temporal features need to be discarded from the cell state. The output value of the forget gate is between 0 and 1, where 0 means completely forgotten and 1 means completely retained; Calculate the input gate and candidate cell states. The input gate determines how many new temporal features are written into the cell state in the combination of the hidden state and the input information at the current moment, and the candidate cell state provides new temporal features; Update the cell state through the forget gate and input gate, retaining and discarding the temporal features; Calculate the output gate and hidden state. The output gate determines which parts of the cell state will be used as the output of the temporal features; the hidden state is the output of the current moment and is also the temporal feature passed to the next moment.

[0019] Preferably, a ResTransformerAtt model is constructed, specifically: The normalized global feature information fc and local feature information res will be merged in one dimension to obtain the joint feature output of different signal modulation types, thereby constructing the ResTransformerAtt model.

[0020] Preferably, based on the ResTransformerAtt model, the recognition of low signal-to-noise ratio modulated signals that fuse global-local information is realized, and the specific process is: The joint feature output of different signal modulation types is converted into an output form directly related to the modulation recognition task through the linear layer fc1 to obtain a normalized classification result; the classification result is a probability distribution table, whose columns represent the number of IQ data in each batch input into the ResTransformerAtt model; the number of elements contained in each row is the number of modulation types to be recognized, and each element represents a modulation type; the value of each element is between [0,1] and the sum of all elements is 1; the recognition type corresponding to the element with the largest value in each row is the final prediction result of the IQ data of the input ResTransformerAtt model corresponding to the row in the probability distribution table.

[0021] Preferably, the ResNet network includes 6 residual modules, 1 flattening layer and 2 fully connected layer modules; The Transformer network includes two forward convolution modules, a maximum pooling layer and a long short-term memory network model.

[0022] A low signal-to-noise ratio modulation signal recognition system integrating global-local information, comprising: The ResNet network processing module is used to input the IQ data of signals of different modulation types into the ResNet network, first pass through the CBAM module to obtain the weighted feature x, and then extract the local feature information res through the ResNet network; The Transformer network processing module is used to extract the time series features of the IQ data of signals of different modulation types through the LSTM module, and then extract the global features of the IQ data of signals of different modulation types through the Transformer network coding module encoder, and obtain the global feature information fc after normalization; The model building module is used to merge the local feature information res and the global feature information fc to obtain the joint feature output of different signal modulation types, thereby constructing the ResTransformerAtt model. The signal recognition module is used to recognize low signal-to-noise ratio modulated signals that integrate global-local information based on the ResTransformerAtt model.

[0023] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention provides a low signal-to-noise ratio modulation signal recognition method and system that integrates global-local information. The spatial relationship of the signal is learned through the local perception of the convolution kernel of ResNet, and the system has a strong ability to extract local context features. However, the local characteristics of the convolution layer limit the network from capturing global information. Transformer has a strong ability to model global information and can better handle long-range dependent features in long time series. Therefore, based on the above analysis, the present invention designs a modulation signal recognition model based on ResNet and Transformer, and uses ResNet and Transformer to design and learn global-local context to strengthen feature representation, which is of great significance in improving the stability of the modulation mode recognition algorithm and the accuracy of feature extraction. The presence of noise will bring random errors, and will also cover up the characteristics of the signal, causing the accuracy of the signal to decrease, affecting the effect of signal processing. The CBAM attention mechanism can enable the model to focus on more critical information in feature extraction, and eliminate and suppress noise and irrelevant information. At the same time, when there is noise or multipath effect in the signal, the LSTM module can also filter out irrelevant disturbances and retain important time series information. Therefore, introducing CBAM as the input of the ResNet model can effectively suppress noise. LSTM has a strong ability to process sequence data and can effectively capture long-term dependencies in data through its "memory" mechanism (cell state and hidden state). This is particularly important for modulation recognition tasks in wireless communications, because the recognition of modulation types not only depends on the signal characteristics of the current moment, but also involves contextual information of the previous and next moments. At the same time, when there is noise or multipath effect in the signal, the LSTM module can filter out irrelevant disturbances and retain important time series information. Therefore, introducing the LSTM module as the input of the Transformer model can simultaneously learn the timing features and global patterns in the signal, significantly enhancing the ability of feature extraction. The simulation experiment verifies the modulation type recognition effect of the model based on the above-mentioned improved ResNet and Transformer combination on signals in low signal-to-noise ratio and complex channel environments. The results show that the model has strong robustness and has high recognition accuracy under low signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is the ResTransformerAtt model structure diagram; Figure 2 This is the structure diagram of the convolutional attention model; Figure 3 This is the structure diagram of the channel attention model; Figure 4 This is the structure diagram of the spatial attention model; Figure 5 This is the LSTM module structure diagram; Figure 6 This is the structural diagram of the ResNet model; Figure 7 This is the structural diagram of the residual module in the ResNet model; Figure 8 The structure of the Transformer model; Fig. 9 is the overall recognition rate of the ResTransformerAtt model at different signal-to-noise ratios; Fig.10 The confusion matrix of the signal modulation type recognition result of the ResTransformerAtt model; Fig.11 This is the confusion matrix when the ResTransformerAtt model has the highest recognition rate for signal modulation type. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below in conjunction with specific embodiments, which are intended to explain the present invention rather than to limit it.

