CN109890043B - Wireless signal noise reduction method based on generative countermeasure network - Google Patents

Wireless signal noise reduction method based on generative countermeasure network Download PDF

Info

Publication number
CN109890043B
CN109890043B CN201910150001.7A CN201910150001A CN109890043B CN 109890043 B CN109890043 B CN 109890043B CN 201910150001 A CN201910150001 A CN 201910150001A CN 109890043 B CN109890043 B CN 109890043B
Authority
CN
China
Prior art keywords
wireless signal
noise reduction
gan
time
network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910150001.7A
Other languages
Chinese (zh)
Other versions
CN109890043A (en
Inventor
陈晋音
成凯回
郑海斌
蒋焘
宣琦
杨东勇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN201910150001.7A priority Critical patent/CN109890043B/en
Publication of CN109890043A publication Critical patent/CN109890043A/en
Application granted granted Critical
Publication of CN109890043B publication Critical patent/CN109890043B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Noise Elimination (AREA)

Abstract

The invention discloses a wireless signal noise reduction method based on a generative countermeasure network, which comprises the following steps: constructing a wireless signal noise reduction network GAN-G and a wireless signal noise reduction discrimination network GAN-D, wherein the GAN-G comprises an LSTM, a convolution layer and a time attention module, the input of the GAN-G is an original wireless signal added with noise, and the output of the GAN-G is a noise-reduced wireless signal; the GAN-D comprises an LSTM and a full connection layer, an original wireless signal and a noise-reduced wireless signal are input, and a judgment result is output; training the GAN-G and the GAN-D by adopting an antagonistic training strategy to obtain a wireless signal noise reduction model and a wireless signal noise reduction discrimination model; and extracting a wireless signal noise reduction model to process the wireless signal to be noise reduced to obtain the noise-reduced wireless signal. The method for carrying out noise reduction processing on the wireless signal data can effectively improve the accuracy of the identification of the wireless signal modulation type in a low signal-to-noise ratio interval.

