Space micro-motion target echo separation method and system based on complex value mask learning network

Abstract

The invention discloses a spatial micro-motion target echo separation method based on a complex value mask learning network, and mainly solves the problems that an existing spatial micro-motion target echo separation method needs manual parameter adjustment, the separation precision is low, and parameter setting excessively depends on expert experience. The implementation scheme comprises the steps of generating a data set and dividing the data set into a training set and a test set; a complex value mask learning network comprising an encoder, a separation mask learning device and a decoder is constructed, the encoder is used for extracting features of the input one-dimensional echo signals, the separation mask learning device is used for generating signal separation masks, and the decoder is used for processing the separated signals and generating final output; iteratively training the complex value mask learning network by using the training set; and inputting the test set into the trained network to obtain a rigid body component and a micro-motion component after echo separation. The method can effectively improve the echo separation precision of the micro-motion target, has higher separation efficiency and application value, and can be used for micro-motion component feature extraction and spatial micro-motion target recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar, and further relates to a method and system for separating echo signals of spatial micro-motion targets, which can be used for feature extraction of micro-motion components and identification of spatial micro-motion targets. Background Art

[0002] Micro-motion refers to vibrations, rotations, swings, etc. generated by local or overall parts of a target in addition to the movement of the target body. Micro-motion of a target will generate sidebands around the main Doppler spectrum corresponding to the body motion, that is, generate the micro-Doppler effect, thus interfering with the main body imaging result and affecting the accurate estimation of micro-motion parameters. Therefore, there is an urgent need to design an effective method for separating echo signals of micro-motion targets to achieve high-resolution focused imaging of the main body echo and effective extraction of micro-Doppler features.

[0003] The prior art usually uses the basis function decomposition method or the modal decomposition method to realize the separation of micro-motion multi-component signals.

[0004] Xueru Bai et al. proposed the CEMD method for signal decomposition and discrimination in their published paper "Imaging of Micromotion Targets WithRotating Parts Based on Empirical-Mode Decomposition". The method decomposes the signal into a series of complex-valued intrinsic mode functions (IMFs), calculates the number of signal zero-crossings of each IMF according to the different frequency characteristics of the IMFs corresponding to rigid body parts and micro-motion parts, and selects an appropriate threshold to distinguish the IMFs of rigid body and micro-motion. Although this method can achieve a better micro-Doppler m-D separation effect than traditional methods, it requires manual adjustment of the zero-crossing threshold of the IMF. When there are many micro-motion components, the separation result will produce breakpoints at the intersection points of the time-frequency diagram, which is not conducive to micro-Doppler feature extraction.

[0005] The patent document with publication number CN 117538840 A discloses a method for separating multi-component micro-Doppler signals based on a second-order complex differential operator. First, a second-order complex differential operator and a null space operator are constructed, and the micro-Doppler signal components are projected into the null space; secondly, signal components are extracted respectively based on the iterative reweighted least squares (IRLS), Tikhonov regularization, and fixed-point iteration principle, and the second-order complex differential operator is regularized to optimize the multi-component micro-Doppler signal separation model; finally, each component signal is separated from the multi-component micro-Doppler signal in turn. Although this method can extract component signals with high signal-to-noise ratio from incomplete data and solve the problem of separating multi-component micro-Doppler signals due to random data loss, it requires manual setting of parameters when solving two optimization problems of signal components and the second-order complex differential operator, and the steps of the algorithm are more, and the computational complexity is higher. Summary of the Invention

[0006] The object of the present invention is to propose a method for separating the echoes of spatial micro-motion targets based on a complex-valued mask learning network, so as to solve the problems of the prior art that manual parameter adjustment is required, the separation efficiency is low, and the parameter setting is too dependent on expert experience, and improve the accuracy of echo separation.

[0007] The technical idea for achieving the object of the present invention is: by constructing a data-driven deep learning network, improving the computing efficiency, and realizing automatic learning of optimal parameters; by extracting high-dimensional features in the echoes and generating signal separation masks, the accuracy of echo separation is improved.

[0008] According to the above idea, the technical solution of the present invention includes the following:

[0009] 1. A method for separating the echoes of spatial micro-motion targets based on a complex-valued mask learning network, characterized by including

[0010] (1) Generating a data set and dividing it into a training set and a test set;

[0011] (2) Constructing a complex-valued mask learning network including an encoder, a separation mask learner, and a decoder:

[0012] The encoder is used to extract the features of the input one-dimensional echo signal;

[0013] The separation mask learner includes a pre-module, a middle module, and a post-module, and is used to generate a signal separation mask;

[0014] The decoder is used to process the separated signal and generate a final output;

[0015] (3) Iteratively training the complex-valued mask learning network using the training set;

[0016] (4) Inputting the test set into the trained network to obtain the rigid body component and the micro-motion component after echo separation.