[0026] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0027] A modulated signal recognition model based on deep residual network (ResNet) and Transformer model is proposed. It is planned to combine the ResNet network with convolutional block attention module (CBAM) and the Transformer network with long short-term memory (LSTM) to design a model for learning global-local context features. The features extracted by the two deep learning models are used to strengthen the feature representation, thereby improving the accuracy of signal recognition and generalization application ability, and completing stable communication recognition of multiple types of signals in complex electromagnetic environments.

[0028] 1. Training and testing datasets RML2016.10a The dataset includes 11 signal types, 8 digital signals: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK, and 3 analog signals: AM-DSB, AM-SSB, and WBFM. The dataset contains a total of 220,000 data, with signal-to-noise ratios (SNR) ranging from -20dB to +18dB. Each signal has 128 complex floating-point time IQ samples and is generated in a harsh simulated propagation environment, which is corrupted by AWGN, multipath fading, sampling rate offset, and center frequency offset, similar to the actual environment. The specific information of the dataset is shown in Table 1 below:

[0029] Among them, 1000 modulation signal samples of a single modulation mode and a signal-to-noise ratio are generated, each sample has data of two signals, I and Q, that is, in-phase and orthogonal signal data, each sample contains 8-16 code elements, and each signal contains 128 sampling points.

[0030] 2. Construction of modulation signal recognition model The ResTransformerAtt model structure of the experimental design is as follows Figure 1 As shown in Figure 1, the IQ data of signals of different modulation types are input into the designed ResNet network and Transformer network in batches to jointly extract features.

[0031] In the ResNet network, the input IQ data first passes through the CBAM module, such as Figure 2 As shown in Figure 1. This module includes a channel attention module and a spatial attention module. The obtained comprehensive attention weighted feature x is then passed through ResNet to obtain the normalized information res extracted by ResNet that pays more attention to the local features of the signal modulation type. Res is input into the Dropout layer to alleviate overfitting and simplify the network.

[0032] Among them, the channel attention module and spatial attention module structures in the CBAM module are as follows: Figure 3 and Figure 4 shown. (1) Channel attention module: The channel attention module focuses on the channel information in the feature map to learn the importance weight of each channel. The module structure of the channel attention module is as follows: Figure 3 shown.

