Electromagnetic signal blind source separation method of dual-path recurrent neural network

Abstract

The invention discloses an electromagnetic signal blind source separation method of a dual-path recurrent neural network, and belongs to the technical field of information. According to the method, the signals are sliced and integrated according to the index of the maximum cross-correlation value, automatic alignment of the signals is achieved, interference caused by signal asynchronization is effectively reduced, and the adaptive capacity and robustness of an algorithm are enhanced; according to the method, through the advantages that the multipath recurrent neural network captures the complex linear relation between signals and learns the signal time dependence relation, the nonlinear mixing problem can be better solved, and electromagnetic signals with time delay and dynamic change characteristics can be better adapted; according to the method, a bidirectional long-short-term memory network with residual connection and element-level multiplication and connection operation are used, the long-term dependency relationship in signals is effectively captured, the signal separation effect is adjusted and optimized through self-adaptive weight, and the signal separation precision is improved.

Description

Technical Field

[0001] The invention belongs to the field of information technology, and in particular relates to an electromagnetic signal blind source separation method of a dual-path recursive neural network. Background Art

[0002] With the rapid development of information technology, the electromagnetic environment has become increasingly complex, which makes the signal susceptible to interference and superposition of other signals during transmission. Against this background, blind source signal separation technology came into being, aiming to recover each independent source signal from the observed mixed signal, and then extract useful signals and remove the influence of interference.

[0003] At present, research in the field of blind source separation has made some progress and developed a variety of algorithms, such as independent component analysis (ICA) and second-order blind identification (SOBI). These algorithms use different strategies to achieve blind separation of signals. However, blind source separation technology still faces many challenges in practical applications, such as non-Gaussianity of signals, nonlinear mixing of signals, noise interference, etc. In addition, the performance of the algorithm may be affected by signal characteristics and environmental conditions, and further research and improvement are needed to improve its robustness and practicality.

[0004] Method 1: Patent "A method for blind source separation of single-channel radar signals under time-frequency aliasing" (CN116400298A) Figure 1 As shown, the method first performs Fourier transform on the received single-channel time-frequency mixed radar signal, converts the time domain signal into a frequency domain signal, and performs frequency band division; then, the frequency domain signal after frequency band division is inverse Fourier transformed to construct a virtual multi-channel time domain signal, and performs eigenvalue decomposition on its covariance matrix to calculate the whitening matrix; initializes and uses the blind source separation matrix to perform source estimation; Fourier transforms the separated signal, normalizes the amplitude to obtain the frequency domain signal, and sets the frequency band division factor; finally, the optimal frequency band division factor is used to perform frequency band division to obtain the final blind source separation result. The invention has the advantages of simple method, high efficiency in separating radar signals, and good separation performance.

[0005] Method 2: Figure 2 It is "Single-channel aliased electromagnetic signal separation method and device based on deep learning" (CN116680521A). Figure 2As shown, the method first obtains a single-channel aliased electromagnetic signal, and then inputs 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 the signal is encoded, the separation network in the signal processing model obtains an estimated mask of the single-channel electromagnetic signal based on the signal encoding, and further enables the decoder to output the signal after the single-channel aliased electromagnetic signal is separated 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 large batches of data through deep learning, and still has a good separation effect even when the signal is subject to strong noise interference.

[0006] At present, existing technical solutions can achieve blind source separation of complex electromagnetic signals to a certain extent, but are limited by the strong dependence of conventional blind source separation methods on prior knowledge and signal sparsity, and there is no efficient neural network method to effectively separate aliased complex electromagnetics. Therefore, the existing technology has certain shortcomings:

[0007] (1) The aforementioned method realizes multi-domain conversion of signals and constructs virtual multi-channel data to complete signal separation. However, in practice, due to the strong dependence on prior knowledge and signal sparsity, when the source signal is severely aliased in the time-frequency domain, its blind source separation effect is not good.

[0008] (2) The aforementioned method 2 is mainly based on blind source signal separation based on deep learning. This method requires a rich data set to learn sufficient signal features when implementing the network model. In addition, its signal separation is constructed based on the data set, and is not universal when separating measured signals, and cannot cover the separation of signals of all types and parameters. Summary of the invention

[0009] In view of the above technical problems existing in the prior art, the present invention proposes a blind source separation method of electromagnetic signals based on a dual-path recursive neural network, which has a reasonable design, overcomes the shortcomings of the prior art, and has good effects.

[0010] In order to achieve the above object, the present invention adopts the following technical solution:

[0011] A method for blind source separation of electromagnetic signals using a dual-path recursive neural network comprises the following steps:

[0012] Step 1: Data preprocessing;

[0013] Step 2: Separate the blind sources of electromagnetic signals through a dual-path recurrent neural network.

[0014] Preferably, step 1 specifically includes the following steps:

[0015] Step 1.1: Signal noise reduction processing;

[0016] Step 1.2: Time domain signal data alignment based on cross-correlation.

[0017] Preferably, step 1.2 specifically includes the following steps:

[0018] Step 1.2.1: cross-correlation calculation;

[0019] Use the cross-correlation function to calculate the similarity between two signals; according to the definition of cross-correlation, the dot product of the two signals is calculated in the time domain and then the results are summed;

[0020] Step 1.2.2: Delay estimation;

[0021] Find the maximum value from the cross-correlation function. This position usually corresponds to the best alignment point between the two signals. The time difference corresponding to this maximum value is the estimated delay.

[0022] Step 1.2.3: Signal alignment;

[0023] According to the estimated time delay, the signal is shifted to align the two signals in time, including signal truncation and padding;

[0024] Step 1.2.4: Verify and adjust;

[0025] Visually check the aligned signals to ensure that the alignment is performed correctly. If the alignment is not satisfactory, return to the delay estimation step and use a more sophisticated estimation method or adjust parameters.

[0026] Preferably, step 2 specifically includes the following steps:

[0027] Step 2.1: Segmentation processing;

[0028] Step 2.2: Block processing;

[0029] Step 2.3: overlap-add processing;

[0030] Step 2.4: Loss function calculation.

[0031] Preferably, in step 2.1, the specific contents are as follows:

[0032] This stage splits the long sequence input data into overlapping small blocks of size K and concatenates all the blocks into a 3-D tensor; the input sequence data is represented by W∈R N×L, where N represents the feature dimension and L represents the sequence length; to make each element of the input sequence appear the same number of times in a block, W is divided into blocks of length K and block shift P, so the number of times each element of the input sequence appears in a block is K / P; zero extension is used to process the first and last blocks; finally, S blocks are obtained, and each block of data is represented by D s ∈R N×K Indicates that all blocks form a 3-D tensor T∈R N×K×S .

[0033] Preferably, step 2.2 specifically includes the following steps:

[0034] Step 2.2.1: Input the divided data T into B consecutively stacked dual-path recurrent neural networks for processing, where b represents the bth dual-path recurrent neural network processing block, T b represents the input data of the bth processing block. Each dual-path recurrent neural network processing block consists of two parts: intra-block processing and inter-block processing, and the input and output dimensions are consistent. The intra-block processing is bidirectionally applied to the second dimension of the input data, that is, within each block of the S blocks;

[0035] U b =[f b (T b [:,:,i]),i=1,...,S](1)

[0036] Among them, U b ∈R H×K×S represents the output of the recurrent neural network, T b [:,:,i]∈R N×K represents the sequence defined by block i;

[0037] Step 2.2.2: Use a linear fully connected layer to transform U b Convert back to T b , after which layer normalization is used;

[0038]

[0039] in, is the conversion feature, and are the weights and biases of the FC (Fully Connected Layer), Indicates that block i is in U b Layer normalization;

[0040] Step 2.2.3: Apply residual connection between LN (Layer Normalization) output and block input;

[0041] Step 2.2.4: As the input of the inter-block RNN ​​submodule, the inter-block RNN ​​is applied to the third dimension of the input data, and linear full connection, layer normalization and residual connection are also used in inter-block processing.

[0042] Preferably, in step 2.3, the specific contents are as follows:

[0043] At the output of the last dual-path recurrent neural network block Apply the overlap-add method to restore the source signal waveform.

[0044] Preferably, step 2.4 specifically includes the following steps:

[0045] Step 2.4.1: Extract features of short-term and long-term multi-segment signals and introduce uPIT loss function and ID loss function to ensure the stability of the entire generated signal;

[0046] Step 2.4.2: To separate multiple signals, the maximum number of possible signals is set as C, and silence detection is performed on the separated C signals, and redundant signals are deleted;

[0047] Step 2.4.3: For various correspondences between the generated signal and the original signal, the loss function is calculated separately through the signal-to-noise ratio loss (SNR Loss) of PIT (Permutation Invariant Training), and only the optimal case is taken. The order of the generated signal is ignored when calculating the loss function.

[0048] Beneficial technical effects brought by the present invention:

[0049] This method slices and integrates the signal according to the index of the maximum cross-correlation value to achieve automatic alignment of the signal, effectively reduce the interference caused by signal asynchrony, and enhance the adaptability and robustness of the algorithm; this method uses the advantages of multipath recurrent neural network to capture the complex linear relationship between signals and learn the time dependency relationship of signals, which can better handle nonlinear mixing problems and better adapt to electromagnetic signals with time delay and dynamic change characteristics; this method uses a bidirectional long short-term memory network with residual connection and element-level multiplication and connection operations to effectively capture the long-term dependency in the signal, optimize the signal separation effect through adaptive weight adjustment, and improve the accuracy of signal separation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of a blind source separation method for single-channel radar signals under time-frequency aliasing;

[0051] Figure 2 This is a flow chart of a single-channel aliased electromagnetic signal separation method based on deep learning;

[0052] Figure 3 It is a flow chart of the electromagnetic signal blind source separation algorithm based on dual-path recurrent neural network;

[0053] Figure 4 It is a flow chart of the segmentation process;

[0054] Figure 5 It is the DPRNN block flow chart;

[0055] Figure 6 It is a flow chart of overlap-add process;

[0056] Figure 7 This is the flow chart of the loss function. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0058] The electromagnetic signal blind source separation method based on dual-path recursive neural network proposed in the present invention can effectively solve the problems described in the above method. First, the data is aligned to reduce the interference caused by signal asynchrony. Secondly, a bidirectional long short-term memory network with residual connection and element-level multiplication and connection operations are used to effectively capture the long-term dependency in the signal, optimize the signal separation effect through adaptive weight adjustment, and improve the accuracy of signal separation.

[0059] The present invention proposes a blind source signal separation method using a dual-path recurrent neural network (DPRNN) as the basic model structure. The RNN is optimized by DPRNN so that it can model long signal sequences. The method includes the following steps: (1) data preprocessing; (2) segment processing; (3) DPRNN block processing; (4) overlap-add processing; (5) loss function calculation. The overall process is shown in Figure 3:

[0060] 1. Data preprocessing

[0061] 1.1 Signal noise reduction processing

[0062] Filter and denoise the data to remove noise and unnecessary peaks, and remove abnormal values ​​or outliers in the signal through methods such as Hampel filter.

[0063] 1.2 Time Domain Signal Data Alignment Based on Cross-Correlation

[0064] 1.2.1 Cross-correlation calculation

[0065] Use the cross-correlation function to calculate the similarity between two signals. According to the definition of cross-correlation, the dot product of the two signals is calculated in the time domain and then the results are summed.

[0066] 1.2.2 Delay Estimation

[0067] Find the position of the maximum value from the cross-correlation function. This position usually corresponds to the best alignment point between the two signals. The time difference corresponding to this maximum value is the estimated delay.

[0068] 1.2.3 Signal Alignment

[0069] According to the estimated time delay, the signal is shifted to align the two signals in time, including signal truncation and padding, such as inserting zero values.

[0070] 1.2.4 Verification and Adjustment

[0071] Visually inspect the aligned signals to ensure that the alignment is performed correctly. If the alignment is not satisfactory, return to the delay estimation step and use a more sophisticated estimation method or adjust parameters.

[0072] 2 Model description

[0073] 2.1DPRNN Network

[0074] The dual-channel recurrent neural network divides the long sequence input into smaller blocks and iteratively applies intra-block and inter-block operations, where the input length of each operation can be proportional to the square root of the original sequence length. The DPRNN model consists of three stage groups: segmentation, block processing, and overlap addition.

[0075] 2.1.1 Segmentation process

[0076] This stage splits the long sequence input data into overlapping small blocks of size K and concatenates all the blocks into a 3-D tensor. The input sequence data is represented as W∈R N×L , where N represents the feature dimension (depending on the dimension of the encoder part) and L represents the sequence length (the length of the time step). In order to make each element of the input sequence appear the same number of times in a block, W is divided into blocks of length K and hop size P, so the number of times each element of the input sequence appears in a block is K / P. Zero-padding is used to process the first and last blocks. Finally, S blocks are obtained, and each block of data is represented by D s ∈R N×K Indicates that all blocks form a 3-D tensor T∈R N×K×S ,like Figure 4 shown.

[0077] 2.1.2 Block Processing

[0078] The divided data T is input into B consecutively stacked DPRNNs for processing, and b is used to represent the bth DPRNN processing block. One of the DPRNNs is as follows: Figure 5 As shown, Tb Represents the input data of the bth processing block. Each DPRNN processing block consists of two parts: intra-block processing and inter-block processing, and the input and output dimensions are consistent. The intra-block processing is bidirectionally applied to the second dimension of the input data, that is, within each block of the S blocks.

[0079] U b =[f b (T b [:,:,i]),i=1,...,S](1)

[0080] Among them U b ∈R H×K×S represents the output of RNN, T b [:,:,i]∈R N×K Represents the sequence defined by block i. Using a linear fully connected layer, U b Convert back to T b , followed by layer normalization.

[0081]

[0082] in is the conversion feature, and are the weights and biases of the FC layer, Indicates that block i is in U b Layer Normalization.

[0083] Then a residual connection is applied between the LN layer output and the block input, after which, As the input of the inter-block RNN ​​submodule, the inter-block RNN ​​is applied to the third dimension of the input data, and linear full connection, layer normalization and residual connection are also used in inter-block processing.

[0084] 2.1.3 Overlap-add process

[0085] The overlap-add method is applied to the output of the last DPRNN block to restore the source signal waveform. Figure 6 shown.

[0086] 2.2DPRNN Network Improvement

[0087] 2.2.1 Improved loss function

[0088] In this method, PIT-SNR Loss is used as the training loss and evaluation indicator. First, the features of short-term and long-term multi-segment signals are extracted and uPIT Loss and ID Loss are introduced to ensure the stability of the entire generated signal. In order to achieve the separation of multiple (2-N) signals, the maximum number of possible signals is set to C. Finally, silence detection is performed on the separated C signals and redundant signals are deleted. Finally, the SNR Loss trained with PIT permutation invariance is used to calculate the Loss for various corresponding relationships between the generated signal and the original signal. Only the optimal case is taken to ensure that the order of the generated signal is ignored when calculating the Loss. Figure 7 shown.

[0089] The present invention proposes an efficient signal time domain alignment method based on the cross-correlation method, which slices and integrates the signal according to the index of the maximum cross-correlation value to achieve automatic alignment of the signal, effectively reduce the interference caused by signal asynchrony, and enhance the adaptability and robustness of the algorithm.

[0090] The present invention proposes a dual-parallel recurrent neural network architecture, which uses a bidirectional long short-term memory network with residual connections and element-level multiplication and connection operations to effectively capture the long-term dependencies in the signal, optimize the signal separation effect through adaptive weight adjustment, and improve the accuracy of signal separation.

[0091] This method integrates the signal slices according to the index of the maximum cross-correlation value, realizes the automatic alignment of the signal, effectively reduces the interference caused by signal asynchrony, and enhances the adaptability and robustness of the algorithm. This method can better handle nonlinear mixing problems and better adapt to electromagnetic signals with time delay and dynamic change characteristics by capturing the complex linear relationship between signals and learning the time dependency relationship of signals through multipath recurrent neural networks; this method uses a bidirectional long short-term memory network with residual connections and element-level multiplication and connection operations to effectively capture the long-term dependency in the signal, optimize the signal separation effect through adaptive weight adjustment, and improve the accuracy of signal separation.

[0092] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for blind source separation of electromagnetic signals using a dual-path recursive neural network, characterized in that: The steps include: Step 1: Data preprocessing; Step 2: Separate the blind sources of electromagnetic signals through a dual-path recurrent neural network.

2. The method for blind source separation of electromagnetic signals using a dual-path recursive neural network according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Signal noise reduction processing; Step 1.2: Time domain signal data alignment based on cross-correlation.

3. The method for blind source separation of electromagnetic signals using a dual-path recursive neural network according to claim 2, characterized in that: Step 1.2 specifically includes the following steps: Step 1.2.1: cross-correlation calculation; Use the cross-correlation function to calculate the similarity between two signals; according to the definition of cross-correlation, the dot product of the two signals is calculated in the time domain and then the results are summed; Step 1.2.2: Delay estimation; Find the maximum value from the cross-correlation function. This position usually corresponds to the best alignment point between the two signals. The time difference corresponding to this maximum value is the estimated delay. Step 1.2.3: Signal alignment; According to the estimated time delay, the signal is shifted to align the two signals in time, including signal truncation and padding; Step 1.2.4: Verify and adjust; Visually inspect the aligned signals to ensure that the alignment was performed correctly; If the alignment effect is not ideal, you need to return to the delay estimation step and use a more sophisticated estimation method or adjust parameters.

4. The method for blind source separation of electromagnetic signals using a dual-path recursive neural network according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Segmentation processing; Step 2.2: Block processing; Step 2.3: overlap-add processing; Step 2.4: Loss function calculation.

5. The method for blind source separation of electromagnetic signals using a dual-path recursive neural network according to claim 4, characterized in that: In step 2.1, the specific contents are as follows: This stage splits the long sequence input data into overlapping small blocks of size K and concatenates all the blocks into a 3-D tensor; the input sequence data is represented by W∈R N×L , where N represents the feature dimension and L represents the sequence length. To make each element of the input sequence appear the same number of times in a block, W is divided into blocks of length K and block shift P. Therefore, the number of times each element of the input sequence appears in a block is K / P. Use zero extension to process the first and last blocks; finally get S blocks, each block of data is processed with D s ∈R N ×K Indicates that all blocks form a 3-D tensor T∈R N×K×S .

6. The method for blind source separation of electromagnetic signals using a dual-path recursive neural network according to claim 4, characterized in that: Step 2.2 specifically includes the following steps: Step 2.2.1: Input the divided data T into B consecutively stacked dual-path recurrent neural networks for processing, where b represents the bth dual-path recurrent neural network processing block, T b represents the input data of the bth processing block. Each dual-path recurrent neural network processing block consists of two parts: intra-block processing and inter-block processing, and the input and output dimensions are consistent. The intra-block processing is bidirectionally applied to the second dimension of the input data, that is, within each block of the S blocks; U b =[f b (T b [:,:,i]),i=1,...,S] (1); Among them, U b ∈R H×K×S represents the output of the recurrent neural network, T b [:,:,i]∈R N×K represents the sequence defined by block i; Step 2.2.2: Use a linear fully connected layer to transform U b Convert back to T b , after which layer normalization is used; in, is the conversion feature, and are the weights and biases of the fully connected layer, Indicates that block i is in U b Layer normalization; Step 2.2.3: Apply a residual connection between the normalization layer output and the block input; Step 2.2.4: As the input of the inter-block RNN ​​submodule, the inter-block RNN ​​is applied to the third dimension of the input data, and linear full connection, layer normalization and residual connection are also used in inter-block processing.

7. The method for blind source separation of electromagnetic signals using a dual-path recursive neural network according to claim 4, characterized in that: In step 2.3, the specific contents are as follows: At the output of the last dual-path recurrent neural network block Apply the overlap-add method to restore the source signal waveform.

8. The method for blind source separation of electromagnetic signals using a dual-path recursive neural network according to claim 4, characterized in that: Step 2.4 specifically includes the following steps: Step 2.4.1: Extract features of short-term and long-term multi-segment signals and introduce uPIT loss function and ID loss function to ensure the stability of the entire generated signal; Step 2.4.2: To separate multiple signals, the maximum number of possible signals is set as C, and silence detection is performed on the separated C signals, and redundant signals are deleted; Step 2.4.3: For various correspondences between the generated signal and the original signal, calculate the loss function separately through the signal-to-noise ratio loss of permutation-invariant training, take only the optimal case, and ignore the order of the generated signal when calculating the loss function.

Images