A QAM signal modulation recognition method based on deep learning channel self-attention mechanism

By introducing deep learning channel self-attention mechanism convolutional neural network in QAM signal recognition, the problem of signal characteristics difference and feature fusion difficulty is solved, efficient QAM signal recognition and accuracy improvement are achieved, and maintenance and calculation process is simplified. .

CN115589349BActive Publication Date: 2024-11-05THE FIFTH RES INST OF TELECOMM SCI & TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211205706.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-30
Publication Date
2024-11-05
Estimated Expiration
2042-09-30

AI Technical Summary

Technical Problem

The prior art is difficult to effectively combine deep learning technology with radio signals, especially in QAM signal modulation recognition, with large differences in signal characteristics and difficult feature fusion, resulting in low recognition accuracy and computational efficiency.

Method used

A channel self-attention mechanism convolutional neural network based on deep learning is designed. By extracting and fusing the I-channel data, Q-channel data and instantaneous phase of QAM signals, the data set for QAM signals are constructed and the channels are used. The self-attention mechanism is trained and predicted to realize signal recognition.

Benefits of technology

It improves the recognition accuracy and computing efficiency of QAM signals, can maintain efficient recognition in a low signal-to-noise ratio environment, simplifies the maintenance and iteration process, and reduces the computational complexity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115589349B_ABST
    Figure CN115589349B_ABST
Patent Text Reader

Abstract

The present invention provides a QAM signal modulation recognition method based on deep learning channel self-attention mechanism, including: accessing QAM signal and other signals from a sensor; calculating the instantaneous phase of I-channel data and Q-channel data of the signal to obtain the instantaneous phase of the signal; constructing a QAM data set using the I-channel data, Q-channel data and instantaneous phase; training the QAM data set using a channel self-attention mechanism convolutional neural network designed for the QAM signal to obtain a model weight file; performing uniform segmentation processing and instantaneous phase calculation on the I-channel data and Q-channel data of the predicted signal respectively to obtain M signal prediction samples; using the channel self-attention mechanism convolutional neural network loaded with the model weight file to predict the M signal prediction samples to obtain M prediction results; analyzing the M prediction results to obtain a recognition result. The present invention has low computational consumption, fast processing speed and high recognition accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a QAM signal modulation recognition method based on a deep learning channel self-attention mechanism. Background Art

[0002] The development of human society has always been accompanied by the progress of communication technology. From beacon fires to satellite reconnaissance, from wild geese to the era of the Internet, human communication technology has developed rapidly and made qualitative leaps, and communication means have become increasingly rich, efficient and complex. The purpose of communication is to transmit information quickly and accurately. Looking back at the brief history of the development of communication technology, although it is only more than a hundred years, the development of communication technology can be said to be changing with each passing day. After more than a hundred years of development, communication technology has undergone a transition from analog to digital. In the process of the development of communication technology, in order to improve the quality and efficiency of communication, more efficiently utilize channel capacity, and meet the needs of different users, communication signals adopt different modulation systems. The modulation of communication signals can be divided into analog modulation and digital modulation. Analog modulation can be divided into amplitude modulation and frequency-phase modulation; the modulation methods of digital signals are more diverse, which can be divided into amplitude shift keying (ASK), frequency shift keying (FSK), phase shift keying (PSK), and quadrature amplitude modulation (QAM). Each modulation method can be further divided, for example, quadrature amplitude modulation can be divided into 8QAM, 16QAM, 32QAM, 64QAM, 128QAM, 256QAM, etc.

[0003] In recent years, deep learning technology has matured in the field of image processing and has been applied to more and more industries, achieving good results. Combining deep learning technology with radio signal detection has become a trend, but the signal characteristics of different modulation methods vary greatly, and multiple signal features need to be combined to complete signal modulation recognition. How to combine the signal's own characteristics with deep learning has attracted the attention of more and more researchers. Summary of the invention

[0004] The present invention aims to provide a QAM signal modulation recognition method based on deep learning channel self-attention mechanism. According to the characteristics of QAM signal, this method designs a convolutional neural network with channel self-attention mechanism, and realizes modulation mode recognition of QAM category signal by fusing and extracting signal features.

[0005] The present invention provides a QAM signal modulation recognition method based on deep learning channel self-attention mechanism, comprising the following steps:

[0006] Step 1: Access signals from sensors, including QAM signals and other signals, where the length of the signal is L;

[0007] Step 2: Calculate the instantaneous phase of the I-channel data and the Q-channel data of the signal to obtain the instantaneous phase of the signal;

[0008] Step 3, construct a QAM data set using the I-channel data, Q-channel data and instantaneous phase of the signal combined with the signal modulation type;

[0009] Step 4: Use the channel self-attention mechanism convolutional neural network designed for QAM signals to train the QAM data set. Stop training after the loss value converges to obtain the model weight file.

