CN-121997249-A - Modulation signal identification method based on convolution attention and multidimensional feature fusion

CN121997249ACN 121997249 ACN121997249 ACN 121997249ACN-121997249-A

Abstract

The invention discloses a modulating signal identification method based on convolution attention and multidimensional feature fusion, which solves the problems that different signals in the same signal family often show similar features under the condition of low signal-to-noise ratio in the prior art, so that a model is difficult to effectively distinguish, the identification accuracy is reduced, the higher identification accuracy of modulating signal identification under the condition of low signal-to-noise ratio is realized, and the calculation complexity is lower. The method comprises the steps of firstly, carrying out pre-training on an IQ and AP sequence, then carrying out recognition by utilizing a pre-training signal recognition network, converting the IQ and AP sequence into an IQ and AP feature map by a preprocessing module, extracting local and global features from the IQ and AP feature map by a feature extraction module to form a path feature map, carrying out weighted fusion on the two paths of features by a WAFF module, and finally outputting a classification result by a classification head according to the fused features.

Inventors

JING DAN
CHEN JIYAO
ZHANG MING
LI XINCHEN
LI ZHENYU
GUO LIANG
WANG XUAN
ZHANG YAN
XING MENGDAO

Assignees

西安电子科技大学

Dates

Publication Date: 20260508
Application Date: 20251225

Claims (10)

1. A modulated signal recognition method based on convolution attention and multi-dimensional feature fusion, comprising: acquiring an original complex signal to be identified, performing data processing on the original complex signal to obtain a normalized IQ sequence, and obtaining an AP sequence according to the normalized IQ sequence; The IQ sequence and the AP sequence are subjected to signal recognition according to a pre-trained signal recognition network to obtain a recognition result, wherein the signal recognition network comprises a preprocessing module, a feature extraction module, a WAFF module and a classification head, the preprocessing module is used for carrying out preliminary shape change on the IQ sequence and the AP sequence to obtain a processed IQ feature map and an AP feature map, the feature extraction module is used for carrying out local detail feature and global feature extraction on the processed IQ feature map and the AP feature map to obtain an IQ path feature map and an AP path feature map, the WAFF module is used for carrying out weighted fusion on the IQ path feature map and the AP path feature map to obtain a fusion feature, and the classification head is used for mapping the fusion feature to a category prediction to obtain a classification result.
2. The method for identifying modulated signals based on convolution attention and multidimensional feature fusion according to claim 1, wherein the data processing is performed on the original complex signal to obtain a normalized IQ sequence, and the AP sequence is obtained according to the normalized IQ sequence, including: extracting a real part from the original complex signal as the I sequence signal and an imaginary part as the Q sequence signal; Respectively carrying out normalization processing on the I sequence signal and the Q sequence signal to obtain the normalized IQ sequence; and converting the normalized IQ sequence into an AP sequence.
3. The method for identifying the modulation signal based on the combination of the convolution attention and the multi-dimensional characteristics according to claim 1, wherein the preprocessing module comprises a first convolution layer, a second convolution layer, a third convolution layer and a channel splicing layer; The preprocessing module is configured to perform preliminary shape change on the IQ sequence and the AP sequence to obtain a processed IQ feature map and an AP feature map, and includes: Inputting the normalized IQ sequence into the first convolution layer, the second convolution layer and the third convolution layer respectively for multi-scale feature extraction to obtain corresponding first features, second features and third features, wherein the convolution kernels of the first convolution layer and the third convolution layer are different, and the second convolution layer is a maximum pooling layer; Fusing the first feature, the second feature and the third feature through the channel splicing layer to form an IQ feature map; And converting the IQ feature map into an amplitude feature map and a phase feature map through polar coordinate transformation, and splicing the amplitude feature map and the phase feature map to obtain an AP feature map.
4. The method for identifying the modulated signal based on the fusion of the convolution attention and the multidimensional feature according to claim 1, wherein the feature extraction module comprises a first stage module, a second stage module, a third stage module and a fourth stage module which are sequentially connected; The feature extraction module is configured to perform local detail feature extraction and global feature extraction on the processed IQ feature map and the AP feature map to obtain an IQ path feature map and an AP path feature map, and includes: the first stage module comprises a first downsampling block and a first depth separable convolution module, wherein the first stage module performs local space-time feature extraction on the processed IQ feature map and the AP feature map to obtain an IQ path primary feature map and an AP path primary feature map; The second stage module comprises a second downsampling block and a second depth separable convolution module, wherein the second depth separable unit is used for carrying out multi-scale local feature extraction on the IQ path primary feature map and the AP path primary feature map to obtain an IQ path intermediate feature map and an AP path intermediate feature map; The third stage module comprises a third downsampling block and a first proxy attention module, wherein the third stage module captures global context dependence of the IQ path intermediate feature map and the AP path intermediate feature map by introducing an attention mechanism to obtain an IQ path perception feature map and an AP path perception feature map; the fourth stage module comprises a fourth downsampling block and a second agent attention module, and further refines global features of the IQ path perception feature map and the AP path perception feature map on the highest semantic level to obtain an IQ path advanced feature map and an AP path advanced feature map.
5. The method for identifying a modulated signal based on convolution attention and multi-dimensional feature fusion according to claim 4, wherein the first depth separable convolution module comprises a first layer normalization layer, a first depth separable convolution layer, a first multi-layer perceptron layer and a first residual error connection layer; The first stage module performs local space-time feature extraction on the processed IQ feature map and the AP feature map to obtain an IQ path primary feature map and an AP path primary feature map, including: inputting the initial downsampling characteristics output by the first downsampling block into the first layer normalization layer for normalization processing; Inputting the normalized features into the first depth separable convolution layer, and sequentially carrying out depth-wise convolution and point-wise convolution to extract local spatial features; inputting the local spatial features into the first multi-layer perceptron layer, and carrying out nonlinear transformation on channel dimensions to obtain the features after nonlinear transformation; and adding the nonlinear transformed features with the initial downsampled features through the first residual error connection layer to obtain an IQ path primary feature map and an AP path primary feature map.
6. The method for identifying a modulated signal based on convolution attention and multi-dimensional feature fusion of claim 4, wherein said first proxy attention module comprises a second normalization layer, a second depth separable convolution layer, a second multi-layer perceptron layer, and a second residual connection layer; The third stage module captures global context dependence of the IQ path intermediate feature map and the AP path intermediate feature map by introducing an attention mechanism to obtain an IQ path aware feature map and an AP path aware feature map, including: inputting the high-layer downsampling characteristics output by the third downsampling block into the second normalization layer for normalization processing; Inputting the normalized features into the agent attention calculation layer, and obtaining a global context enhanced feature map by generating agent vectors, agent-key value attention calculation and query-agent attention calculation; inputting the feature map enhanced by the global context into the second depth separable convolution layer, and carrying out local feature recovery and enhancement to obtain local enhancement features; And adding the local enhancement feature and the high-layer downsampling feature through the second residual error connecting layer to obtain an IQ path perception feature map and an AP path perception feature map.
7. The method for identifying a modulated signal based on convolution attention and multi-dimensional feature fusion according to claim 1, wherein said WAFF module comprises a local attention branch, a global attention branch, a weight generation layer and a weighted fusion layer; The WAFF module performs weighted fusion on the IQ path feature map and the AP path feature map to obtain a fusion feature, including: Adding the IQ path feature map and the AP path feature map to obtain initial fusion features; inputting the initial fusion feature into the local attention branch and the global attention branch simultaneously, and respectively extracting local attention weight and global attention weight; adding the local attention weight and the global attention weight, and generating an attention weight map through a Sigmoid activation function; Performing self-adaptive weighted fusion on the IQ path feature map and the AP path feature map according to the attention weight map by using the weighted fusion layer to obtain weighted fusion features; And adding the weighted fusion feature and the initial fusion feature through residual connection to obtain a final fusion feature.
8. The method for identifying modulated signals based on convolution attention and multi-dimensional feature fusion according to claim 1, wherein said classification header comprises a global averaging pooling layer and a fully connected classification layer; the classification head is used for mapping the fusion characteristic to the category prediction to obtain a classification result, and comprises the following steps: Inputting the fusion features into the global average pooling layer, and compressing space dimensions to obtain feature vectors; Inputting the feature vector into the fully-connected classification layer, and outputting probability distribution corresponding to each modulation class; And outputting the category with the highest probability as a final recognition result.
9. The method for identifying a modulated signal based on convolution attention and multi-dimensional feature fusion according to claim 1, wherein the training process of the signal identification network comprises: Dividing a signal data set comprising a plurality of modulation types and signal to noise ratios into a training set, a verification set and a test set in proportion; Model training is carried out by adopting a cross entropy loss function and AdamW optimizers; The cosine annealing strategy is used in the training process to adjust the learning rate and Dropout is set to prevent overfitting.
10. The method of claim 9, wherein the signal data set comprises an rml2016.10a data set, and wherein the modulation type comprises at least one of 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK, AM-DSB, AM-SSB, WBFM.