[0017] Furthermore, the pre-module is composed of a normalization layer and a convolutional layer connected together. The convolutional kernel size of this convolutional layer is 3×3, the number of channels is 64, and the stride is 1, and it is used to adjust the representation form of the encoded features.

[0018] Furthermore, the middle module includes: eight convolutional layers, eight PReLU activation layers, eight normalization layers, and four depthwise separable convolutional layers, and their connection relationship is:

[0019] The first convolutional layer → the first PReLU activation layer → the first normalization layer → the first depthwise separable convolutional layer → the second PReLU activation layer → the second normalization layer → the second convolutional layer → the third convolutional layer → the third PReLU activation layer → the third normalization layer → the second depthwise separable convolutional layer → the fourth PReLU activation layer → the fourth normalization layer → the fourth convolutional layer → the fifth convolutional layer → the fifth PReLU activation layer → the fifth normalization layer → the third depthwise separable convolutional layer → the sixth PReLU activation layer → the sixth normalization layer → the sixth convolutional layer → the seventh convolutional layer → the seventh PReLU activation layer → the seventh normalization layer → the fourth depthwise separable convolutional layer → the eighth PReLU activation layer → the eighth normalization layer → the eighth convolutional layer are connected in sequence. This intermediate module is used to increase the receptive field so that the network can learn information in a longer time range.

[0020] Furthermore, the post-module is composed of a PReLU activation layer, a convolutional layer, and a Sigmoid activation layer connected. The convolutional kernel size of this convolutional layer is 3×3, the number of channels is 1, and the stride is 1. It is used to map the feature map output by the previous layer to between (0, 1) to generate a mask for weighting and encoding features, controlling the retention or suppression of different information.

[0021] 2. A spatial micro-motion target echo separation system based on a complex-valued mask learning network, characterized by including

[0022] A dataset generation module, used to generate a dataset containing corresponding echo signal components through a micro-motion signal model, and randomly divide it into a training set and a test set;

[0023] A network construction module, used to construct a complex-valued mask learning network including an encoder, a separation masker, and a decoder, and separate the echo signals in the dataset into micro-motion signal components and rigid body signal components;

[0024] A network training module, used to train the complex-valued mask learning network using the training set;

[0025] A network testing module, used to input the test set into the trained complex-valued mask learning network to obtain the echo separation result.

[0026] 3. An electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the program, it implements the steps of the above-mentioned spatial micro-motion target echo separation method.

[0027] 4. A non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to make the computer execute any step of the above-mentioned spatial micro-motion target echo separation method.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] First, the present invention preferably designs a complex-valued mask learning network, and obtains the separated micro-motion component and rigid body component by extracting the features in the echo and automatically generating an echo separation mask. Compared with the prior art, the separation accuracy of the echo is improved;

[0030] Second, the present invention preferably adopts a data-driven network structure, and automatically learns the network parameters through the backpropagation of the deep network. Compared with the prior art echo separation method of the iterative algorithm, it avoids manual parameter tuning, reduces the dependence on expert knowledge, improves the efficiency of the algorithm, and reduces the computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flowchart for implementing the method for separating the echo of a spatial micro-motion target based on the complex-valued mask learning network of the present invention;

[0032] Figure 2 is Figure 1 the separation network structure diagram in

[0033] Figure 3 is a block diagram of the system structure for separating the echo of a spatial micro-motion target based on the complex-valued mask learning network of the present invention;

[0034] Figure 4 is a block diagram of the structure of the electronic device of the present invention;

[0035] Figure 5 is a comparison diagram of the separation results between the present invention and the prior echo separation method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0037] Embodiment 1: A method for separating the echo of a spatial micro-motion target based on a complex-valued mask learning network.

[0038] Referring to Figure 1 , the implementation steps of this embodiment include the following:

[0039] Step 1, generate a data set, and divide it into a training set and a test set.

[0040] In the MATLAB simulation software, a motion model of the micro-motion target is constructed using the geometric characteristics of the micro-motion signal, and the initial parameters are set: signal carrier frequency 5 GHz, pulse repetition frequency 512 Hz, included angle θ = 38°;

[0041] According to the initial parameters, the Doppler frequency of the target micro-motion is calculated to obtain the one-dimensional simulation echo data s(t) with different rotation speeds and rotation radii:

[0042]

[0043] where σ is the intensity of the target scattering point, R(t) is the distance from the target scattering point to the radar, and f c is the carrier frequency;

[0044] For each echo data, the corresponding echo signal component is used as a label to form a data set;

[0045] The data set is randomly divided into a training set and a test set.

[0046] Step two, construct a complex-valued mask learning network.

[0047] Reference Figure 2 , the implementation of this step includes the following:

[0048] 2.1) Set the encoder: Select a one-dimensional convolutional module with a convolutional kernel size of 3×3, an output channel number of 128, and a stride of 1 to form an encoder, which is used to map the original signal to a high-dimensional space for feature extraction;

[0049] 2.2) Construct a separation mask learning module composed of a cascaded pre-module, middle module, and post-module, where:

[0050] The pre-module is composed of a normalization layer and a convolutional layer connected. The convolutional kernel size of this convolutional layer is 3×3, the number of channels is 64, and the stride is 1, which is used to adjust the representation form of the encoded features;

[0051] The intermediate module includes eight convolutional layers, eight PReLU activation layers, eight normalization layers, and four depth-wise separable convolutional layers. The connection relationship is as follows: the first convolutional layer, the first PReLU activation layer, the first normalization layer, the first depth-wise separable convolutional layer, the second PReLU activation layer, the second normalization layer, the second convolutional layer, the third convolutional layer, the third PReLU activation layer, the third normalization layer, the second depth-wise separable convolutional layer, the fourth PReLU activation layer, the fourth normalization layer, the fourth convolutional layer, the fifth convolutional layer, the fifth PReLU activation layer, the fifth normalization layer, the third depth-wise separable convolutional layer, the sixth PReLU activation layer, the sixth normalization layer, the sixth convolutional layer, the seventh convolutional layer, the seventh PReLU activation layer, the seventh normalization layer, the fourth depth-wise separable convolutional layer, the eighth PReLU activation layer, the eighth normalization layer, and the eighth convolutional layer are connected in sequence. The parameters of each layer are set as follows:

[0052] The convolution kernel size of the first convolution layer, the third convolution layer, the fifth convolution layer, and the seventh convolution layer are all 1×1, the number of channels are all 256, and the step size is all 1;

[0053] The convolution kernel size of the second convolution layer, the fourth convolution layer, the sixth convolution layer, and the eighth convolution layer are all 1×1, the number of channels is 128, and the stride is 1;

[0054] The convolution kernel size of the first depthwise separable convolution layer is 3×3, the number of channels is 256, the stride is 1, and the dilation coefficient is 1;

[0055] The second depth-wise separable convolutional layer has a convolution kernel size of 3×3, a channel number of 256, a stride of 1, and a dilation factor of 2;

[0056] The convolution kernel size of the third depthwise separable convolutional layer is 3×3, the number of channels is 256, the stride is 1, and the expansion coefficient is 4;

[0057] The fourth depthwise separable convolutional layer has a convolution kernel size of 3×3, a number of channels of 256, a stride of 1, and a dilation coefficient of 8;

[0058] The post-module is composed of a PReLU activation layer, a convolutional layer, and a Sigmoid activation layer. The convolutional layer has a convolution kernel size of 3×3, a channel number of 1, and a stride of 1. It is used to maintain information flow during feature transformation, improve training stability, and output a separation mask.

[0059] 2.3) Set up the decoder: A one-dimensional convolution module with a convolution kernel size of 3×3, 128 output channels, and a stride of 1 is selected to form the decoder. It is used to perform feature mapping, restore the masked encoded features to the original dimension, obtain the separated signal, and output it.

[0060] 2.4) Cascade the encoder, the separation mask learner module, and the decoder in sequence to form a complex-valued mask learning network.

[0061] Step 3: Iteratively train the complex-valued mask learning network using the training set.

[0062] 3.1) In this example, set the learning rate of the network during training to 0.0001, the weight decay coefficient to 0.00001, define the entire network parameter set θ based on the complex-valued mask learning network and initialize it. Take a single echo signal sample s from the training set and input it into the complex-valued mask learning network for forward propagation, outputting the micro-motion component p and the rigid body component q of the echo signal sample s.

[0063] 3.2) Set the loss function of the network

[0064]

[0065] where is the mean squared error of the micro-motion signal, p i is the i-th micro-motion component output in the echo signal sample, is the micro-motion component corresponding to p i in the training set label;

[0066] is the mean squared error of the rigid body signal, q i is the i-th rigid body component output in the echo signal sample, is the rigid body component corresponding to q i in the training set label;

[0067] 3.3) Obtain the gradient grad of the loss function with respect to the network parameter set θ through backpropagation:

[0068]

[0069] 3.4) According to the gradient grad, use the Adam optimizer to update the network parameter set θ to obtain the currently updated parameter set: θ' = θ - η × grad, where η is the learning rate;

[0070] 3.5) Repeat steps 3.1) - 3.4) until the loss function converges to obtain the trained complex-valued mask learning network.

[0071] Step 4: Input the test set into the trained network to obtain the rigid body component and the micro-motion component after echo separation.

[0072] Example 2: A spatial micro-motion target echo separation system based on a complex-valued mask learning network.

