Multi-component LFM signal decomposition method, system and equipment

Abstract

The invention discloses a multi-component LFM signal decomposition method, system and device, and relates to the technical field of signal processing, and the method comprises the steps: obtaining a multi-component LFM signal, and building an LFM signal decomposition network model; the LFM signal decomposition network model comprises an encoder, a signal feature extraction module and a decoder; inputting the multi-component LFM residual signal data into the LFM signal decomposition network model, and encoding the residual signal data by using a one-dimensional convolutional layer in an encoder; after coded data passes through a normalization layer, feature information is extracted through a one-dimensional convolution layer; sequentially performing data modeling and feature extraction in a signal feature extraction module, converting a two-dimensional matrix into a one-dimensional sequence, further extracting features, and inputting the features into a decoder to obtain a real part and an imaginary part of an LFM signal component; according to the method, a dual-path recurrent neural network module is introduced to design a multi-component LFM signal decomposition network model, so that the precision of multi-component LFM signal decomposition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a multi-component LFM signal decomposition method, system and equipment. Background Art

[0002] As the observation scenes and electromagnetic environments become increasingly complex, broadband receivers often receive multiple radiation source signals in radar reconnaissance systems. These radiation source signals overlap in the time domain, forming multi-component signals. In actual reconnaissance, linear frequency modulation (LFM) signals are widely used, and effective analysis of multi-component LFM signals in complex and dense signal environments is of great practical significance. The effective separation of multi-component LFM signals is the basis and key to signal component parameter estimation.

[0003] With the continuous development of deep learning, deep learning theory is increasingly being applied to the field of signal processing. In signal separation applications, existing methods mainly train models based on data sets generated by specific system parameters, such as sampling frequency, which makes the network model sensitive to system parameters and difficult to be widely used in practice. At the same time, deep learning-based methods require the number of signal components to be known, which cannot be met in practical applications, severely limiting the practical application of network models.

[0004] Although the current methods such as time-frequency analysis have solved the problem of separating multi-component LFM signals, the separation accuracy of such methods is low when the noise is large. At the same time, the cross terms generated by multi-component signals will interfere with signal separation and parameter estimation, affecting the accuracy of multi-component LFM signal separation. Summary of the invention

[0005] In view of the shortcomings of the prior art that the cross terms generated by multi-component signals will interfere with signal separation and parameter estimation, thereby affecting the accuracy of multi-component LFM signal separation, the present invention proposes a multi-component LFM signal decomposition method, system and device. By acquiring a multi-component LFM signal data set, a multi-component LFM signal decomposition network is constructed using a dual-path recurrent neural network module to achieve multi-component LFM signal decomposition, thereby solving the problems of the prior art that the separation accuracy decreases when the noise is large and the cross terms generated by the multi-component signals will interfere with signal separation and parameter estimation.

[0006] A multi-component LFM signal decomposition method comprises the following steps:

[0007] Acquire multi-component LFM signal data to be decomposed, and convert it into multi-component LFM residual signal data;

[0008] Building an LFM signal decomposition network model; wherein the LFM signal decomposition network model includes an encoder, a signal feature extraction module and a decoder; the signal feature extraction module includes a segmentation module and a dual-path recurrent neural network module;

[0009] The multi-component LFM residual signal data is input into the LFM signal decomposition network model, and the residual signal data is encoded and extracted by an encoder to obtain feature information of the residual signal data; the one-dimensional long sequence in the feature information of the residual signal data is segmented by a segmentation module to form a two-dimensional data matrix; data modeling is performed on the two-dimensional data matrix by a dual-path recurrent neural network module, and feature information is extracted from the modeled data through a two-dimensional complex domain convolution layer, and the two-dimensional matrix in the feature information is transformed into a one-dimensional sequence by an overlap-addition method; feature extraction is performed on the one-dimensional sequence, and the extracted features are input into the one-dimensional complex domain transposed convolution layer of the decoder to obtain the real part and the imaginary part of the LFM signal component.

