CN115442191B - Communication signal noise reduction method and system based on relative average generation countermeasure network - Google Patents

Communication signal noise reduction method and system based on relative average generation countermeasure network Download PDF

Info

Publication number
CN115442191B
CN115442191B CN202211387887.5A CN202211387887A CN115442191B CN 115442191 B CN115442191 B CN 115442191B CN 202211387887 A CN202211387887 A CN 202211387887A CN 115442191 B CN115442191 B CN 115442191B
Authority
CN
China
Prior art keywords
signal
noise reduction
loss
noise
clean
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
CN202211387887.5A
Other languages
Chinese (zh)
Other versions
CN115442191A (en
Inventor
彭亮
范有臣
胡豪杰
方胜良
刘涵
温晓敏
徐照菁
马昭
王孟涛
程东航
王梦阳
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.)
Peoples Liberation Army Strategic Support Force Aerospace Engineering University
Original Assignee
Peoples Liberation Army Strategic Support Force Aerospace Engineering University
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 Peoples Liberation Army Strategic Support Force Aerospace Engineering University filed Critical Peoples Liberation Army Strategic Support Force Aerospace Engineering University
Priority to CN202211387887.5A priority Critical patent/CN115442191B/en
Publication of CN115442191A publication Critical patent/CN115442191A/en
Application granted granted Critical
Publication of CN115442191B publication Critical patent/CN115442191B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/08—Modifications for reducing interference; Modifications for reducing effects due to line faults ; Receiver end arrangements for detecting or overcoming line faults
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00—Modulated-carrier systems
    • H04L27/0008—Modulated-carrier systems arrangements for allowing a transmitter or receiver to use more than one type of modulation
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Power Engineering (AREA)
  • Noise Elimination (AREA)

Abstract

The invention relates to the technical field of radio communication, and particularly discloses a communication signal noise reduction method and a system based on a relative average generation countermeasure network, wherein the method comprises the steps of obtaining a clean signal, and obtaining a corresponding noisy signal based on the clean signal; constructing a training set and a test set based on the clean signal and the noisy signal corresponding to the clean signal; constructing a noise reduction model for generating a countermeasure network based on relative average; training a noise reduction model for generating a countermeasure network based on relative average based on the training set to extract a trained generator; reducing the noisy signal to a clean signal based on the trained generator; the method realizes the end-to-end de-noising of the radio signal, and retains the essential characteristics of the signal after the de-noising, which has important significance for the subsequent processing of signal modulation identification, signal demodulation and the like.

Description

Communication signal noise reduction method and system based on relative average generation countermeasure network
Technical Field
The invention relates to the technical field of radio communication, in particular to a communication signal noise reduction method and system based on a relative average generation countermeasure network.
Background
The premise of the radio signal modulation and identification as signal demodulation is the important research content in the communication field, and plays an important role in civil and military signal communication; however, due to the complexity and variability of the signal transmission environment and the problems of the receiving device, the received signal always has a certain amount of noise, which brings difficulties to the modulation identification and signal demodulation of the received signal; the effective means for solving the problem is to reduce the noise of the received signal, and the noise reduction algorithm which can reduce the noise of the signal and simultaneously reserve the essential characteristics of the signal has important significance for realizing the modulation recognition and accurate demodulation of the received signal.
The signal denoising is to extract a clean signal from a noisy signal, which is a difficult problem in the field of signal processing, and the effective denoising of the signal has important significance for signal communication; most of the existing signal denoising methods are methods based on signal processing, such as a denoising method based on wavelet decomposition, a denoising method based on Empirical Mode Decomposition (EMD), and the like; the denoising effect of wavelet decomposition is closely related to the number of decomposition layers and the selection of a threshold, the denoising effect is influenced by too few decomposition layers, signal distortion is caused by too many decomposition layers, and the selection of the optimal decomposition layers and the threshold is usually determined by experience; the noise reduction algorithm based on the EMD is superior to the noise reduction algorithm based on the EMD, but has the problems of modal aliasing, end point effect and the like.
In recent years, deep learning techniques have been widely used in the fields of speech processing, image processing, natural language processing, and the like; with the development of the technology, the neural network is also applied to the field of signal processing, such as signal modulation identification, signal data enhancement and the like; signal noise reduction has been a difficult problem in the field of signal processing, and some scholars have attempted to solve this problem using neural networks;
the method comprises the following steps that (1) an IAFNet which is published by Haiwang Wang et al on IEEE Communications Letters, wherein Few-Shot Learning for Modulation Recognition in an underlying water impulse Noise [ J ] carries out pulse Noise preprocessing on signals before signal Modulation Recognition, and under the condition of less marked samples, signal features are more effectively extracted through denoising and task driving, so that the Modulation Recognition precision of Underwater acoustic signals is improved; plum and bin et al disclose an underwater acoustic communication signal noise reduction algorithm [ J ] based on RCGAN on the electronic newspaper, which uses a relative condition generation confrontation network algorithm to perform noise reduction processing on an underwater acoustic communication signal, and applies the algorithm from an emulation signal to an actual signal through transfer learning, wherein the two methods have large network parameters, and a good effect can be obtained by performing fine tuning for many times; weijiang Zhu et al, in IEEE Transactions on Geoscience and remove Sensing, using Deep Neural Networks [ J ], denoise Seismic signals by Deep learning, use Seismic images as data sets, and process the Seismic signals by Convolutional Neural Networks (CNN), wherein the method does not cause large deviation to amplitude, time and phase information after Denoising the signals, but cannot realize perfect separation of the signals and noise; in a Denoising of Radio Frequency Partial Discharge signaling Using Artificial Neural Network [ J ] published by air Abbas Soltani et al in engines, an Artificial Neural Network (ANN) is used for Denoising a measured Radio Frequency signal emitted by a Partial Discharge source and converting the Denoising problem into a curve fitting problem.
Disclosure of Invention
In view of the above problems, an object of the present invention is to provide a communication signal noise reduction method based on a relative average generation countermeasure network, which takes a time domain waveform of a radio receiving signal as a processing object, and can map the waveform of the noise signal to a clean signal waveform, so as to implement end-to-end noise reduction of the radio signal, and retain the essential characteristics of the signal after noise reduction, which is significant for subsequent processing such as signal modulation identification and signal demodulation.
It is a second object of the present invention to provide a communication signal noise reduction system that generates a countermeasure network based on relative averaging.
The first technical scheme adopted by the invention is as follows: a method for reducing noise in a communication signal based on a relative average generation countermeasure network, comprising the steps of:
s100: acquiring a clean signal, and acquiring a corresponding noisy signal based on the clean signal; constructing a training set and a test set based on the clean signal and the corresponding noisy signal;
s200: constructing a noise reduction model for generating a countermeasure network based on relative average;
s300: training a noise reduction model for generating a countermeasure network based on relative average based on the training set to extract a trained generator;
s400: and reducing the noisy signal into a clean signal based on the trained generator.
Preferably, the noise reduction model for generating the countermeasure network based on the relative average in the step S200 includes an arbiter and a generator;
the generator and the discriminator both comprise two fully-connected layers and two bidirectional long-time and short-time memory models; the discriminator is a discriminator for relatively averagely generating the countermeasure network.
Preferably, the number of hidden layer nodes in the bidirectional long-and-short term memory model layer is 128.
Preferably, the training of the noise reduction model based on the relative average generation countermeasure network in step S300 includes the following sub-steps:
s310: inputting the signal with noise in the training set into a generator, outputting a de-noising signal, inputting a clean signal and the de-noising signal into a discriminator, calculating discrimination loss, and updating the discriminator based on the discrimination loss;
s320: calculating a signal content loss and a countermeasure loss through the denoised signal and the clean signal, calculating a total loss of the generator based on the signal content loss and the countermeasure loss, and updating the generator according to the total loss;
s330: and repeating the steps S310 to S320, and iterating for N times in this way to finish the training of the noise reduction model.
Preferably, the discriminant loss is calculated by the following formula:
Figure 131655DEST_PATH_IMAGE001
in the formula (I), the compound is shown in the specification,
Figure 449635DEST_PATH_IMAGE002
loss of relative discriminators; />
Figure 682646DEST_PATH_IMAGE003
Averaging all data in the small batch;
Figure 945131DEST_PATH_IMAGE004
the output when the input of the relative discriminator is a real signal; />
Figure 455878DEST_PATH_IMAGE005
Is the output of the phase discriminator when the input is the generated signal.
Preferably, the signal content loss is calculated by the following formula:
Figure 752998DEST_PATH_IMAGE006
in the formula (I), the compound is shown in the specification,
Figure 829014DEST_PATH_IMAGE007
is a loss of signal content; />
Figure 844375DEST_PATH_IMAGE008
Is the mean square error between the denoised signal and the clean signal; />
Figure 893233DEST_PATH_IMAGE009
Is the absolute value error between the denoised signal and the clean signal. />
Preferably, the challenge loss is calculated by the following formula:
Figure 917297DEST_PATH_IMAGE010
in the formula (I), the compound is shown in the specification,
Figure 291777DEST_PATH_IMAGE011
to combat the loss; />
Figure 466538DEST_PATH_IMAGE003
Averaging all data in the small batch;
Figure 519420DEST_PATH_IMAGE004
the output when the input of the relative discriminator is a real signal; />
Figure 791132DEST_PATH_IMAGE005
The relative discriminator input is the output when the signal is generated.
Preferably, before extracting the trained generator in step S300, the method further includes evaluating the trained noise reduction model based on the test set.
Preferably, the evaluating the trained noise reduction model based on the test set comprises:
inputting the signal with noise in the test set into a generator after training to obtain a de-noising signal; if the signal-to-noise ratio of the de-noising signal is increased by more than a threshold value compared with the signal-to-noise ratio of the signal with noise, judging that the performance of the generator is good, and extracting the trained generator; if the signal-to-noise ratio of the denoised signal does not increase more than a threshold compared to the signal-to-noise ratio of the noisy signal, the training is re-performed.
The second technical scheme adopted by the invention is as follows: a communication signal noise reduction system based on a relative average generation countermeasure network comprises a data set construction module, a noise reduction model construction module, a training module and a noise reduction module;
the data set construction module is used for acquiring a clean signal and obtaining a corresponding noisy signal based on the clean signal; constructing a training set and a test set based on the clean signal and the corresponding noisy signal;
the noise reduction model building module is used for building a noise reduction model for generating a countermeasure network based on relative average;
the training module is used for training a noise reduction model for generating a confrontation network based on relative average based on the training set so as to extract a trained generator;
and the noise reduction module is used for reducing the noisy signals into clean signals based on the trained generator.
The beneficial effects of the above technical scheme are that:
(1) The communication signal noise reduction method based on the relative average generation countermeasure network disclosed by the invention takes the radio receiving signal time domain waveform as a processing object, can map the noise signal waveform to a clean signal waveform, realizes the end-to-end noise reduction of the radio signal, and retains the essential characteristics of the signal after noise reduction, which has important significance for subsequent processing such as signal modulation identification, signal demodulation and the like.
(2) Firstly, training a noise reduction model based on a relative average generation countermeasure network, inputting noisy signal data in a training set into a generator, and outputting a generated signal by the generator; then the generated signal and the corresponding clean signal are sequentially input into a discriminator, the discriminator discriminates the input signal and judges whether the input signal is the signal generated by the generator or the clean signal without noise; and after continuous iterative training, the self optimization of the noise reduction model is realized by using an error back-propagation algorithm, and after the training of the noise reduction model is finished, the signal with noise can be restored into a clean signal by using the generator.
Drawings
Fig. 1 is a schematic flow chart of a method for reducing noise of a communication signal of a countermeasure network based on relative average generation according to an embodiment of the present invention;
FIG. 2 is a waveform diagram of a clean signal provided by one embodiment of the present invention;
FIG. 3 is a waveform diagram of a noisy signal provided by an embodiment of the present invention;
FIG. 4 is a diagram illustrating a noise reduction model for generating a countermeasure network based on relative averaging according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a generator network architecture provided by one embodiment of the present invention;
FIG. 6 is a diagram illustrating a network structure of a discriminator according to an embodiment of the present invention;
FIG. 7 is a comparison graph of noise reduction performance for different numbers of hidden nodes according to an embodiment of the present invention;
fig. 8 is a comparison graph of time domain waveforms before and after noise reduction of an 8PSK signal according to an embodiment of the present invention;
FIG. 9 is a comparison graph of noise reduction performance of different methods provided by an embodiment of the present invention;
fig. 10 is a schematic structural diagram of a communication signal noise reduction system for generating a countermeasure network based on relative averaging according to an embodiment of the present invention.
Detailed Description
The embodiments of the present invention will be described in further detail with reference to the drawings and examples. The following detailed description of the embodiments and the accompanying drawings are provided to illustrate the principles of the invention and are not intended to limit the scope of the invention, which is defined by the claims, i.e., the invention is not limited to the preferred embodiments described.
In the description of the present invention, it is to be noted that, unless otherwise specified, "a plurality" means two or more; the terms "first," "second," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance; the specific meaning of the above terms in the present invention can be understood as appropriate to those of ordinary skill in the art.
Example one
As shown in fig. 1, an embodiment of the present invention provides a method for reducing noise of a communication signal of a countermeasure network based on relative average generation, comprising the following steps:
s100: acquiring a clean signal, and acquiring a corresponding noisy signal based on the clean signal; constructing a training set and a test set based on the clean signal and the noisy signal corresponding to the clean signal;
signals in a wireless channel are transmitted by electromagnetic wave propagation in space, and some unwanted signals, collectively referred to as noise (noise), exist in the channel; noise is an interference present in a channel, which not only limits the transmission rate of information, but also causes signal distortion; in the time domain, the received signal
Figure 805356DEST_PATH_IMAGE012
Expressed as a clean signal->
Figure 405096DEST_PATH_IMAGE013
Sum noise signal
Figure 347423DEST_PATH_IMAGE014
The received signal is represented by the following formula:
Figure 847543DEST_PATH_IMAGE015
the invention uses GNU Radio to simulate clean signals without noise by adopting eight modulation modes (8 PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM and QPSK modulation modes), and the sampling frequency is 1MHZ, so as to obtain clean signal samples as shown in figure 2.
In order to simulate the clean signal being contaminated by noise, gaussian white noise with different signal-to-noise ratios is randomly added to the clean signal, thereby generating a corresponding noisy signal sample as shown in fig. 3, wherein the signal-to-noise ratio of the noisy signal is between-8 dB and 10dB, and the interval is 2dB.
A clean signal and a noisy signal with noise added to the clean signal constitute a data, the noisy signal is input, the clean signal corresponds to a label, a total of 80000 signal data are in a data set formed by the clean signal and the noisy signal corresponding thereto, and data information of the data set is shown in table 1.
TABLE 1 data information of data set
Figure 563827DEST_PATH_IMAGE016
The data set is divided into 4:1, dividing the signal into a training set and a test set, wherein the training set comprises 64000 signal data, and the test set comprises 16000 signal data; the training set contains noisy signals
Figure 913512DEST_PATH_IMAGE017
And a corresponding clean signal->
Figure 45547DEST_PATH_IMAGE018
。
S200: constructing a noise reduction model based on a relative average generation countermeasure network (RaGAN);
noise in a wireless channel can interfere radio signal transmission, and an original useful signal can be further restored by reducing noise components in a received signal; the purpose of denoising is to filter noise from the noise-contaminated signal as much as possible (i.e., denoise signal), thereby minimizing the expected error between the clean signal and the denoised signal; the traditional signal noise reduction method is used for separating signals from noise, and a large amount of manual experience is usually needed; and the dependence on human experience can be reduced by carrying out noise reduction through a deep learning technology, and the method is widely applied to speech enhancement and image denoising.
As shown in FIG. 4, the noise reduction model for generating the countermeasure network based on relative averaging constructed by the present invention includes a Discriminator (D) and a Generator (G), which are trained alternately to solve the minimum and maximum problems of the countermeasure; the idea of this model is that the purpose of the training generator G is to fool the trained discriminator D to distinguish between the generated signal and the clean signal; by this method, the generator can learn to create a solution that is highly close to the clean signal, and is difficult to distinguish by the discriminator D, thereby achieving the purpose of denoising the noisy signal.
The application of the bidirectional long-and-short-term memory model (BilSTM) on the neural network well solves the problems of gradient loss and gradient explosion in the original Recurrent Neural Network (RNN) and has good performance in processing time sequence data, so that the BiLSTM is used as a core to construct a generator and a discriminator.
As shown in fig. 5, the generator comprises two fully connected layers and two layers of BiLSTM, and both the input and output dimensions of the generator are designed to be 128; the input dimensionality in the BilSTM layer is 1, the number of nodes of a hidden layer is 128, the Dropout factor is 0.6, and after the output of a full connection layer, a Leaky ReLU function with the activation factor of 0.2 is adopted to carry out nonlinear activation on the output; after the generator inputs data, firstly, layer Normalization (Layer Normalization) is carried out on the data, then, waveform characteristics and time characteristics of a noisy signal are extracted through full-connection mapping and a BilSTM Layer to achieve the purpose of signal denoising, the data is subjected to standardized reduction before being output to ensure the generalization capability of a model and accelerate the convergence of the model, and finally, a denoised signal is output through full-connection Layer mapping
Figure 291852DEST_PATH_IMAGE019
I.e. is->
Figure 644948DEST_PATH_IMAGE019
Is the output of the generator.
As shown in fig. 6, the discriminator is also composed of two fully-connected layers and two layers of BiLSTM, with an input dimension of 128 and an output dimension of 1; the input dimension in the BilSTM layer is 1, the number of nodes in the hidden layer is 128, and the Dropout factor is 0.6; the method comprises the steps that the layer standardization is carried out on data input by a discriminator, a Dropout layer with a factor of 0.6 is introduced after the data is output by a full connection layer to prevent model overfitting, the difference of waveforms in real signal data and generated signal data is extracted through a BilSTM layer, and then a judgment result is output through the mapping of the full connection layer.
The invention uses a discriminator for a relative average generative countermeasure network (RaGAN) whose discriminator D estimates that an input signal is a clean signal for a standard generative countermeasure network
Figure 812756DEST_PATH_IMAGE018
Or generates a signal->
Figure 545219DEST_PATH_IMAGE019
The probability of (d); while a discriminator that generates a countering network on a relative average will attempt to predict a clean signal ≥>
Figure 114872DEST_PATH_IMAGE018
Than generates a signal->
Figure 373291DEST_PATH_IMAGE019
A more realistic probability;
define the relative arbiter as
Figure 559552DEST_PATH_IMAGE020
The criterion decider can be denoted as->
Figure 830128DEST_PATH_IMAGE021
Wherein->
Figure 254287DEST_PATH_IMAGE022
Is a sigmoid function, is greater than>
Figure 355711DEST_PATH_IMAGE023
Is the untransformed arbiter output; then relative arbiter>
Figure 294848DEST_PATH_IMAGE020
Expressed by the following formula:
Figure 103535DEST_PATH_IMAGE024
in the formula (I), the compound is shown in the specification,
Figure 647780DEST_PATH_IMAGE004
the output when the input of the relative discriminator is a real signal; />
Figure 513580DEST_PATH_IMAGE022
Is sigmoid function;
Figure 877697DEST_PATH_IMAGE025
the output of the discriminator which is not transformed when the real signal is input; />
Figure 21233DEST_PATH_IMAGE003
Averaging all data in the small batch;
Figure 419985DEST_PATH_IMAGE026
the output of the discriminator is not converted when the generated signal is input.
The number of hidden layer nodes used in the BilSTM layer in the generator and the discriminator can influence the noise reduction effect of the model, and in order to find the optimal solution point number, the invention carries out experimental comparison on the BilSTM by using different numbers of hidden layer nodes; the comparison file results are shown in FIG. 7;
as can be seen from FIG. 7, the noise reduction performance of the model is closely related to the number of nodes of the hidden layer in the BilSTM, and the noise reduction performance of the model is improved along with the increase of the number of the used nodes; when the number of the nodes of the hidden layer is 128, the noise reduction performance of the model reaches the highest; however, when the number of nodes in the hidden layer exceeds 128, the effect of improving the model performance cannot be achieved by increasing the number of nodes due to the overlarge model parameter quantity.
S300: training a noise reduction model based on a relative average generation countermeasure network based on a training set to extract a trained generator;
the training set is input into a noise reduction model for generating a countermeasure network based on relative average, and specific parameters of training are shown in table 2.
TABLE 2 noise reduction model training parameters
Figure 863211DEST_PATH_IMAGE027
The specific process of training the noise reduction model based on the relative average generation countermeasure network based on the training set comprises the following steps:
s310: the noisy signal in the training set
Figure 511361DEST_PATH_IMAGE028
The signal is input into a generator and output to obtain a de-noising signal->
Figure 661851DEST_PATH_IMAGE029
Will make the clean signal->
Figure 607721DEST_PATH_IMAGE030
And de-noising signal>
Figure 614991DEST_PATH_IMAGE029
Inputting the result into a discriminator, and calculating the discrimination loss->
Figure 953699DEST_PATH_IMAGE031
And updating the discriminator;
wherein the loss function of the relative arbiter is defined as follows:
Figure 173459DEST_PATH_IMAGE001
in the formula (I), the compound is shown in the specification,
Figure 278294DEST_PATH_IMAGE002
loss of relative discriminators; />
Figure 659728DEST_PATH_IMAGE003
Averaging all data in the small batch;
Figure 16891DEST_PATH_IMAGE004
the output when the input of the relative discriminator is a real signal; />
Figure 774763DEST_PATH_IMAGE005
The relative discriminator input is the output when the signal is generated.
S320: by de-noising the signal
Figure 999683DEST_PATH_IMAGE029
And a clean signal->
Figure 552018DEST_PATH_IMAGE030
Evaluating the loss of signal content>
Figure 865319DEST_PATH_IMAGE032
And against loss>
Figure 426882DEST_PATH_IMAGE033
Based on the loss of content >>
Figure 506308DEST_PATH_IMAGE032
And against loss>
Figure 495124DEST_PATH_IMAGE033
Calculate total loss of generator->
Figure 295721DEST_PATH_IMAGE034
And based on the total loss->
Figure 129816DEST_PATH_IMAGE034
Updating the generator G;
wherein the overall loss function of the generator is represented as a weighted sum of the signal content loss and the countervailing loss, the loss function being represented by the following formula:
Figure 329329DEST_PATH_IMAGE035
in the formula (I), the compound is shown in the specification,
Figure 161150DEST_PATH_IMAGE036
to the total loss of the generator; />
Figure 917884DEST_PATH_IMAGE007
Is a loss of signal content; />
Figure 818320DEST_PATH_IMAGE011
To combat the loss; />
Figure 875268DEST_PATH_IMAGE037
To balance the coefficients of the different loss terms.
In order to better enable the generator to carry a noise signal
Figure 205887DEST_PATH_IMAGE017
Reverting to a clean signal>
Figure 981076DEST_PATH_IMAGE018
The present invention expresses the content loss function as:
Figure 419623DEST_PATH_IMAGE006
in the formula (I), the compound is shown in the specification,
Figure 331078DEST_PATH_IMAGE007
is a loss of signal content; />
Figure 832598DEST_PATH_IMAGE008
For de-noising a signal->
Figure 829504DEST_PATH_IMAGE019
And the clean signal->
Figure 71742DEST_PATH_IMAGE018
Mean square error between; />
Figure 103283DEST_PATH_IMAGE009
For de-noising a signal->
Figure 775704DEST_PATH_IMAGE019
And the clean signal->
Figure 56644DEST_PATH_IMAGE018
Absolute value error therebetween; wherein the content of the first and second substances,
Figure 672931DEST_PATH_IMAGE038
Figure 293399DEST_PATH_IMAGE039
in the formula (I), the compound is shown in the specification,nfor the length of a single sample,
Figure 402300DEST_PATH_IMAGE040
is a clean signal>
Figure 373799DEST_PATH_IMAGE018
To (1) aiA sampling point is selected and/or determined>
Figure 957839DEST_PATH_IMAGE018
Is a clean signal; />
Figure 963973DEST_PATH_IMAGE041
For de-noising a signal->
Figure 978196DEST_PATH_IMAGE019
To (1) aiAnd (4) sampling points.
Penalty function of generator
Figure 171411DEST_PATH_IMAGE042
Expressed as:
Figure 559143DEST_PATH_IMAGE010
in the formula (I), the compound is shown in the specification,
Figure 419782DEST_PATH_IMAGE011
to combat the loss; />
Figure 604907DEST_PATH_IMAGE003
Averaging all data in the small batch;
Figure 285418DEST_PATH_IMAGE004
the output when the input of the relative discriminator is a real signal; />
Figure 211262DEST_PATH_IMAGE005
The relative discriminator input is the output when the signal is generated.
S330: repeating the steps S310 to S320, and iterating for N times in this way, wherein the learning rate of each iteration is 0.001; the number of iterations N is, for example, 100, and the learning rate is halved after 50 iterations, and Adam is used as the optimizer.
The initial learning rates of a generator and a discriminator in the noise reduction model for generating the countermeasure network based on the relative average are both set to be 0.001, and the learning rate is halved after 50 times of training; in the training process, the noise reduction model is continuously optimized in an iterative mode by calculating a loss function, and after the noise reduction model is trained, the final weight models of the generator and the discriminator are stored, namely the training of the relatively average generation countermeasure network is completed; and after the noise reduction model of the countermeasure network is generated on the basis of relative average and is trained well, extracting the trained generator in the noise reduction model.
Further, in one embodiment, the evaluation of the trained noise reduction model based on the test set comprises:
inputting the signal with noise in the test set into a generator after training to obtain a noise-removed signal; if the signal-to-noise ratio of the de-noising signal is increased by more than a threshold value (SNR) compared with the signal-to-noise ratio of the noise-carrying signal, judging that the performance of the generator is good, and extracting the generator which is trained; if the signal-to-noise ratio of the denoised signal does not increase more than a threshold compared to the signal-to-noise ratio of the noisy signal, the training is re-performed.
S400: the noisy signal is reduced to a clean signal based on the trained generator.
In order to verify the noise reduction performance of the method, the method is tested on a trained generator based on a test set, and the noise reduction performance of a RaGAN noise reduction model, wavelet decomposition, empirical Mode Decomposition (EMD) and a standard GAN noise reduction model in the invention on signals is compared under different input signal-to-noise ratios; adopting the output signal-to-noise ratio as a judgment standard of the noise reduction performance in the comparison of the noise reduction performance of different noise reduction models to the signal;
Figure 926408DEST_PATH_IMAGE043
in the formula (I), the compound is shown in the specification,
Figure 282434DEST_PATH_IMAGE044
to output a signal-to-noise ratio; />
Figure 450241DEST_PATH_IMAGE045
Is the effective power of the signal; />
Figure 320721DEST_PATH_IMAGE046
Is the effective power of the noise.
The input signal-to-noise ratio of the concentrated signal with noise is tested in the range of [ -8, 10] dB, and the step length is 2dB; fig. 8 shows a comparison graph of time domain waveforms of 8PSK signals before and after denoising in different methods (RaGAN denoising model, wavelet decomposition, empirical Mode Decomposition (EMD), and standard GAN denoising model in the present invention).
As can be seen from fig. 9, the noise reduction performance of the method of the present invention is significantly improved compared to the two conventional methods, and is also improved relative to the standard GAN noise reduction performance when the input signal is above-2 dB; under the condition of 0dB signal-to-noise ratio, the output signal-to-noise ratio is improved by about 4dB relative to the wavelet decomposition and EMD noise reduction method; therefore, the method (namely the RaGAN noise reduction model) provided by the invention has the advantage that the signal noise reduction performance is remarkably improved.
The invention provides a noise reduction model based on a relative average generation countermeasure network aiming at the reduction of the receiving quality of radio communication signals, wherein the model takes a BilSTM as a core to construct a generator and a discriminator, and uses a weighting loss function to train, so as to reduce the signals with noise into clean signals; compared with several signal noise reduction methods, the method has the advantages that the signal noise reduction performance is remarkably improved, and the validity of the model is proved through verification on a simulation signal data set (namely a data set consisting of clean signal samples and noisy signal samples).
Example two
As shown in fig. 10, an embodiment of the present invention provides a communication signal noise reduction system for generating a countermeasure network based on relative averaging, including a data set building module, a noise reduction model building module, a training module, and a noise reduction module;
the data set construction module is used for acquiring a clean signal and obtaining a signal with noise corresponding to the clean signal based on the clean signal; constructing a training set and a test set based on the clean signal and the noisy signal corresponding to the clean signal;
the noise reduction model building module is used for building a noise reduction model for generating a countermeasure network based on relative average;
the training module is used for training a noise reduction model for generating a countermeasure network based on relative average based on the training set so as to extract a trained generator;
and the noise reduction module is used for reducing the noisy signal into a clean signal based on the trained generator.
In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the units is only one logical division, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.
The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.
In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit.
The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: various media capable of storing program codes, such as a U disk, a removable hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (8)

1. A method for reducing noise in a communication signal based on a relative average generated countermeasure network, comprising the steps of:
s100: acquiring a clean signal, and acquiring a corresponding noisy signal based on the clean signal; constructing a training set and a test set based on the clean signal and the noisy signal corresponding to the clean signal;
s200: constructing a noise reduction model for generating a countermeasure network based on relative average;
s300: training a noise reduction model for generating a countermeasure network based on relative average based on the training set to extract a trained generator;
s400: reducing the noisy signal to a clean signal based on the trained generator;
wherein the training of the noise reduction model based on the relative average generated countermeasure network in step S300 includes the following sub-steps:
s310: inputting the signal with noise in the training set into a generator, outputting a de-noising signal, inputting a clean signal and the de-noising signal into a discriminator, calculating discrimination loss, and updating the discriminator based on the discrimination loss; the discriminant loss is calculated by the following formula:
Figure FDA0004019087030000011
in the formula (I), the compound is shown in the specification,
Figure FDA0004019087030000012
loss of relative discriminators; e [. C]Averaging all data in the small batch; d RaGAN (x r ,x f ) The output when the input of the relative discriminator is a real signal; d RaGAN (x f ,x r ) The input of the relative discriminator is the output when the signal is generated;
s320: calculating a signal content loss and a countermeasure loss through the denoised signal and the clean signal, calculating a total loss of the generator based on the signal content loss and the countermeasure loss, and updating the generator according to the total loss;
s330: and repeating the steps S310 to S320, and iterating for N times in this way to finish the training of the noise reduction model.
2. The method for reducing noise of a communication signal according to claim 1, wherein the step S200 of generating a noise reduction model of the countermeasure network based on the relative average includes an arbiter and a generator;
the generator and the discriminator both comprise two fully-connected layers and two bidirectional long-time and short-time memory models; the discriminator is a discriminator for generating a countermeasure network relatively averagely.
3. The method of reducing noise in a communication signal of claim 2, wherein the number of hidden layer nodes in the bidirectional long-term memory model layer is 128.
4. The method of claim 1, wherein the signal content loss is calculated by the following equation:
L X ＝(L MSE +L 1 )/2
in the formula, L X Is a loss of signal content; l is MSE Is the mean square error between the denoised signal and the clean signal; l is a radical of an alcohol 1 Is the absolute value error between the denoised signal and the clean signal.
5. The method of claim 1, wherein the penalty loss is calculated by the following equation:
Figure FDA0004019087030000021
in the formula (I), the compound is shown in the specification,
Figure FDA0004019087030000022
to combat the loss; e [. C]Averaging all data in the small batch; d RaGAN (x r ,x f ) The output when the input of the relative discriminator is a real signal; d RaGAN (x f ,x r ) The relative discriminator input is the output when the signal is generated.
6. The method of claim 1, wherein before extracting the trained generator in step S300, the method further comprises evaluating the trained noise reduction model based on a test set.
7. The method of claim 6, wherein evaluating the trained noise reduction model based on the test set comprises:
inputting the signal with noise in the test set into a generator after training to obtain a de-noising signal; if the signal-to-noise ratio of the de-noising signal is increased by more than a threshold value compared with the signal-to-noise ratio of the signal with noise, judging that the performance of the generator is good, and extracting the trained generator; if the signal-to-noise ratio of the denoised signal does not increase more than a threshold compared to the signal-to-noise ratio of the noisy signal, the training is re-performed.
8. A communication signal noise reduction system based on a relative average generation countermeasure network is characterized by comprising a data set construction module, a noise reduction model construction module, a training module and a noise reduction module;
the data set construction module is used for acquiring a clean signal and obtaining a corresponding noisy signal based on the clean signal; constructing a training set and a test set based on the clean signal and the noisy signal corresponding to the clean signal;
the noise reduction model building module is used for building a noise reduction model for generating a countermeasure network based on relative average;
the training module is used for training a noise reduction model for generating a countermeasure network based on relative average based on the training set so as to extract a trained generator;
the noise reduction module is used for reducing the signal with noise into a clean signal based on the trained generator;
wherein the training module performs the following operations:
s310: inputting the signal with noise in the training set into a generator, outputting a de-noising signal, inputting a clean signal and the de-noising signal into a discriminator, calculating discrimination loss, and updating the discriminator based on the discrimination loss; the discriminant loss is calculated by the following formula:
Figure FDA0004019087030000023
in the formula (I), the compound is shown in the specification,
Figure FDA0004019087030000024
loss of relative discriminators; e [. C]For performing average operation on all data in small batchMaking; d RaGAN (x r ,x f ) The output when the input of the relative discriminator is a real signal; d RaGAN (x f ,x r ) The input of the relative discriminator is the output when the signal is generated;
s320: calculating a signal content loss and a countering loss through the denoised signal and the clean signal, calculating a total loss of the generator based on the signal content loss and the countering loss, and updating the generator according to the total loss;
s330: and repeating the steps S310 to S320, and iterating for N times in this way to finish the training of the noise reduction model.
CN202211387887.5A 2022-11-08 2022-11-08 Communication signal noise reduction method and system based on relative average generation countermeasure network Active CN115442191B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211387887.5A CN115442191B (en) 2022-11-08 2022-11-08 Communication signal noise reduction method and system based on relative average generation countermeasure network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211387887.5A CN115442191B (en) 2022-11-08 2022-11-08 Communication signal noise reduction method and system based on relative average generation countermeasure network

Publications (2)

Publication Number Publication Date
CN115442191A CN115442191A (en) 2022-12-06
CN115442191B true CN115442191B (en) 2023-03-24

Family

ID=84252818

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211387887.5A Active CN115442191B (en) 2022-11-08 2022-11-08 Communication signal noise reduction method and system based on relative average generation countermeasure network

Country Status (1)

Country Link
CN (1) CN115442191B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116250844B (en) * 2023-03-03 2024-04-26 山东大学 Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network
CN115967451B (en) * 2023-03-13 2023-06-27 微网优联科技(成都)有限公司 Wireless router signal processing method and device and wireless router using same
CN116167465A (en) * 2023-04-23 2023-05-26 杭州经纬信息技术股份有限公司 Solar irradiance prediction method based on multivariate time series ensemble learning

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109890043A (en) * 2019-02-28 2019-06-14 浙江工业大学 A kind of wireless signal noise-reduction method based on production confrontation network
WO2020076217A1 (en) * 2018-10-10 2020-04-16 Telefonaktiebolaget Lm Ericsson (Publ) Method for peak to average power reduction of dm-rs signals
CN111067507A (en) * 2019-12-26 2020-04-28 常熟理工学院 Electrocardiosignal denoising method based on generation of countermeasure network and strategy gradient
CN112466320A (en) * 2020-12-12 2021-03-09 中国人民解放军战略支援部队信息工程大学 Underwater acoustic signal noise reduction method based on generation countermeasure network
CN114446314A (en) * 2021-12-31 2022-05-06 中国人民解放军陆军工程大学 Voice enhancement method for deeply generating confrontation network

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020076217A1 (en) * 2018-10-10 2020-04-16 Telefonaktiebolaget Lm Ericsson (Publ) Method for peak to average power reduction of dm-rs signals
CN109890043A (en) * 2019-02-28 2019-06-14 浙江工业大学 A kind of wireless signal noise-reduction method based on production confrontation network
CN111067507A (en) * 2019-12-26 2020-04-28 常熟理工学院 Electrocardiosignal denoising method based on generation of countermeasure network and strategy gradient
CN112466320A (en) * 2020-12-12 2021-03-09 中国人民解放军战略支援部队信息工程大学 Underwater acoustic signal noise reduction method based on generation countermeasure network
CN114446314A (en) * 2021-12-31 2022-05-06 中国人民解放军陆军工程大学 Voice enhancement method for deeply generating confrontation network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
一种基于生成对抗网络的语音去噪方法;王华锋等;《北方工业大学学报》;20191015(第05期);全文 *

Also Published As

Publication number Publication date
CN115442191A (en) 2022-12-06

Similar Documents

Publication Publication Date Title
CN115442191B (en) Communication signal noise reduction method and system based on relative average generation countermeasure network
CN107845389B (en) Speech enhancement method based on multi-resolution auditory cepstrum coefficient and deep convolutional neural network
CN110619885B (en) Method for generating confrontation network voice enhancement based on deep complete convolution neural network
CN111653288B (en) Target person voice enhancement method based on conditional variation self-encoder
CN112464837B (en) Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples
CN109890043B (en) Wireless signal noise reduction method based on generative countermeasure network
CN110456332B (en) Underwater acoustic signal enhancement method based on automatic encoder
CN108922513B (en) Voice distinguishing method and device, computer equipment and storage medium
CN111564160B (en) Voice noise reduction method based on AEWGAN
CN109841206A (en) A kind of echo cancel method based on deep learning
CN113094993B (en) Modulation signal denoising method based on self-coding neural network
CN112466320B (en) Underwater sound signal noise reduction method based on generation countermeasure network
CN110867181A (en) Multi-target speech enhancement method based on SCNN and TCNN joint estimation
CN108777146A (en) Speech model training method, method for distinguishing speek person, device, equipment and medium
CN112615804B (en) Short burst underwater acoustic communication signal modulation identification method based on deep learning
CN111968666B (en) Hearing aid voice enhancement method based on depth domain self-adaptive network
CN111723701A (en) Underwater target identification method
CN116403590B (en) Bionic signal processing method based on wavelet transformation and generation countermeasure network
CN110808057A (en) Voice enhancement method for generating confrontation network based on constraint naive
CN109767781A (en) Speech separating method, system and storage medium based on super-Gaussian priori speech model and deep learning
CN112086100B (en) Quantization error entropy based urban noise identification method of multilayer random neural network
CN111899750B (en) Speech enhancement algorithm combining cochlear speech features and hopping deep neural network
CN111159891A (en) Single-sample polarization filter coefficient prediction method based on DNN neural network
CN107045874A (en) A kind of Non-linear Speech Enhancement Method based on correlation
CN112906591A (en) Radar radiation source identification method based on multi-stage jumper residual error network

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