[0033] The channel attention module consists of the following steps: a. Average pooling: average the elements in each channel of the input feature to get the average value of each channel. The vector obtained in this way can reflect the global features of each channel. b. Max pooling: Extract the global maximum value of each channel of the input feature and capture the most significant features of the channel. The vector obtained in this way can reflect the most significant features of each channel. c. Shared multi-layer perceptron: The results of maximum pooling and average pooling are transformed nonlinearly through a shared multi-layer perceptron to learn the weight of each channel. It contains 2 fully connected layers: the first fully connected layer compresses the last dimension of the vector obtained by average pooling and maximum pooling (usually reduced to 1 / 4 of the original). In this fully connected layer, the ReLU activation function is usually used to introduce nonlinearity; the second fully connected layer maps the last dimension of the compressed vector back to the original dimension to generate attention weights. In this fully connected layer, the Sigmoid activation function is usually used to limit the attention weights between 0 and 1. d. Addition operation (+): Add the results of the maximum pooling and average pooling branches after processing by the shared multi-layer perceptron, fuse the two statistical information, and obtain the comprehensive weight of each channel.

[0034] e. Through the Sigmoid activation function: Apply the Sigmoid activation function to the result of the addition to map the channel weights to the range of (0,1). The channel weights are normalized through the Sigmoid function so that they can be directly used to adjust the strength of the input features.

[0035] f. Output channel attention weights: Channel attention weights are the importance scores of each channel and are used to reweight the input features to highlight features that are useful for classification or object detection tasks.

[0036] (2) Spatial attention module: The spatial attention module focuses on the spatial information in the feature map to learn the importance weight of each spatial position. The module structure of the spatial attention module is as follows: Figure 4 shown.

[0037] The spatial attention module consists of the following steps: a. Maximum pooling and average pooling: Perform maximum pooling and average pooling operations on the input features respectively to obtain two different spatial information vectors. b. Stacking: Stack the two vectors obtained by maximum pooling and average pooling to form a new vector. This vector contains different pooling information for each spatial position. c. Convolutional layer: Use a convolutional layer to fuse the information of maximum pooling and average pooling, learn the relationship between spatial features, and generate an attention weight for each position. d. Through the Sigmoid activation function: Use the Sigmoid activation function to normalize the convolutional layer output to the (0,1) range and generate attention weights for each spatial position. Each value represents the importance of the corresponding spatial position.

[0038] e. Output spatial attention weight: Weight each spatial position of the input feature, highlight the key spatial positions, suppress unimportant areas, make the model more focused on specific spatial areas, and improve the effects of tasks such as classification and object detection On the Transformer network, the IQ data of signals of different modulation types first pass through the LSTM module to capture the time series features between the data, and then the time series features are input into the Transformer encoding module coder to extract the global features of the signal, and obtain feature information codex that focuses more on the global modulation type of the signal. Codex passes through a linear layer Fc to obtain normalized feature information fc that focuses more on the global modulation type of the signal. The linear layer Fc includes a normalization layer and a linear layer. fc then merges the extracted features with res in dimension 1 to obtain the joint feature output of the extracted signal modulation type, forming the ResTransformerAtt model.

[0039] Based on the ResTransformerAtt model, the recognition of low signal-to-noise ratio modulated signals that integrate global and local information is realized. The specific process is as follows: The linear layer fc1 is used to convert the joint feature output of different signal modulation types into an output form directly related to the modulation recognition task, and a normalized classification result is obtained; the classification result is a probability distribution table, whose columns represent the number of IQ data input into the ResTransformerAtt model in each batch; the number of elements contained in each row is the number of modulation types that need to be recognized, and each element represents a modulation type; The value of each element is between [0,1] and the sum of all elements is 1; the recognition type corresponding to the element with the largest value in each row is the final prediction result of the IQ data of the input ResTransformerAtt model corresponding to the row in this probability distribution table.

[0040] The core operation of the linear layer fc1 is a linear transformation , the linear layer transforms the input high-dimensional feature vector X into a vector of length C, i.e., output Z, through the weight matrix W and bias b; each output value Z represents the category score or probability of the model for the corresponding category.

[0041] The main functional networks ResNet and Transformer in the ResTransformerAtt model are introduced as follows: The structure of the ResNet model is as follows Figure 6 As shown in Figure 2, the model consists of 6 residual modules (ReStk0-ReStk5), 1 flattening layer (Flat) and 2 fully connected layer modules (fc3 and fc5).

