CN116933066A - Wireless signal single-channel blind source separation method and device based on deep neural network - Google Patents

Wireless signal single-channel blind source separation method and device based on deep neural network Download PDF

Info

Publication number
CN116933066A
CN116933066A CN202310833146.3A CN202310833146A CN116933066A CN 116933066 A CN116933066 A CN 116933066A CN 202310833146 A CN202310833146 A CN 202310833146A CN 116933066 A CN116933066 A CN 116933066A
Authority
CN
China
Prior art keywords
complex
signal
signals
training
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.)
Pending
Application number
CN202310833146.3A
Other languages
Chinese (zh)
Inventor
于淼
郭鹏程
顾淼淼
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.)
National University of Defense Technology
Original Assignee
National University of Defense Technology
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 National University of Defense Technology filed Critical National University of Defense Technology
Priority to CN202310833146.3A priority Critical patent/CN116933066A/en
Publication of CN116933066A publication Critical patent/CN116933066A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2134—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on separation criteria, e.g. independent component analysis
    • G06F18/21345—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on separation criteria, e.g. independent component analysis enforcing sparsity or involving a domain transformation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G06F18/15—Statistical pre-processing, e.g. techniques for normalisation or restoring missing data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2131—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on a transform domain processing, e.g. wavelet transform
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Signal Processing (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Digital Transmission Methods That Use Modulated Carrier Waves (AREA)

Abstract

The invention discloses a method and a device for separating a wireless signal single-channel blind source based on a deep neural network, which specifically comprise the following steps: in the training phase, creating a signal data set, wherein the data set comprises a training set and a verification set; building a double-path complex time domain convolution network; creating a self-defined network training objective function, and training and verifying the dual-path complex time domain convolution network by using the obtained signal data set to obtain a trained dual-path complex time domain convolution network; in the application stage, acquiring a time-frequency aliasing signal to be separated through a single antenna; the time-frequency aliasing signals are subjected to sectional arrangement to obtain complex-form aliasing signals with fixed lengths; and separating a plurality of signals from the complex form aliasing signals by using the trained double-path complex time domain convolution network, and sending the signals to a demodulator for demodulation to recover bit data of the separated signals. The invention can realize the effective separation of the mixed signals with the same frequency at the same time under the condition of receiving the single antenna, and improves the anti-interference capability of wireless communication.

Description

Wireless signal single-channel blind source separation method and device based on deep neural network
Technical Field
The invention relates to the technical field of signal processing, in particular to a wireless signal single-channel blind source separation method based on a deep neural network.
Background
Due to the growing mobile applications, many wireless devices will typically access the same spectrum resources at the same time, which makes collisions and collisions in the wireless medium unavoidable. To overcome these drawbacks, single-channel blind source separation is a promising technique, especially in the limited electromagnetic spectrum. The cognitive radio technology can alleviate the problem of low spectrum utilization rate in a dynamic spectrum access mode, but still cannot essentially meet spectrum resources required by future users. As a technique that can increase the utilization of the system band, single-channel blind source separation has been a research hotspot because it can recover the source signal from the mixed signal with minimal a priori knowledge and less antenna resources.
Single-channel blind source separation is the process of recovering a multi-dimensional source signal from a one-dimensional mixed signal. Compared with common blind source separation, single-channel blind source separation has only one path of observation signal, which is a pathological problem and is extremely difficult to solve. The most advanced single channel blind source separation scheme in current wireless communications can be divided into several aspects: single-channel blind source separation based on virtual multi-channel, which converts single channel into multi-channel blind source separation by utilizing wavelet decomposition, wavelet packet decomposition, empirical Mode Decomposition (EMD) and other decomposition methods; single-channel blind source separation based on basis function training is carried out, and a classical blind source separation optimization algorithm is utilized to reconstruct a source signal so as to obtain a basis function; single-channel blind source separation based on transform domain filtering is carried out, and the sparsity of source signals in time-frequency domain, cyclic domain, wavelet domain and other domains is utilized to filter the superimposed signals; based on model parameter estimation and reconstructed single-channel blind source separation, the required source signals are reconstructed by estimating signal model parameters. However, existing single channel blind source separation schemes are difficult to achieve excellent performance due to excessive computational complexity, especially at low signal-to-noise ratios (SNR).
To address the above drawbacks, deep learning has been used in single-channel blind source separation, particularly in the field of separating speech waveforms, which is mainly focused on the time-frequency domain or directly in the time domain. Representative methods of the time-frequency domain include deep cluster learning, permutation-invariant training, and deep attraction networks. Representative time domain networks include Long and Short Term Memory (LSTM) recursive networks, convolutional Recursive Networks (CRNs), and the like. Recent advances in voice separation networks have also shown great potential in the field of wireless communications. However, there is still a great difference between the communication and voice fields. First, the communication bits are compressed very tightly, typically using a small number of algebraic operations, and lacking much inherent redundancy or context, while sequences in the speech domain can exhibit relatively long recognizable formants to synthesize words and phrases. Second, the impact of information on speech processing mostly occurs on a larger time scale, while the meaning of symbols in communication usually occurs on a smaller time scale (e.g., microsecond or nanosecond). Third, wireless communication systems typically employ complex baseband representations of I/Q data, while speech signals are typically represented in real form only in the time domain. These limitations result in the fact that signal separation networks in the speech domain cannot be used directly for the separation of communication signals.
The present invention is therefore directed to achieving deep learning based single channel blind source separation of wireless signals. In civil communication, the invention can separate out two time-frequency aliasing signals on the same frequency band, increases the information transmission rate under the condition of not increasing the cost of spectrum resources, improves the spectrum utilization efficiency, and has great significance for relieving the tense spectrum resources. In the field of military communication anti-interference, blind source separation can realize separation of a target signal and an interference signal, and weak signal enhancement and communication anti-interference are realized.
Disclosure of Invention
The invention aims to provide a method for separating a wireless signal single-channel blind source based on a deep neural network, which can effectively separate simultaneous same-frequency mixed signals under the condition of single antenna reception, thereby improving the anti-interference capability of wireless communication.
The technical solution for realizing the purpose of the invention is as follows: a wireless signal single-channel blind source separation method based on a deep neural network comprises a training stage and an application stage, wherein:
the training phase comprises the following steps:
a1, creating a data set of time-frequency aliasing signals, wherein the data set comprises communication signals and interference signals, and dividing the data set into a training set and a verification set;
A2, constructing a double-path complex time domain convolution network, wherein the double-path complex time domain convolution network comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
step A3, creating a self-defined network training objective function;
step A4, training and verifying the dual-path complex time domain convolution network obtained in the step A2 by using the signal data set obtained in the step A1 based on a network training objective function to obtain a trained dual-path complex time domain convolution network;
the application stage comprises the following steps:
step B1, acquiring a time-frequency aliasing signal to be separated through a single antenna;
step B2, carrying out sectional arrangement on the time-frequency aliasing signals to obtain complex-form aliasing signals with fixed lengths;
step B3, separating a plurality of signals from complex-form aliasing signals by using a trained double-path complex time domain convolution network, and completing blind separation of the aliasing signals;
and B4, sending the separated signal obtained in the step B3 to a demodulator for demodulation, and recovering bit data of the separated signal.
Further, the time-frequency aliasing communication signal in the step A1 is as follows:
where x (t) is a time-frequency aliased communication signal, s 1 (t),...,s λ (t)∈C 1×T Represents lambda source signals, T represents signal length, a i The unknown mixing coefficient is shown, and n (t) is gaussian white noise.
Further, the construction of the data set in step A1 may be performed in any of two ways:
the first way is: the method comprises the steps of generating communication signals and interference signals in a plurality of different modulation modes by adopting computer software, wherein the modulation modes are digital modulation or analog modulation, and the interference signals comprise digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb-shaped spectrum interference signals, broadband blocking interference and sweep frequency interference;
based on the generated communication signals and interference signals, randomly mixing the communication signals and the interference signals to form an aliasing signal, adding noise to the aliasing signal to generate a time-frequency aliasing communication signal, wherein the SNR (signal-to-noise ratio) is-5 dB to 20dB, and dividing a data set of the time-frequency aliasing communication signal into a training set and a verification set according to the ratio of 80% to 20% according to the network training requirement;
in a second mode, the actual communication equipment and the interference equipment are sampled to obtain a data set of the time-frequency aliasing communication signals, and the data set of the time-frequency aliasing communication signals is divided into a training set and a verification set according to the network training requirements and the ratio of 80% to 20%.
Further, the step A2 of constructing a dual-path complex time domain convolutional network includes a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module, which specifically includes the following steps:
the double-path complex time domain convolution network adopts two parallel links, which respectively correspond to two separated signals, and each path comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
the complex domain feature extraction module adopts a two-stage complex domain coding structure and comprises two complex convolution layers and a complex normalization layer; the real part and the imaginary part of the convolution output of the first layer are spliced in a new dimension, the complex normalized layer is processed, then the second layer of complex convolution is carried out, and all operations of the complex domain feature extraction module are carried out on the real part and the imaginary part of the complex value signal simultaneously;
the complex domain signal separation module consists of a plurality of complex depth convolution modules and complex LSTM modules, wherein the complex LSTM modules are arranged in the middle of the separation module and are responsible for extracting the correlation before and after signals; the complex depth convolution module adopts a jump connection path, the input and the output of each complex depth convolution module are added in the real part and the imaginary part respectively, the complex depth convolution module enters the next complex depth convolution module, and the dimensionality of the signal input and the signal output is kept unchanged; the inner part of the convolution kernel adopts the form of expansion convolution, the convolution kernel of each complex depth convolution module presents exponential increment, and the corresponding convolution kernel sizes of the V complex depth convolution modules are 1,2 V-1 The method comprises the steps of carrying out a first treatment on the surface of the Each link is provided with 2V complex depth convolution modules, and symmetrical structures are shown on two sides of a complex LSTM;
in the complex domain signal recovery module, the output of the complex domain signal separation module is firstly subjected to complex 1×1 convolution, then the real part and the imaginary part are respectively subjected to sigmoid activation, and a mask matrix is obtained by splicing in a new dimension; the mask and the first layer of rewinding output in the complex domain feature extraction module have the same dimension, and the two obtain the recovered signal high-dimensional representation through dot multiplication; finally, performing the rewinding operation again to reduce the dimension of the high-dimension representation, namely recovering the estimated signal.
Further, the network training objective function in step A3 adopts a weighted combination of the scale invariant signal-to-noise ratio SI-SNR and the mean square error MSE as a model training loss function, which is specifically as follows:
the scale invariant signal to noise ratio SI-SNR is expressed as:
wherein s andrepresenting the source signal and the estimated signal,<·,·>representing the dot product of two vectors, I.I 2 Representing a binary norm;
mean square error MSE is defined as
Where T represents the signal length, SI-SNR and MSE are calculated and averaged for the real and imaginary parts of the signal, respectively, i.e.:
SI-SNR ave ＝mean(SI-SNR real +SI-SNR imag )
MSE ave ＝mean(MSE real +MSE imag )
wherein SI-SNR ave Representing average scale invariant signal-to-noise ratio, SI-SNR real Representing real scale invariant signal-to-noise ratio, SI-SNR imag Representing the imaginary scale-invariant signal-to-noise ratio, MSE ave Mean squared error, MSE real Representing real mean square error, MSE imag Representing the imaginary mean square error;
finally, the Loss function Loss of the first link of the network l The definition is as follows:
Loss l ＝SI-SNR avel +αMSE avel ,l＝1,2
wherein α represents a training coefficient, and α=1 is set; SI-SNR avel Representing the average scale invariant signal-to-noise ratio of the first path, MSE avel Indicating the mean squared error of the first path.
Further, in step A4, training and verifying the dual-path complex time domain convolutional network obtained in step A2 by using the signal data set obtained in step A1, specifically as follows:
grouping and packaging data of a training set and a verification set, wherein the sequence combination is mini_batch, the mini_batch is sent into a double-path complex time domain convolution network, network parameters are set, an Adam optimizer is adopted, the initial learning rate is 0.001, and when the loss of the verification set is not reduced in continuous 3 rounds of training, the learning rate is halved; the total training adopts M rounds, M is a positive integer set according to actual conditions, gradient clipping with the maximum L2 norm of 5 is applied during training, model training is stopped in advance, and training is stopped when the loss of the continuous 6 rounds of verification sets is no longer reduced; after the training set and the verification set are finished, the model stores the optimal parameter model.
The device is used for realizing the single-channel blind source separation method based on the double-path complex time domain convolution network, and comprises a training module and an application module, wherein:
the training module comprises a data set creating unit, a convolution network building unit, an objective function creating unit and a network training unit, wherein:
the data set creating unit creates a data set of the time-frequency aliasing signal, which comprises a communication signal and an interference signal, and divides the data set into a training set and a verification set;
the convolution network construction unit is used for constructing a double-path complex time domain convolution network and comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
the target function creating unit creates a self-defined network training target function;
the network training unit is used for training and verifying the dual-path complex time domain convolutional network obtained by the convolutional network building unit by using the signal data set obtained by the data set building unit based on the network training target function to obtain a trained dual-path complex time domain convolutional network;
the application module comprises a signal acquisition unit, a time-frequency aliasing signal arrangement unit, a blind separation unit and a demodulation unit, wherein:
The signal acquisition unit acquires a time-frequency aliasing signal to be separated through a single antenna;
the time-frequency aliasing signal arrangement unit is used for carrying out sectional arrangement on the time-frequency aliasing signals to obtain complex-form aliasing signals with fixed lengths;
the blind separation unit separates a plurality of signals from complex-form aliasing signals by using a trained double-path complex time domain convolution network to finish blind separation of the aliasing signals;
and the demodulation unit is used for sending the separated signal obtained by the blind separation unit to a demodulator for demodulation and recovering bit data of the separated signal.
Compared with the prior art, the invention has the remarkable advantages that: (1) The dual-path complex time domain convolution network can directly separate wireless communication signals, fully utilizes the correlation between signal I/Q data in network training, and can directly demodulate the separated signal complex form; (2) In order to build the double-path complex time domain convolution network, a series of novel network modules are adopted, including complex coding and decoding, complex LSTM, complex depth convolution modules and the like, and the high-dimensional characteristics of signals are fully extracted; (3) Simulation results show that the network has good signal separation performance, can realize effective separation of simultaneous same-frequency mixed signals under the condition of single antenna reception, and improves the anti-interference capability of wireless communication.
Drawings
Fig. 1 is a flow chart of a method for separating a blind source of a wireless signal single channel based on a deep neural network.
Fig. 2 is a schematic diagram of a dual path complex time domain convolutional network in accordance with the present invention.
FIG. 3 is a schematic diagram of a complex domain feature extraction module according to the present invention.
FIG. 4 is a schematic diagram of a complex depth convolution module according to the present invention.
Fig. 5 is a schematic diagram of correlation coefficients of BPSK and QPSK separation results at different signal-to-noise ratios according to the present invention.
Fig. 6 is a schematic diagram of bit error rate of BPSK and QPSK separation results at different signal-to-noise ratios according to the present invention.
Detailed Description
Aiming at the situations that the traditional single-channel separation method is poor in effect and the existing voice separation network is not adaptive to communication signals, the invention establishes a neural network capable of meeting complex communication signal separation, the network can fully utilize the correlation of signal I/Q data, and the signals can be directly expressed in complex form after being separated by the network, thereby being more beneficial to subsequent demodulation and being superior to the traditional separation method and the existing separation network model in performance.
The invention discloses a wireless signal single-channel blind source separation method based on a deep neural network, which comprises a training stage and an application stage, wherein:
the training phase comprises the following steps:
A1, creating a data set of time-frequency aliasing signals, wherein the data set comprises communication signals and interference signals, and dividing the data set into a training set and a verification set;
a2, constructing a double-path complex time domain convolution network, wherein the double-path complex time domain convolution network comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
step A3, creating a self-defined network training objective function;
step A4, training and verifying the dual-path complex time domain convolution network obtained in the step A2 by using the signal data set obtained in the step A1 based on a network training objective function to obtain a trained dual-path complex time domain convolution network;
the application stage comprises the following steps:
step B1, acquiring a time-frequency aliasing signal to be separated through a single antenna;
step B2, carrying out sectional arrangement on the time-frequency aliasing signals to obtain complex-form aliasing signals with fixed lengths;
step B3, separating a plurality of signals from complex-form aliasing signals by using a trained double-path complex time domain convolution network, and completing blind separation of the aliasing signals;
and B4, sending the separated signal obtained in the step B3 to a demodulator for demodulation, and recovering bit data of the separated signal.
As a specific example, the time-frequency aliasing communication signal in step A1 is as follows:
Where x (t) is a time-frequency aliased communication signal, s 1 (t),...,s λ (t)∈C 1×T Represents lambda source signals, T represents signal length, a i The unknown mixing coefficient is shown, and n (t) is gaussian white noise.
As a specific example, the construction of the data set in step A1 takes any of the following two ways:
the first way is: the method comprises the steps of generating communication signals and interference signals in various modulation modes by adopting computer software, wherein the modulation modes are digital modulation signals or analog modulation signals, and the interference signals comprise digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb-shaped spectrum interference signals, broadband blocking interference signals and sweep-frequency interference signals. Among them, digital modulation includes MPSK (multi-system phase shift keying), MASK (multi-system amplitude shift keying), MFSK (multi-system frequency shift keying), MQAM (multi-system quadrature amplitude modulation), CPSK, OQPSK, OFDM, GMSK, DPSK, TCM, CPFSK. The analog modulation includes AM, FM, PM, DSB, VSB, SSB.
Based on the generated communication signals and interference signals, randomly mixing the communication signals and the interference signals to form an aliasing signal, adding noise to the aliasing signal to generate a time-frequency aliasing communication signal, wherein the SNR (signal-to-noise ratio) is-5 dB to 20dB, and dividing a data set of the time-frequency aliasing communication signal into a training set and a verification set according to the ratio of 80% to 20% according to the network training requirement;
In a second mode, the actual communication equipment and the interference equipment are sampled to obtain a data set of the time-frequency aliasing communication signals, and the data set of the time-frequency aliasing communication signals is divided into a training set and a verification set according to the network training requirements and the ratio of 80% to 20%.
As a specific example, the building a dual-path complex time domain convolutional network described in step A2 includes a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module, which specifically includes:
the double-path complex time domain convolution network adopts two parallel links, which respectively correspond to two separated signals, and each path comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
the complex domain feature extraction module adopts a two-stage complex domain coding structure and comprises two complex convolution layers and a complex normalization layer; the real part and the imaginary part of the convolution output of the first layer are spliced in a new dimension, the complex normalized layer is processed, then the second layer of complex convolution is carried out, and all operations of the complex domain feature extraction module are carried out on the real part and the imaginary part of the complex value signal simultaneously;
the complex domain signal separation module consists of a plurality of complex depth convolution modules and complex LSTM modules, wherein the complex LSTM modules are arranged in the middle of the separation module and are responsible for extracting the correlation before and after signals; the complex depth convolution module adopts a jump connection path, the input and the output of each complex depth convolution module are added in the real part and the imaginary part respectively, the complex depth convolution module enters the next complex depth convolution module, and the dimensionality of the signal input and the signal output is kept unchanged; the inner part of the convolution kernel adopts the form of expansion convolution, the convolution kernel of each complex depth convolution module presents exponential increment, and the corresponding convolution kernel sizes of the V complex depth convolution modules are 1,2 V-1 To more fully capture the high-dimensional characteristics of the signal; each link is provided with 2V complex depth convolution modules, and symmetrical structures are shown on two sides of a complex LSTM;
in the complex domain signal recovery module, the output of the complex domain signal separation module is firstly subjected to complex 1×1 convolution, then the real part and the imaginary part are respectively subjected to sigmoid activation, and a mask matrix is obtained by splicing in a new dimension; the mask and the first layer of rewinding output in the complex domain feature extraction module have the same dimension, and the two obtain the recovered signal high-dimensional representation through dot multiplication; finally, performing the rewinding operation again to reduce the dimension of the high-dimension representation, namely recovering the estimated signal.
As a specific example, the network training objective function described in step A3 uses a weighted combination of the scale-invariant signal-to-noise ratio SI-SNR and the mean square error MSE as the model training loss function, specifically as follows:
the scale invariant signal to noise ratio SI-SNR is expressed as:
wherein s andrepresenting the source signal and the estimated signal,<·,·>representing the dot product of two vectors, I.I 2 Representing a binary norm;
mean square error MSE is defined as
Where T represents the signal length, SI-SNR and MSE are calculated for the real and imaginary parts of the signal, respectively, i.e.:
SI-SNR ave ＝mean(SI-SNR real +SI-SNR imag )
MSE ave ＝mean(MSE real +MSE imag )
wherein SI-SNR ave Representing average scale invariant signal-to-noise ratio, SI-SNR real Representing real scale invariant signal-to-noise ratio, SI-SNR imag Representing the imaginary scale-invariant signal-to-noise ratio, MSE ave Mean squared error, MSE real Representing real mean square error, MSE imag Representing the imaginary mean square error;
finally, the Loss function Loss of the first link of the network l The definition is as follows:
Loss l ＝SI-SNR avel +αMSE avel ,l＝1,2
wherein α represents a training coefficient, and α=1 is set; SI-SNR avel Representing the average scale invariant signal-to-noise ratio of the first path, MSE avel Indicating the mean squared error of the first path.
As a specific example, the training and verification of the dual-path complex time domain convolutional network obtained in step A2 by using the signal data set obtained in step A1 in step A4 is specifically as follows:
grouping and packaging data of a training set and a verification set, wherein the sequence combination is mini_batch, the mini_batch is sent into a double-path complex time domain convolution network, network parameters are set, an Adam optimizer is adopted, the initial learning rate is 0.001, and when the loss of the verification set is not reduced in continuous 3 rounds of training, the learning rate is halved; the total training adopts M rounds, M is a positive integer set according to actual conditions, gradient clipping with the maximum L2 norm of 5 is applied during training, model training is stopped in advance, and training is stopped when the loss of the continuous 6 rounds of verification sets is no longer reduced; after the training set and the verification set are finished, the model stores the optimal parameter model.
The invention discloses a single-channel blind source separation device based on a double-path complex time domain convolution network, which is used for realizing the single-channel blind source separation method based on the double-path complex time domain convolution network, and comprises a training module and an application module, wherein:
the training module comprises a data set creating unit, a convolution network building unit, an objective function creating unit and a network training unit, wherein:
the data set creating unit creates a data set of the time-frequency aliasing signal, which comprises a communication signal and an interference signal, and divides the data set into a training set and a verification set;
the convolution network construction unit is used for constructing a double-path complex time domain convolution network and comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
the target function creating unit creates a self-defined network training target function;
the network training unit is used for training and verifying the dual-path complex time domain convolutional network obtained by the convolutional network building unit by using the signal data set obtained by the data set building unit based on the network training target function to obtain a trained dual-path complex time domain convolutional network;
the application module comprises a signal acquisition unit, a time-frequency aliasing signal arrangement unit, a blind separation unit and a demodulation unit, wherein:
The signal acquisition unit acquires a time-frequency aliasing signal to be separated through a single antenna;
the time-frequency aliasing signal arrangement unit is used for carrying out sectional arrangement on the time-frequency aliasing signals to obtain complex-form aliasing signals with fixed lengths;
the blind separation unit separates a plurality of signals from complex-form aliasing signals by using a trained double-path complex time domain convolution network to finish blind separation of the aliasing signals;
and the demodulation unit is used for sending the separated signal obtained by the blind separation unit to a demodulator for demodulation and recovering bit data of the separated signal.
The invention will be described in further detail with reference to the accompanying drawings and specific examples.
Examples
With reference to fig. 1, the invention provides a single-channel blind source separation method based on a double-path complex time domain convolution network, which comprises the following steps:
step 1: acquiring a single-channel time-frequency aliasing communication signal;
step 2: creating a signal data set according to the aliasing signals obtained in the step 1, wherein the data set comprises a training set and a verification set;
step 3: the method comprises the steps of constructing a double-path complex time domain convolution network, wherein the double-path complex time domain convolution network comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
step 4: creating a self-defined network training objective function;
Step 5: training and verifying the dual-path complex time domain convolution network obtained in the step 3 by using the signal data set obtained in the step 2 to obtain two separated communication signal waveforms, and completing blind separation of aliasing signals;
step 6: and (5) sending the separated waveform obtained in the step (5) to a demodulator for demodulation, and recovering bit data of the separated signal.
Further, the time-frequency aliasing signal in step 1 is shown in the following formula:
wherein x (t) is an aliasing signal, s 1 (t),...,s λ (t)∈C 1×T Represents lambda source signals, T represents signal length, a ij The unknown mixing coefficient is shown, and n (t) is gaussian white noise.
Further, the construction process of the data set in step 2 is as follows:
MATLAB is adopted to generate a plurality of different modulation signals, such as BPSK, QPSK, 8PSK, 16QAM, 64QAM and the like, 1000 modulation signals are used, the signal length is 1024, the real part and the imaginary part of a signal sample are respectively extracted and stored in a new one-dimension, and all the signals are packed together to form a signal set.
Two different modulation modes, such as BPSK and QPSK, are selected, signals are mixed randomly to form an aliasing signal, noise is added, the signal-to-noise ratio SNR is between-5 dB and 20dB, and an aliasing signal set is divided into a training set and a verification set according to the ratio of 80% to 20% according to the network training requirement.
Further, the dual-path complex time domain convolutional network in step 3 includes a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module, which specifically includes the following steps:
the dual-path complex time domain convolution network adopts two parallel links, which respectively correspond to two separated signals. Each path comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module.
The complex domain feature extraction module adopts a two-stage complex domain coding structure and comprises two complex convolution layers and a complex normalization layer. The real part and the imaginary part of the convolution output of the first layer are spliced in a new dimension, the complex normalization layer is processed, then the second layer of complex convolution is carried out, and all operations are carried out on the real part and the imaginary part of the complex value signal simultaneously.
The complex domain signal separation module consists of a plurality of complex depth convolution modules and complex LSTM modules, wherein the complex LSTM is positioned in the middle of the separation module and is responsible for extracting the correlation before and after signals. The complex depth convolution module adopts a jump connection path, the input and the output of each module are added in the real part and the imaginary part respectively, and the complex depth convolution module enters the next depth convolution module, and the dimensionality of the signal input and the signal output is kept unchanged. The inner part of the convolution kernel adopts the form of expansion convolution, the convolution kernel of each complex depth convolution module presents exponential increment, and the corresponding convolution kernel sizes of the V depth convolution modules are 1,2 V-1 So as to more fully capture the high-dimensional characteristics of the signal. Each link has 2V complex depth convolution modules, which exhibit a symmetrical structure on both sides of the complex LSTM.
In the complex domain signal recovery module, the output of the separation module is firstly subjected to complex 1×1 convolution, then the real part and the imaginary part are respectively subjected to sigmoid activation, and a mask matrix can be obtained by splicing in a new dimension. The mask and the first layer of the convolution output in the complex domain feature extraction module have the same dimension, and the two obtain the recovered signal high-dimensional representation through dot multiplication. Finally, the estimated signal can be recovered by performing the rewinding operation again to reduce the dimension of the high-dimension representation.
Further, for the network training objective function in step 4, a weighted combination of scale-invariant signal-to-noise ratio (SI-SNR) and Mean Square Error (MSE) is used as a model training loss function, which is specifically as follows:
the SI-SNR can be expressed as:
wherein s andrepresenting the source signal and the estimated signal.<·,·>Representing the dot product of two vectors, I.I 2 Representing a binary norm, MSE is defined as
T denotes the signal length, SI-SNR and MSE are calculated for the real and imaginary parts of the signal, respectively, i.e.:
SI-SNR ave ＝mean(SI-SNR real +SI-SNR imag )
MSE ave ＝mean(MSE real +MSE imag )
finally, the loss function of the network first link may be defined as:
Loss l ＝SI-SNR avel +αMSE avel ,l＝1,2
where α represents a training coefficient, and α=1 is set.
Further, the training and verification of the network using the signal data set in step 5 is specifically as follows:
and (3) grouping and packaging data of the training set and the verification set, combining sequences into mini_batch, sending the mini_batch into a neural network, setting network parameters, adopting an Adam optimizer, setting the initial learning rate to be 0.001, and halving the learning rate when the loss of the verification set is not reduced in continuous 3 rounds of training. The total training adopts M rounds, M is determined according to actual conditions, gradient clipping with maximum L2 norm of 5 is applied during training, model training adopts advanced stopping, and training is stopped when the loss of the verification set of 6 rounds is no longer reduced. After the training set and the verification set are finished, the model stores the optimal parameter model.
Further, in step 6, the obtained separated waveform is sent to a demodulator to demodulate and recover the bit data of the separated signal, which is specifically as follows:
after the mixed signals of BPSK and QPSK are input into a network for separation, two separated waveforms are obtained, the two separated waveforms are respectively demodulated to recover bit data of signals, and the separated bit error rate can be obtained by comparing the bit data with the bit data of a source signal and can be used as a system separation performance evaluation index.
The embodiment also relates to a single-channel blind source separation device based on the double-path complex time domain convolution network, which is used for realizing the single-channel blind source separation method based on the double-path complex time domain convolution network, and the device comprises a training module and an application module, wherein:
The training module comprises a data set creating unit, a convolution network building unit, an objective function creating unit and a network training unit, wherein:
the data set creation unit creates a data set of time-frequency aliasing signals, including communication signals and interference signals, and includes a training set and a verification set;
the convolution network construction unit is used for constructing a double-path complex time domain convolution network and comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
the target function creating unit creates a self-defined network training target function;
the network training unit is used for training and verifying the dual-path complex time domain convolutional network obtained by the convolutional network building unit by using the signal data set obtained by the data set building unit based on the network training target function to obtain a trained dual-path complex time domain convolutional network;
the application module comprises a signal acquisition unit, a time-frequency aliasing signal arrangement unit, a blind separation unit and a demodulation unit, wherein:
the signal acquisition unit acquires a time-frequency aliasing signal to be separated through a single antenna;
the time-frequency aliasing signal arrangement unit is used for carrying out sectional arrangement on the time-frequency aliasing signals to obtain complex-form aliasing signals with fixed lengths;
The blind separation unit separates a plurality of signals from complex-form aliasing signals by using a trained double-path complex time domain convolution network to finish blind separation of the aliasing signals;
and the demodulation unit is used for sending the separated signal obtained by the blind separation unit to a demodulator for demodulation and recovering bit data of the separated signal.
The technical scheme of the invention is described in detail below with reference to the accompanying drawings:
fig. 2 is a schematic diagram of a dual-path complex time domain convolutional network according to the present invention, showing the overall process of mixed signal from input network to complex codec, complex deep convolutional, complex LSTM processing, mask generation, and signal demodulation.
Generating various types of time-frequency aliasing signals, such as BPSK, QPSK, 8PSK, 16QAM, 64QAM and the like, wherein the forms are shown in the formula (1):
wherein x (t) is an aliasing signal, s 1 (t),...,s λ (t)∈C 1×T Represents lambda source signals, T represents signal length, a ij The unknown mixing coefficient is shown, and n (t) is gaussian white noise. 1000 modulated signals of each type, a signal length 1024, the real part and the imaginary part of the signal sample are extracted and stored in a new dimension respectively, all the signals are packed together to form a signal set, the signal storage format is (N, 2, L), N represents the number of samples, 2 represents the two dimensions of the real part and the imaginary part, and L represents the signal length.
Two different modulation modes, such as BPSK and QPSK, are selected, signals are randomly mixed into an aliasing signal, noise is added, and the aliasing signal is randomly added according to the signal-to-noise ratio of-5 dB to 20dB, so that an aliasing data set with noise is generated. According to the network training requirement, the aliasing noisy data set is divided into a training set and a verification set according to the proportion of 80% and 20%.
The training set and verification set signals (mixed signals and source signals) are packed and input into a network, and firstly enter a complex domain feature extraction module.
As shown in fig. 3, the complex domain feature extraction module employs two-stage coding, including two complex convolution layers and a complex normalization layer,all operations are performed simultaneously on the real and imaginary parts of the complex-valued signal. The complex signal input can be expressed as x=x r +jx i ,x∈C 1×T First the signal enters the first convolution layer containing M convolution kernels, the convolution kernel structure can be expressed as w=w r +jw i ,w∈C 1×M ，w r And w i Representing the real and imaginary parts of the complex convolution, respectively, the convolution kernel being of size k, the first layer convolution output may be represented as
F out ＝(x r *w r -x i *w i )+j(x r *w i +x i *w r ) (2)
The real and imaginary parts of the first layer convolution output will be spliced in the new dimension and then will enter the complex normalization layer, which can be expressed in terms of
LN rr ＝LN r Re(F out ),LN ri ＝LN r Im(F out ) (3)
LN ir ＝LN i Re(F out ),LN ii ＝LN i Im(F out ) (4)
cLN out ＝(LN rr -LN ii )+j(LN ri +LN ir ) (5)
Wherein LN r And LN i Representing the real and imaginary parts of the normalized layer, LN ri Representing real normalization of the imaginary part of the signal cLN out Representing the normalized output. cLN out And splicing the real part and the imaginary part on the new dimension again, and performing a re-convolution operation which comprises N convolution kernels and is similar to the first layer to obtain complex domain feature extraction output.
The signal enters a complex domain signal separation module after complex domain feature extraction, the complex domain signal separation module consists of a plurality of complex depth convolution modules and complex LSTM modules, and the complex LSTM is positioned in the middle of the separation module and is responsible for extracting the correlation before and after the signal. The complex depth convolution module adopts a jump connection path, and the input and output of each module are respectively in real partAnd the imaginary part is added and enters the next deep convolution module, and the dimensions of the signal input and output remain unchanged. The inner part of the convolution kernel adopts the form of expansion convolution, the convolution kernel of each complex depth convolution module presents exponential increment, and the corresponding convolution kernel sizes of the V depth convolution modules are 1,2 V-1 So as to more fully capture the high-dimensional characteristics of the signal. Each link has 2V complex depth convolution modules, which exhibit a symmetrical structure on both sides of the complex LSTM.
The complex depth convolution module is shown in fig. 4, after complex signals are input, the real part and the imaginary part are respectively subjected to convolution operation as shown in formula (2), and then the real part and the imaginary part of the output are respectively subjected to real LeakyReLU activation operation as shown in formula (6):
Where α ε R is a trainable scale factor in order to prevent gradient disappearance. After activation, the real and imaginary parts enter the complex normalization layer, which can be expressed as:
cgLN out ＝(gLN r (f r )-gLN i (f i ))+j(gLN r (f i )+gLN i (f r )) (7)
e (f) and Var (f) represent signal characteristics f ε C 1×T Mean and variance of (c), μ, v ε R 1×1 Is a trainable parameter, f r ∈R 1×T And f i ∈R 1×T Representing the real and imaginary parts of f. Subsequently, cgLN out The complex D-Conv processing is performed twice, the convolution kernel of the D-Conv is S, and the S size grows exponentially along with different complex depth convolution modules, so that the signal characteristics are captured in multiple scales. D-Conv convolution is then performed again with real LeakyReLU activation, complex normalization and rewinding lamination, and the output and input are stored with the same dimension and added as the next complex numberThe input of the depth convolution module.
The complex LSTM is in the middle of the separation module and is responsible for extracting the correlation before and after the signal. Similar to the deconvolution operation, the outputs L of the complex LSTM out Can be expressed as
L rr ＝LSTM r (x r )；L ri ＝LSTM r (x i )
L ir ＝LSTM i (x r )；L ii ＝LSTM i (x i )
L out ＝(L rr -L ii )+j(L ri +L ir )
Wherein LSTM r And LSTM i Representing real and imaginary operations of complex LSTM, L ri Representing a real LSTM operation on the imaginary part of the signal.
After being processed by 2V deep convolution modules and CLSTM in the separation module, the signals enter a complex domain signal recovery module. The output of the separation module is denoted f sepout ，f sepout Firstly, a complex number 1 multiplied by 1 convolution operation is carried out, the number of convolution kernels is M, the number of convolution kernels is consistent with the number of first-layer convolution kernels in a complex number domain feature extraction module, and convolution output is expressed as P l ∈C M×T L=1, 2, where l represents the link number, T represents the signal length, for P l The real part and the imaginary part of the target source signal are respectively subjected to sigmoid activation function operation, and then are spliced in a new dimension to obtain mask output, wherein the mask vector of the target source signal can be expressed by CRM:
wherein s is lr Sum s li Representing real and imaginary parts of the source signal of the first link, y lr And y li Representing real and imaginary representations of the mixed signal, each estimated source signal representation may be defined asThe concrete steps are as follows:
the mask output and the first layer of rewinding output in the complex domain feature extraction module have the same dimension, and the mask output and the first layer of rewinding output are multiplied by the point of the root of the common-node-alpha so as to obtain the high-dimensional representation of the restored signal. Re-pairingThe estimated signal can be recovered by performing a complex convolution operation>Expressed as:
wherein D-conv represents a complex one-dimensional convolution, the real part and the imaginary part of the signal are respectively from M dimension to 1 dimension, and k represents the convolution kernel size.
The network training objective function adopts a weighted combination of a scale invariant signal-to-noise ratio (SI-SNR) and a Mean Square Error (MSE) as a model training loss function, and the method concretely comprises the following steps:
The SI-SNR can be expressed as:
wherein s andrepresenting source signals and estimatesAnd (5) counting signals.<·,·>Representing the dot product of two vectors, I.I 2 Representing a binary norm, MSE is defined as
T denotes the signal length, SI-SNR and MSE are calculated for the real and imaginary parts of the signal, respectively, i.e.:
SI-SNR ave ＝mean(SI-SNR real +SI-SNR imag )
MSE ave ＝mean(MSE real +MSE imag )
finally, the loss function of the network first link may be defined as:
Loss l ＝SI-SNR avel +αMSE avel ,l＝1,2
where α represents a training coefficient, and α=1 is set.
And (3) grouping and packaging data of the training set and the verification set, sending the data into a neural network, setting network parameters, adopting an Adam optimizer, setting the initial learning rate to be 0.001, and halving the learning rate when the loss of the verification set is not reduced in continuous 3 rounds of training. The total training adopts M rounds, M is determined according to actual conditions, gradient clipping with maximum L2 norm of 5 is applied during training, model training adopts advanced stopping, and training is stopped when the loss of the verification set of 6 rounds is no longer reduced. After the training set and the verification set are finished, the model stores the optimal parameter model.
After the mixed signals of BPSK and QPSK are input into a network for separation, two separated waveforms are obtained, the two separated waveforms are respectively demodulated to recover bit data of signals, and the separated bit error rate can be obtained by comparing the bit data with the bit data of a source signal and can be used as a system separation performance evaluation index. Fig. 5 shows pearson correlation coefficients of BPSK and QPSK separated signals with the source signal at different signal-to-noise ratios, and it can be seen that the correlation coefficients increase stepwise with increasing signal-to-noise ratio. Fig. 6 shows the BPSK and QPSK separation error rates at different signal-to-noise ratios, and for convenience of testing, the signal is not added with channel coding, and it can be seen that the dual-path complex time domain convolutional network has good signal separation capability.
In summary, the invention can realize the effective separation of the simultaneous same-frequency mixed signals under the condition of single antenna receiving, and can improve the anti-interference capability of wireless communication; the invention also provides a new communication multiplexing method, which provides a solving method for relieving the contradiction between increasingly stressed frequency spectrum resources and frequency demand.

Claims (7)

1. The wireless signal single-channel blind source separation method based on the deep neural network is characterized by comprising a training stage and an application stage, wherein:
the training phase comprises the following steps:
a1, creating a data set of time-frequency aliasing signals, wherein the data set comprises communication signals and interference signals, and dividing the data set into a training set and a verification set;
a2, constructing a double-path complex time domain convolution network, wherein the double-path complex time domain convolution network comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
step A3, creating a self-defined network training objective function;
step A4, training and verifying the dual-path complex time domain convolution network obtained in the step A2 by using the signal data set obtained in the step A1 based on a network training objective function to obtain a trained dual-path complex time domain convolution network;
the application stage comprises the following steps:
Step B1, acquiring a time-frequency aliasing signal to be separated through a single antenna;
step B2, carrying out sectional arrangement on the time-frequency aliasing signals to obtain complex-form aliasing signals with fixed lengths;
step B3, separating a plurality of signals from complex-form aliasing signals by using a trained double-path complex time domain convolution network, and completing blind separation of the aliasing signals;
and B4, sending the separated signal obtained in the step B3 to a demodulator for demodulation, and recovering bit data of the separated signal.
2. The method for single-channel blind source separation of wireless signals based on a deep neural network according to claim 1, wherein the time-frequency aliasing communication signals in step A1 are represented by the following formula:
where x (t) is a time-frequency aliased communication signal, s 1 (t),...,s λ (t)∈C 1×T Represents lambda source signals, T represents signal length, a i The unknown mixing coefficient is shown, and n (t) is gaussian white noise.
3. The single-channel blind source separation method based on the double-path complex time domain convolution network according to claim 1, wherein the construction of the data set in step A1 adopts any one of the following two modes:
the first way is: the method comprises the steps of generating communication signals and interference signals in a plurality of different modulation modes by adopting computer software, wherein the modulation modes are digital modulation or analog modulation, and the interference signals comprise digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb-shaped spectrum interference signals, broadband blocking interference and sweep frequency interference;
Based on the generated communication signals and interference signals, randomly mixing the communication signals and the interference signals to form an aliasing signal, adding noise to the aliasing signal to generate a time-frequency aliasing communication signal, wherein the SNR (signal-to-noise ratio) is-5 dB to 20dB, and dividing a data set of the time-frequency aliasing communication signal into a training set and a verification set according to the ratio of 80% to 20% according to the network training requirement;
in a second mode, the actual communication equipment and the interference equipment are sampled to obtain a data set of the time-frequency aliasing communication signals, and the data set of the time-frequency aliasing communication signals is divided into a training set and a verification set according to the network training requirements and the ratio of 80% to 20%.
4. The single-channel blind source separation method based on the double-path complex time domain convolution network according to claim 1, wherein the constructing of the double-path complex time domain convolution network in the step A2 comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module, and is specifically as follows:
the double-path complex time domain convolution network adopts two parallel links, which respectively correspond to two separated signals, and each path comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
The complex domain feature extraction module adopts a two-stage complex domain coding structure and comprises two complex convolution layers and a complex normalization layer; the real part and the imaginary part of the convolution output of the first layer are spliced in a new dimension, the complex normalized layer is processed, then the second layer of complex convolution is carried out, and all operations of the complex domain feature extraction module are carried out on the real part and the imaginary part of the complex value signal simultaneously;
the complex domain signal separation module consists of a plurality of complex depth convolution modules and complex LSTM modules, wherein the complex LSTM modules are arranged in the middle of the separation module and are responsible for extracting the correlation before and after signals; the complex depth convolution module adopts a jump connection path, the input and the output of each complex depth convolution module are added in the real part and the imaginary part respectively, the complex depth convolution module enters the next complex depth convolution module, and the dimensionality of the signal input and the signal output is kept unchanged; the inner part of the convolution kernel adopts the form of expansion convolution, the convolution kernel of each complex depth convolution module presents exponential increment, and the corresponding convolution kernel sizes of the V complex depth convolution modules are 1,2 V-1 The method comprises the steps of carrying out a first treatment on the surface of the Each link is provided with 2V complex depth convolution modules, and symmetrical structures are shown on two sides of a complex LSTM;
in the complex domain signal recovery module, the output of the complex domain signal separation module is firstly subjected to complex 1×1 convolution, then the real part and the imaginary part are respectively subjected to sigmoid activation, and a mask matrix is obtained by splicing in a new dimension; the mask and the first layer of rewinding output in the complex domain feature extraction module have the same dimension, and the two obtain the recovered signal high-dimensional representation through dot multiplication; finally, performing the rewinding operation again to reduce the dimension of the high-dimension representation, namely recovering the estimated signal.
5. The single-channel blind source separation method based on the double-path complex time domain convolution network according to claim 1, wherein the network training objective function in the step A3 adopts a weighted combination of a scale-invariant signal-to-noise ratio SI-SNR and a mean square error MSE as a model training loss function, specifically as follows:
the scale invariant signal to noise ratio SI-SNR is expressed as:
wherein s andrepresenting the source signal and the estimated signal,<·,·>representing the dot product of two vectors, I.I 2 Representing a binary norm;
mean square error MSE is defined as
Where T represents the signal length, SI-SNR and MSE are calculated and averaged for the real and imaginary parts of the signal, respectively, i.e.:
SI-SNR ave ＝mean(SI-SNR real +SI-SNR imag )
MSE ave ＝mean(MSE real +MSE imag )
wherein SI-SNR ave Representing the averageScale-invariant signal-to-noise ratio, SI-SNR real Representing real scale invariant signal-to-noise ratio, SI-SNR imag Representing the imaginary scale-invariant signal-to-noise ratio, MSE ave Mean squared error, MSE real Representing real mean square error, MSE imag Representing the imaginary mean square error;
finally, network NoLoss function of link->The definition is as follows:
wherein α represents a training coefficient, and α=1 is set;indicate->Road average scale invariant signal-to-noise ratio, +.>Indicate->Mean squared error of the path.
6. The single-channel blind source separation method based on the dual-path complex time domain convolution network according to claim 1, wherein the training and verification of the dual-path complex time domain convolution network obtained in step A2 by using the signal data set obtained in step A1 in step A4 is specifically as follows:
Grouping and packaging data of a training set and a verification set, wherein the sequence combination is mini_batch, the mini_batch is sent into a double-path complex time domain convolution network, network parameters are set, an Adam optimizer is adopted, the initial learning rate is 0.001, and when the loss of the verification set is not reduced in continuous 3 rounds of training, the learning rate is halved; the total training adopts M rounds, M is a positive integer set according to actual conditions, gradient clipping with the maximum L2 norm of 5 is applied during training, model training is stopped in advance, and training is stopped when the loss of the continuous 6 rounds of verification sets is no longer reduced; after the training set and the verification set are finished, the model stores the optimal parameter model.
7. A single-channel blind source separation device based on a double-path complex time domain convolution network, which is characterized in that the device is used for realizing the single-channel blind source separation method based on the double-path complex time domain convolution network according to any one of claims 1 to 6, and the device comprises a training module and an application module, wherein:
the training module comprises a data set creating unit, a convolution network building unit, an objective function creating unit and a network training unit, wherein:
the data set creating unit creates a data set of the time-frequency aliasing signal, which comprises a communication signal and an interference signal, and divides the data set into a training set and a verification set;
The convolution network construction unit is used for constructing a double-path complex time domain convolution network and comprises a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;
the target function creating unit creates a self-defined network training target function;
the network training unit is used for training and verifying the dual-path complex time domain convolutional network obtained by the convolutional network building unit by using the signal data set obtained by the data set building unit based on the network training target function to obtain a trained dual-path complex time domain convolutional network;
the application module comprises a signal acquisition unit, a time-frequency aliasing signal arrangement unit, a blind separation unit and a demodulation unit, wherein:
the signal acquisition unit acquires a time-frequency aliasing signal to be separated through a single antenna;
the time-frequency aliasing signal arrangement unit is used for carrying out sectional arrangement on the time-frequency aliasing signals to obtain complex-form aliasing signals with fixed lengths;
the blind separation unit separates a plurality of signals from complex-form aliasing signals by using a trained double-path complex time domain convolution network to finish blind separation of the aliasing signals;
and the demodulation unit is used for sending the separated signal obtained by the blind separation unit to a demodulator for demodulation and recovering bit data of the separated signal.
CN202310833146.3A 2023-07-07 2023-07-07 Wireless signal single-channel blind source separation method and device based on deep neural network Pending CN116933066A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310833146.3A CN116933066A (en) 2023-07-07 2023-07-07 Wireless signal single-channel blind source separation method and device based on deep neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310833146.3A CN116933066A (en) 2023-07-07 2023-07-07 Wireless signal single-channel blind source separation method and device based on deep neural network

Publications (1)

Publication Number Publication Date
CN116933066A true CN116933066A (en) 2023-10-24

Family

ID=88374949

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310833146.3A Pending CN116933066A (en) 2023-07-07 2023-07-07 Wireless signal single-channel blind source separation method and device based on deep neural network

Country Status (1)

Country Link
CN (1) CN116933066A (en)

Similar Documents

Publication Publication Date Title
CN109672639B (en) Signal demodulation method based on machine learning
CN113630130B (en) End-to-end digital communication demodulation method
CN111444805B (en) Improved multi-scale wavelet entropy digital signal modulation identification method
CN113206813B (en) OFDM system peak-to-average power ratio suppression method based on improved selective mapping method
CN111277530B (en) Modulation-demodulation method for carrier system keying
Zhou et al. Distributed automatic modulation classification based on cyclic feature via compressive sensing
CN111368979A (en) PAPR suppression method for MIMO-OFDM system
CN114745230A (en) OTFS signal receiving and recovering method based on deep neural network structure
CN114422311B (en) Signal modulation recognition method and system combining deep neural network and expert priori features
CN110290083B (en) Multi-carrier differential chaotic system demodulation method based on low-rank matrix estimation
CN112003804A (en) Multipath multivariate differential chaotic shift keying iterative receiving method
Ghauri KNN based classification of digital modulated signals
US8189709B2 (en) Frequency domain equalization method for continuous phase modulated signals
Zhou et al. Wavelet cyclic feature based automatic modulation recognition using nonuniform compressive samples
Abdelmutalab et al. Automatic modulation classification using polynomial classifiers
Pavlov et al. Deep learning application for classification of SEFDM signals
CN116933066A (en) Wireless signal single-channel blind source separation method and device based on deep neural network
CN113242201B (en) Wireless signal enhanced demodulation method and system based on generation classification network
CN115392325A (en) Multi-feature noise reduction modulation identification method based on cycleGan
CN115150230A (en) Orthogonal time-frequency space modulation system and method for improving spectrum efficiency
Huynh-The et al. Deep learning-based automatic modulation classification for wireless ofdm communications
Alnajjar et al. Automatic modulation classification in deep learning
CN115086127B (en) Improved MPPSK signal demodulation system based on stack DSAE network
KR102670053B1 (en) Automatic Modulation Classification Scheme Based on Deep Learning Technique
Zheng et al. Deep learning-aided receiver against nonlinear distortion of hpa in ofdm systems

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