Description

Wireless signal noise reduction method based on generative countermeasure network
Technical Field
The invention belongs to the application of a deep learning method in the field of wireless signal identification, and particularly relates to a wireless signal noise reduction method based on a generative countermeasure network.
Background
The deep learning technology relies on strong feature learning ability, and is widely applied to various fields, mainly including computer vision, natural language processing, complex network analysis, wireless signal analysis and the like. Deep learning is a method for performing characterization learning on data in machine learning, and features of the data are extracted through a huge neural network, so that a mechanism of a human brain is simulated to explain the data. Typical deep neural networks include convolutional neural networks and cyclic neural networks, wherein the convolutional neural networks are widely applied to image classification tasks and target detection tasks by virtue of strong feature extraction performance of the convolutional neural networks. In addition, the long-short term memory network can better handle functions such as natural language translation and speech recognition by learning the time series information of the sequence data. As a more advanced technique in the field of deep learning in recent years, the generative countermeasure network (GAN) has achieved good results in many fields, such as picture synthesis, picture completion, text generation, and video generation, by virtue of its excellent ability to learn mapping relationships. The generative confrontation network comprises two depth model structures: a Generator model (Generator) and a Discriminator model (Discriminator). For a given piece of information, the generator model can map the information into a certain feature space through a certain mapping relation; the discriminator model mainly judges whether the generated data and the real data are true or false. Currently, deep learning techniques have been applied to wireless signal processing tasks and achieve better results. The wireless signal processing mainly comprises the tasks of signal detection, signal modulation type identification, signal demodulation and the like. The automatic identification technology of the modulation type of the wireless signal mainly realizes the intelligent receiving and processing of the modulation signal. Signal modulation can be divided into analog modulation and digital modulation, in the form of different modulated signals. What is called analog modulation using analog signals; the use of digital signal modulation is referred to as digital modulation.
Different modulation types are used for communication signals for different application needs. Although the wireless signal modulation technology is mature, when the wireless signal contains more noise, the identification of the wireless signal modulation type becomes difficult, and the existing identification methods are affected by different degrees, so that the identification accuracy cannot be ensured. Therefore, it is important to design an efficient wireless signal noise reduction method. According to the research of related patents and papers, there have been some patents and papers that propose different denoising methods, such as patent "a bayesian wavelet packet denoising method suitable for non-gaussian signal", which proposes a bayesian wavelet packet denoising method suitable for non-gaussian signal, by performing discrete wavelet packet decomposition on a signal time sequence and determining the optimal decomposition layer number, and then performing inverse transformation on the wavelet coefficients of real signals in each decomposition layer obtained by estimation to realize reconstruction of signals, thereby obtaining denoised signals. The patent of application of Mallat algorithm for optimizing threshold based on genetic algorithm to denoising heart sound signals proposes a method for optimizing threshold by utilizing genetic algorithm to realize signal denoising, and denoising the signals by utilizing threshold optimized by genetic algorithm and using Mara algorithm to divide high frequency band and low frequency band. There is a certain applicability of these methods.
Disclosure of Invention
The invention aims to provide a wireless signal noise reduction method based on a generative countermeasure network, which is characterized in that a generator model based on a long-time memory network is designed, a time attention mechanism is introduced to improve the effect of generating a noise reduction signal, a discriminator model based on the long-time memory network is designed, and the accuracy of the generation model is improved through two classifications of true and false signals, so that the high-efficiency noise reduction processing aiming at the noise wireless signal is realized. The method for carrying out noise reduction processing on the wireless signal data can effectively improve the accuracy of the identification of the wireless signal modulation type in a low signal-to-noise ratio interval. The efficiency of wireless signal modulation type identification is greatly improved.
In order to achieve the purpose, the invention provides the following technical scheme:
a wireless signal noise reduction method based on a generative countermeasure network comprises the following steps:
constructing a wireless signal noise reduction network and a wireless signal noise reduction judging network, wherein the wireless signal noise reduction network comprises a long-time memory neural network, a convolution layer and a time attention module, the input of the wireless signal noise reduction network is a low signal-to-noise ratio wireless signal added with noise, and the output of the wireless signal noise reduction network is a wireless signal subjected to noise reduction;
the wireless signal noise reduction judging network comprises a long-time memory neural network and a full connection layer, the input of the wireless signal noise reduction judging network is connected with the output of the wireless signal noise reduction network, the input data is an original wireless signal without noise and a wireless signal after noise reduction, and the output is a judging result of the wireless signal after noise reduction and the original wireless signal without noise;
training a wireless signal noise reduction network and a wireless signal noise reduction judging network by adopting a confrontation training strategy to obtain a wireless signal noise reduction model and a wireless signal noise reduction judging model;
and extracting a wireless signal noise reduction model to process the wireless signal to be noise reduced to obtain the noise-reduced wireless signal.
The low signal-to-noise ratio wireless signal is taken from original wireless signal data added with noise, the size of a low signal-to-noise ratio wireless signal sample is time x 2, wherein time represents a wireless signal sampling node, and 2 represents the characteristic dimension of the wireless signal.