[0042] The residual module is used to extract the modulation type features of the input signal, and the flattening layer converts the multi-dimensional features output by the pooling layer into a one-dimensional vector so as to be connected to the fully connected layer. These two fully connected layers are used to improve the accuracy of feature extraction and the ability of network learning. fc3 contains a linear layer, a SELU activation layer and a Dropout layer. fc5 contains only a linear layer.

[0043] The structure of the residual module is as follows Figure 7 As shown in the figure, it consists of two identical residual submodules. The input first passes through the first residual submodule. First, it passes through a convolution layer Conv1 with a convolution kernel size of 1×2 and a step size of 1 to obtain an output feature x1, then passes through a convolution layer Conv1d with a convolution kernel size of 3×2 and a step size of 1, then passes through a ReLU activation layer, and then passes through a convolution layer Conv1d with a convolution kernel size of 3×2 and a step size of 1. The output feature obtained is added to x1 to obtain a new feature, recorded as x2. x2 then passes through a ReLU layer to accelerate the training process and reduce the problem of gradient disappearance, and obtains the optimized feature x3, which is input into the next residual submodule. After the feature extraction process is the same as the above submodule, the feature x5 obtained is passed through a pooling layer pool to obtain the final feature extracted by a complete residual module. Because the experiment needs to extract some obvious features from the data, the maximum pooling method is selected to remove redundant information, remove noise, and better retain texture features. The structure of the Transformer model of the experimental design is as follows Figure 8 As shown. The input first passes through two forward convolution modules, each module contains a convolution layer Conv1 with a convolution kernel size of 2×2 and a stride of 1, and a ReLU activation layer. Then it passes through a maximum pooling layer maxpool1d layer. The maxpool1d layer has great advantages in problems with time series characteristics such as speech recognition and text classification, and can better extract important information. In addition, the maxpool1d layer is more robust to noise and small changes and has better robustness. After the input signal is pooled, it is immediately input into the Long Short-Term Memory (LSTM) module, as shown Figure 5As shown. LSTM introduces a special storage unit and gating mechanism to more effectively capture and process long-term dependencies in sequence data. The obtained features are then input into a linear layer and a Dropout layer in turn to alleviate overfitting. The output feature x is then input into the encoder module encoder of the Transformer. At this time, the input feature dimension is 128, the number of self-attention heads of the encoding module is 2, and the number of encoding layers is 1. The features obtained after the encoding model are shaped to prepare for the subsequent input into the linear layer to complete feature extraction. First, its 0th dimension and 1st dimension are exchanged, and then its 1st dimension and 2nd dimension are flattened into one dimension. Finally, the shaped features are output through a linear layer fc1 to output the modulation type features of the signal finally extracted by the Transformer model. The linear layer fc1 contains a linear layer, a ReLU activation layer, and a Dropout layer.

[0044] Experimental Results The data set is divided into a training set and a test set, where the training set data accounts for 80% of the total data. The training set reads 512 data each time, and the content of each data is the IQ two-way signal data of a signal of a certain modulation type under a certain signal-to-noise ratio.

[0045] like Fig. 9 As shown in the figure, when the signal-to-noise ratio (SNR) is greater than 0dB, the recognition rate of the modulation type of the signal is high, reaching 88.21%. This shows that the ResTransformerAtt model can dynamically adjust the extraction of key features of the signal, thereby extracting more accurate and discriminative feature representations.

[0046] like Fig.10 The figure shows the overall recognition confusion matrix of different modulation signals. It can be seen from the figure that the overall recognition rates of AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, PAM4 and QAM16 are all above 60% under low signal-to-noise ratio.

[0047] The horizontal axis of the confusion matrix represents the predicted signal modulation category, the vertical axis represents the actual signal modulation category, the numbers on the diagonal are the probability of correctly predicting the signal modulation category, and the numbers in the other squares are the probability of the signal corresponding to the vertical axis at this moment being incorrectly predicted as a signal of other modulation types, that is, the degree of confusion. For example, the probability of correctly predicting 8PSK is 0.54, the probability of incorrectly predicting it as QAM16 is 0.03, QAM64 is 0.02, QPSK is 0.03, and there is no confusion with other remaining signals.