[0010] Step 5: Evenly split the I-channel data and Q-channel data of the signal to be predicted into L-length segments, and calculate the instantaneous phase of the segmented I-channel data and Q-channel data to obtain M signal prediction samples;

[0011] Step 6: Use the channel self-attention mechanism convolutional neural network loaded with the model weight file to predict the M signal prediction samples to obtain M prediction results;

[0012] Step 7: Analyze the M prediction results and select the type of result with the most predictions as the signal recognition result.

[0013] Furthermore, the QAM signals include six types of QAM signals, namely 8QAM, 16QAM, 32QAM, 64QAM, 128QAM and 256QAM, and the number of QAM signals is N; the other signals refer to signals of other modulation modes except QAM modulation mode, and the number of other signals is 3*N.

[0014] Furthermore, the method for calculating the instantaneous phase in step 2 is:

[0015]

[0016] Wherein, t is the time, I(t) represents the I-channel data of the signal, Q(t) represents the Q-channel data of the signal; Φ(t) represents the instantaneous phase of the signal.

[0017] Furthermore, the channel self-attention mechanism convolutional neural network described in step 4 is specifically as follows:

[0018] The three cov_block modules are used to extract the features of I-channel data, Q-channel data and instantaneous phase respectively, and then the three cov_block modules are used to further extract the features to obtain three feature matrices corresponding to the three types of input data;

[0019] The three feature matrices are fused using feature concatenation to obtain a fused feature matrix;

[0020] The fused feature matrix is ​​firstly extracted using six tem_block modules connected in sequence to obtain a feature vector; a channel self-attention mechanism is embedded in the tem_block module;

[0021] A 1*1 convolution kernel is used to reduce the dimension of the feature vector, and finally the feature vector after the dimension reduction operation is compressed through the Reshape operation.

[0022] Furthermore, the cov_block module is as follows:

[0023] The cov_block module first extracts features from the input through n k*k convolution kernels, and then extracts features again using n 1*1 convolution kernels. The feature matrices of the two outputs are added together, batch normalized, and finally the feature matrix is ​​output.

[0024] Furthermore, the tem_block module is as follows:

[0025] The team_block module first uses the cov_block module to extract features from the input feature matrix, then uses the attention module attention_block to process the features, and finally uses maximum pooling to compress the features.

[0026] Furthermore, the attention module attention_block is as follows:

[0027] The attention_block module first uses m*2 1*1 convolution kernels to extract features from the input feature matrix, where m is the number of channels of the input feature matrix, and obtains feature matrix 1;

[0028] Perform global average pooling on feature matrix 1, and then perform full connections of m / 4 and m in sequence to obtain channel weight features;

[0029] Multiply feature matrix 1 with channel weight feature to obtain feature matrix 2;

[0030] Use m 1*1 convolution kernels to extract features from feature matrix 2, and output the final feature matrix after standardization.

[0031] Furthermore, Hard_swish activation is used when performing full connection of m.

[0032] Furthermore, in step 5, when the I-channel data and Q-channel data of the signal to be predicted are evenly divided into pieces of length L, the pieces of length less than L are padded with zeros.

[0033] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0034] 1. In terms of computational time, compared with the traditional modulation recognition method, the present invention takes very short computational time, and the data processing time for different modulation modes and different signal-to-noise ratios is basically the same; in traditional modulation recognition, there are many types of signal feature extraction methods, the process is complicated, and it is difficult to implement effective fusion methods for different features. The computational time is different when processing different situations, and the algorithm processing time is uncontrollable.

[0035] 2. In terms of recognition accuracy, because the present invention uses a convolutional neural network with an embedded attention mechanism, it has strong feature fusion and extraction capabilities, excellent high-order semantic expression capabilities, and high recognition accuracy. Even under low signal-to-noise conditions, it still has good results. However, after the traditional method performs feature processing on the data, the features are associated with rules, the feature fusion efficiency is low, and the expression ability is poor. When processing data in a complex environment, the recognition accuracy cannot be guaranteed.

[0036] 3. Compared with the general deep learning modulation recognition, the present invention designs a convolutional neural network with an embedded attention mechanism for QAM signals, which can efficiently extract features of the signal and the intrinsic correlation between features. The features of different channels can reflect part of the features of the signal data, and the use of channel attention can suppress useless features. The general deep learning network has limited ability to fuse business data, and its feature expression ability is not as strong as the network model with the attention mechanism. After verification by actual data, the network structure proposed by the present invention has a higher recognition accuracy than the general network model.