In the invention, the obtained wireless signal noise reduction model has the main function of carrying out noise reduction processing on the low signal-to-noise ratio wireless signal added with noise, namely reducing the noise contained in the sample under the condition of keeping the data information of the original wireless signal to the maximum extent.
Wherein, the wireless signal noise reduction network includes:
LSTM1, LSTM2, convolutional layer, LSTM3 and LSTM4 which are connected in sequence;
time attention module alpha1: LSTM1 outputs a feature layer with the size of time x 32, a feature layer with the size of time x 2 is obtained through LSTM _ A1, the transposition of the feature layer and an input low signal-to-noise ratio wireless signal are subjected to matrix multiplication to obtain a feature layer with the size of time x time, the feature layer is input into LSTM _ B1 to obtain a time attention weight W with the size of time x 321Feeding back the temporal attention weight to LSTM1 changes the weight at the time of calculating the output; the specific mode is as follows:
Figure BDA0001981249800000041
f1＝W1*f1
wherein the content of the first and second substances,
Figure BDA0001981249800000042
representing a matrix multiplication operation, f1Feature layers output for LSTM1 with time x 32, fA1The feature layer output for LSTM _ a1, size time x 2,
Figure BDA0001981249800000043
denotes fA1Transposed matrix of (1), F1To obtain a feature layer of size time, W1The temporal attention weight output for LSTM _ B1 is time x 32, which represents the multiplication of the corresponding elements of the two matrices.
Time attention module alpha2: LSTM2 outputs feature layers of size time x 64,obtaining a feature layer with the size of time x 2 through LSTM _ A2, carrying out matrix multiplication on the transpose of the feature layer and an input low signal-to-noise ratio wireless signal to obtain a feature layer with the size of time x time, inputting the feature layer into LSTM _ B2 to obtain a time attention weight W with the size of time 642Feeding back the temporal attention weight to LSTM2 changes the weight of the calculated output; the specific mode is as follows:
Figure BDA0001981249800000044
f2＝W2*f2
wherein the content of the first and second substances,
Figure BDA0001981249800000045
representing a matrix multiplication operation, f2Feature layers output for LSTM2 with time x 64, fA2The feature layer output for LSTM _ a2, size 128 x 2,
Figure BDA0001981249800000046
denotes fA2Transposed matrix of (1), F2To obtain a feature layer of size time, W2The temporal attention weight output for LSTM _ B2 is time x 64, which represents the multiplication of the corresponding elements of the two matrices.
Time attention module alpha3: LSTM3 outputs a feature layer with the size of time x 32, obtains a feature layer with the size of time x 2 through LSTM _ A3, performs matrix multiplication on the transpose of the feature layer and an input low signal-to-noise ratio wireless signal to obtain a feature layer with the size of time x time, inputs the feature layer into LSTM _ B3, and obtains a time attention weight W with the size of time x 323The temporal attention weight is fed back to LSTM3 to change its weight in the following way:
Figure BDA0001981249800000051
f3＝W3*f3
wherein the content of the first and second substances,
Figure BDA0001981249800000052
representing a matrix multiplication operation, f3Feature layers output for LSTM3 with time x 32, fA3The feature layer output for LSTM _ a3, size time 32,
Figure BDA0001981249800000053
denotes fA3Transposed matrix of (1), F3To obtain a feature layer of size time, W3The temporal attention weight output for LSTM _ B3 is time x 32, which represents the multiplication of the corresponding elements of the two matrices.
The wireless signal noise reduction discrimination network includes:
the LSTM is used for extracting the time sequence characteristics of the input wireless signals;
the first full connection layer is used for synthesizing the time sequence characteristics of the LSTM output;
and the second full connection layer is used as an output layer and used for integrating the characteristics output by the first full connection layer and outputting a classification result.
The input of the wireless signal noise reduction discrimination model GAN-D is an original wireless signal without noise and a wireless signal after noise reduction, and the output is a result of judging the wireless signal after noise reduction and the original wireless signal without noise. A discriminator model based on a long-time and short-time memory network is designed, and accuracy of the model is improved through two classifications of over-true and false signals, so that efficient noise reduction processing for noise wireless signals is achieved.
Defining the class mark of the wireless signal after noise reduction as 0 and the class mark of the original wireless signal data as 1 when training the parameters of the GAN-D model; the wireless signal noise reduction discrimination model GAN-D has the main functions of performing difference judgment on the distribution of the wireless signals subjected to noise reduction, namely judging whether the distribution of the wireless signals subjected to noise reduction is consistent with the distribution of the original wireless signals or not, and feeding back the parameters of the wireless signal noise reduction model GAN-G according to the judgment result.
Specifically, the specific process of training is as follows:
(a) pre-training a wireless signal noise reduction network and a wireless signal noise reduction discrimination network;
(b) the method comprises the following steps of training a pre-trained wireless signal noise reduction network GAN-G and a wireless signal noise reduction discrimination network GAN-D in an antagonistic manner, wherein the specific process comprises the following steps:
fixing parameters of the GAN-G, inputting the wireless signals subjected to noise reduction into the GAN-D, and training the parameters of the GAN-D; fixing parameters of the GAN-D, inputting a low signal-to-noise ratio wireless signal into the GAN-G, and training the parameters of the GAN-G;
(c) repeating the steps (a) and (b), alternately training parameters of the GAN-G and the GAN-D until the two realize G-D Nash equilibrium or reach the maximum training times T2;
(d) and (4) adding a time attention module consisting of a long-time memory network and then training the GAN-G, namely repeating the steps (a) to (c), so that the GAN-G more efficiently extracts the time sequence characteristics of the wireless signals and generates the wireless signals after noise reduction.
During the countertraining, the process of training the wireless signal noise reduction network GAN-G is as follows:
inputting the wireless signal output by the GAN-G after noise reduction into the GAN-D, and performing cross entropy calculation loss on the obtained output and 0; inputting an original wireless signal into the GAN-D, and performing cross entropy calculation loss on the obtained output and 1;
the optimization goals of the process are as follows:
Figure BDA0001981249800000061
wherein D (-) represents a wireless signal noise reduction discrimination model GAN-D, G (-) represents a wireless signal noise reduction model GAN-G, xsignalRepresenting the data of the original wireless signal and,
Figure BDA0001981249800000062
representing low signal-to-noise ratio wireless signal data, E (-) representing mathematical expectation, x-pdata (x)signal) Representing x samples from an original wireless signal data set xsignal，
Figure BDA0001981249800000063
Representing x-samples from a low signal-to-noise ratio wireless signal dataset
Figure BDA0001981249800000064
During the countertraining, the process of training the wireless signal noise reduction network GAN-D is as follows:
inputting a low signal-to-noise ratio wireless signal into a wireless signal noise reduction model GAN-G, inputting the obtained noise-reduced wireless signal into a wireless signal noise reduction discrimination model GAN-D, performing cross entropy calculation loss on the obtained output, and training parameters of GAN-G by minimizing the loss feedback:
the optimization goals of the process are as follows:
Figure BDA0001981249800000071
wherein D (-) represents a wireless signal noise reduction discrimination model GAN-D, G (-) represents a wireless signal noise reduction model GAN-G,
Figure BDA0001981249800000072
representing low signal-to-noise ratio wireless signal data, E (-) representing a mathematical expectation,
Figure BDA0001981249800000073
representing x-samples from a low signal-to-noise ratio wireless signal dataset
Figure BDA0001981249800000074
Through the training strategy, the wireless signal noise reduction model GAN-G can be used as a filter of a wireless signal, noise in the wireless signal can be better filtered, and information contained in the original wireless signal is retained to the maximum extent. By using the generative countermeasure network, the output of the noise reduction model GAN-G is closer to the distribution of real wireless signals, the accuracy of the wireless signal noise reduction model is improved, the noise in the wireless signals with low signal-to-noise ratio can be better filtered, and the signal data in the original signal data can be retained to the maximum extent. The method for carrying out noise reduction processing on the wireless signal data can effectively improve the accuracy of the identification of the wireless signal modulation type in a low signal-to-noise ratio interval.
The invention provides a wireless signal noise reduction method based on a generative confrontation network, which utilizes a long-time memory network to learn the time sequence characteristics of a wireless signal, utilizes the excellent capability of learning a mapping relation of the generative confrontation network to carry out noise reduction processing on the wireless signal, and introduces a time attention module into a generative model in the generative confrontation network to improve the efficiency of a wireless signal noise reducer and has universality.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
Fig. 1 is a schematic structural diagram of a wireless signal noise reduction method based on a generative countermeasure network with time 128 according to an embodiment;
FIGS. 2(a), 2(b), and 2(c) show the time attention module α provided in the embodiment1Time attention module alpha2Time attention module alpha3Schematic structural diagram of (a);
fig. 3 is a wireless signal modulation type classifier based on deep learning provided by an embodiment;
fig. 4 is a comparison of the wireless signal debugging and identification accuracy before and after the wireless signal noise reduction method based on the generative countermeasure network according to the embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
The embodiment provides a wireless signal noise reduction method based on a generative countermeasure network, which utilizes a wireless signal noise reduction model to perform noise reduction processing on a wireless signal with a low signal-to-noise ratio. Through the training method shown in FIG. 1, a wireless signal noise reduction model GAN-G and a wireless signal noise reduction discrimination model GAN-D can be obtained.
The wireless signal noise reduction model GAN-G can filter noise in the wireless signal with low signal to noise ratio to obtain the wireless signal with high signal to noise ratio. A radio noise reduction model GAN-G is constructed using a long-term memory neural network (LSTM), a convolutional layer (Conv), and a temporal attention module. Taking time 128 as an example, the specific structure is shown in fig. 1: the size of an input original wireless signal sample is 128 x 2, wherein 128 represents a sampling time node of the wireless signal, and 2 represents a characteristic dimension of the wireless signal; and carrying out noise adding processing on the original wireless signal data to obtain the wireless signal data with low signal-to-noise ratio. Inputting the obtained low signal-to-noise ratio wireless signal data into a wireless signal noise reduction model, obtaining a feature layer with the size of 128 x 32 after passing through an LSTM1, and passing through a time attention module alpha1Performing time attention on the feature layers calculated in the LSTM1, more fully extracting the features on the wireless signal data time sequence, obtaining the feature layer with the size of 128 x 64 through the LSTM2, and performing time attention through a time attention module alpha2More fully extract the characteristics of the wireless signal data time sequence, and use the step length of 1 and the size of [5,5,1 ]]The convolution check of 128 × 64 performs convolution operation on the feature layer of 128 × 64 to extract local features of the wireless signal, so as to obtain the feature layer with the size of 128 × 64, obtain the feature layer with the size of 128 × 32 through the LSTM3, and pass through the time attention module alpha3The time sequence characteristics of the wireless signal data are extracted more fully, and output with the size of 128 x 2 is obtained through the LSTM4, and the output is the wireless signal data after noise reduction.
