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

Abstract

This invention discloses a method and apparatus for single-channel blind source separation of wireless signals based on deep neural networks. Specifically, in the training phase, a signal dataset is created, including a training set and a validation set; a dual-path complex time-domain convolutional network is constructed; a custom network training objective function is created; the obtained signal dataset is used to train and validate the dual-path complex time-domain convolutional network, resulting in a trained dual-path complex time-domain convolutional network; in the application phase, the time-frequency aliased signal to be separated is acquired through a single antenna; the time-frequency aliased signal is segmented and processed to obtain a fixed-length complex-form aliased signal; using the trained dual-path complex time-domain convolutional network, multiple signals are separated from the complex-form aliased signal and sent to a demodulator for demodulation to recover the bit data of the separated signal. This invention can effectively separate simultaneously and in-frequency mixed signals under single-antenna reception conditions, improving the anti-interference capability of wireless communication.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, and particularly relates to a wireless signal single-channel blind source separation method based on a deep neural network. BACKGROUND

[0002] Due to the growing mobile applications, many wireless devices will usually access the same spectrum resources at the same time, which makes the collision and conflict in the wireless medium inevitable. In order to overcome these shortcomings, single-channel blind source separation is a very promising technology, especially under the limited electromagnetic spectrum. Cognitive radio technology can alleviate the problem of low spectrum utilization through dynamic spectrum access, but it still cannot essentially meet the spectrum resources required by future users. As a technology that can increase the utilization rate of system frequency band, single-channel blind source separation has always been a research hotspot because it can recover the source signal from the mixed signal with the least prior knowledge and fewer antenna resources.

[0003] Single-channel blind source separation is a process of recovering multi-dimensional source signals from one-dimensional mixed signals. Compared with ordinary blind source separation, single-channel blind source separation only has one observation signal, which is a pathological problem and extremely difficult to solve. The most advanced single-channel blind source separation scheme in current wireless communication 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 using decomposition methods such as wavelet decomposition, wavelet packet decomposition and empirical mode decomposition (EMD); single-channel blind source separation based on basis function training, which uses classical blind source separation optimization algorithm to reconstruct the source signal to obtain the basis function; single-channel blind source separation based on transform domain filtering, which uses the sparsity of source signals in time-frequency domain, cyclic domain, wavelet domain and other domains to filter the superimposed signal; single-channel blind source separation based on model parameter estimation and reconstruction, which estimates the signal model parameters and then reconstructs the required source signal. However, due to the excessively high computational complexity, especially under the condition of low signal-to-noise ratio (SNR), the existing single-channel blind source separation scheme is difficult to achieve excellent performance.

[0004] To solve the above-mentioned shortcomings, deep learning has been used in single-channel blind source separation, especially in the field of separating speech waveforms, and such methods mainly focus on the time-frequency domain or directly in the time domain. Representative methods in the time-frequency domain include deep clustering learning, permutation invariant training, and deep attractor networks. Representative time-domain networks include long short-term memory (LSTM) recurrent networks, convolutional recurrent networks (CRN), etc. The latest progress in speech separation networks also shows great potential in the field of wireless communication. However, there are still great differences between the communication and speech fields. First, communication information bits are compressed very tightly, usually using a small number of algebraic operations, and lack many inherent redundancies or contexts, while in the speech domain, relatively long identifiable formants appear in sequences to synthesize words and phrases. Second, the impact of information on speech processing mostly occurs on a larger time scale, while the symbolic meaning in communication usually occurs on a smaller time scale (such as microseconds or nanoseconds). Third, wireless communication systems usually use complex baseband representations of I / Q data, while speech signals are usually represented in real numbers in the time domain. These limitations result in the fact that signal separation networks in the speech domain cannot be directly used for the separation of communication signals.

[0005] Therefore, the present application is dedicated to realizing wireless signal single-channel blind source separation based on deep learning. In civil communication, the present application can separate two time-frequency mixed signals in the same frequency band, increase the information transmission rate without increasing the cost of spectrum resources, and improve the spectrum utilization efficiency, which is of great significance to alleviate the tension of spectrum resources. In the field of military communication anti-jamming, blind source separation can separate target signals and interference signals, realize weak signal enhancement and communication anti-jamming. SUMMARY

[0006] The purpose of the present application is to provide a wireless signal single-channel blind source separation method based on deep neural networks, which realizes effective separation of simultaneous and same frequency mixed signals under the condition of single antenna reception, thereby improving the anti-jamming capability of wireless communication.

[0007] The technical solution for achieving the purpose of the present application is: a wireless signal single-channel blind source separation method based on deep neural networks, including a training phase and an application phase, wherein:

[0008] The training phase includes the following steps:

[0009] Step A1, creating a data set of time-frequency mixed signals, including communication signals and interference signals, and dividing the data set into a training set and a validation set;

[0010] Step A2, building a double-path complex time-domain convolutional network, including a complex feature extraction module, a complex signal separation module, and a complex signal recovery module;

[0011] Step A3, creating a customized network training objective function;

[0012] Step A4, based on the network training objective function, using the signal data set obtained in step A1, training and verifying the double-path complex time domain convolution network obtained in step A2, to obtain a trained double-path complex time domain convolution network;

[0013] The application stage includes the following steps:

[0014] Step B1, obtaining a time-frequency mixed signal to be separated through a single antenna;

[0015] Step B2, segmenting and arranging the time-frequency mixed signal to obtain a complex mixed signal of fixed length;

[0016] Step B3, using the trained double-path complex time domain convolution network to separate multiple signals from the complex mixed signal, to complete blind separation of the mixed signal;

[0017] Step B4, sending the separated signal obtained in step B3 to a demodulator for demodulation to recover the bit data of the separated signal.