[0010] Furthermore, the obtaining of the multi-component LFM signal data to be decomposed specifically comprises the following steps:

[0011] The multi-component LFM signal model is expressed as:

[0012]

[0013] Among them, g k (t) represents the kth LFM signal, K is the number of LFM signal components, is additive Gaussian white noise, σ 2 represents the noise variance, represents a Gaussian distribution, rect(·) represents a rectangular pulse function, A k ,t k , ΔT k 、f k and γ k Respectively represent the amplitude, time center, time width, center frequency and modulation rate of the kth signal;

[0014] Represent the multi-component LFM signal in discrete form:

[0015]

[0016] n=-(N-1) / 2,-(N-1) / 2+1,...,(N-1) / 2-1,(N-1) / 2

[0017] Where N is the signal length, a k,1 and a k,2They represent the center angular frequency and angular modulation frequency corresponding to unit sampling, and the center frequency and modulation frequency are calculated as f k =a k,1 f s / (2π) and γ k =a k,2 f s 2 / π,f s is the system sampling frequency;

[0018] Convert the discrete form of the multi-component LFM signal into a vector form:

[0019]

[0020] Where s is the observation vector, i.e., the multi-component LFM signal to be decomposed; g k represents the kth LFM signal component; ω represents the noise vector.

[0021] Furthermore, the multi-component LFM signal data is converted into multi-component LFM residual signal data, and the conversion process specifically includes the following steps:

[0022] The multi-component LFM residual signal vector is represented as s k , the initial value is the signal to be decomposed, that is, s1 = s;

[0023] The LFM signal component with the largest amplitude is separated from the residual signal vector using a deep learning network. Then update the residual signal:

[0024]

[0025] Repeat the decomposition until the decomposition times reaches the upper limit K or the residual signal component s k+1 The energy is less than the noise energy, and a multi-component LFM residual signal is obtained; the multi-component LFM residual signal s is set k The length of is N=512, and the number of components K of each signal is randomly taken as an integer value in the interval [0,5].

[0026] Furthermore, the segmentation module is used to segment the one-dimensional long sequence in the feature information of the residual signal data to form a two-dimensional data matrix; which specifically includes the following steps:

[0027] The one-dimensional data sequence with the number of channels C=64 and the length N=512 is framed in a manner of frame length 2P=200 and step length P=100;

[0028] Each frame is treated as a separate data block, and these data blocks are stacked together to form a two-dimensional data matrix.

[0029] Furthermore, the dual-path recurrent neural network module includes four dual-path recurrent neural network sub-modules, each of which includes an intra-block calculation unit and an inter-block calculation unit; data modeling is completed by inputting the two-dimensional data matrix into the four dual-path recurrent neural network sub-modules and iteratively performing four intra-block calculations and inter-block calculations.

[0030] Furthermore, the method of transforming the two-dimensional matrix in the feature information into a one-dimensional sequence by using an overlap-add method specifically includes the following steps:

[0031] The overlapping parts of two adjacent data blocks in the two-dimensional data matrix are merged by averaging;

[0032] The two-dimensional data matrix is ​​transformed into a one-dimensional sequence by merging all the data blocks.

[0033] Furthermore, the intra-block computing unit and the inter-block computing unit both include a bidirectional long short-term memory network BI-LSTM, a fully connected layer and a normalization layer; wherein the computing process of the intra-block computing unit includes the following steps:

[0034] The frame matrix contained in the two-dimensional data matrix is ​​input into the intra-block computing unit, and the intra-frame data modeling is performed through the bidirectional long short-term memory network BI-LSTM;

[0035] The modeled data is restored to its dimension through a fully connected layer;

[0036] The data of the restored dimension is passed through the normalization layer and the residual learning strategy to obtain the output data calculated within the block;

[0037] The calculation process of the inter-block calculation unit includes the following steps:

[0038] The output data of the intra-block calculation is transposed and then input into the inter-block calculation module, and the inter-frame data modeling is performed through the bidirectional long short-term memory network BI-LSTM;

[0039] The modeled data is restored to its dimension through a fully connected layer;

[0040] The data of the restored dimension is passed through the normalization layer and the residual learning strategy to obtain the output data of the inter-block calculation.