The structure of the wireless signal noise reduction discrimination model GAN-D is shown in FIG. 1: the size of the input signal is 128 x 2, a feature layer with the size of 128 x 64 is obtained by extracting features on a wireless signal time sequence through an LSTM network, a feature layer with the size of 64 is obtained by using a full-connection FC to synthesize the features of the previous layer, and an output layer with the size of 1, namely, the non-normalized confidence coefficient, is obtained by using the full-connection FC. And performing cross entropy calculation loss on the obtained non-normalized confidence coefficient 0 or 1 matrix, and quantitatively calculating the difference between the denoised wireless signal and the original wireless signal. The wireless signal noise reduction discrimination model GAN-D inputs an original wireless signal and a noise-reduced wireless signal, and outputs a judgment result of the noise-reduced wireless signal.
And performing noise reduction processing on the wireless signal with low signal-to-noise ratio through the wireless signal noise reduction model GAN-G, and taking the output noise-reduced wireless signal and the original wireless signal as the input of the wireless signal noise reduction discrimination model GAN-D. And adjusting parameters of the GAN-G and the GAN-D through the differential judgment of the GAN-D on the noise reduction wireless signals generated by the GAN-G. The game of the GAN-D and GAN-G enables the noise reduction wireless signals output by the GAN-G to be closer to the distribution of the original wireless signals, and the high efficiency and the accuracy of the noise reducer are improved. The method for carrying out noise reduction processing on the wireless signal data can effectively improve the accuracy of the identification of the wireless signal modulation type in a low signal-to-noise ratio interval. The efficiency of wireless signal modulation type identification is greatly improved.
In order to avoid model collapse of the generative countermeasure network in the training process and simultaneously more accurately reduce noise of the wireless signal with low signal to noise ratio, the wireless signal noise reduction method based on the generative countermeasure network adopts the following training method:
pre-training GAN-D and GAN-D: for original wireless signal xsignalProcessing the added noise to obtain a low signal-to-noise ratio wireless signal added with noise
Figure BDA0001981249800000101
Low signal-to-noise ratio wireless signal
Figure BDA0001981249800000102
The output is the wireless signal after noise reduction as the input of the GAN-G
Figure BDA0001981249800000103
Will reduce the wireless signal after the noise
Figure BDA0001981249800000104
And corresponding original wireless signal xsignalAnd outputting the signal serving as the input of the GAN-D, namely judging the difference between the noise-reduced wireless signal and the original wireless signal, and finishing the pre-training when the iteration times (epochs) reach a preset value T1.
Antagonistic training of GAN-G and GAN-D: the specific process is as follows:
(1) fixing parameters of wireless signal noise reduction model GAN-G, and inputting wireless signal with low signal-to-noise ratio
Figure BDA0001981249800000105
Obtaining the wireless signal after noise reduction in a wireless signal noise reduction model GAN-G
Figure BDA0001981249800000106
Will reduce the wireless signal after the noise
Figure BDA0001981249800000107
And corresponding original wireless signal xsignalAnd the obtained output and a1 matrix or a 0 matrix with the same size are subjected to cross entropy calculation loss as the input of a wireless signal noise reduction discrimination model GAN-D, parameters of the training GAN-D are fed back through a minimized loss function, so that the wireless signal noise reduction discrimination model can accurately discriminate the noise-reduced wireless signal from the original wireless signal, the class mark of the noise-reduced wireless signal in the training process is set to be 0, and the class mark of the corresponding original wireless signal is set to be 1.
The optimization target in the training process is as follows:
Figure BDA0001981249800000108
wherein D (-) represents a wireless signal noise reduction discrimination model GAN-D, G (-) represents a wireless signal noise reduction model GAN-G, xsignalRepresenting the data of the original wireless signal and,
Figure BDA0001981249800000109
representing low signal-to-noise ratio wireless signal data, E (-) representing mathematical expectation, x-pdata (x)signal) Representing x samples from an original wireless signal data set xsignal，
Figure BDA0001981249800000111
Representing x-samples from a low signal-to-noise ratio wireless signal dataset
Figure BDA0001981249800000112
(2) Fixing the parameters of a wireless signal noise reduction discrimination model GAN-D to reduce the signal-to-noise ratio of the wireless signal
Figure BDA0001981249800000113
Inputting the data into a wireless signal noise reduction model GAN-G to obtain noise-reduced wireless signal data
Figure BDA0001981249800000114
And inputting the wireless signal subjected to noise reduction into a wireless signal noise reduction discrimination model GAN-D, carrying out cross entropy calculation loss on the obtained output, wherein 1 represents the class mark of the wireless signal subjected to noise reduction and output by GAN-G, and the parameters of GAN-D are trained by minimizing the loss feedback.
The optimization goals of the process are as follows:
Figure BDA0001981249800000115
wherein D (-) represents a wireless signal noise reduction discrimination model GAN-D, G (-) represents a wireless signal noise reduction model GAN-G,
Figure BDA0001981249800000116
representing low signal-to-noise ratio wireless signal data, E (-) representing a mathematical expectation,
Figure BDA0001981249800000117
representing x-samples from a low signal-to-noise ratio wireless signal dataset
Figure BDA0001981249800000118
(3) And (3) repeating the steps (1) to (2) until the wireless signal noise reduction model GAN-G and the wireless signal noise reduction discrimination model GAN-D realize Nash equilibrium or reach a preset iteration number.
There are times when nash equalization is implemented:
Figure BDA0001981249800000119
Figure BDA00019812498000001110
through the countermeasure training of the GAN-D and the GAN-G, the optimization target is realized, so that the wireless signals generated by the GAN-G after noise reduction are closer to the distribution of the original signals, namely, the noise reduction is realized on the basis of reserving the information of the original signals to the maximum extent.
(4) The time attention module alpha shown in figures 2(a), 2(b) and 2(c) is introduced on the basis of GAN-G1Time attention module alpha2Time attention module alpha3And retraining the wireless signal noise reduction network. By introducing the time attention module consisting of the LSTM and the full connection, the characteristics on the wireless signal time sequence are more fully extracted, so that the wireless signal noise reduction model GAN-G can better learn the time sequence characteristics of the wireless signal, and the accuracy of wireless signal noise reduction is improved.
The time attention module has the following specific structure:
time attention module alpha1: LSTM1 outputs a feature layer with size of 128 x 32, obtains a feature layer with size of 128 x 2 through LSTM _ A1, performs matrix multiplication operation on the transpose of the feature layer and an input low signal-to-noise ratio wireless signal to obtain a feature layer with size of 128 x 128, inputs the feature layer into LSTM _ B1 to obtain a time attention weight W with size of 128 x 321Feeding back the temporal attention weight to LSTM1 changes the weight when calculating the output, as follows:
Figure BDA0001981249800000121
f1＝W1*f1
wherein the content of the first and second substances,
Figure BDA0001981249800000122
representing a matrix multiplication operation, f1Feature layer output for LSTM1, size 128 x 32, fA1The feature layer output for LSTM _ a1, size 128 x 2,
Figure BDA0001981249800000123
denotes fA1Transposed matrix of (1), F1To obtain a feature layer of size 128 x 128, W1The temporal attention weight output for LSTM _ B1, with a size of 128 x 32, represents the multiplication of the corresponding elements of the two matrices.
Module alpha for temporal attention2: LSTM2 outputs feature layer with size of 128 x 64, obtains feature layer with size of 128 x 2 through LSTM _ A2, the transposition of the feature layer and the input low signal-to-noise ratio wireless signal are processed by matrix multiplication to obtain feature layer with size of 128 x 128, the feature layer is input into LSTM _ B2 to obtain time attention weight W with size of 128 x 642Feeding back the temporal attention weight to LSTM2 changes the weight of the calculated output in the following way:
Figure BDA0001981249800000124
f2＝W2*f2
wherein the content of the first and second substances,
Figure BDA0001981249800000131
representing a matrix multiplication operation, f2Feature layer output for LSTM2, size 128 x 64, fA2The feature layer output for LSTM _ a2, size 128 x 2,
Figure BDA0001981249800000132
denotes fA2Transposed matrix of (1), F2To obtain a feature layer of size 128 x 128, W2The temporal attention weight output for LSTM _ B2, size 128 x 64, represents the multiplication of the corresponding elements of the two matrices.
Module alpha for temporal attention3: LSTM3 outputs feature layer with size of 128X 32, feature layer with size of 128X 2 is obtained through LSTM _ A3, the transposition of the feature layer and the input low signal-to-noise ratio wireless signal are subjected to matrix multiplication to obtain feature layer with size of 128X 128, the feature layer is input into LSTM _ B3 to obtain time attention weight W with size of 128X 323The temporal attention weight is fed back to LSTM3 to change its weight in the following way:
Figure BDA0001981249800000133
f3＝W3*f3
wherein
Figure BDA0001981249800000134
Representing a matrix multiplication operation, f3Feature layer output for LSTM3, size 128 x 32, fA3The feature layer output for LSTM _ a3, size 128 x 32,
Figure BDA0001981249800000135
denotes fA3Transposed matrix of (1), F3To obtain a feature layer of size 128 x 128, W3The temporal attention weight output for LSTM _ B3, with a size of 128 x 32, represents the multiplication of the corresponding elements of the two matrices.
By designing the training method, the trained wireless signal noise reduction model GAN-G can be used as a noise filter of a wireless signal, and noise filtering can be performed on the basis of reserving original signal data information to the maximum extent. The efficiency of the wireless signal noise reducer is improved by introducing the time attention module; the accuracy of the wireless signal noise reducer is improved through the countermeasure training of the wireless signal noise reduction model GAN-G and the wireless signal noise reduction discrimination model GAN-D.
Specific experiments are as follows:
the basic cases of the data set include: (1) the raw wireless signal data set has 132000 training samples and 33000 test samples, each sample being a matrix of size [128,2 ]; (2) the data sets can be divided into 11 types according to different wireless signal modulation types, each type is divided equally, namely 12000 samples exist in each type in a training set, and 3000 samples exist in each type in a testing set; (3) each modulation type signal has 20 different signal-to-noise ratios (the signal-to-noise ratio has an even number of-20 to 18). (4) Gaussian white noise is added into the wireless telecommunication signal with the signal-to-noise ratio of 18DB to obtain a low signal-to-noise ratio wireless signal data set corresponding to the Gaussian white noise.
Case of wireless signal modulation type identifier:
the deep learning based wireless signal modulation type recognizer, consisting of LSTM and full concatenation, is trained using the raw wireless signal data set described above.
The specific structure of the wireless signal modulation type identifier is shown in fig. 3: the data of the wireless signal with the size of 128 × 2 is input, the feature layer with the size of 128 × 128 is obtained through the LSTM, the feature layer with the size of 128 × 32 is obtained through the LSTM, and the output layer with the size of 11 is obtained through the full connection FC. I.e. 11 represents a modulation type of class 11 in the wireless signal data set.
As shown in fig. 4, using the wireless signal modulation type identifier, the modulation type identification results of the wireless signal with low snr and the wireless signal after noise reduction are respectively input and compared:
and training a wireless signal noise reduction model by using the original training set with the signal-to-noise ratio of 18DB and the low signal-to-noise ratio wireless signal added with noise. And inputting the wireless signals with the signal-to-noise ratio lower than-10 DB in the original wireless signal test set into a wireless signal noise reduction model GAN-G to obtain corresponding noise-reduced wireless signal data.
Inputting wireless signals with signal-to-noise ratio of-12 to-20 DB in an original wireless signal data set into a wireless signal modulation type identifier to obtain the identification result of the wireless signal modulation type before noise reduction; and inputting the wireless signal data subjected to noise reduction into a wireless signal modulation type identifier to obtain the identification result of the wireless signal modulation type subjected to noise reduction.
Experimental results show that the accuracy of identifying the modulation type of the wireless signal in a low signal-to-noise ratio interval can be effectively improved by carrying out noise reduction processing on the wireless signal data by the method. The efficiency of wireless signal modulation type identification is greatly improved. The wireless signal noise reduction model achieves the expected effect.
The above-mentioned embodiments are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions, equivalents, etc. made within the scope of the principles of the present invention should be included in the scope of the present invention.

