Single-channel time-frequency aliasing interference signal elimination method and system based on deep learning

Abstract

The present invention discloses a method and system for eliminating single-channel time-frequency aliasing interference signals based on deep learning. The method is as follows: step one, obtaining a plurality of mixed signals received by a single channel, each mixed signal containing a target signal and its same-frequency interference signal; step two, constructing an LSTM neural network including a dual attention mechanism suitable for signal separation; step three, using the mixed signals received by the plurality of single channels as network input data, training the constructed LSTM neural network, and obtaining a neural network model for eliminating single-channel time-frequency aliasing interference signals; step four, using the actual mixed signal with the target signal as the input of the neural network model for eliminating single-channel time-frequency aliasing interference signals, thereby outputting the required interference-free target signal. The present invention does not require precise parameter estimation of the target signal and the interference signal, is plug-and-play, and learning and training are not limited by signal style, and has wide applicability.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing technology, and specifically relates to a single-channel time-frequency aliasing interference signal elimination method and system based on deep learning. Background Art

[0002] In recent years, human society has entered a period of rapid development in the information age. In wireless communication systems, with the continuous deployment and use of various electromagnetic devices, the communication environment is becoming increasingly complex, and the electromagnetic spectrum is becoming increasingly crowded. Observation signals received by non-cooperative third parties are often mixed signals with noise or aliasing of co-frequency signals. In such cases, complex algorithms are required to extract the characteristic information of the target signal. Especially in situations with low signal-to-noise ratios or severe aliasing of co-frequency signals, the target signal is severely corrupted. Directly processing the mixed signal in these situations does not effectively resolve the target signal. Therefore, research on blind signal processing techniques to extract the target signal from the mixed signal is crucial. The current advantages of single-channel co-channel interference elimination are low cost, requiring only one signal transmission channel, and high operability, facilitating flexible future functional upgrades. Therefore, research on single-channel co-channel interference elimination is of practical significance. However, due to the limited prior information available in single-channel conditions, no effective solution currently exists. Based on this, the present invention proposes a single-channel time-frequency aliasing interference signal elimination method and system based on deep learning. Summary of the Invention

[0003] In response to the above-mentioned defects in the prior art, the present invention provides a single-channel time-frequency aliasing interference signal elimination method and system based on deep learning. The present invention does not require precise parameter estimation of the target signal and the interference signal, is plug-and-play, and learning and training are not restricted by signal style, and has wide applicability.

[0004] The present invention adopts the following technical solutions:

[0005] In one aspect, the present invention provides a single-channel time-frequency aliasing interference signal elimination method based on deep learning, comprising the following steps:

[0006] Step 1: Acquire multiple mixed signals received by single channels as network input data, where each mixed signal used as network input data contains a target signal and its co-frequency interference signal;

[0007] Step 2: Build an LSTM neural network with a dual attention mechanism suitable for signal separation;

[0008] Step 3: Using the mixed signals received by the above multiple single channels as network input data, the constructed LSTM neural network with a dual attention mechanism suitable for signal separation is trained to obtain a neural network model for eliminating single-channel time-frequency aliasing interference signals;

[0009] Step 4: The actual mixed signal with the target signal is used as the input of the neural network model for eliminating the single-channel time-frequency aliasing interference signal, thereby outputting the required interference-free target signal.

[0010] Preferably, in step 1, obtaining a mixed signal received by multiple single channels is achieved by the following operations:

[0011] The target communication signal and the co-frequency interference signal are sampled respectively. The modulation patterns of the target communication signal and the co-frequency interference signal can be selected according to actual needs to obtain N target signal sampling sequences x = [x1, x2, ..., x M ],in i=1,2,...,M, M is the number of sampling points; obtain N co-frequency interference signal sampling sequences y=[y1,y2,...,y M ],in i=1, 2, ..., M, where M is the number of sampling points. Thus, the samples of the target communication signal and the co-frequency interference signal are obtained.

[0012] The co-frequency interference signal is processed, and z=x+a*y is set, where a is a random number greater than 0, to obtain a sample set of N mixed signals.

[0013] Cross-merge the I-path data and Q-path data of the mixed signal z into a column vector (I-path represents the real part of the signal, Q-path represents the imaginary part of the signal), and construct a matrix with 2M rows and 1 column This is a mixed signal superimposed with the same-frequency interference signal, which serves as the data input for neural network training.

[0014] Preferably, in step 2, an LSTM neural network including a dual attention mechanism suitable for signal separation is constructed, which is implemented by the following operations:

[0015] Given n driving sequences (input features):

[0016] X=(x 1 ,x 2 ,...,x n ) T =(x1,x2,...,x T )∈R n*T

[0017] Where T represents the time step and n represents the dimension of the input feature. Through the input attention mechanism, the encoder generates new input The encoder architecture uses an LSTM. In the decoder, another LSTM network layer decodes the encoder information and selects the corresponding hidden layer states through a temporal attention mechanism, thereby obtaining the final, uninterrupted target signal sequence. The hidden layer dimension of the encoder and decoder is 64, the model time step is T = 10, the learning rate is set to 0.01, and the number of hidden units is set to 128.

[0018] The specific implementation process of the input attention mechanism in the LSTM neural network encoder of the dual attention mechanism is as follows:

[0019] Given input X=(x1,x2,...,x T )∈R n*T , x t ∈R n , at time t, the encoder converts x t Mapping to h t :f1(h t-1 , x t ), h t ∈R m represents the state of the encoder hidden layer at time t, m represents the dimension of the hidden layer, and f(.) is a nonlinear activation function. Construct the input attention layer and convert the previous hidden layer state h t-1 And the cell state of the LSTM unit (cell state can also be called the cell state of LSTM) state s t-1 As input to this layer we get:

[0020]

[0021] Among them, v e ∈R T , W e ∈R T*2m , U e ∈R T*T are parameters that need to be learned.

[0022] Use a one-dimensional Gaussian function to balance the weights:

[0023] make Where x represents μ1 represents the mean, σ1 2 Represents variance.

[0024] The calculation method of mean is: ∑ represents the summation symbol, x i represents the i-th value, and n represents the total number of values.

[0025] How to calculate variance: x irepresents the i-th value, μ1 represents the mean, and n represents the total number of values.

[0026] Input the output of the attention layer Input to the softmax layer to get To ensure that the sum of all attention weights is 1, Indicates the importance of the kth input feature at time t. After obtaining the attention weight, the driving sequence can be adaptively extracted At this time, the state of the updated hidden layer is

[0027] The specific implementation process of the temporal attention mechanism in the LSTM neural network decoder of the dual attention mechanism is as follows:

[0028] Similar to the attention layer in the encoder, the attention layer in the decoder also needs to store the previous hidden layer state d t-1 and the cell state s′ of the LSTM unit t-1 The output of this layer is obtained as the input of this layer:

[0029]

[0030] where v d ∈R m , W d ∈R m*2p , U d ∈R m*m are parameters that need to be learned.

[0031] Use a one-dimensional Gaussian function to balance the weights:

[0032] make Where y represents μ2 represents the mean, σ2 2 Represents variance.

[0033] The calculation method of mean is: ∑ represents the summation symbol, y i represents the i-th value, and n represents the total number of values.

[0034] How to calculate variance: y i represents the i-th value, μ2 represents the mean, and n represents the total number of values.

[0035] Through the softmax layer, we get the i-th encoder hidden state h i Importance of the final result The decoder sums all the encoder hidden states according to the weights to obtain the vector Note t It is different at different time steps.

[0036] After obtaining the vector, combine it with the target sequence to obtain:

[0037]

[0038] Among them, w T , is the parameter to be learned. Update decoder hidden state f2 is a nonlinear activation function. The vector output obtained after passing through the decoder is the target signal output by the model.

[0039] Preferably, in step three, the constructed LSTM neural network including a dual attention mechanism suitable for signal separation is trained, and the following operation is adopted to implement it: using the mixed signals received by the above-mentioned multiple single channels as network input data, the constructed LSTM neural network including a dual attention mechanism suitable for signal separation is trained to obtain a neural network model for eliminating single-channel time-frequency aliasing interference signals. Preferably, in step four, the actual mixed signal with the target signal is used as the input of the neural network model for eliminating single-channel time-frequency aliasing interference signals, thereby outputting the required interference-free target signal, and the following operation is adopted to implement it:

[0040] The actual data is sampled to obtain a sequence s(n), where n = 0, 1, 2, ..., M-1. The I-channel data and Q-channel data of the sequence s(n) are cross-merged into a 2M-row column vector. The processed column vector is used as the input of the trained neural network, and the output obtained is the interference-free target communication signal.

[0041] On the other hand, the present invention also discloses a single-channel time-frequency aliasing interference signal elimination system based on deep learning, which includes the following modules based on the above method:

[0042] Data acquisition module: acquires multiple mixed signals received by single channels as network input data. Each mixed signal used as network input data contains the target signal and its co-frequency interference signal.

[0043] Network building module: used to build an LSTM neural network with a dual attention mechanism suitable for signal separation;

[0044] Network training module: used to use the mixed signals received by the above multiple single channels as network input data to train the constructed LSTM neural network with a dual attention mechanism suitable for signal separation, and obtain a neural network model for eliminating single-channel time-frequency aliasing interference signals;

[0045] Target signal output module: used to take the actual mixed signal with the target signal as the input of the neural network model used for single-channel time-frequency aliasing interference signal elimination, thereby outputting the required interference-free target signal.

[0046] The present invention does not require precise parameter estimation of the target signal and the interference signal, is plug-and-play, and learning and training are not restricted by signal patterns, thus having wide applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flowchart of a method for eliminating single-channel time-frequency aliasing interference signals based on deep learning according to an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of a mixed signal generation process according to an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of the input attention mechanism flow in an LSTM neural network including a dual attention mechanism suitable for signal separation according to one embodiment of the present invention;

[0050] Figure 4 A schematic diagram of a flow chart of a temporal attention mechanism in an LSTM neural network including a dual attention mechanism suitable for signal separation according to an embodiment of the present invention;

[0051] Figure 5 This is a block diagram of a single-channel time-frequency aliasing interference signal elimination system based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below with reference to specific embodiments, and the specific methods and approaches of the present invention will be described in detail with reference to the accompanying drawings. It should be noted that the embodiments described below are only a part of the present invention and do not represent all embodiments.

[0053] Figure 1 This is a flow chart of a method for eliminating single-channel time-frequency aliasing interference signals based on deep learning according to an embodiment of the present invention. Figure 1 , the specific steps of this embodiment are as follows:

[0054] S101, obtaining multiple mixed signals received by single channels as network input data, each mixed signal used as network input data includes a target signal and a co-channel interference signal thereof.

[0055] In step S101, a mixed signal includes a target signal and its co-frequency interference signal. The target signal and the interference signal can be any digital modulation signal, including BPSK, QPSK, 8PSK, QAM, 2FSK, and the like. To ensure that the target signal and the interference signal are co-frequency, the time and frequency of both signals are the same when the signals are collected. Therefore, the mixed signal data used in this embodiment ensures the versatility of the embodiment and is applicable to eliminating co-frequency interference between any two different digital modulation signals.

[0056] S102: Construct an LSTM neural network suitable for signal separation that includes a dual attention mechanism. The neural network constructed in this embodiment fully utilizes the advantages of recurrent neural networks, capable of memorizing input information from the previous time, processing inputs of arbitrary length, maintaining a model shape that does not change with increasing input length, and sharing weights over time. The use of the dual attention mechanism can capture more contextual and spatial information, thereby improving the accuracy and inference speed of the model.

[0057] S103 uses the mixed signals received from the multiple single channels as network input data to train the constructed LSTM neural network with a dual attention mechanism for signal separation, thereby obtaining a neural network model for eliminating single-channel time-frequency aliasing interference signals. The process of training the LSTM neural network with a dual attention mechanism is the process of going from a mixed signal with interference to a target signal without interference.

[0058] S104: The actually sampled mixed signal with the target signal is used as an input of a neural network model for eliminating the single-channel time-frequency aliasing interference signal, thereby outputting the required interference-free target signal.

[0059] The following describes each step in more detail.

[0060] In step S101 of this embodiment, the specific implementation method of generating the mixed signal is as follows: Figure 2 .

[0061] The target communication signal and the co-frequency interference signal are sampled respectively. The modulation patterns of the target communication signal and the co-frequency interference signal can be selected according to actual needs to obtain N target signal sampling sequences x = [x1, x2, ..., x M ],in i=1,2,...,M, M is the number of sampling points; obtain N co-frequency interference signal sampling sequences y=[y1,y2,...,y M ],in i=1, 2, ..., M, where M is the number of sampling points. Thus, the samples of the target communication signal and the co-frequency interference signal are obtained.

[0062] The co-channel interference signal is processed, and z = x + a * y, where a is a random number greater than 0, to obtain a sample set of N mixed signals. When reproducing this embodiment, the value of a can also be adjusted based on the interference-to-signal ratio between the target signal and the co-channel interference signal, further exploring the effectiveness of eliminating single-channel time-frequency aliasing interference signals at different interference-to-signal ratios.

[0063] Cross-merge the I-channel data and Q-channel data of the mixed signal z into a column vector, and construct a matrix with 2M rows and 1 column This is a mixed signal superimposed with the same-frequency interference signal, which serves as the data input for neural network training.

[0064] In step S102 of this embodiment, the specific model structure of the LSTM neural network with dual attention mechanism is as follows: Figure 3 、 Figure 4 As shown. Given n driving sequences (input features): X=(x 1 ,x 2 ,...,x n ) T =(x1,x2,...,x T )∈R n*T , T represents the time step, n represents the dimension of the input feature. Through the input attention mechanism, the encoder generates new input In this embodiment, the encoder architecture uses an LSTM. In the decoder, another LSTM network layer is used to decode the encoder information and select the corresponding hidden layer states through a temporal attention mechanism to obtain the final interference-free target signal sequence. In this embodiment, the hidden layer dimension of the encoder and decoder is 64, the model time step T = 10, the learning rate is set to 0.01, and the number of hidden units is set to 128.

[0065] The input attention mechanism in the LSTM neural network encoder of the dual attention mechanism is as follows Figure 3 As shown. Given input X=(x1,x2,...,x T )∈R n*T , x t ∈R n , at time t, the encoder converts x t Mapping to h t :f1(h t-1 , x t ), h t ∈R m represents the state of the encoder hidden layer at time t, m represents the dimension of the hidden layer, and f(.) is a nonlinear activation function. Construct the input attention layer and convert the previous hidden layer state h t-1 and the cell state s of the LSTM unit t-1As input to this layer we get:

[0066]

[0067] Among them, v e ∈R T , W e ∈R T*2m , U e ∈R T*T are parameters that need to be learned.

[0068] If the attention mechanism in the model gives extremely high weight to certain parts or features while giving extremely low weight to other parts, the model learning process will be unstable and difficult to converge to a good solution. It may also be due to excessive focus on certain features and neglect of overall balance, resulting in good performance on training data but poor generalization ability on unseen data, resulting in overfitting. To avoid this phenomenon, this embodiment adopts the idea of ​​Gaussian blur, using a one-dimensional Gaussian function to balance the weights.

[0069] make Where x represents μ1 represents the mean, σ1 2 Represents variance.

[0070] The calculation method of mean is: ∑ represents the summation symbol, x i represents the i-th value, and n represents the total number of values.

[0071] How to calculate variance: x i represents the i-th value, μ1 represents the mean, and n represents the total number of values.

[0072] Input the output of the attention layer Go to the softmax layer to get To ensure that the sum of all attention weights is 1, Indicates the importance of the kth input feature at time t. After obtaining the attention weight, the driving sequence can be adaptively extracted At this time, the state of the updated hidden layer is

[0073] The temporal attention mechanism in the LSTM neural network decoder with dual attention mechanism is as follows Figure 4 As shown. Similar to the attention layer in the encoder, the attention layer of the decoder also needs to take the previous hidden layer state d t-1 and the cell state s′ of the LSTM unit t-1 The output of this layer is obtained as the input of this layer:

[0074]

[0075] Among them, v d ∈R m , W d ∈R m*2p , U d ∈R m*m are parameters that need to be learned.

[0076] Use a one-dimensional Gaussian function to balance the weights:

[0077] make Where y represents μ2 represents the mean, σ2 2 Represents variance.

[0078] The calculation method of mean is: ∑ represents the summation symbol, y i represents the i-th value, and n represents the total number of values.

[0079] How to calculate variance: y i represents the i-th value, μ2 represents the mean, and n represents the total number of values.

[0080] Through the softmax layer, we get the i-th encoder hidden state h i Importance of the final result The decoder sums all the encoder hidden states according to the weights to obtain the vector Note t It is different at different time steps.

[0081] After obtaining the vector, combine it with the target sequence to obtain: where w T , is the parameter to be learned. Update decoder hidden state f2 is a nonlinear activation function. The vector output obtained after passing through the decoder is the target signal output by the model.

[0082] In step S103 of this embodiment, the constructed LSTM neural network including the dual attention mechanism is trained using mixed signals received by multiple single channels until a trained neural network is obtained, and the trained neural network is used as a network model for eliminating single-channel time-frequency aliasing interference signals.

[0083] In step S104 of this embodiment, the sampled mixed signal containing the target signal is used as input to a neural network model for eliminating single-channel time-frequency aliasing interference signals, thereby outputting the desired interference-free target signal. The actual data is sampled to obtain a sequence s(n), where n = 0, 1, 2, ..., M-1. The I-channel data and Q-channel data of sequence s(n) are cross-merged into a 2M-row column vector. This processed column vector is used as input to the trained neural network, and the resulting output is the interference-free target communication signal.

[0084] like Figure 5 As shown, this embodiment discloses a system for eliminating single-channel time-frequency aliasing interference signals based on deep learning, including the following modules:

[0085] Data acquisition module: Acquires multiple mixed signals received by single channels as network input data. Each mixed signal used as network input data contains the target signal and its co-frequency interference signal. Sample the target communication signal and the co-frequency interference signal respectively. The modulation style of the target communication signal and the co-frequency interference signal can be selected according to actual needs to obtain N target signal sampling sequences x = [x1, x2, ..., x M ],in M is the number of sampling points; obtain N co-frequency interference signal sampling sequences y=[y1,y2,...,y M ],in M is the number of sampling points. Thus, the samples of the target communication signal and the co-channel interference signal are obtained. The co-channel interference signal is processed, and z = x + a * y, a is a random number greater than 0, and a sample set of N mixed signals is obtained. The I-channel data and Q-channel data of the mixed signal z are cross-merged into a column vector to construct a matrix with 2M rows and 1 column. This is a mixed signal superimposed with the same-frequency interference signal, which serves as the data input of the neural network.

[0086] Network building module: used to build an LSTM neural network with a dual attention mechanism suitable for signal separation. Given n driving sequences (input features): X = (x 1 ,x 2 ,...,x n ) T =(x1,x2,...,x T )∈R n*T , T represents the time step, n represents the dimension of the input feature. Through the input attention mechanism, the encoder generates new input In this module, the encoder architecture uses an LSTM. In the decoder, another LSTM network layer is used to decode the encoder information and select the corresponding hidden layer states through a temporal attention mechanism to obtain the final interference-free target signal sequence. In this module, the hidden layer dimension of the encoder and decoder is 64, the model time step T = 10, the learning rate is set to 0.01, and the number of hidden units is set to 128.

[0087] Network training module: Use the mixed signals received by multiple single channels to train the constructed LSTM neural network containing the dual attention mechanism until a trained neural network is obtained. The trained neural network is used as the network model for eliminating single-channel time-frequency aliasing interference signals.

[0088] The target signal output module uses the sampled mixed signal containing the target signal as input to a neural network model used to eliminate single-channel time-frequency aliasing interference signals, thereby outputting the desired interference-free target signal. The actual data is sampled to obtain a sequence s(n), where n = 0, 1, 2, ..., M-1. The I-channel and Q-channel data of sequence s(n) are cross-merged into a 2M-row column vector. This processed column vector serves as the input to the trained neural network, and the resulting output is the interference-free target communication signal.

[0089] The specific relevant contents of the system of this embodiment can be executed with reference to the above method embodiment, and the example explanations of the functions performed by each unit are consistent with the example explanations of the above method embodiment.

[0090] In summary, the present invention provides a novel and practical method and system that effectively solves the problem of co-channel interference between signals in single-channel mixed signal blind source separation, and demonstrates the technical and economic advantages of improved efficiency and reduced costs. By implementing the specific embodiments described above, those skilled in the art will clearly understand the scope and potential applications of the present invention.

[0091] Finally, the scope of the present invention is not limited to the above-described embodiments. All changes and modifications made without departing from the spirit and essential features of the present invention shall be deemed to fall within the scope of protection of the present invention. Applicants reserve the right to further clarify and define the claims in subsequent processes.

Claims

1. A single-channel time-frequency aliasing interference signal elimination method based on deep learning, characterized in that: The steps include: Step 1: Obtain multiple mixed signals received by single channels, each mixed signal contains the target signal and its co-channel interference signal; Step 2: Build an LSTM neural network with a dual attention mechanism suitable for signal separation; Step 3: Using the mixed signals received by the multiple single channels as network input data, the constructed LSTM neural network is trained to obtain a neural network model for eliminating single-channel time-frequency aliasing interference signals; Step 4: Use the actual mixed signal with the target signal as the input of the neural network model for eliminating the single-channel time-frequency aliasing interference signal, and output the interference-free target signal; In step 1, a mixed signal received by multiple single channels is obtained, as follows: Sample the target communication signal and the same-frequency interference signal respectively to obtain N target signal sampling sequences ,in, , , M is the number of sampling points; obtain N co-frequency interference signal sampling sequences ,in, , , M is the number of sampling points, for The real part of is the imaginary unit, for The imaginary part of for The real part of for The imaginary part of Process the co-frequency interference signal, set z = x + a * y, a is a random number greater than 0, and obtain a sample set of N mixed signals; Cross-merge the I-channel data and Q-channel data of the mixed signal z into a column vector, and construct a matrix with 2M rows and 1 column ,This is a mixed signal with the same-frequency interference signal superimposed; In step 2: Given n driving sequences: Where T represents the time step, n represents the dimension of the input feature; through the input attention mechanism, the encoder generates new input , the encoder architecture uses LSTM; in the decoder, another LSTM network layer is used to decode the encoder information, and the corresponding hidden layer state is selected through the temporal attention mechanism to obtain the final interference-free target signal sequence; The specific implementation process of the input attention mechanism in the LSTM neural network encoder of the dual attention mechanism is as follows: Given input , , at time t, the encoder will Mapped to : , represents the state of the encoder hidden layer at time t, m represents the dimension of the hidden layer, Is a nonlinear activation function; construct the input attention layer and convert the previous hidden layer state And the cell state of the LSTM unit As input to this layer we get: in, is the kth input drive sequence, , , are the parameters that need to be learned; Use one-dimensional Gaussian function to balance the weights: make , where x represents , represents the mean, represents variance; The mean is calculated as follows: , ∑ represents the summation symbol, represents the i-th value, and n represents the total number of values; How to calculate variance: , represents the i-th value, represents the mean, and n represents the total number of values; The output of the attention layer Input to the softmax layer to get To ensure that the sum of all attention weights is 1, Indicates the importance of the k-th input feature at time t; after obtaining the attention weight, the driving sequence is adaptively extracted , then update the state of the hidden layer to ; The specific implementation process of the temporal attention mechanism in the LSTM neural network decoder of the dual attention mechanism is as follows: The decoder's attention layer needs to take the previous hidden layer state And the cell state of the LSTM unit The output of this layer is obtained as the input of this layer: in, is the hidden state of the encoder, , , p is the size of the hidden state, are the parameters that need to be learned; Use one-dimensional Gaussian function to balance the weights: make , where y represents , represents the mean, represents variance; The mean is calculated as follows: , ∑ represents the summation symbol, represents the i-th value, and n represents the total number of values; How to calculate variance: , represents the i-th value, represents the mean, and n represents the total number of values; Through the softmax layer, we get the hidden state of the i-th encoder Importance of the final result , the decoder sums all the encoder hidden states according to the weights to obtain the vector ; After obtaining the vector, combine it with the target sequence to obtain: in, , b is the parameter to be learned; using the calculated , update the decoder hidden state , It is a nonlinear activation function; the vector output obtained after passing through the decoder is the target signal output by the model; Step 4 is as follows: Sampling the actual data, we get the sequence , , convert the sequence The I-channel data and Q-channel data are cross-merged into a 2M-row column vector, and the processed column vector is used as the input of the trained neural network, and the output obtained is the interference-free target communication signal.

2. A single-channel time-frequency aliasing interference signal elimination system based on deep learning, based on the method of claim 1, characterized in that: Includes the following modules: Data acquisition module: acquires multiple mixed signals received by single channels, each mixed signal contains the target signal and its co-frequency interference signal; Network building module: Build an LSTM neural network with a dual attention mechanism suitable for signal separation; Network training module: using the mixed signals received by the multiple single channels as network input data, training the constructed LSTM neural network to obtain a neural network model for eliminating single-channel time-frequency aliasing interference signals; Target signal output module: The actual mixed signal with the target signal is used as the input of the neural network model for eliminating the single-channel time-frequency aliasing interference signal, and the interference-free target signal is output.

Images