Description

Modulation signal identification method based on convolution attention and multidimensional feature fusion Technical Field The invention relates to the technical field of deep learning, in particular to a modulation signal identification method based on convolution attention and multidimensional feature fusion. Background As an important link between signal detection and demodulation, automatic Modulation Recognition (AMR) is used to provide modulation information of wireless signals, which plays a key role in intelligent wireless communication. Conventional modulation identification techniques can be broadly divided into two categories, likelihood-based (LB) and feature-based (FB) approaches. The LB method calculates the matching degree of the signal and different modulation modes based on Bayesian theory. In contrast, FB methods distinguish modulation types by manually performing feature extraction, in combination with traditional machine learning classifiers. Deep learning is capable of automatically extracting features from received signals and classifying them, and has made great progress in the field of automatic modulation recognition in recent years, such as Convolutional Neural Networks (CNNs) and long-term memory networks (LSTM). CNNs are typically used to extract spatial features of data, while LSTM is good at modeling time series data. Self-attention mechanisms and transducers have also shown great potential in the field of radio signal classification as new approaches in the field of natural language processing and computer vision. The transducer is capable of extracting global features containing spatiotemporal information through a multi-headed self-attention Mechanism (MHSA). In addition, most of the existing modulation recognition methods only adopt a single dimension of a signal as an input of a model, and feature expression of the modulation signal in different dimensions cannot be fully captured. In contrast, the multi-dimensional input signal has better recognition capability and stronger robustness. Therefore, the invention designs a deep learning model based on agent attention mechanism and multidimensional feature fusion, which uses convolution and attention to extract the multidimensional features of the input signals, and then uses an attention fusion module to weight and fuse the information from different dimensionalities. The influence of different input signal dimensions can be flexibly adjusted by the model through weighted fusion, so that the richness and the accuracy of the feature expression under the low signal-to-noise ratio are improved. Both the traditional likelihood-based and feature-based modulation recognition methods rely on a large amount of a priori knowledge, and it is difficult to achieve fast and accurate modulation recognition in modern complex wireless communication environments. The deep learning-based automatic modulation recognition technique has several drawbacks, firstly CNN is generally used to extract spatial features of data, but lacks attention to signal temporal features, while LSTM is good at modeling time-series data, but cannot accelerate in parallel, and is computationally inefficient in the face of large data volumes. This limitation makes it difficult for these deep learning models to fully capture the overall characteristics of the signal when processing complex modulated signals, resulting in limited robustness and generalization capabilities in practical applications, thereby affecting their recognition performance. Secondly, compared with CNN and LSTM, the transducer captures global features by using a self-attention mechanism, but local feature modeling is omitted, and the transducer has higher computational complexity, large resource occupation and great limitation in deployment in edge equipment with limited resources. Finally, many approaches rely solely on IQ features of the domain signal itself, ignoring the potential information of the signal in other dimensions. Such single-dimensional input limits the overall modeling of the signal features by the model, resulting in poor recognition performance in complex signal environments. Particularly, under the condition of low signal-to-noise ratio, different signals (QAM 16/QAM 64, 8 PSK/BPSK) in the same signal family often show similar characteristics, so that the models are difficult to effectively distinguish, and the recognition accuracy is reduced. Disclosure of Invention The invention provides a modulating signal identification method based on convolution attention and multidimensional feature fusion, which solves the problems that different signals in the same signal family often show similar features under the condition of low signal-to-noise ratio in the prior art, so that a model is difficult to effectively distinguish, the identification accuracy is reduced, the higher identification accuracy of modulating signal identification under the condition of low signal-to-noise ratio