Claims (5)

1. A wireless signal noise reduction method based on a generative countermeasure network comprises the following steps:
constructing a wireless signal noise reduction network and a wireless signal noise reduction judging network, wherein the wireless signal noise reduction network comprises a long-time memory neural network, a convolution layer and a time attention module, the input of the wireless signal noise reduction network is a low signal-to-noise ratio wireless signal added with noise, and the output of the wireless signal noise reduction network is a wireless signal subjected to noise reduction;
the wireless signal noise reduction judging network comprises a long-time memory neural network and a full connection layer, the input of the wireless signal noise reduction judging network is connected with the output of the wireless signal noise reduction network, the input data is an original wireless signal without noise and a wireless signal after noise reduction, and the output is a judging result of the wireless signal after noise reduction and the original wireless signal without noise;
training a wireless signal noise reduction network and a wireless signal noise reduction judging network by adopting a confrontation training strategy to obtain a wireless signal noise reduction model and a wireless signal noise reduction judging model;
extracting a wireless signal noise reduction model to process a wireless signal to be noise reduced to obtain a noise-reduced wireless signal;
(a) pre-training a wireless signal noise reduction network and a wireless signal noise reduction discrimination network;
(b) the method comprises the following steps of training a wireless signal noise reduction network GAN-G and a wireless signal noise reduction discrimination network GAN-D after pre-training, wherein the specific process comprises the following steps:
fixing parameters of the GAN-G, inputting the wireless signals subjected to noise reduction into the GAN-D, and training the parameters of the GAN-D; fixing parameters of the GAN-D, inputting a low signal-to-noise ratio wireless signal into the GAN-G, and training the parameters of the GAN-G;
(c) repeating the steps (a) and (b), alternately training parameters of the GAN-G and the GAN-D until the two realize G-D Nash equilibrium or reach the maximum training times T2;
(d) and (4) adding a time attention module consisting of a long-time memory network and then training the GAN-G, namely repeating the steps (a) to (c), so that the GAN-G more efficiently extracts the time sequence characteristics of the wireless signals and generates the wireless signals after noise reduction.
2. The method of wireless signal noise reduction based on a generative countermeasure network as recited in claim 1, wherein the wireless signal noise reduction network comprises:
LSTM1, LSTM2, convolutional layer, LSTM3 and LSTM4 which are connected in sequence;
time attention module alpha1: LSTM1 outputs a feature layer with the size of time x 32, a feature layer with the size of time x 2 is obtained through LSTM _ A1, the transposition of the feature layer and an input low signal-to-noise ratio wireless signal are subjected to matrix multiplication to obtain a feature layer with the size of time x time, the feature layer is input into LSTM _ B1 to obtain a time attention weight W with the size of time x 321Feeding back the temporal attention weight to LSTM1 changes the weight at the time of calculating the output;
time attention module alpha2: LSTM2 outputs a feature layer with the size of time x 64, obtains a feature layer with the size of time x 2 through LSTM _ A2, performs matrix multiplication on the transpose of the feature layer and an input low signal-to-noise ratio wireless signal to obtain a feature layer with the size of time x time, inputs the feature layer into LSTM _ B2 and obtains a time attention weight W with the size of time x 642Feeding back the temporal attention weight to LSTM2 changes the weight of the calculated output;
time attention module alpha3: LSTM3 output rulerObtaining a feature layer with the size of time x 32 through LSTM _ A3, performing matrix multiplication on the transpose of the feature layer and an input low signal-to-noise ratio wireless signal to obtain a feature layer with the size of time x 2, inputting the feature layer into LSTM _ B3 to obtain a time attention weight W with the size of time x 323The temporal attention weight is fed back into LSTM3 to change its weight.
3. The method of claim 1, wherein the wireless signal noise reduction discrimination network comprises:
the LSTM is used for extracting the time sequence characteristics of the input wireless signals;
the first full connection layer is used for synthesizing the time sequence characteristics of the LSTM output;
and the second full connection layer is used as an output layer and used for integrating the characteristics output by the first full connection layer and outputting a classification result.
4. The method as claimed in claim 1, wherein the training of the wireless signal noise reduction network GAN-G during the countermeasure training comprises:
inputting the wireless signal output by the GAN-G after noise reduction into the GAN-D, and performing cross entropy calculation loss on the obtained output and 0; inputting an original wireless signal into the GAN-D, and performing cross entropy calculation loss on the obtained output and 1;
the optimization goals of the process are as follows:
Figure FDA0003289598170000031
wherein D (-) represents the wireless signal noise reduction discrimination network GAN-D, G (-) represents the wireless signal noise reduction network GAN-G, xsignalRepresenting the data of the original wireless signal and,
Figure FDA0003289598170000032
representing low signal-to-noise ratio wireless signalsData, E (-) represents the mathematical expectation, x-pdata (x)signal) Representing x samples from an original wireless signal data set xsignal，
Figure FDA0003289598170000033
Representing x-samples from a low signal-to-noise ratio wireless signal dataset
Figure FDA0003289598170000034
5. The method as claimed in claim 1, wherein the training of the wireless signal noise reduction discrimination network GAN-D during the countermeasure training comprises:
inputting a low signal-to-noise ratio wireless signal into a wireless signal noise reduction network GAN-G, inputting the obtained noise-reduced wireless signal into a wireless signal noise reduction discrimination network GAN-D, performing cross entropy calculation loss on the obtained output, and training parameters of the GAN-G by minimizing the loss feedback:
the optimization goals of the process are as follows:
Figure FDA0003289598170000035
wherein D (-) represents a wireless signal noise reduction discrimination network GAN-D, G (-) represents a wireless signal noise reduction network GAN-G,
Figure FDA0003289598170000036
representing low signal-to-noise ratio wireless signal data, E (-) representing a mathematical expectation,
Figure FDA0003289598170000037
representing x-samples from a low signal-to-noise ratio wireless signal dataset
Figure FDA0003289598170000038
CN201910150001.7A 2019-02-28 2019-02-28 Wireless signal noise reduction method based on generative countermeasure network Active CN109890043B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910150001.7A CN109890043B (en) 2019-02-28 2019-02-28 Wireless signal noise reduction method based on generative countermeasure network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910150001.7A CN109890043B (en) 2019-02-28 2019-02-28 Wireless signal noise reduction method based on generative countermeasure network