[0041] Furthermore, it also includes constructing a loss function, and training the LFM signal decomposition network model using a gradient descent algorithm according to the loss function, wherein the minimum mean square error between the LFM signal component estimation value and the LFM signal component true value is used as the loss function of the LFM signal decomposition network model, which is expressed as:

[0042]

[0043] Among them, g(n) and are the true value of the LFM signal component and the estimated value of the LFM signal component respectively.

[0044] The present invention also proposes a multi-component LFM signal decomposition system, comprising:

[0045] An acquisition module, used for acquiring multi-component LFM signal data to be decomposed, and converting it into multi-component LFM residual signal data;

[0046] A model building module, used to build an LFM signal decomposition network model; wherein the LFM signal decomposition network model includes an encoder, a signal feature extraction module and a decoder; the signal feature extraction module includes a segmentation module and a dual-path recurrent neural network module;

[0047] The decomposition module is used to input the multi-component LFM residual signal data into the LFM signal decomposition network model, encode and extract the residual signal data through an encoder to obtain feature information of the residual signal data; use the segmentation module to segment the one-dimensional long sequence in the feature information of the residual signal data to form a two-dimensional data matrix; perform data modeling on the two-dimensional data matrix through a dual-path recurrent neural network module, extract feature information from the modeled data through a two-dimensional complex domain convolution layer, and transform the two-dimensional matrix in the feature information into a one-dimensional sequence by overlapping and adding; extract features from the one-dimensional sequence, input the extracted features into the one-dimensional complex domain transposed convolution layer of the decoder, and obtain the real and imaginary parts of the LFM signal components.

[0048] The present invention also proposes a multi-component LFM signal decomposition computer device, comprising: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the multi-component LFM signal decomposition method when executing the computer program.

[0049] The present invention provides a multi-component LFM signal decomposition method, which has the following beneficial effects:

[0050] The present invention introduces a dual-path recurrent neural network module to design a multi-component LFM signal decomposition network model, which effectively utilizes the correlation of data within a block and the correlation between data blocks; uses a segmentation module to cut long sequences in signal data, and inputs the cut two-dimensional data matrix into the dual-path recurrent neural network module for data modeling, thereby realizing local and global modeling of the data, thereby obtaining the connection between data within and between frames in the two-dimensional data matrix, and improving the accuracy of multi-component LFM signal decomposition; thereby effectively improving the estimation accuracy of multi-component LFM signal parameters in complex environments, and providing accurate signal parameters for applications such as radar imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is an overall block diagram of multi-component LFM signal decomposition in an embodiment of the present invention;

[0052] Figure 2 This is a diagram of a multi-component LFM signal decomposition network structure in an embodiment of the present invention;

[0053] Figure 3 It is a structural diagram of a segmentation module in an embodiment of the present invention;

[0054] Figure 4 Schematic diagram of testing the real and imaginary parts of a multi-component LFM signal in an embodiment of the present invention;

[0055] Figure 5 Schematic diagram of testing the real and imaginary parts of LFM signal components 1, 2, and 3 in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0057] The present invention proposes a multi-component LFM signal decomposition method, which specifically includes the following steps:

[0058] S1: Construct a multi-component LFM signal model; assume that the multi-component LFM signal model can be expressed as:

[0059]

[0060] Among them, g k (t) represents the kth LFM signal, K is the number of LFM signal components, is additive Gaussian white noise, σ 2 represents the noise variance. rect(·) represents the rectangular pulse function, A k ,t k , ΔT k 、f k and γ k Respectively represent the amplitude, time center, time width, center frequency and modulation rate of the kth signal. In order to make the data set not affected by the actual system sampling frequency, the multi-component LFM signal is represented as a discrete form:

[0061]

[0062] n=-(N-1) / 2,-(N-1) / 2+1,...,(N-1) / 2-1,(N-1) / 2

[0063] Where N is the signal length, a k,1 and a k,2 They represent the corresponding center angular frequency and angular modulation frequency under unit sampling respectively. The actual physical center frequency and modulation frequency are calculated as f k =a k,1 f s / (2π) and γ k =a k,2 f s 2 / π, where f s is the system sampling frequency.

[0064] The discrete form of the multi-component LFM signal is organized into a vector form:

[0065]

[0066] Among them, s is the observation vector, that is, the multi-component LFM signal to be decomposed, g k represents the kth LFM signal component, and ω represents the noise vector. The LFM signal decomposition process can be modeled as a regression problem, that is, solving the vector g from the multi-component LFM signal s k .

[0067] According to deep learning theory, a deep learning network can be designed to describe the signal decomposition problem, and the problem can be solved by minimizing the network's loss function using the training set.

[0068] S2: Obtain a multi-component LFM signal dataset. Assume that the multi-component LFM residual signal vector is represented as s k , the initial value is the signal to be decomposed, that is, s1 = s. The LFM signal component with the largest amplitude is separated from the residual signal vector using a deep learning network. Then update the residual signal:

[0069]

[0070] Repeat the above decomposition until the decomposition times reaches the upper limit K or the residual signal component s k+1 The energy is less than the noise energy.

[0071] Set the multi-component LFM residual signal s k The length of is N = 512, and the number of components K of each signal is randomly integer in the interval [0,5]. The signal parameters are randomly selected in the empirical value interval and follow a uniform distribution in the training set: A k ~U(0,1),n k ~U(100,400),ΔN k ~U(50,512), a k,1 ~U(-1,1), ak,2 ~U(-0.05,0.05), signal-to-noise ratio SNR~U(-10dB,20dB), thereby generating a multi-component LFM signal and its corresponding maximum amplitude LFM signal component.

[0072] Finally, 100k multi-component LFM signals and their corresponding maximum amplitude LFM signal components are generated, thereby obtaining a multi-component LFM signal data set. In order to better perform subsequent network training and performance optimization, the data set is normalized to the range of [0,1] using the minimax criterion.

[0073] S3: Use the dual-path recurrent neural network module to build a multi-component LFM signal decomposition network; Figure 2 As shown in Figure 2, a dual-path recurrent neural network module (DPRNNB) is introduced to design a multi-component LFM signal decomposition network. k All information will be k The real and imaginary parts of the signal are regarded as two independent channels as observation data and input into the multi-component LFM signal decomposition network. The network consists of an encoder, a signal feature extraction module and a decoder, which is composed of a convolutional neural network (CNN) and a long short-term memory neural network (LSTM). The core of the network is a DPRNNB composed of two LSTMs. DPRNNB divides long sequence data into overlapping frame data blocks, and uses DPRNNB to perform intra-block and inter-block calculations on the above data blocks, thereby realizing local and global modeling of the data.

[0074] Will s k After inputting into the network, the encoder is first used to encode the data through a one-dimensional convolutional layer (1-DConv). Figure 2 As shown in the figure, the encoded data passes through the normalization layer (LN) and then the feature information is extracted by 1-DConv. To solve the problem that LSTM cannot model long sequences, this network uses a segmentation module to cut the long sequence, thereby dividing the one-dimensional long sequence into a two-dimensional short sequence frame matrix. The specific segmentation process is as follows Figure 3 As shown, a one-dimensional data sequence with a channel number of C=64 and a length of N=512 is divided into frames with a frame length of 2P=200 and a step length of P=100. Each frame is treated as a separate data block and these data blocks are stacked together to form a two-dimensional data matrix, where S=8 represents the number of divisible blocks of the sequence.

[0075] The two-dimensional data matrix is ​​input into DPRNNB for data modeling. The two-dimensional data matrix is ​​obtained by overlapping the one-dimensional data sequence. The data in the frame matrix is ​​not only connected within the frame, but also between frames. The commonly used network structure only considers the connection of data within the frame when processing the frame matrix, but ignores the connection of data between frames. The DPRNNB used in this network consists of two modules: intra-block calculation and inter-block calculation, so as to make full use of the connection between data within and between frames. DPRNNB adopts LSTM modeling to effectively utilize the correlation of data within the block and the correlation between data blocks. The correlation of data within the block is reflected in the temporal order of data, and the correlation between data blocks is reflected in the fact that the second half of the previous block is the same as the first half of the next block during the block division process. The network structure contains 4 DPRNNBs to iteratively perform 4 intra-block calculations and inter-block calculations. The modeled data needs to go through a two-dimensional complex domain convolution layer (2-DConv) to extract features again and use overlap-addition to achieve waveform reconstruction. Overlap-addition is the inverse process of the segmentation module. The overlapping parts of two adjacent data blocks in the two-dimensional data matrix will be merged in an average manner. After all blocks are merged, the two-dimensional matrix is ​​transformed into a one-dimensional sequence. The generated one-dimensional sequence is further extracted through 1-DConv and activation function PReLU, and finally the decoder of the one-dimensional complex domain transposed convolution layer (1-DtransposedConv) is used to obtain the LFM signal component g k The real and imaginary parts of .

[0076] The proposed DPRNNB structure is as follows Figure 2 As shown in the figure, the frame matrix is ​​first used as the input data for intra-block calculation, and bidirectional LSTM (BI-LSTM) is used to model the intra-frame data. Since the use of BI-LSTM will cause the data dimension to change, a fully connected layer is added after BI-LSTM to restore the data dimension, and then the output data of the intra-block calculation is obtained through LN and residual learning strategy. After the intra-frame data modeling, the inter-frame data modeling is performed. The inter-frame data modeling first needs to transpose the output data of the intra-block calculation and use it as the input data. Finally, the final output data can be obtained through the same network structure as the intra-block calculation.

[0077] S4: Construct a loss function and use the gradient descent algorithm to train the network. Since the signal decomposition task is a regression problem, the minimum mean square error between the estimated value of the LFM signal component and the true value of the LFM signal component is used as the network loss function:

[0078]

[0079] Among them, g(n) and is the true value of the LFM signal component and the estimated value of the LFM signal component. The network model is trained based on supervised learning. The loss function minimization problem is solved by gradient descent and its variant optimization algorithm. The gradient of the network parameters is calculated based on the back propagation algorithm to update the network parameters. NvidiaGe Force GTX 3090 GPU is used for acceleration during the training process.

[0080] Multi-component LFM signal decomposition network test, realize multi-component LFM signal decomposition; input the new multi-component LFM signal into the trained multi-component LFM signal decomposition network model, obtain the estimated value of the LFM signal component with the largest energy, and calculate the residual signal, which is input into the network model again, and repeat the above operation until the decomposition times reaches K=5 or the residual signal energy is less than the surrounding noise energy.

[0081] In order to verify the beneficial effects of the present invention, the following test experiments were carried out: Figure 4 is the multi-component LF M signal to be decomposed, Figure 4 (a) and (b) are schematic diagrams of the real and imaginary parts of the multi-component LFM signal to be decomposed, respectively, which includes three LFM signal components and noise, and the SNR is 10dB. Figure 5 are the estimated results and true values ​​of the three components, Figure 5 (a) and (b) are schematic diagrams of the real and imaginary parts of LFM signal component 1, respectively. Figure 5 (c) and (d) are schematic diagrams of the real and imaginary parts of LFM signal component 2, respectively. Figure 5 (e) and (f) are schematic diagrams of the real and imaginary parts of LFM signal component 3, respectively. It can be seen that the estimated values ​​and true values ​​of the three components are highly similar. Therefore, this method can effectively realize the decomposition of multi-component LFM signals.

[0082] Based on the same inventive concept, the present invention also proposes a multi-component LFM signal decomposition system, comprising:

[0083] The acquisition module is used to acquire the multi-component LFM signal data to be decomposed and convert it into multi-component LFM residual signal data.

[0084] The model building module is used to build an LFM signal decomposition network model; wherein the LFM signal decomposition network model includes an encoder, a signal feature extraction module and a decoder; the signal feature extraction module includes a segmentation module and a dual-path recurrent neural network module.

[0085] The decomposition module is used to input the multi-component LFM residual signal data into the LFM signal decomposition network model, encode and extract the residual signal data through the encoder to obtain the feature information of the residual signal data; use the segmentation module to segment the one-dimensional long sequence in the feature information of the residual signal data to form a two-dimensional data matrix; use the dual-path recurrent neural network module to perform data modeling on the two-dimensional data matrix, extract feature information from the modeled data through a two-dimensional complex domain convolution layer, and transform the two-dimensional matrix in the feature information into a one-dimensional sequence by using the overlap-add method; extract features from the one-dimensional sequence, input the extracted features into the one-dimensional complex domain transposed convolution layer of the decoder, and obtain the real and imaginary parts of the LFM signal components.

[0086] The present invention also provides a multi-component LFM signal decomposition computer device, comprising: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the multi-component LFM signal decomposition method when executing the computer program.

[0087] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-component LFM signal decomposition method, characterized in that: The following steps are involved: Acquire multi-component LFM signal data to be decomposed, and convert it into multi-component LFM residual signal data; Building an LFM signal decomposition network model; wherein the LFM signal decomposition network model includes an encoder, a signal feature extraction module and a decoder; the signal feature extraction module includes a segmentation module and a dual-path recurrent neural network module; The multi-component LFM residual signal data is input into the LFM signal decomposition network model, and the residual signal data is encoded and extracted by an encoder to obtain feature information of the residual signal data; a segmentation module is used to segment a one-dimensional long sequence in the feature information of the residual signal data to form a two-dimensional data matrix; data modeling is performed on the two-dimensional data matrix by a dual-path recurrent neural network module, and feature information is extracted from the modeled data through a two-dimensional complex domain convolution layer, and the two-dimensional matrix in the feature information is transformed into a one-dimensional sequence by an overlap-addition method; feature extraction is performed on the one-dimensional sequence, and the extracted features are input into a one-dimensional complex domain transposed convolution layer of a decoder to obtain the real part and imaginary part of the LFM signal component.

2. A multi-component LFM signal decomposition method according to claim 1, characterized in that: The step of obtaining the multi-component LFM signal data to be decomposed specifically includes the following steps: The multi-component LFM signal model is expressed as: Among them, g k (t) represents the kth LFM signal, K is the number of LFM signal components, is additive Gaussian white noise, σ 2 represents the noise variance, represents a Gaussian distribution, rect(·) represents a rectangular pulse function, A k ,t k , ΔT k 、f k and γ k Respectively represent the amplitude, time center, time width, center frequency and modulation rate of the kth signal; Represent the multi-component LFM signal in discrete form: n=-(N-1) / 2,-(N-1) / 2+1,...,(N-1) / 2-1,(N-1) / 2 Where N is the signal length, a k,1 and a k,2 They represent the center angular frequency and angular modulation frequency corresponding to unit sampling, and the center frequency and modulation frequency are calculated as f k =a k,1 f s / (2π) and γ k =a k,2 f s 2 / π,f s is the system sampling frequency; Convert the discrete form of the multi-component LFM signal into a vector form: Where s is the observation vector, i.e., the multi-component LFM signal to be decomposed; g k represents the kth LFM signal component; ω represents the noise vector.

3. A multi-component LFM signal decomposition method according to claim 2, characterized in that: The multi-component LFM signal data is converted into multi-component LFM residual signal data, and the conversion process specifically includes the following steps: The multi-component LFM residual signal vector is represented as s k , the initial value is the signal to be decomposed, that is, s1 = s; The LFM signal component with the largest amplitude is separated from the residual signal vector using a deep learning network. Then update the residual signal: Repeat the decomposition until the decomposition times reaches the upper limit K or the residual signal component s k+1 The energy is less than the noise energy, and a multi-component LFM residual signal is obtained; the multi-component LFM residual signal s is set k The length of is N=512, and the number of components K of each signal is randomly taken as an integer value in the interval [0,5].

4. A multi-component LFM signal decomposition method according to claim 1, characterized in that: The segmentation module is used to segment the one-dimensional long sequence in the feature information of the residual signal data to form a two-dimensional data matrix; It specifically includes the following steps: The one-dimensional data sequence with the number of channels C=64 and the length N=512 is framed in a manner of frame length 2P=200 and step length P=100; Each frame is treated as a separate data block, and these data blocks are stacked together to form a two-dimensional data matrix.

5. The multi-component LFM signal decomposition method according to claim 1, characterized in that: The dual-path recurrent neural network module includes four dual-path recurrent neural network sub-modules, each of which includes an intra-block calculation unit and an inter-block calculation unit; the data modeling is completed by inputting the two-dimensional data matrix into the four dual-path recurrent neural network sub-modules and iteratively performing four intra-block calculations and inter-block calculations.

6. A multi-component LFM signal decomposition method according to claim 1, characterized in that: The method of transforming the two-dimensional matrix in the feature information into a one-dimensional sequence by using the overlap-add method specifically includes the following steps: The overlapping parts of two adjacent data blocks in the two-dimensional data matrix are merged by averaging; The two-dimensional data matrix is ​​transformed into a one-dimensional sequence by merging all the data blocks.

7. A multi-component LFM signal decomposition method according to claim 5, characterized in that: The intra-block computing unit and the inter-block computing unit both include a bidirectional long short-term memory network BI-LSTM, a fully connected layer and a normalization layer; wherein the computing process of the intra-block computing unit includes the following steps: The frame matrix contained in the two-dimensional data matrix is ​​input into the intra-block computing unit, and the intra-frame data modeling is performed through the bidirectional long short-term memory network BI-LSTM; The modeled data is restored to its dimension through a fully connected layer; The data of the restored dimension is passed through the normalization layer and the residual learning strategy to obtain the output data calculated within the block; The calculation process of the inter-block calculation unit includes the following steps: The output data of the intra-block calculation is transposed and then input into the inter-block calculation module, and the inter-frame data modeling is performed through the bidirectional long short-term memory network BI-LSTM; The modeled data is restored to its dimension through a fully connected layer; The data of the restored dimension is passed through the normalization layer and the residual learning strategy to obtain the output data of the inter-block calculation.

8. A multi-component LFM signal decomposition method according to claim 2, characterized in that: The method also includes constructing a loss function, and training the LFM signal decomposition network model using a gradient descent algorithm according to the loss function; wherein the minimum mean square error between the estimated value of the LFM signal component and the true value of the LFM signal component is used as the loss function of the LFM signal decomposition network model, which is expressed as: Among them, g(n) and are the true value of the LFM signal component and the estimated value of the LFM signal component respectively.

9. A multi-component LFM signal decomposition system, characterized in that: include: An acquisition module, used for acquiring multi-component LFM signal data to be decomposed, and converting it into multi-component LFM residual signal data; A model building module, used to build an LFM signal decomposition network model; wherein the LFM signal decomposition network model includes an encoder, a signal feature extraction module and a decoder; the signal feature extraction module includes a segmentation module and a dual-path recurrent neural network module; The decomposition module is used to input the multi-component LFM residual signal data into the LFM signal decomposition network model, encode and extract the residual signal data through an encoder to obtain feature information of the residual signal data; use the segmentation module to segment the one-dimensional long sequence in the feature information of the residual signal data to form a two-dimensional data matrix; perform data modeling on the two-dimensional data matrix through a dual-path recurrent neural network module, extract feature information from the modeled data through a two-dimensional complex domain convolution layer, and transform the two-dimensional matrix in the feature information into a one-dimensional sequence by overlapping and adding; extract features from the one-dimensional sequence, input the extracted features into the one-dimensional complex domain transposed convolution layer of the decoder, and obtain the real and imaginary parts of the LFM signal components.

10. A computer device for decomposing a multi-component LFM signal, characterized in that: include: A memory, a processor and a computer program stored in the memory, wherein the processor implements the steps of the multi-component LFM signal decomposition method according to any one of claims 1 to 8 when executing the computer program.

Images