[0048] like Fig.11The figure shows the confusion matrix when the model recognition rate reaches the highest. ResTransformerAtt reaches the best when SNR=12dB. Except for QAM16, QAM64 and WBFM signals, the recognition rates of other signals are all above 95%.

[0049] References are as follows: [1] Kim K and Polydoros A. Digital modulation classification: theBPSK versus QPSK case[C] / / MILCOM 88, 21st Century Military Communications-What's Possible? Conference record. Military Communications Conference. SanDiego:1988, 2: 431-436. [2] Boiteau D, Martret C L. A general maximum likelihood framework for modulation classification[C] / / Proceedings of the 1998 IEEE InternationalConference on Acoustics, Speech and Signal Processing, ICASSP' 98(Cat.No.98CH36181). Seattle: IEEE, 1998, 2165-2168. [3] Nandi AK, Azzouz E E. Algorithms for automatic modulationrecognition of communication signals[J]. IEEE Transactions on communications,1998, 46(4): 431-436. [4] Reichert J. Automatic classification of communication signals using higher order statistics[C]. IEEE International Conference on Acoustics,1992:221-224. The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any ordinary technician in the industry can smoothly implement the present invention as shown in the drawings and described above. However, any equivalent changes, modifications and evolutions made by technicians familiar with the profession without departing from the scope of the technical solution of the present invention using the technical content disclosed above are all equivalent embodiments of the present invention. At the same time, any equivalent changes, modifications and evolutions made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A method for identifying low signal-to-noise ratio modulated signals by integrating global and local information, characterized in that: include, The IQ data of signals of different modulation types are input into the ResNet network, firstly passed through the CBAM module to obtain the weighted feature x, and then extracted through the ResNet network to obtain the local feature information res; The IQ data of signals of different modulation types are input into the Transformer network. The LSTM module is used to extract the temporal features. Then, the encoder module of the Transformer network is used to extract the global features of the IQ data of signals of different modulation types. After normalization, the global feature information fc is obtained. The local feature information res and the global feature information fc are combined to obtain the joint feature output of different signal modulation types, thereby constructing a modulation signal recognition model; Based on the modulation signal recognition model, the recognition of low signal-to-noise ratio modulation signals that integrate global and local information is achieved.

2. The method for identifying low signal-to-noise ratio modulated signals by integrating global and local information according to claim 1, characterized in that: The IQ data of signals of different modulation types are input into the ResNet network to extract the local feature information res, specifically: On the ResNet network, the IQ data of signals of different modulation types first pass through the CBAM module to obtain the weighted feature x of the comprehensive attention weight, and then pass through ResNet to extract the local feature information res; the local feature information res is input into the Dropout layer to alleviate the overfitting phenomenon, and the simplified local feature information res is obtained; The CBAM module includes a channel attention module and a spatial attention module; The comprehensive attention weight is obtained by multiplying the attention weights of the channel attention module and the spatial attention module.

3. The method for identifying low signal-to-noise ratio modulation signals by integrating global and local information according to claim 2, characterized in that: The process of acquiring the attention weight of the channel attention module is: The elements in each channel of the IQ data features of the input signals of different modulation types are averaged to obtain the average value of each channel, thereby obtaining a global average pooling vector reflecting each channel; Extract the global maximum value in each channel of the IQ data features of signals of different modulation types, so as to obtain the most significant maximum pooling feature vector reflecting each channel; The maximum pooling vector and the global average pooling vector are nonlinearly transformed through a shared multi-layer perceptron to generate the attention weights of each channel of the maximum pooling vector and the global average pooling vector respectively; Add the attention weights of each channel of the maximum pooling vector and the global average pooling vector, fuse the two statistical information, and get the comprehensive weight of each channel; The comprehensive weight is input into the Sigmoid activation function for normalization to generate the attention weight of the channel attention module.

4. The method for identifying low signal-to-noise ratio modulated signals by integrating global and local information according to claim 2, characterized in that: The process of acquiring the attention weight of the spatial attention module is: Perform maximum pooling and average pooling operations on the IQ data of input signals of different modulation types to obtain two different spatial information vectors; Stack two different spatial information vectors to obtain a vector containing different pooling information for each spatial position; Use a convolutional layer to fuse the vectors containing different pooled information at each spatial position, learn the relationship between spatial features, and generate an attention weight for each position; The attention weight generated at each position is input into the Sigmoid activation function for normalization to generate the attention weight of the spatial attention module.

5. The method for identifying low signal-to-noise ratio modulation signals by integrating global and local information according to claim 1, characterized in that: The IQ data of signals of different modulation types are input into the Transformer network to extract the global feature information fc, which is: On the Transformer network, the IQ data of signals of different modulation types are first subjected to time series feature extraction by the LSTM module, and then subjected to the Transformer network coding module coder to extract the global features of the IQ data of signals of different modulation types to obtain the global feature information codex. The global feature information codex is subjected to a linear layer Fc to obtain the normalized global feature information fc. The linear layer Fc includes a layer normalization layer and a linear layer.

6. The method for identifying low signal-to-noise ratio modulation signals by integrating global and local information according to claim 5, characterized in that: The process of extracting time series features by the LSTM module is as follows: The IQ data of signals of different modulation types are used as the input sequence of the LSTM module. The input sequence is controlled by the forget gate, input gate and output gate in the LSTM module. The features of some sequences are discarded from the cell state, the features of some new sequences are written into the cell state, and the features of some sequences are output to the hidden state. The cell state in the LSTM module will gradually contain the local and global feature information of the sequence. After the LSTM module processes the entire input sequence of IQ data of signals of different modulation types, the extracted hidden state is used as the representation of the timing feature.

7. The method for identifying low signal-to-noise ratio modulation signals by integrating global and local information according to claim 1, characterized in that: The modulation signal recognition model is constructed, specifically: The normalized global feature information fc and local feature information res will be merged in one dimension to obtain the joint feature output of different signal modulation types, thereby constructing a modulated signal recognition model.

8. The method for identifying low signal-to-noise ratio modulation signals by integrating global and local information according to claim 1, characterized in that: The basic modulation signal recognition model realizes the recognition of low signal-to-noise ratio modulation signals that integrate global-local information. The specific process is as follows: The joint feature output of different signal modulation types is converted into an output form directly related to the modulation recognition task through the linear layer fc1 to obtain a normalized classification result; the classification result is a probability distribution table, whose columns represent the number of IQ data in each batch of input modulation signal recognition model; the number of elements contained in each row is the number of modulation types to be recognized, and each element represents a modulation type; the value of each element is between [0,1] and the sum of all elements is 1; the recognition type corresponding to the element with the largest value in each row is the final prediction result of the IQ data of the input modulation signal recognition model corresponding to the row in the probability distribution table.

9. The method for identifying low signal-to-noise ratio modulation signals by integrating global and local information according to claim 1, characterized in that: The ResNet network includes 6 residual modules, 1 flattening layer and 2 fully connected layer modules; The Transformer network includes two forward convolution modules, a maximum pooling layer and a long short-term memory network model.

10. A low signal-to-noise ratio modulated signal recognition system integrating global-local information, based on a low signal-to-noise ratio modulated signal recognition method integrating global-local information according to any one of claims 1 to 9, characterized in that: include, The ResNet network processing module is used to input the IQ data of signals of different modulation types into the ResNet network, first pass through the CBAM module to obtain the weighted feature x, and then extract the local feature information res through the ResNet network; The Transformer network processing module is used to input the IQ data of signals of different modulation types into the Transformer network, firstly extract the time series features through the LSTM module, and then extract the global features of the IQ data of signals of different modulation types through the Transformer network coding module encoder, and obtain the global feature information fc after normalization; The model building module is used to merge the local feature information res and the global feature information fc to obtain the joint feature output of different signal modulation types, thereby constructing a modulated signal recognition model. The signal recognition module is used to realize the recognition of low signal-to-noise ratio modulation signals that integrate global and local information based on the modulation signal recognition model.

Images