[0018] Further, the time-frequency mixed communication signal in step A1 is as follows:

[0019]

[0020] In the formula, x(t) is a time-frequency mixed communication signal, s1(t),..., s λ (t)∈C 1×T represent λ source signals, T represents the signal length, a i represents unknown mixing coefficients, and n(t) represents Gaussian white noise.

[0021] Further, the construction of the data set in step A1 adopts any one of the following two ways:

[0022] The first way is to generate communication signals and interference signals of multiple different modulation modes by computer software, the modulation mode being digital modulation or analog modulation, and the interference signal including a digital modulation interference signal, an analog modulation interference signal, a noise frequency modulation interference signal, a noise amplitude modulation interference signal, a single tone interference signal, a multi-tone interference signal, a comb spectrum interference signal, a wideband blocking interference, and a sweep frequency interference;

[0023] Based on the generated communication signals and interference signals, the communication signals and interference signals are randomly mixed to form a mixed signal, noise is added to the mixed signal to generate a time-frequency mixed communication signal, the signal-to-noise ratio SNR is -5dB to 20dB, and the data set of the time-frequency mixed communication signal is divided into a training set and a verification set according to the network training requirements in a ratio of 80% and 20%;

[0024] Secondly, the actual communication device and the interference device are sampled to obtain a data set of the time-frequency aliasing communication signal, and the data set of the time-frequency aliasing communication signal is divided into a training set and a verification set according to a proportion of 80% and 20% according to network training requirements.

[0025] Further, the dual-path complex time domain convolution network built in step A2 includes a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module, and specifically as follows.

[0026] The dual-path complex time domain convolution network adopts two parallel links corresponding to two separated signals, and each path includes a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module.

[0027] The complex domain feature extraction module adopts a two-level complex domain coding structure, including two complex convolution layers and a complex normalization layer; the real part and the imaginary part output by the first layer convolution are spliced in a new dimension, and after being processed by the complex normalization layer, the second layer complex convolution is performed, and all operations of the complex domain feature extraction module are performed on the real part and the imaginary part of the complex value signal at the same time.

[0028] The complex domain signal separation module is composed of multiple complex depth convolution modules and a complex LSTM module, the complex LSTM module is arranged 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 skip connection path, the input and output of each complex depth convolution module are added in the real part and the imaginary part respectively, and then enter the next complex depth convolution module, and the dimension of the signal input and output remains unchanged; the convolution kernel inside adopts an inflation convolution form, the convolution kernel size of each complex depth convolution module presents an exponential increase, and the convolution kernel sizes corresponding to V complex depth convolution modules are 1, 2,..., 2 V-1 ; there are 2V complex depth convolution modules on each link, and the complex LSTM presents a symmetrical structure on both sides;

[0029] In the complex domain signal recovery module, the output of the complex domain signal separation module is first subjected to a complex 1x1 convolution, and then the real part and the imaginary part are subjected to sigmoid activation respectively, and a mask matrix is obtained by splicing in a new dimension; the mask and the output of the first layer complex convolution in the complex domain feature extraction module have the same dimension, and the recovered signal high-dimensional representation is obtained by point multiplication; finally, a complex convolution operation is performed again to reduce the dimension of the high-dimensional representation, that is, to recover the estimated signal.

[0030] Further, the network training target function in 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, and specifically as follows.

[0031] The scale-invariant signal-to-noise ratio (SI-SNR) is represented as:

[0032]

[0033]

[0034]

[0035] where s and denote the source signal and the estimated signal, <·,·> denotes the dot product of two vectors, and ||·||2 denotes the 2-norm;

[0036] The mean square error (MSE) is defined as

[0037]

[0038] where T denotes the signal length, and the SI-SNR and MSE are calculated for the real and imaginary parts of the signal, respectively, and averaged, i.e.,

[0039] SI-SNR ave = mean(SI-SNR real + SI-SNR imag )

[0040] MSE ave = mean(MSE real + MSE imag )

[0041] where SI-SNR ave denotes the average scale-invariant signal-to-noise ratio, SI-SNR real denotes the real scale-invariant signal-to-noise ratio, SI-SNR imag denotes the imaginary scale-invariant signal-to-noise ratio, MSE ave denotes the average mean square error, MSE real denotes the real mean square error, and MSE imag denotes the imaginary mean square error;

[0042] Finally, the loss function Loss l for the lth link of the network is defined as:

[0043] Loss l = SI-SNR avel + aMSE avel , l = 1, 2

[0044] where a denotes the training coefficient, which is set to a = 1; SI-SNR avel denotes the lth path average scale-invariant signal-to-noise ratio, and MSE avel denotes the lth path average mean square error.

[0045] Further, step A4 uses the signal data set obtained in step A1 to train and verify the double-path complex time domain convolution network obtained in step A2, as follows:

[0046] The data in the training set and the verification set are grouped and packaged, and the sequence is combined into a mini batch, which is sent to the double-path complex time domain convolution network. The network parameters are set, the Adam optimizer is used, the initial learning rate is 0.001, and when the verification set loss no longer decreases in the next 3 rounds of training, the learning rate is halved; the total training uses M rounds, M is a positive integer set according to the actual situation, gradient clipping with a maximum L2 norm of 5 is applied during training, and the model training uses early stopping, which stops training when the verification set loss no longer decreases for 6 consecutive rounds; after the model is trained with the training set and the verification set, the optimal parameter model is saved.

[0047] A single-channel blind source separation device based on a double-path complex time domain convolution network, which is used to implement the single-channel blind source separation method based on the double-path complex time domain convolution network, the device comprises a training module and an application module, wherein:

[0048] The training module comprises a data set creation unit, a convolution network building unit, a target function creation unit, and a network training unit, wherein:

[0049] The data set creation unit creates a data set of time-frequency mixed signals, including communication signals and interference signals, and divides the data set into a training set and a verification set;

[0050] The convolution network building unit builds a double-path complex time domain convolution network, including a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module;

[0051] The target function creation unit creates a custom network training target function;

[0052] The network training unit uses the signal data set obtained by the data set creation unit to train and verify the double-path complex time domain convolution network obtained by the convolution network building unit based on the network training target function, and obtains the trained double-path complex time domain convolution network;

[0053] The application module comprises a signal acquisition unit, a time-frequency mixed signal arrangement unit, a blind separation unit, and a demodulation unit, wherein:

[0054] The signal acquisition unit acquires the time-frequency mixed signals to be separated through a single antenna;

[0055] The time-frequency mixed signal arrangement unit segments and arranges the time-frequency mixed signals to obtain complex mixed signals of a fixed length;

[0056] The blind separation unit separates a plurality of signals from the complex mixed signal by using the trained double-path complex time domain convolution network, and completes blind separation of the mixed signal.

[0057] The demodulation unit sends the separated signal obtained by the blind separation unit into a demodulator for demodulation, so as to recover the bit data of the separated signal.

[0058] Compared with the prior art, the present application has the following advantages: (1) the double-path complex time domain convolution network can directly separate the wireless communication signal, and the correlation between the signal I / Q data is fully utilized in network training, and the separated signal can be directly demodulated in 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 encoding and decoding, complex LSTM, complex deep convolution module, etc., so that the high-dimensional features of the signal are fully extracted; (3) simulation results show that the network has good signal separation performance, and can effectively separate the simultaneous and same frequency mixed signals under the condition of single antenna reception, thereby improving the anti-interference ability of wireless communication. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The flowchart of the wireless signal single-channel blind source separation method based on the deep neural network of the present application.

[0060] Figure 2 The schematic diagram of the double-path complex time domain convolution network of the present application.

[0061] Figure 3 The schematic diagram of the complex domain feature extraction module of the present application.

[0062] Figure 4 The schematic diagram of the complex deep convolution module of the present application.

[0063] Figure 5 The schematic diagram of the correlation coefficient of the BPSK and QPSK separation results under different signal-to-noise ratios of the present application.

[0064] Figure 6 The schematic diagram of the bit error rate of the BPSK and QPSK separation results under different signal-to-noise ratios of the present application. DETAILED DESCRIPTION

[0065] In view of the poor effect of the traditional single-channel separation method and the inadaptation of the existing speech separation network to communication signals, the present application establishes a neural network that can meet the separation of complex communication signals. The network can fully utilize the correlation of the signal I / Q data, the separated signal can be directly expressed in complex form after being separated by the network, which is more conducive to subsequent demodulation, and the performance is better than that of the traditional separation method and the existing separation network model.

[0066] The application discloses a wireless signal single-channel blind source separation method based on a deep neural network.

[0067] The training stage comprises the following steps:

[0068] Step A1, a data set of time-frequency mixed signals is created, including communication signals and interference signals, and the data set is divided into a training set and a verification set;

[0069] Step A2, a double-path complex time domain convolution network is built, including a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module;

[0070] Step A3, a custom network training target function is created;

[0071] Step A4, based on the network training target function, the double-path complex time domain convolution network obtained in step A2 is trained and verified by using the signal data set obtained in step A1, so that a trained double-path complex time domain convolution network is obtained;

[0072] The application stage comprises the following steps:

[0073] Step B1, a time-frequency mixed signal to be separated is obtained through a single antenna;

[0074] Step B2, the time-frequency mixed signal is segmented and arranged to obtain a complex mixed signal with a fixed length;

[0075] Step B3, the trained double-path complex time domain convolution network is used to separate multiple signals from the complex mixed signal, so that blind separation of the mixed signal is completed;

[0076] Step B4, the separated signal obtained in step B3 is sent to a demodulator for demodulation, so that bit data of the separated signal is recovered.

[0077] As a specific example, the time-frequency mixed communication signal in step A1 is shown in the following formula:

[0078]

[0079] In the formula, x(t) is the time-frequency mixed communication signal, s1(t),..., s λ (t)∈C 1×T represent λ source signals, T represents the signal length, a i represents unknown mixing coefficients, and n(t) represents Gaussian white noise.

[0080] As a specific example, the data set in step A1 is constructed in any one of the following two ways:

[0081] The first mode: a computer software is used to generate a plurality of communication signals with different modulation modes and interference signals. The modulation modes are digital modulation signals or analog modulation signals. The interference signals include 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 spectrum interference signals, wideband blocking interference signals and sweep frequency interference signals. The digital modulation includes MPSK (M-ary phase shift keying), MASK (M-ary amplitude shift keying), MFSK (M-ary frequency shift keying), MQAM (M-ary quadrature amplitude modulation), CPSK, OQPSK, OFDM, GMSK, DPSK, TCM and CPFSK. The analog modulation includes AM, FM, PM, DSB, VSB and SSB.

[0082] Based on the generated communication signals and interference signals, the communication signals and the interference signals are randomly mixed to form a mixed signal, noise is added to the mixed signal to generate a time-frequency mixed communication signal, the signal-to-noise ratio SNR is -5dB to 20dB, and according to the network training requirements, the data set of the time-frequency mixed communication signal is divided into a training set and a verification set according to the proportion of 80% and 20%;

[0083] The second mode: the actual communication equipment and the interference equipment are sampled to obtain a data set of a time-frequency mixed communication signal, and according to the network training requirements, the data set of the time-frequency mixed communication signal is divided into a training set and a verification set according to the proportion of 80% and 20%.

[0084] As a specific example, the step A2 of building a double-path complex time domain convolution network includes a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module, and the details are as follows:

[0085] The double-path complex time domain convolution network adopts two parallel links, which correspond to two separated signals respectively. Each path includes a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module.

[0086] The complex domain feature extraction module adopts a two-level complex domain coding structure, which includes two complex convolution layers and one complex normalization layer. The real part and the imaginary part of the output of the first layer convolution are spliced in a new dimension, and then the second layer complex convolution is performed after the complex normalization layer processing. All operations of the complex domain feature extraction module are performed on the real part and the imaginary part of the complex value signal at the same time.

[0087] The complex domain signal separation module is composed of multiple complex depth convolution modules and a complex LSTM module, the complex LSTM module is arranged 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 skip connection path, the input and output of the complex depth convolution module are added in the real part and the imaginary part respectively each time, and then enter the next complex depth convolution module, the dimension of the signal input and output remains unchanged; the convolution kernel adopts an expanded convolution form, the size of the convolution kernel of each complex depth convolution module increases exponentially, the size of the convolution kernel corresponding to the V complex depth convolution modules is 1, 2,..., 2 V-1 , so as to more fully capture the high-dimensional characteristics of the signal; there are 2V complex depth convolution modules on each link, and the complex LSTM presents a symmetrical structure on both sides;

[0088] In the complex domain signal recovery module, the output of the complex domain signal separation module first passes through a complex 1*1 convolution, and then the real part and the imaginary part are activated by sigmoid respectively, and the mask matrix is obtained by splicing in the new dimension; the mask and the output of the first layer complex convolution in the complex domain feature extraction module have the same dimension, and the recovered signal high-dimensional representation is obtained by point multiplication; finally, a complex convolution operation is performed again to reduce the dimension of the high-dimensional representation, that is, to recover the estimated signal.

[0089] As a specific example, the network training objective function described in step A3 adopts a weighted combination of scale invariant signal-to-noise ratio SI-SNR and mean square error MSE as the model training loss function, which is as follows:

[0090] The scale invariant signal-to-noise ratio SI-SNR is represented as:

[0091]

[0092]

[0093]

[0094] Where s and represent the source signal and the estimated signal, <·,·> represents the dot product of two vectors, and ||·||2 represents the two norm;

[0095] The mean square error MSE is defined as

[0096]

[0097] Where T represents the signal length, and the SI-SNR and MSE of the real part and the imaginary part of the signal are calculated respectively, that is:

[0098] SI-SNR ave = mean(SI-SNR real +SI-SNRimag )

[0099] MSE ave =mean(MSE real +MSE imag )

[0100] wherein, SI-SNR ave represents the average scale invariant signal-to-noise ratio, SI-SNR real represents the real part scale invariant signal-to-noise ratio, SI-SNR imag represents the imaginary part scale invariant signal-to-noise ratio, MSE ave represents the average mean square error, MSE real represents the real part mean square error, MSE imag represents the imaginary part mean square error.

[0101] Finally, the loss function Loss l of the network of the lth link is defined as:

[0102] Loss l =SI-SNR avel +αMSE avel ,l=1,2

[0103] wherein, α represents a training coefficient, and is set as α=1; SI-SNR avel represents the average scale invariant signal-to-noise ratio of the lth path, MSE avel represents the average mean square error of the lth path.

[0104] As a specific example, the signal data set obtained in step A1 is used to train and verify the dual-path complex time domain convolution network obtained in step A2, and the specific steps are as follows:

[0105] The data of the training set and the verification set are grouped and packaged, and the sequence is combined into a mini_batch, which is sent into the dual-path complex time domain convolution network. Network parameters are set, and the Adam optimizer is adopted with an initial learning rate of 0.001. When the loss of the verification set no longer decreases in the next 3 rounds of training, the learning rate is halved. A total of M rounds of training are adopted, M is a positive integer set according to the actual situation. During the training, the gradient clipping with a maximum L2 norm of 5 is applied. The model training adopts early stopping, and the training is stopped when the loss of the verification set no longer decreases for 6 consecutive rounds. After the model is trained by the training set and the verification set, the optimal parameter model is saved.

[0106] The application discloses a single-channel blind source separation device based on a dual-path complex time domain convolution network.

[0107] The training module comprises a data set creation unit, a convolution network building unit, a target function creation unit and a network training unit, wherein:

[0108] The data set creation unit creates a data set of time-frequency aliasing signals, including communication signals and interference signals, and divides the data set into a training set and a validation set.

[0109] The convolution network building unit builds a double-path complex time-domain convolution network, including a complex domain feature extraction module, a complex domain signal separation module and a complex domain signal recovery module.

[0110] The target function creation unit creates a custom network training target function.

[0111] The network training unit trains and validates the double-path complex time-domain convolution network obtained by the convolution network building unit based on the network training target function and the signal data set obtained by the data set creation unit, to obtain a trained double-path complex time-domain convolution network.

[0112] The application module comprises a signal acquisition unit, a time-frequency aliasing signal arrangement unit, a blind separation unit and a demodulation unit, wherein:

[0113] The signal acquisition unit acquires the time-frequency aliasing signal to be separated through a single antenna.

[0114] The time-frequency aliasing signal arrangement unit segments and arranges the time-frequency aliasing signal to obtain a complex form aliasing signal of a fixed length.

[0115] The blind separation unit separates a plurality of signals from the complex form aliasing signal by using the trained double-path complex time-domain convolution network, to complete blind separation of the aliasing signal.

[0116] The demodulation unit sends the separated signal obtained by the blind separation unit to a demodulator for demodulation, to recover the bit data of the separated signal.

[0117] The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0118] Embodiment

[0119] In combination Figure 1 , the application provides a single-channel blind source separation method based on a double-path complex time-domain convolution network, comprising the following steps:

[0120] Step 1: acquiring a single-channel time-frequency aliasing communication signal;

[0121] Step 2: creating a signal data set according to the aliasing signal obtained in step 1, the data set comprising a training set and a validation set;

[0122] Step 3: Build a double-path complex time domain convolution network, including a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module.

[0123] Step 4: Create a custom network training objective function.

[0124] Step 5: Train and verify the double-path complex time domain convolution network obtained in Step 3 using the signal dataset obtained in Step 2, obtain the separated two-way communication signal waveform, and complete the blind separation of the mixed signal.

[0125] Step 6: Send the separated waveform obtained in Step 5 to the demodulator for demodulation to recover the bit data of the separated signal.

[0126] Further, the time-frequency mixed signal in Step 1 is as follows:

[0127]

[0128] where x(t) is the mixed signal, s1(t),...,s λ (t)∈C 1×T represent λ source signals, T represents the signal length, a ij unknown mixing coefficients, and n(t) represents Gaussian white noise.

[0129] Further, the data set construction process in Step 2 is as follows:

[0130] Generate various different modulation signals such as BPSK, QPSK, 8PSK, 16QAM, 64QAM, etc. using MATLAB, 1000 of each modulation signal, signal length 1024, extract and store the real and imaginary parts of the signal samples in a new one-dimensional, and pack all the signals together to form a signal set.

[0131] Select two different modulation methods, such as BPSK and QPSK, randomly mix the signals to form mixed signals, add noise, and the signal-to-noise ratio SNR is from -5dB to 20dB. According to the network training requirements, divide the mixed signal set into training set and validation set according to the proportion of 80% and 20%.

[0132] Further, the double-path complex time domain convolution 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 are as follows:

[0133] The double-path complex time domain convolution network uses two parallel links, corresponding to two separated signals. Each path includes a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module.

[0134] The complex domain feature extraction module adopts a two-stage complex domain coding structure, including two complex convolution layers and a complex normalization layer. The real part and the imaginary part of the output of the first layer convolution are spliced in a new dimension, and after being processed by the complex normalization layer, the second layer complex convolution is performed. All operations are performed on the real part and the imaginary part of the complex value signal at the same time.

[0135] The complex domain signal separation module is composed of multiple complex depth convolution modules and a complex LSTM module. The complex LSTM is 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 skip connection path. The input and output of each module are added in the real part and the imaginary part respectively, and then enter the next depth convolution module. The dimensions of the signal input and output remain unchanged. The convolution kernel inside adopts an expanded convolution form. The convolution kernel size of each complex depth convolution module presents an exponential increase, and the convolution kernel size corresponding to V depth convolution modules is 1, 2,..., 2 V-1 , so as to more fully capture the high-dimensional features of the signal. Each link has 2V complex depth convolution modules, which present a symmetrical structure on both sides of the complex LSTM.

[0136] In the complex domain signal recovery module, the output of the separation module is first subjected to a complex 1x1 convolution, and then the real part and the imaginary part are subjected to sigmoid activation respectively. The splicing in the new dimension can obtain a mask matrix. The mask has the same dimension as the output of the first layer complex convolution in the complex domain feature extraction module. The two are multiplied by point to obtain the high-dimensional representation of the recovered signal. Finally, a complex convolution operation is performed again to reduce the dimension of the high-dimensional representation, and the estimated signal can be recovered.

[0137] 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 the model training loss function, as follows:

[0138] The SI-SNR can be expressed as:

[0139]

[0140]

[0141]

[0142] where s and represent the source signal and the estimated signal. <·,·> represents the dot product of two vectors, ||·||2 represents the two norm, and MSE is defined as

[0143]

[0144] T represents the signal length, and the SI-SNR and MSE are calculated for the real part and the imaginary part of the signal respectively, i.e.

[0145] SI-SNR ave = mean(SI-SNR real + SI-SNR imag )

[0146] MSE ave = mean(MSE real + MSE imag )

[0147] Finally, the loss function of the network of the lth link can be defined as:

[0148] Loss l = SI-SNR avel + a MSE avel , l = 1, 2

[0149] Wherein, a represents a training coefficient, and a = 1 is set.

[0150] Further, the signal data set in step 5 is used to train and verify the network, and the specific steps are as follows:

[0151] The data of the training set and the verification set are grouped and packaged, and the sequence combination is a mini batch, which is sent to the neural network. Network parameters are set, Adam optimizer is used, the initial learning rate is 0.001, and the learning rate is halved when the verification set loss is no longer reduced in the next 3 rounds of training. A total of M rounds of training are used, and M is determined according to the actual situation. During the training, the gradient clipping with the maximum L2 norm of 5 is applied. The model training adopts early stopping, and the training is stopped when the verification set loss is no longer reduced for 6 consecutive rounds. After the model is trained by the training set and the verification set, the optimal parameter model is saved.

[0152] Further, in step 6, the obtained separated waveforms are sent to the demodulator to demodulate and recover the bit data of the separated signals, and the specific steps are as follows:

[0153] After the mixed signals of BPSK and QPSK are input into the network for separation, two separated waveforms are obtained, and the bit data of the signals can be recovered by demodulating them respectively. By comparing with the bit data of the source signals, the separation error rate can be obtained, which can be used as an evaluation index of the system separation performance.

[0154] The embodiment also relates to 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 the device comprises a training module and an application module, and wherein:

[0155] The training module comprises a data set creation unit, a convolution network building unit, a target function creation unit and a network training unit, and wherein:

[0156] a data set creation unit, which creates a data set of the time-frequency mixed signals, including the communication signals and the interference signals, and includes the data set into a training set and a verification set;

[0157] a convolution network building unit, which builds a double-path complex time-domain convolution network, including a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module;

[0158] a target function creation unit, which creates a custom network training target function;

[0159] a network training unit, which trains and verifies the double-path complex time-domain convolution network obtained by the convolution network building unit based on the network training target function and the signal data set obtained by the data set creation unit, to obtain a trained double-path complex time-domain convolution network;

[0160] The application module includes a signal acquisition unit, a time-frequency mixed signal arrangement unit, a blind separation unit, and a demodulation unit, wherein:

[0161] The signal acquisition unit acquires the time-frequency mixed signals to be separated through a single antenna;

[0162] The time-frequency mixed signal arrangement unit arranges the time-frequency mixed signals in segments to obtain complex mixed signals of a fixed length;

[0163] The blind separation unit separates a plurality of signals from the complex mixed signals by using the trained double-path complex time-domain convolution network, to complete the blind separation of the mixed signals;

[0164] The demodulation unit sends the separated signals obtained by the blind separation unit to a demodulator for demodulation, to recover the bit data of the separated signals.

[0165] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings:

[0166] Figure 2 The double-path complex time-domain convolution network is a schematic diagram of the present application, which shows the whole process of the mixed signals from inputting into the network to complex encoding and decoding, complex deep convolution, complex LSTM processing, mask generation, and signal demodulation.

[0167] A plurality of types of time-frequency mixed signals are generated, such as BPSK, QPSK, 8PSK, 16QAM, 64QAM, and the like, which are shown in formula (1):

[0168]

[0169] In the formula, x(t) is the mixed signal, s1(t),..., s λ (t)∈C 1×Trepresents λ source signals, T represents the signal length, a ij represents unknown mixing coefficients, n(t) represents Gaussian white noise. Each modulation signal is 1000, the signal length is 1024, the real and imaginary parts of the signal samples are extracted and stored in a new one-dimensional, and all signals are packaged together to form a signal set. The signal storage format is (N, 2, L), N represents the number of samples, 2 represents the real and imaginary parts of two dimensions, and L represents the signal length.

[0170] Two different modulation methods are selected, such as BPSK and QPSK, and the signals are randomly mixed into aliasing signals, and noise is added. The noise is randomly added according to the signal-to-noise ratio from -5dB to 20dB to generate an aliasing noisy data set. According to the network training requirements, the aliasing noisy data set is divided into a training set and a validation set according to the proportions of 80% and 20%.

[0171] The training set and validation set signals (mixed signals and source signals) are packaged and input into the network, and first enter the complex domain feature extraction module.

[0172] As Figure 3 mentioned, the complex domain feature extraction module adopts two-level coding, including two complex convolution layers and one complex normalization layer, and all operations are performed on the real and imaginary parts of the complex value signal at the same time. The complex signal input can be represented as x = x r +jx i ,x∈C 1×T , first the signal enters the first complex convolution layer containing M convolution kernels, the convolution kernel structure can be represented as w = w r +jw i ,w∈C 1×M , w r and w i represent the real and imaginary parts of the complex convolution, respectively, and the size of the convolution kernel is k. The first layer convolution output can be represented as

[0173] F out =(x r *w r -x i *w i )+j(x r *w i +x i *w r ) (2)

[0174] The real and imaginary parts of the first layer convolution output are spliced in a new dimension, and then enter the complex normalization layer, which can be represented as

[0175] LN rr =LN r Re(F out ),LN ri =LNr Im(F out ) (3)

[0176] LN ir = LN i Re(F out ), LN ii = LN i Im(F out ) (4)

[0177] cLN out = (LN rr - LN ii ) + j(LN ri + LN ir ) (5)

[0178] where LN r and LN i represent the real part and the imaginary part of the normalized layer, LN ri represents the real normalization operation on the imaginary part of the signal, cLN out represents the normalized output. cLN out will splice the real part and the imaginary part again in the new dimension, and perform a complex convolution operation again, containing N convolution kernels, the complex convolution operation is similar to the first layer, to obtain the complex domain feature extraction output.

[0179] After the signal passes through the complex domain feature extraction, it enters the complex domain signal separation module, which is composed of multiple complex deep convolution modules and complex LSTM modules. The complex LSTM is in the middle of the separation module and is responsible for extracting the correlation before and after the signal. The complex deep convolution module adopts a skip connection path. The input and output of each module are added in the real part and the imaginary part respectively, and enter the next deep convolution module. The dimension of the signal input and output remains unchanged. The convolution kernel inside adopts the form of dilated convolution. The size of the convolution kernel of each complex deep convolution module increases exponentially, and the size of the convolution kernel corresponding to V deep convolution modules is 1, 2,..., 2 V-1 , so as to more fully capture the high-dimensional features of the signal. Each link has 2V complex deep convolution modules, which present a symmetrical structure on both sides of the complex LSTM.

[0180] The complex deep convolution module is shown in Figure 4 After the complex signal input, the real part and the imaginary part are subjected to convolution operation respectively, as shown in formula (2), and then the real part and the imaginary part of the output are subjected to real LeakyReLU activation operation respectively, as shown in formula (6):

[0181]

[0182] Here, α∈R is a trainable scaling factor designed to prevent gradient vanishing. After activation, the real and imaginary parts enter a complex normalization layer, which can be represented as:

[0183] cgLN out =(gLN r (f r )-gLN i (f i ))+j(gLN r (f i )+gLN i (f r (7)

[0184]

[0185] E(f) and Var(f) represent the signal characteristics f∈C 1×T The mean and variance, μ,ν∈R 1×1 These are trainable parameters, f r ∈R 1×T and f i ∈R 1×T Let f represent the real and imaginary parts. Then, cgLN out The process involves two complex D-Conv processes. The kernel size of the D-Conv is S, and the size of S increases exponentially with different complex depth convolution modules to capture signal features at multiple scales. After the D-Conv convolution, the process goes through real LeakyReLU activation, complex normalization, and complex convolution layers. The output retains the same dimension as the input and is added together to serve as the input for the next complex depth convolution module.

[0186] The complex LSTM is located in the middle of the separation module and is responsible for extracting the correlation between the preceding and following signals. Similar to complex convolution, the output L of the complex LSTM... out It can be represented as

[0187] L rr =LSTM r (x r );L ri =LSTM r (x i )

[0188] L ir =LSTM i (x r );L ii =LSTM i (x i )

[0189] L out =(L rr -Lii )+j(L ri +L ir )

[0190] where LSTM r and LSTM i denote the real and imaginary part operations of complex LSTM, L ri denotes the real LSTM operation on the imaginary part of the signal.

[0191] After processing by 2V deep convolution modules and CLSTM in the separation module, the signal enters the complex domain signal recovery module. The output of the separation module is denoted as f sepout , f sepout is first subjected to a complex 1x1 convolution operation, the number of convolution kernels is M, which is consistent with the number of the first layer complex convolution kernels in the complex domain feature extraction module, and the convolution output is denoted as P l ∈C M×T , l = 1, 2, where l denotes the link number and T denotes the signal length. The real and imaginary parts of P l are subjected to sigmoid activation function operations, respectively, and then spliced in the new dimension to obtain the mask output. The mask vector of the target source signal can be denoted as CRM:

[0192]

[0193] where s lr and s li denote the real and imaginary part representations of the lth link source signal, y lr and y li denote the real and imaginary part representations of the mixed signal, and each estimated source signal representation can be defined as Specifically:

[0194]

[0195] The mask output has the same dimension as the first layer complex convolution output in the complex domain feature extraction module, and the recovered signal high-dimensional representation is obtained by point multiplication between the two. Then, the recovered signal is subjected to a complex convolution operation to recover the estimated signal , which is denoted as:

[0196]

[0197] where D-conv denotes complex one-dimensional convolution, the signal real and imaginary parts are converted from M dimensions to 1 dimension, respectively, and k denotes the convolution kernel size.

[0198] The network training objective function adopts 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, which is specifically as follows:

[0199] The SI-SNR can be expressed as:

[0200]

[0201]

[0202]

[0203] where s and denote the source signal and the estimated signal. <·,·> denotes the two vector dot product, ||·||2 denotes the two norm, and MSE is defined as

[0204]

[0205] T denotes the signal length, and the SI-SNR and MSE are calculated for the real part and the imaginary part of the signal, respectively, i.e.:

[0206] SI-SNR ave = mean(SI-SNR real + SI-SNR imag )

[0207] MSE ave = mean(MSE real + MSE imag )

[0208] Finally, the loss function of the network for the lth link can be defined as:

[0209] Loss l = SI-SNR avel + aMSE avel , l = 1, 2

[0210] where a denotes the training coefficient, and a = 1 is set.

[0211] The data of the training set and the validation set are grouped and packaged, sent into the neural network, the network parameters are set, the Adam optimizer is used, the initial learning rate is 0.001, and when the loss of the validation set no longer decreases in the next 3 rounds of training, the learning rate is halved. A total of M rounds of training are used, and M is determined according to the actual situation. During the training, the gradient clipping with the maximum L2 norm of 5 is applied, the model training is stopped in advance, and when the loss of the validation set no longer decreases for 6 consecutive rounds, the training is stopped. After the model is trained on the training set and the validation set, the optimal parameter model is saved.

[0212] After the mixed signal of BPSK and QPSK is input into the network for separation, two separated waveforms are obtained, and the bit data of the signals can be recovered by demodulating the two separated waveforms respectively; by comparing the bit data with the bit data of the source signals, the separation error rate can be obtained, which can be used as a system separation performance evaluation index. Figure 5 The Pearson correlation coefficient of the separated signals of BPSK and QPSK and the source signals under different signal-to-noise ratios is shown, and it can be seen that the correlation coefficient gradually increases with the increase of the signal-to-noise ratio. Figure 6 The separation error rate of BPSK and QPSK under different signal-to-noise ratios is shown, and in order to facilitate testing, the signal is not added channel coding, and it can be seen that the dual-path complex time domain convolution network has good signal separation capability.

[0213] In conclusion, the application can realize effective separation of simultaneous and same frequency mixed signals under the condition of single antenna reception, and can improve the anti-interference ability of wireless communication; the application also provides a new communication multiplexing method, which provides a solution to relieve the contradiction between the increasingly tense spectrum resource and the demand for frequency use.

Claims

1. A method for wireless signal single-channel blind source separation based on deep neural networks, characterized in that, Comprising a training phase and an application phase, wherein: The training phase comprises the following steps: Step A1, creating a data set of time-frequency mixed signals, including communication signals and interference signals, and dividing the data set into a training set and a validation set; Step A2, building a double-path complex time domain convolution network, including a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module, as follows: The double-path complex time domain convolution network adopts two parallel links, corresponding to two separated signals, and each path includes 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-level complex domain coding structure, including two complex convolution layers and a complex normalization layer; the real part and the imaginary part of the first layer convolution output are spliced in a new dimension, and after processing by the complex normalization layer, the second layer complex convolution is performed; all operations of the complex domain feature extraction module are performed on the real part and the imaginary part of the complex value signal simultaneously; The complex domain signal separation module is composed of multiple complex depth convolution modules and a complex LSTM module, the complex LSTM module is arranged 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 skip connection path, the input and output of the complex depth convolution module are added in the real part and the imaginary part respectively each time, and enters the next complex depth convolution module, the dimension of the signal input and output remains unchanged; the convolution kernel adopts an inflation convolution form, the convolution kernel size of each complex depth convolution module presents an exponential increase, The convolution kernel size corresponding to the complex depth convolution module is ; each link has The complex depth convolution modules present a symmetrical structure on both sides of the complex LSTM. In the complex domain signal recovery module, the output of the complex domain signal separation module is first subjected to a complex convolution, and then the real part and the imaginary part are subjected to sigmoid activation respectively, and the mask matrix is obtained by splicing in the new dimension; the mask and the output of the first layer complex convolution in the complex domain feature extraction module have the same dimension, and the two are multiplied to obtain the high-dimensional representation of the recovered signal; finally, a complex convolution operation is performed again to reduce the dimension of the high-dimensional representation, that is, to recover the estimated signal; Step A3, creating a custom network training objective function; Step A4, based on the network training objective function, using the signal data set obtained in step A1, training and verifying the double-path complex time domain convolution network obtained in step A2, to obtain the trained double-path complex time domain convolution network; The application phase comprises the following steps: Step B1, obtaining the time-frequency mixed signals to be separated through a single antenna; Step B2, segmenting and organizing the time-frequency mixed signals to obtain complex mixed signals of fixed length; Step B3, using the trained double-path complex time domain convolution network to separate multiple signals from the complex mixed signals, to complete the blind separation of the mixed signals; Step B4, sending the separated signals obtained in step B3 into a demodulator for demodulation to recover the bit data of the separated signals.

2. The method of claim 1, wherein the deep neural network-based wireless signal single-channel blind source separation method is characterized by, The time-frequency mixed communication signal in step A1 is as follows: ； wherein is a time-frequency aliasing communication signal, denotes source signals, denotes the signal length, denotes unknown mixing coefficients, denotes a Gaussian white noise. 3.The deep neural network-based wireless signal single-channel blind source separation method according to claim 1, wherein, The construction of the data set in step A1 adopts any one of the following two methods: The first method: using computer software to generate communication signals and interference signals of multiple different modulation modes, the modulation mode being digital modulation or analog modulation, and the interference signals including 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 spectrum interference signals, wideband blocking interference, and sweep frequency interference; Based on the generated communication signals and interference signals, the communication signals and interference signals are randomly mixed to form mixed signals, and noise is added to the mixed signals to generate time-frequency mixed communication signals, with a signal-to-noise ratio SNR of -5dB to 20dB; according to the network training requirements, the data set of time-frequency mixed communication signals is divided into a training set and a validation set in a ratio of 80% and 20%; The second method: sampling actual communication equipment and interference equipment to obtain a data set of time-frequency mixed communication signals, and dividing the data set of time-frequency mixed communication signals into a training set and a validation set in a ratio of 80% and 20% according to the network training requirements.

4. The method of claim 1, wherein, The network training objective function in step A3 adopts a weighted combination of scale invariant signal-to-noise ratio SI-SNR and mean square error MSE as the model training loss function, as follows: The scale-invariant signal-to-noise ratio (SI-SNR) is expressed as: ； wherein and denote the source signal and the estimated signal, denotes the two vector dot products, denotes the two norm; The mean square error (MSE) is defined as ； where denotes the signal length, the SI-SNR and the MSE are computed for the real and imaginary part of the signal, respectively, and averaged, i.e. ； ； wherein, denotes the average scale-invariant signal-to-noise ratio, denotes the real scale-invariant signal-to-noise ratio, denotes the imaginary scale-invariant signal-to-noise ratio, denotes the average mean squared error, denotes the real mean squared error, denotes the imaginary mean squared error; Finally, the network's first loss function for a link is defined as: ； wherein, denotes a training coefficient, set ; denotes the road average scale-invariant signal-to-noise ratio, denotes the road average mean squared error. 5.The deep neural network-based wireless signal single-channel blind source separation method according to claim 1, wherein, Step A4 uses the signal data set obtained in step A1 to train and verify the double-path complex time domain convolution network obtained in step A2, as follows: The data of the training set and the verification set are grouped and packaged, and the sequence is combined into a mini batch, which is sent to the double-path complex time domain convolution network. The network parameters are set, and the Adam optimizer is used with an initial learning rate of 0.

001. When the loss of the verification set no longer decreases in three consecutive training rounds, the learning rate is halved. A total of M rounds of training are used, where M is a positive integer set according to the actual situation. Gradient clipping with a maximum L2 norm of 5 is applied during training. The model training uses early stopping, which stops training when the loss of the verification set no longer decreases for six consecutive rounds. After the model is trained on the training set and the verification set, the optimal parameter model is saved.

6. A wireless signal single-channel blind source separation apparatus based on a deep neural network, characterized by, The device is used to implement the deep neural network-based wireless signal single-channel blind source separation method of any one of claims 1-5. The device includes a training module and an application module, wherein: The training module includes a data set creation unit, a convolution network building unit, a target function creation unit, and a network training unit, wherein: The data set creation unit creates a data set of time-frequency mixed signals, including communication signals and interference signals, and divides the data set into a training set and a verification set; The convolution network building unit builds a double-path complex time domain convolution network, including a complex domain feature extraction module, a complex domain signal separation module, and a complex domain signal recovery module; The target function creation unit creates a custom network training target function; The network training unit uses the signal data set obtained by the data set creation unit to train and verify the double-path complex time domain convolution network obtained by the convolution network building unit based on the network training target function, and obtains the trained double-path complex time domain convolution network; The application module includes a signal acquisition unit, a time-frequency mixed signal arrangement unit, a blind separation unit, and a demodulation unit, wherein: The signal acquisition unit acquires the time-frequency mixed signals to be separated through a single antenna; The time-frequency mixed signal arrangement unit segments and arranges the time-frequency mixed signals to obtain complex mixed signals of a fixed length; The blind separation unit uses the trained double-path complex time domain convolution network to separate multiple signals from the complex mixed signals, completing the blind separation of the mixed signals; The demodulation unit sends the separated signals obtained by the blind separation unit to a demodulator for demodulation to recover the bit data of the separated signals.

Images