[0037] 4. The present invention is easy to update and maintain, has strong feasibility, and has a controllable iteration cycle. When encountering emergencies, traditional modulation type recognition will add strategies for specific noise or interference encountered, but this method has high maintenance costs, unstable iteration cycles, and the calculation code will become more and more complicated, and the maintainability is not high; while in the present invention, only the encountered burst signal needs to be added to the data set for training, and the obtained model can be replaced, which is convenient for maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0039] Figure 1This is a flowchart of a QAM signal modulation recognition method based on a deep learning channel self-attention mechanism in an embodiment of the present invention.

[0040] Figure 2 This is a model training flowchart of a convolutional neural network with a channel self-attention mechanism in an embodiment of the present invention.

[0041] Figure 3 This is a flow chart of modulation identification in an embodiment of the present invention.

[0042] Figure 4 Schematic diagram of the structure of the convolutional neural network with channel self-attention mechanism in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions 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 part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0044] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Example

[0046] like Figure 1 , Figure 2 , Figure 3 As shown, this embodiment proposes a QAM signal modulation recognition method based on deep learning channel self-attention mechanism, comprising the following steps:

[0047] Step 1. Access signals from the sensor, including QAM signals and other signals, and the length of the signal is L. In this embodiment, the QAM signals include six types of QAM signals, namely 8QAM, 16QAM, 32QAM, 64QAM, 128QAM and 256QAM, and the number of QAM signals is N (generally N≥10000); the other signals refer to signals of other modulation modes except QAM modulation mode, and the number of other signals is 3*N; L=1024.

[0048] Step 2: Calculate the instantaneous phase of the I-channel data and the Q-channel data of the signal to obtain the instantaneous phase of the signal; the method for calculating the instantaneous phase is:

[0049]

[0050] Wherein, t is the time, I(t) represents the I-channel data of the signal, Q(t) represents the Q-channel data of the signal; Φ(t) represents the instantaneous phase of the signal.

[0051] Step 3, construct a QAM data set using the I-channel data, Q-channel data and instantaneous phase of the signal combined with the signal modulation type;

[0052] Step 4: Use the channel self-attention mechanism convolutional neural network designed for QAM signals to train the QAM data set. Stop training after the loss value converges to obtain the model weight file.

[0053] The channel self-attention mechanism convolutional neural network designed in this embodiment is as follows Figure 4 As shown, the three inputs of the channel self-attention mechanism convolutional neural network are the I-channel data, Q-channel data and instantaneous phase of the signal, where the size of the I-channel data is 1*1024*1, the size of the Q-channel data is 1*1024*1, and the size of the instantaneous phase is 1*1024*1. The channel self-attention mechanism convolutional neural network includes three basic feature processing modules, namely the cov_block module, the attention module attention_block and the team_block module, as follows:

[0054] (1) The cov_block module first extracts features from the input using n k*k convolution kernels, and then extracts features again using n 1*1 convolution kernels. The feature matrices of the two outputs are added together, and batch normalization is performed before outputting the feature matrix. For example, for (cov_block, 3*3, 16), the value of k is 3 and the value of n is 16.

[0055] (2) The attention module attention_block is as follows:

[0056] The attention_block module first uses m*2 1*1 convolution kernels to extract features from the input feature matrix, where m is the number of channels of the input feature matrix, and obtains feature matrix 1;

[0057] Perform global average pooling on feature matrix 1, and then perform full connection of m / 4 and m in sequence to obtain channel weight features; among them, Hard_swish activation mode is used when performing full connection of m, and the formula is expressed as: y=x*Relu(x+3,max_value=6) / 6.0;

[0058] Multiply feature matrix 1 with channel weight feature to obtain feature matrix 2;

[0059] Use m 1*1 convolution kernels to extract features from feature matrix 2, and output the final feature matrix after standardization.

[0060] (3) The team_block module first uses the cov_block module (cov_block, k*k, t) to extract features from the input feature matrix, then uses the attention module attention_block to process the features, and finally uses maximum pooling to compress the features.

[0061] Therefore, the channel self-attention mechanism convolutional neural network of this embodiment is specifically as follows:

[0062] The features of I-channel data, Q-channel data and instantaneous phase are extracted through three cov_block modules (cov_block, 1*1, 16), and then further extracted through three cov_block modules (cov_block, 1*1, 32), and three 1*1024*32 feature matrices corresponding to the three input data are obtained;

[0063] The three feature matrices are fused using feature concatenation to obtain a fused feature matrix with a size of 1*1024*96;

[0064] The fused feature matrix is ​​first extracted using six tem_block modules connected in sequence to obtain a feature vector; a channel self-attention mechanism is embedded in the tem_block module; specifically: first, a tem_block module with 128 1*1 convolution kernels is used to extract features to obtain a feature vector with a size of 1*1024*128, and then a tem_block module with 128 3*3 convolution kernels is used five times to extract features to obtain a feature vector with a size of 1*32*128.

[0065] Use numb 1*1 convolution kernels to reduce the dimension of the feature vector, and finally compress the feature vector after the dimension reduction operation through the Reshape operation.

[0066] Step 5, the I-channel data and Q-channel data of the predicted signal are evenly divided into pieces of length L, the pieces of length less than L are padded with zeros, and the instantaneous phases of the divided I-channel data and Q-channel data are calculated to obtain M signal prediction samples; the instantaneous phase calculation method refers to step 2 and will not be repeated here.

[0067] Step 6: Use the channel self-attention mechanism convolutional neural network loaded with the model weight file to predict the M signal prediction samples to obtain M prediction results;

[0068] Step 7: Analyze the M prediction results and select the type of result with the most predictions as the signal recognition result.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A QAM signal modulation recognition method based on deep learning channel self-attention mechanism, characterized in that: The steps include: Step 1: Access signals from sensors, including QAM signals and other signals, where the length of the signal is L; Step 2: Calculate the instantaneous phase of the I-channel data and the Q-channel data of the signal to obtain the instantaneous phase of the signal; Step 3, construct a QAM data set using the I-channel data, Q-channel data and instantaneous phase of the signal combined with the signal modulation type; Step 4: Use the channel self-attention mechanism convolutional neural network designed for QAM signals to train the QAM data set. Stop training after the loss value converges to obtain the model weight file. Step 5: Evenly split the I-channel data and Q-channel data of the signal to be predicted into L-length segments, and calculate the instantaneous phase of the segmented I-channel data and Q-channel data to obtain M signal prediction samples; Step 6: Use the channel self-attention mechanism convolutional neural network loaded with the model weight file to predict the M signal prediction samples to obtain M prediction results; Step 7: Analyze the M prediction results, and select the type result with the most predictions as the signal recognition result; The channel self-attention mechanism convolutional neural network described in step 4 is as follows: The three cov_block modules are used to extract the features of I-channel data, Q-channel data and instantaneous phase respectively, and then the three cov_block modules are used to further extract the features to obtain three feature matrices corresponding to the three types of input data; The three feature matrices are fused using feature concatenation to obtain a fused feature matrix; The fused feature matrix is ​​firstly extracted using six tem_block modules connected in sequence to obtain a feature vector; a channel self-attention mechanism is embedded in the tem_block module; A 1*1 convolution kernel is used to reduce the dimension of the feature vector, and finally the feature vector after the dimension reduction operation is compressed through the Reshape operation.

2. The QAM signal modulation recognition method based on deep learning channel self-attention mechanism according to claim 1 is characterized in that: In step 1, the QAM signals include six types of QAM signals, namely 8QAM, 16QAM, 32QAM, 64QAM, 128QAM and 256QAM, and the number of QAM signals is N; the other signals refer to signals of other modulation modes except QAM modulation mode, and the number of other signals is 3*N.

3. The QAM signal modulation recognition method based on deep learning channel self-attention mechanism according to claim 1 is characterized in that: The method for calculating the instantaneous phase in step 2 is: Among them, t is the time, Indicates the I-channel data of the signal, Indicates the Q channel data of the signal; Represents the instantaneous phase of the signal.

4. The QAM signal modulation recognition method based on deep learning channel self-attention mechanism according to claim 1 is characterized in that: The cov_block module is as follows: The cov_block module first extracts features from the input through n k*k convolution kernels, and then extracts features again using n 1*1 convolution kernels. The feature matrices of the two outputs are added together, batch normalized, and finally the feature matrix is ​​output.

5. The QAM signal modulation recognition method based on deep learning channel self-attention mechanism according to claim 4 is characterized in that: The tem_block module is as follows: The team_block module first uses the cov_block module to extract features from the input feature matrix, then uses the attention module attention_block to process the features, and finally uses maximum pooling to compress the features.

6. The QAM signal modulation recognition method based on deep learning channel self-attention mechanism according to claim 5 is characterized in that: The attention module attention_block is as follows: The attention_block module first uses m*2 1*1 convolution kernels to extract features from the input feature matrix, where m is the number of channels of the input feature matrix, and obtains feature matrix 1; Perform global average pooling on feature matrix 1, and then perform full connections of m / 4 and m in sequence to obtain channel weight features; Multiply feature matrix 1 with channel weight feature to obtain feature matrix 2; Use m 1*1 convolution kernels to extract features from feature matrix 2, and output the final feature matrix after standardization.

7. The QAM signal modulation recognition method based on deep learning channel self-attention mechanism according to claim 6 is characterized in that: Hard_swish activation is used when performing full connection of m.

8. The QAM signal modulation recognition method based on deep learning channel self-attention mechanism according to claim 1 is characterized in that: In step 5, when the I-channel data and Q-channel data of the predicted signal are evenly divided into pieces of length L, the pieces of length less than L are padded with zeros.