Publications (2)

Publication Number Publication Date
CN109890043A CN109890043A (en) 2019-06-14
CN109890043B true CN109890043B (en) 2021-12-21

Family

ID=66929872

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910150001.7A Active CN109890043B (en) 2019-02-28 2019-02-28 Wireless signal noise reduction method based on generative countermeasure network

Country Status (1)

Country Link
CN (1) CN109890043B (en)

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110969186B (en) * 2019-10-28 2023-04-07 浙江工业大学 Channel detection-based attack-resisting defense method and device facing wireless signal identification
CN112132266A (en) * 2020-09-23 2020-12-25 中国人民解放军陆军工程大学 Signal modulation identification system and modulation identification method based on convolution cycle network
CN112307926B (en) * 2020-10-26 2022-12-06 西北工业大学 Acoustic passive ship target classification method based on generation countermeasure network
CN112487914B (en) * 2020-11-25 2021-08-31 山东省人工智能研究院 ECG noise reduction method based on deep convolution generation countermeasure network
WO2022151067A1 (en) * 2021-01-13 2022-07-21 Oppo广东移动通信有限公司 Wireless signal noise reduction method and apparatus, device, and storage medium
CN113191321B (en) * 2021-05-21 2022-04-22 电子科技大学 Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network
CN113349800A (en) * 2021-06-10 2021-09-07 重庆工程学院 Brain wave signal noise reduction method based on generating type countermeasure network
CN113626960B (en) * 2021-07-29 2022-06-14 南京信息工程大学 Anti-mode coupling few-mode signal complex format analysis method and device
CN115442191B (en) * 2022-11-08 2023-03-24 中国人民解放军战略支援部队航天工程大学 Communication signal noise reduction method and system based on relative average generation countermeasure network
CN116956201B (en) * 2023-09-19 2023-12-08 成都中轨轨道设备有限公司 Intelligent coupling early warning method for big data decision

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11501152B2 (en) * 2017-05-05 2022-11-15 Intel Corporation Efficient learning and using of topologies of neural networks in machine learning
GB201710877D0 (en) * 2017-07-06 2017-08-23 Nokia Technologies Oy A method and an apparatus for evaluating generative machine learning model
CN108616470A (en) * 2018-03-26 2018-10-02 天津大学 Modulation Signals Recognition method based on convolutional neural networks
CN108877832B (en) * 2018-05-29 2022-12-23 东华大学 Audio tone quality restoration system based on GAN
CN108986797B (en) * 2018-08-06 2021-07-06 中国科学技术大学 Voice theme recognition method and system
CN108986835B (en) * 2018-08-28 2019-11-26 百度在线网络技术（北京）有限公司 Based on speech de-noising method, apparatus, equipment and the medium for improving GAN network

Also Published As

Publication number Publication date
CN109890043A (en) 2019-06-14

Similar Documents

Publication Publication Date Title
CN109890043B (en) Wireless signal noise reduction method based on generative countermeasure network
Zhou et al. Adaptive genetic algorithm-aided neural network with channel state information tensor decomposition for indoor localization
CN110600017B (en) Training method of voice processing model, voice recognition method, system and device
CN107393542B (en) Bird species identification method based on two-channel neural network
CN112464837B (en) Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples
CN112259120B (en) Single-channel human voice and background voice separation method based on convolution cyclic neural network
CN111723701B (en) Underwater target identification method
CN113488060B (en) Voiceprint recognition method and system based on variation information bottleneck
CN112735460B (en) Beam forming method and system based on time-frequency masking value estimation
CN109887047B (en) Signal-image translation method based on generation type countermeasure network
CN111968666B (en) Hearing aid voice enhancement method based on depth domain self-adaptive network
CN111222442A (en) Electromagnetic signal classification method and device
CN112115821B (en) Multi-signal intelligent modulation mode identification method based on wavelet approximate coefficient entropy
CN112992173B (en) Signal separation and denoising method based on improved BCA blind source separation
CN110808057A (en) Voice enhancement method for generating confrontation network based on constraint naive
CN111310719B (en) Unknown radiation source individual identification and detection method
CN115442191A (en) Communication signal noise reduction method and system based on relative average generation countermeasure network
CN111899750A (en) Speech enhancement algorithm combining cochlear speech features and hopping deep neural network
CN111968669B (en) Multi-element mixed sound signal separation method and device
CN116434759B (en) Speaker identification method based on SRS-CL network
CN113095113B (en) Wavelet line spectrum feature extraction method and system for underwater target identification
CN117037824A (en) Data enhancement method and system for acoustic scene classification
CN116738278A (en) Automatic modulation classification method based on wavelet transformation and multi-mode feature fusion
CN114626412A (en) Multi-class target identification method and system for unattended sensor system
CN111785262B (en) Speaker age and gender classification method based on residual error network and fusion characteristics

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant