Single-channel aliased electromagnetic signal separation method and device based on deep learning

Abstract

The present invention discloses a method and device for separating single-channel aliased electromagnetic signals based on deep learning, including: obtaining a single-channel aliased electromagnetic signal; inputting the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, enabling the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enabling the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask. The present invention does not need to obtain prior knowledge of the signal, is suitable for blind signal separation, and extracts common features in a large amount of data through deep learning. Even when the signal is strongly interfered by noise, it still has a good separation effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and particularly relates to a method and device for separating single-channel aliased electromagnetic signals based on deep learning. Background Art

[0002] With the rapid development of modern society and the wide application of electronic technology, the electromagnetic environment has become increasingly complex. In a complex electromagnetic environment, various electronic devices and wireless communication systems are constantly increasing, and electromagnetic radiation is also increasing day by day, which brings great challenges to the operation of communication devices and communication quality. In addition, there are various interference sources and signal sources, such as power equipment, lightning, television, radar, etc., and these interference sources and signal sources will affect the normal operation of communication devices and signal quality.

[0003] To solve this problem, a method capable of separating individual signals from mixed signals needs to be found. Blind source separation can effectively solve the blind separation problem in the field of wireless communication, such as separating signals of different users from a complex wireless channel to achieve multi-user communication. Blind source separation can also be used in military communication and intelligence collection, such as separating different signal sources and information from complex electronic intelligence signals to obtain important intelligence and data. Currently, the main blind separation algorithms include the ICA (Independent Component Analysis) algorithm, the NMF (Nonnegative Matrix Factorization) algorithm, the SCA (Sparse Component Analysis) algorithm, etc., and these algorithms have good effects in solving the problem of mixed communication signals. Among them, the ICA algorithm can separate mixed signals into independent signal components by utilizing the independence characteristics between mixed signals; the NMF algorithm can decompose mixed signals into multiple non-negative signal components by using the method of matrix decomposition; the SCA algorithm obtains the sparse representation of signals through matrix decomposition or optimization solution of signals, and uses these sparse representations to achieve blind separation of mixed signals.

[0004] However, the above-mentioned blind separation algorithms usually process real-valued signals. In a digital communication system, most signals are complex-valued signals, and they are still complex-valued signals after mixing. If these algorithms are directly applied to a digital communication system, the characteristic information in the signals, such as the information jointly expressed by the real part and the imaginary part in complex data, cannot be effectively obtained, resulting in an unsatisfactory separation effect; especially when the signal is severely interfered, these algorithms will even fail and cannot complete the corresponding separation task. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a method and device for separating single-channel aliased electromagnetic signals based on deep learning. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] In a first aspect, the present invention provides a method for separating single-channel aliased electromagnetic signals based on deep learning, including:

[0007] Obtain a single-channel aliased electromagnetic signal;

[0008] Input the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal to obtain a signal encoding. After that, the separation network in the signal processing model obtains an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding. Further, the decoder outputs the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask.

[0009] In an embodiment of the present invention, the step of inputting the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal to obtain a signal encoding. After that, the separation network in the signal processing model obtains an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding. Further, the decoder outputs the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask includes:

[0010] Input the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal to obtain a signal encoding. After that, the separation network in the signal processing model obtains an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding. Further, the decoder calculates the Hadamard product of the signal encoding and the estimated mask to obtain a feature encoding, and then performs waveform reconstruction according to the feature encoding to output the separated signal of the single-channel aliased electromagnetic signal.

[0011] In an embodiment of the present invention, the signal processing model is trained according to the following steps:

[0012] Generate multiple source signals based on different modulation methods and carrier frequencies;

[0013] Mix two source signals with the same frequency and the same modulation or the same frequency and different modulations to obtain multiple observation signals;

[0014] Generate multiple training samples, each training sample including an observation signal and the two source signals that are mixed to generate the observation signal;

[0015] Input a preset number of training samples into the neural network model to be trained, and obtain two predicted separated signals corresponding to the observed signals in each training sample; the neural network model to be trained is a preset initial neural network model;

[0016] Use the two predicted separated signals corresponding to the observed signals in each training sample, the two source signals generating the observed signal, and a preset loss function to calculate the loss value;

[0017] Judge whether the neural network model to be trained converges according to the loss value; if it converges, the neural network model to be trained is a signal processing model that has been trained; otherwise, adjust the network parameters of the neural network model to be trained, and return to the step of inputting a preset number of training samples into the neural network model to be trained.

[0018] In an embodiment of the present invention, the step of mixing two source signals with the same frequency and the same modulation or the same frequency and different modulations to obtain a plurality of observed signals includes:

[0019]

[0020] where s i (t) represents the i-th source signal, a i represents the preset mixing coefficient corresponding to the i-th source signal, n(t) represents additive white Gaussian noise, i = 1, 2, and t represents time.

[0021] In an embodiment of the present invention, the preset loss function is the mean absolute error loss function.

[0022] In an embodiment of the present invention, the encoder includes a first convolutional layer, the separation network includes a layer normalization module, a second convolutional layer, four convolutional modules connected in sequence, a ReLU layer, a third convolutional layer, and a mask activation module, and the decoder includes a transposed convolutional module;

[0023] where the convolutional module includes a multi-stage convolutional unit connected in sequence, and from the first stage to the last stage, the dilation factor between adjacent two convolutional units increases exponentially.

[0024] In an embodiment of the present invention, the first convolutional layer, the second convolutional layer, the third convolutional layer, and the multi-stage convolutional unit all process their own input data according to the following formula:

[0025]

[0026] where Re{·} and Im{·} respectively represent the real part and the imaginary part of the input data of themselves, w(t) represents the weight value of themselves, Y(t) represents the output data of themselves, and t represents time.

[0027] In a second aspect, the present invention provides a single-channel aliased electromagnetic signal separation device based on deep learning, comprising:

[0028] An acquisition module, configured to acquire a single-channel aliased electromagnetic signal;

[0029] An input module, configured to input the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, the separation network in the signal processing model is enabled to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enable the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask.

[0030] In a third aspect, the present invention further provides an electronic device, comprising a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus;

[0031] The memory is used to store a computer program;

[0032] The processor is configured to implement the method steps described in the first aspect when executing the program stored on the memory.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] The present invention provides a method and device for separating single-channel aliased electromagnetic signals based on deep learning. This method does not require prior knowledge of single-channel aliased electromagnetic signals, is suitable for blind signal separation, and extracts common features from a large amount of data through deep learning. Even when the signal is strongly interfered by noise, it still has a good separation effect.

[0035] The following will further elaborate on the present invention in detail with reference to the accompanying drawings and embodiments. Description of the Drawings

[0036] Figure 1 is a flowchart of a method for separating single-channel aliased electromagnetic signals based on deep learning provided by an embodiment of the present invention;

[0037] Figure 2 is a schematic diagram of a method for separating single-channel aliased electromagnetic signals based on deep learning provided by an embodiment of the present invention;

[0038] Figure 3 is a schematic structural diagram of a data processing model provided by an embodiment of the present invention;

[0039] Figure 4 is a schematic structural diagram of a convolutional unit 1-D Conv provided by an embodiment of the present invention;

[0040] Figure 5a It is a schematic diagram of the separation effect of the observed signal with a signal-to-noise ratio of 6 dB provided by an embodiment of the present invention;

[0041] Figure 5b It is a schematic diagram of the separation effect of the observed signal with a signal-to-noise ratio of 8 dB provided by an embodiment of the present invention;

[0042] Figure 5c It is a schematic diagram of the separation effect of the observed signal with a signal-to-noise ratio of 10 dB provided by an embodiment of the present invention;

[0043] Figure 6a It is a schematic diagram of the separation effect of the mixed signal under the FSK+QPSK modulation method provided by an embodiment of the present invention;

[0044] Figure 6b It is a schematic diagram of the separation effect of the mixed signal under the QPSK+8PSK modulation method provided by an embodiment of the present invention;

[0045] Figure 6c It is a schematic diagram of the separation effect of the mixed signal under the QPSK+8PSK modulation method provided by an embodiment of the present invention;

[0046] Figure 7a It is a correlation coefficient curve graph of the separation result provided by an embodiment of the present invention;

[0047] Figure 7b It is a SER curve graph under different signal-to-noise ratios provided by an embodiment of the present invention;

[0048] Figure 8 The present invention provides a single-channel aliased electromagnetic signal separation device based on deep learning provided by an embodiment of the present invention;

[0049] Figure 9 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Specific embodiments

[0050] The following further describes the present invention in detail with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0051] Figure 1 It is a flowchart of a method for separating single-channel aliased electromagnetic signals based on deep learning provided by an embodiment of the present invention, Figure 2 It is a schematic diagram of a method for separating single-channel aliased electromagnetic signals based on deep learning provided by an embodiment of the present invention. As Figure 1-2 shown, an embodiment of the present invention provides a method for separating single-channel aliased electromagnetic signals based on deep learning, including:

[0052] S1. Obtain a single-channel aliased electromagnetic signal;

[0053] S2. Input the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, enable the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding. Further, enable the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask.

[0054] Optionally, in step S2, the step of inputting the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal, obtaining the signal encoding, enabling the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enabling the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask includes:

[0055] Input the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, enable the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding. Further, enable the decoder to calculate the Hadamard product of the signal encoding and the estimated mask to obtain the feature encoding, and then perform waveform reconstruction according to the feature encoding to output the separated signal of the single-channel aliased electromagnetic signal.

[0056] In this embodiment, the pre-trained signal processing model consists of three parts: an encoder, a separation network, and a decoder. The encoder regards the input single-channel aliased electromagnetic signal as composed of L segments with a length of N, and each segment is represented as where k = 1, 2,..., L. After being processed by the encoder, an N-dimensional abstract encoded feature can be obtained It can be expressed by matrix multiplication as:

[0057]

[0058] where, contains N vectors, representing the encoder, and the length of each vector is L. represents the non-linear activation function. Since the input single-channel aliased electromagnetic signal to the encoder is complex, a complex activation function needs to be used to process it. The complex activation function can perform non-linear mapping on complex data. The complex activation function used in the present invention is The mathematical expression is:

[0059]

[0060] Further, after being processed by the encoder, the obtained signal-coded input is separated by the separation network, and the separation network calculates the estimated mask based on the signal coding. The function of the decoder is the opposite of that of the encoder. The decoder calculates the feature coding of the single-channel aliased electromagnetic signal according to the signal coding and the estimated mask:

[0061]

[0062] where denotes the Hadamard product operation, and s mask denotes the estimated mask.

[0063] The decoder can reconstruct the waveform of the single-channel aliased electromagnetic signal according to the abstract features contained in the feature coding:

[0064]

[0065] where denotes the reconstructed signal, h decoder denotes the decoder, which has N vectors of length L. These reconstructed segments are added together to obtain the final waveform.

[0066] Optionally, the above signal processing model is obtained by training according to the following steps:

[0067] Generate multiple source signals based on different modulation methods and carrier frequencies;

[0068] Mix two source signals with the same frequency and the same modulation or the same frequency but different modulations to obtain multiple observed signals;

[0069] Generate multiple training samples, each of which includes an observed signal and the two source signals that are mixed to generate the observed signal;

[0070] Input a preset number of training samples into the neural network model to be trained, and obtain two predicted separated signals corresponding to the observed signal in each training sample; the neural network model to be trained is a preset initial neural network model;

[0071] Use the two predicted separated signals corresponding to the observed signal in each training sample, the two source signals that generate the observed signal, and a preset loss function to calculate the loss value;

[0072] Judge whether the neural network model to be trained converges according to the loss value; if it converges, the neural network model to be trained is the trained signal processing model; otherwise, adjust the network parameters of the neural network model to be trained, and return to the step of inputting a preset number of training samples into the neural network model to be trained.

[0073] Specifically, before training the signal processing model, data for training and testing is first generated. In this embodiment, the number of symbols N for generating data is 128, the oversampling multiple is 8, the modulation method can be FSK, 16QAM or MPSK, the shaping filter type is a root raised cosine filter, and the roll-off factor is 0.6. In this way, multiple source signals can be generated; since the data processing model performs blind separation on the mixed signals in a single channel, after generating multiple source signals, it is also necessary to mix the source signals through a mixing matrix to obtain the observed signals.

[0074] Optionally, the step of mixing two source signals with the same frequency and the same modulation or the same frequency but different modulations to obtain multiple observed signals includes:

[0075]

[0076] where s i (t) represents the i-th source signal, a i represents the preset mixing coefficient corresponding to the i-th source signal, n(t) represents additive white Gaussian noise, i = 1, 2, and t represents time.

[0077] It should be understood that the observed signal is composed of two source signals with the same frequency and the same modulation or the same frequency but different modulations, and there is a time delay between these two source signals. The mixing method of the two is shown as x(t) = As(t) + n(t), and this mathematical expression can be expanded as:

[0078]

[0079] After the generation of the observed signal is completed, the observed signal x(t) and the source signals s1(t), s2(t) used to generate x(t) are jointly stored in a data set as a group of training samples, so as to obtain the data set required for training and testing: D = {(x 1 (t), s 1 (t)), (x 2 (t), s 2 (t)), …, (x i (t), s i (t))}, and the ratio of training, validation, and testing is 8:1:1.

[0080] Figure 3 is a schematic structural diagram of the data processing model provided by the embodiment of the present invention. As Figure 3 shown, in this embodiment, the encoder includes a first convolutional layer, the separation network includes a layer normalization module, a second convolutional layer, four convolutional modules connected in sequence, a ReLU layer, a third convolutional layer, and a mask activation module, and the decoder includes a transposed convolutional module;

[0081] Among them, the convolution module includes multiple levels of convolution units connected in sequence, and from the first level to the last level, the dilation factor between adjacent two convolution units increases exponentially.

[0082] Specifically, the encoder includes a one-dimensional first convolutional layer. The structure of the separation network is shown in the figure. The separation network includes a layer normalization module, a second convolutional layer, four 1-D convolution modules connected in sequence, a ReLU layer, a third convolutional layer, and a mask activation module. Among them, each 1-D convolution module contains multiple convolution units 1-D Conv. In order to enable the network to better capture the long-term dependencies in the input signal, the convolution unit 1-DConv uses the dilated convolution technique, and within the same 1-D convolution module, the dilation factor between adjacent two convolution units 1-D Conv increases exponentially. The depth stacking of the dilated convolution enables the separation network to have the ability to extract long-term dependencies. Before the dilated convolution operation, padding is required for the input to satisfy that the input length and output length of the convolution unit are the same. After passing through the separation network, an estimated mask for the source signal can be obtained.

[0083] The specific structure of the convolution unit 1-D Conv is as Figure 4 shown. Each 1-D Conv unit consists of a 1×1 convolution operation and a DW (Depthwise Convolution) operation, and nonlinear activation and normalization are required between every two convolution operations. As can be seen from Figure 4 , there are two 1×1Conv layers at the output position. One of them serves as the output of the 1-D convolution unit, and the other serves as a skip connection. The 1×1 convolution layer can also be called PW (Pointwise Convolution).

[0084] In addition, please continue to refer to Figure 4 , there is a residual connection from the input to the output in the convolution unit 1-D Conv. The residual connection can effectively solve the problem of network degradation caused by the increase in network depth. For a deep network, its weight matrix is a matrix with a very high dimension, but many of these dimensions do not preserve information, which also leads to a decline in the network's expressive ability. The applicant found during the research process that this is due to the symmetry of the network, and the residual connection can effectively solve this problem. Moreover, since the residual connection connects different layers in the network, the residual connection also enhances the propagation of gradients.

[0085] In this embodiment, stacking 1-D convolution modules multiple times can enhance the network's ability to extract deep features, thereby improving the separation performance of the network. In each 1-D convolution module, each convolutional unit 1-D Conv has an exponentially increasing dilation factor, so the separation network obtains the ability to extract multi-scale features in the input signal. Finally, these features are fused with multi-layer features through skip connections.

[0086] It should be noted that traditional neural networks mainly process real-number data. Therefore, in order to enable the above neural network to process complex-number data in the communication field, this embodiment uses the principle of complex convolution to transform the convolutional layer, so that the neural network can correctly map the features of complex-number data.

[0087] The first convolutional layer, the second convolutional layer, the third convolutional layer, and the multi-stage convolutional unit all process their own input complex-number data according to the following formula:

[0088]

[0089] Furthermore, it can be represented by matrix operations:

[0090]

[0091] Among them, Re{·} and Im{·} respectively represent the real part and the imaginary part of the input data, w(t) represents the self-weight value, Y(t) represents the self-output data, and t represents the time.

[0092] During the training process, the loss function adopted in this embodiment is MAE (Mean Absolute Error). MAE measures the convergence of the neural network to be trained by calculating the absolute value difference between the two predicted separation signals output by the neural network model to be trained and the two source signals of the generated observation signals in the training samples. This loss function is more sensitive to outliers.

[0093] Optionally, the Adam algorithm is used in this embodiment to optimize the loss function of the neural network to be trained, and the model parameters of the neural network to be trained are updated layer by layer through the backpropagation algorithm. In addition, in order to train a neural network with better performance more quickly, the present invention further introduces a learning rate decay mechanism and an early stopping strategy. Specifically, the input of the neural network to be trained is the training sample in the training set D train ={(s i (n),x i (n)),i=1,2,…,N}, the learning rate is initialized to α = 0.01, the preset learning rate decay coefficient γ = 0.1, and from the training set D trainExtract m training samples from it and input them into the neural network to be trained, and calculate the loss value. If the loss value of the preset function does not decrease for two consecutive times, then let α = γα, and return to the step of extracting m training samples and inputting them into the neural network to be trained. Stop training until the loss value of the preset loss function does not decrease for five consecutive times or the number of iterations reaches the preset value, and obtain a trained signal processing model.

[0094] In order to better quantify the separation effect of the signal processing model, the present invention will use two evaluation indicators to evaluate the separation result of the single-channel aliased electromagnetic signal. First is the correlation coefficient, and its mathematical expression is as follows:

[0095]

[0096] The correlation coefficient can well reflect the similarity between the separation result y i (t) and its corresponding actual source signal s j (t). It can be seen from the above formula that 0 ≤ ξ ij ≤ 1. The closer ξ ij is to 1, the closer y i (t) is to s j (t); on the contrary, the closer ξ ij is to 0, the stronger the independence of the two. When y i (t) = λs j (t), ξ ij = 1, indicating that the obtained separation result is exactly the same as the source signal, only with a difference in amplitude. Here, λ represents any real number.

[0097] In the field of communication, the goal of the blind separation task is to obtain the original sequence of the transmitted signal rather than the waveform. Therefore, it is also necessary to use SER (Symbol Error Rate) to measure the accuracy of the separation result. SER represents the error ratio between the separated signal and the transmitted signal, and the formula is as follows:

[0098]

[0099] Among them, N e represents the number of incorrect symbols received at the receiving end, and N s represents the total number of symbols transmitted at the transmitting end.

[0100] Next, the single-channel aliased electromagnetic signal separation method based on deep learning provided by the present invention will be further described through simulation experiments.

[0101] In this embodiment, observation signals are formed by mixing two-by-two multiple single-channel source signals such as FSK, MPSK, and 16QAM. Among them, the sampling frequency of the FSK signal is 480 kHz and the frequency interval is 60 kHz. For the MPSK and 16QAM signals, the symbol rate is 60 kHz and the sampling frequency is also 480 kHz. In addition, shaping filtering is required, and the roll-off coefficient of the shaping filtering is 0.6, and the number of sampling points for each data is 8. After obtaining the source signal, the observation signal is obtained by adding Gaussian white noise.

[0102] Simulation 1: Simulate the separation effect of the mixed source signals of the same modulation method. The simulation results at different signal-to-noise ratios are as Figures 5a-5c shown, where Figures 5a-5c the modulation methods used are all FSK+FSK, Figure 5a , Figure 5b , Figure 5c and the signal-to-noise ratios of Figures 5a-5c are 6 dB, 8 dB, and 10 dB respectively. It can be seen from

[0103] that the waveforms of the predicted separated signals and the corresponding source signals basically coincide, indicating that the signal processing model provided by the present invention can effectively complete the separation of the same modulation signals.

[0103] Simulation 2: Simulate the separation effect of the mixed source signals of different modulation methods. The simulation results under different modulation methods are as Figures 6a-6c shown, where Figure 6a the modulation method of the mixed signal used is FSK+QPSK, Figures 6b-6c and the modulation methods of the mixed signals used are all QPSK+8PSK. It can be seen from Figures 6a-6c that the waveforms of the predicted separated signals and the corresponding source signals basically coincide, indicating that the signal processing model provided by the present invention can effectively complete the separation of different modulation signals.

[0104] Figure 7a is the correlation coefficient curve graph of the separation result provided by the embodiment of the present invention, Figure 7b and Figure 7a is the SER curve graph under different signal-to-noise ratios provided by the embodiment of the present invention. In order to be able to effectively quantify the separation effect, at different signal-to-noise ratios, the correlation coefficient curve of the separation is as Figure 7a shown. It can be seen that the signal processing model can well complete the separation of the mixed signals when SNR>5 dB. In addition, the symbol error rate is also a very important indicator in the communication field. It can be seen from Figure 7b that the network has a low symbol error rate when SNR>5 dB.

[0105] Compared with the existing methods, the single-channel aliased electromagnetic signal separation method based on deep learning provided by the present invention requires less memory and has a faster separation speed.

[0106] Table 1

[0107] Network name Network scale Operating speed Signal processing model 1.65M 66 s / epoch LSTM-TasNet 10.2M 375 s / epoch

[0108] Combined with Figure 7 and Table 1, it can be seen that although the separation performance of the present invention is not much different from that of LSTM-TasNet, the operation speed is 5-6 times faster than that of LSTM-TasNet, and it can well adapt to the low-latency communication scenario.

[0109] In addition, the SER curves of the present invention and LSTM-TasNet at different signal-to-noise ratios are as Figure 8 shown. It can be seen that the SER of the signal separated by FastICA is very high, and the signal cannot be effectively separated, while CSSN and LSTM-TasNet can well complete the task.

[0110] Based on the above simulation results and analysis, the present invention has high separation performance in the case of no channel prior information and SNR>5dB, and can well complete the signal blind separation task, so it is applicable to the communication environment with low latency requirements.

[0111] Figure 8 is a single-channel aliased electromagnetic signal separation device based on deep learning provided by an embodiment of the present invention. As Figure 8 shown, an embodiment of the present invention also provides a single-channel aliased electromagnetic signal separation device based on deep learning, including:

[0112] An acquisition module 810, configured to acquire a single-channel aliased electromagnetic signal;

[0113] An input module 820, configured to input the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, the separation network in the signal processing model is based on the signal encoding. Obtain an estimated mask of the single-channel aliased electromagnetic signal, and further enable the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask.

[0114] An embodiment of the present invention also provides an electronic device, as Figure 9 shown, including a processor 901, a communication interface 902, a memory 903, and a communication bus 904. Among them, the processor 901, the communication interface 902, and the memory 903 communicate with each other through the communication bus 904.

[0115] The memory 903 is used to store a computer program;

[0116] The processor 901, when executing the program stored in the memory 903, implements the following steps:

[0117] Obtain a single-channel aliased electromagnetic signal;

[0118] Input the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, enable the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enable the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask.

[0119] The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, only a thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0120] The communication interface is used for communication between the above electronic device and other devices.

[0121] The memory may include a Random Access Memory (RAM), and may also include a Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0122] The above-mentioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0123] The method provided by the embodiments of the present invention can be applied to an electronic device. Specifically, the electronic device can be: a desktop computer, a portable computer, a smart mobile terminal, a server, etc. There is no limitation here. Any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0124] For the device / electronic device / storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments.

[0125] It should be noted that the device, electronic device, and storage medium of the embodiments of the present invention are respectively the device, electronic device, and storage medium that apply the above-mentioned method for separating single-channel aliased electromagnetic signals based on deep learning. Then, all the embodiments of the above-mentioned method for separating single-channel aliased electromagnetic signals based on deep learning are applicable to the device, electronic device, and storage medium, and can achieve the same or similar beneficial effects.

[0126] It can be seen from the above embodiments that the beneficial effects of the present invention are as follows:

[0127] The present invention provides a method and device for separating single-channel aliased electromagnetic signals based on deep learning. This method does not need to obtain prior knowledge of the single-channel aliased electromagnetic signals, is suitable for blind signal separation, and extracts common features from a large number of data through deep learning. Even when the signal is strongly interfered by noise, it still has a good separation effect.

[0128] In the description of the present invention, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more, unless otherwise specifically defined.

[0129] The description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0130] Although the present application has been described in connection with various embodiments, those skilled in the art will recognize other variations of the disclosed embodiments while practicing the claimed present application by viewing the drawings, the disclosure, and the appended claims.

[0131] The above is a further detailed description of the present invention in connection with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited only to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for separating single-channel aliased electromagnetic signals based on deep learning, characterized in that, Including: Obtaining a single-channel aliased electromagnetic signal; Inputting the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, enabling the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enabling the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask; The encoder includes a first convolutional layer, the separation network includes a layer normalization module, a second convolutional layer, four sequentially connected convolutional modules, a ReLU layer, a third convolutional layer, and a mask activation module, and the decoder includes a transposed convolutional module; Wherein, the convolutional module includes sequentially connected multi-level convolutional units, and from the first level to the last level, the dilation factor between two adjacent convolutional units increases exponentially; The first convolutional layer, the second convolutional layer, the third convolutional layer, and the multi-level convolutional units all process their input data according to the following formula: Wherein, Re{·} and Im{·} respectively represent the real part and the imaginary part of the input data of themselves, w(t) represents the weight value of themselves, Y(t) represents the output data of themselves, and t represents the time.

2. The method for separating single-channel aliased electromagnetic signals based on deep learning according to claim 1, wherein The step of inputting the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, enabling the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enabling the decoder to output the separated signal of the single-channel aliased electromagnetic signal includes: Inputting the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal. After obtaining the signal encoding, enabling the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enabling the decoder to calculate the Hadamard product of the signal encoding and the estimated mask to obtain a feature encoding, and then perform waveform reconstruction according to the feature encoding to output the separated signal of the single-channel aliased electromagnetic signal.

3. The single-channel aliased electromagnetic signal separation method based on deep learning according to claim 1, wherein The signal processing model is obtained by training according to the following steps: Generating a plurality of source signals based on different modulation methods and carrier frequencies; Mixing two source signals with the same frequency and the same modulation or the same frequency but different modulation to obtain a plurality of observed signals; Generating a plurality of training samples, each training sample including an observed signal and the two source signals that are mixed to generate the observed signal; Inputting a preset number of training samples into a neural network model to be trained to obtain two predicted separated signals corresponding to the observed signal in each training sample; the neural network model to be trained is a preset initial neural network model; Calculating a loss value using the two predicted separated signals corresponding to the observed signal in each training sample, the two source signals that generate the observed signal, and a preset loss function; Determine whether the neural network model to be trained converges according to the loss value; if it converges, the neural network model to be trained is the trained signal processing model; otherwise, adjust the network parameters of the neural network model to be trained and return to the step of inputting a preset number of training samples into the neural network model to be trained.

4. The single-channel aliased electromagnetic signal separation method based on deep learning according to claim 3, wherein, The step of mixing two source signals with the same frequency and the same modulation or the same frequency but different modulations to obtain a plurality of observed signals includes: Among them, s i (t) represents the i-th source signal, a i represents the preset mixing coefficient corresponding to the i-th source signal, n(t) represents additive white Gaussian noise, i = 1, 2, and t represents time.

5. The method for separating single-channel aliased electromagnetic signals based on deep learning according to claim 3, characterized in that, The preset loss function is the mean absolute error loss function.

6. A single-channel aliased electromagnetic signal separation device based on deep learning, characterized in that, Including: An acquisition module for acquiring a single-channel aliased electromagnetic signal; An input module for inputting the single-channel aliased electromagnetic signal into a pre-trained signal processing model, so that the encoder in the signal processing model encodes the single-channel aliased electromagnetic signal to obtain a signal encoding, and then enables the separation network in the signal processing model to obtain an estimated mask of the single-channel aliased electromagnetic signal based on the signal encoding, and further enables the decoder to output the separated signal of the single-channel aliased electromagnetic signal based on the signal encoding and the estimated mask; The encoder includes a first convolutional layer, the separation network includes a layer normalization module, a second convolutional layer, four sequentially connected convolutional modules, a ReLU layer, a third convolutional layer, and a mask activation module, and the decoder includes a transposed convolutional module; Wherein, the convolutional module includes a plurality of sequentially connected convolutional units, and from the first stage to the last stage, the dilation factor between two adjacent convolutional units increases exponentially; The first convolutional layer, the second convolutional layer, the third convolutional layer, and the plurality of convolutional units all process their own input data according to the following formula: Where Re{·} and Im{·} respectively represent the real part and the imaginary part of their own input data, w(t) represents their own weight value, Y(t) represents their own output data, and t represents the time.

7. An electronic device, characterized in that, Including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory complete mutual communication through the communication bus; The memory is used for storing a computer program; The processor is used to implement the method steps described in any one of claims 1-5 when executing the program stored on the memory.

Images