[0073] Reference Figure 3 , this example includes: a data set generation module 1, a network construction module 2, a network training module 3, and a network testing module 4.

[0074] The data set generation module is used to generate a data set containing corresponding echo signal components through a micro-motion signal model, and randomly divide it into a training set and a test set; the network construction module is used to construct a complex-valued mask learning network including an encoder, a separation masker, and a decoder, and this complex-valued mask learning network is used to separate the echo signals in the data set into micro-motion signal components and rigid body signal components; the network training module is used to train the complex-valued mask learning network constructed in the network construction module using the training set; the network testing module is used to input the test set into the trained copy mask learning network in the network training module to obtain the echo separation result.

[0075] Embodiment 3, an electronic device and a non-transitory computer-readable storage medium for separating the echoes of spatial micro-motion targets based on a complex-valued mask learning network.

[0076] Reference Figure 4 , in this embodiment, the electronic device includes a bus, a processor, and a memory, and is used to execute all steps of the method for separating the echoes of spatial micro-motion targets of the present invention by running computer program instructions. Exemplarily, the electronic device can be a distributed computing node system, a server group / server, a desktop computer, a laptop computer, etc.

[0077] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, and it can be divided into an address bus, a data bus, and a control bus according to its functions. For the sake of simplicity, although only a thick line is shown in the figure, it does not mean that there is only one bus or one type of bus.

[0078] The processor can be implemented in the form of a Central Processing Unit (CPU), a Micro Processing Unit (MCU), a System on Chip (SoC), or a Field Programmable Gate Array (FPGA), etc.

[0079] The memory includes volatile memory and non-volatile memory. The volatile memory is used for temporarily storing data when the program is running, such as Random Access Memory (RAM); the non-volatile memory is used for data storage, such as Read Only Memory (ROM), flash memory, Hard Disk Drive (HDD), and Solid State Drive (SSD).

[0080] The embodiment of the present invention also provides a computer-readable storage medium for storing program instructions, and the program instructions are run to execute all steps of the method for separating the echoes of spatial micro-motion targets of the present invention.

[0081] The method steps in the above embodiments can be implemented by software or computer code and stored in a recording medium, such as a CDROM, RAM, floppy disk, hard disk, or magneto-optical disk, or stored in a remote recording medium or a non-transitory machine-readable medium, and then loaded into the local recording medium for execution.

[0082] It should be noted that: in the method for separating the echoes of spatially micro-moving targets provided in the embodiments of the present invention, each functional module exists as an independent physical entity, and its dynamic components are implemented in the form of software functional modules. When sold or used as an independent product, it can be stored in a computer-readable storage medium. This storage medium can be a read-only memory, a magnetic disk, an optical disk, or the like.

[0083] The method for separating the echoes of spatially micro-moving targets provided in the embodiments of the present invention can be processed by software and stored on a recording medium suitable for a general-purpose computer, a dedicated processor, or programmable or dedicated hardware such as an ASIC or an FPGA.

[0084] The effects of the present invention can be further illustrated by the following simulation experiments:

[0085] I. Simulation experiment conditions

[0086] The hardware platform for the simulation experiment is an Intel(R) Core i9-10920X CPU and an NVIDIA GeForce RTX3090 GPU; the software platform for the simulation experiment is MATLAB 2021b and Python 3.8.

[0087] In the MATLAB software, the radar carrier frequency is set to 5 GHz, the pulse repetition frequency is set to 512 Hz, the included angle is set to 38°, the micro-motion rotation speed of the target is set to 2π - 7π rad / s, and the micro-motion rotation radius is set to 0.1 - 1.1 m, then a total of 400 echo data are generated.

[0088] II. Simulation experiment content and result analysis

[0089] Under the above simulation conditions, the echo data of the micro-moving target are separated by using the present invention and the existing CEMD method respectively, and the echo separation results of the two methods are output, as Figure 5 shown, which includes the micro-motion component and the rigid body component.

[0090] From Figure 5 it can be observed that the micro-motion component separated by the CEMD method will have curve ambiguity and missing at the intersection points of the curves, and the separation result of the rigid body component is more defective; the micro-motion separation result of the present invention is more complete and clear, the reconstruction effect at the intersection points of the curves is better, and the separation result of the rigid body component is more complete and has better focusing.

[0091] According to the above simulation experiments, the structural similarity SSIM of the micro-motion component image and the rigid body component image of the present invention and the existing CEMD method is calculated respectively, and it is used as the evaluation index to compare the separation performance of the present invention and the existing CEMD method. The results are shown in Table 1:

[0092] Table 1 Comparison of separation performance between the existing CEMD method and the present invention

[0093] Rigid body component structure similarity Micro-motion component structure similarity Existing CEMD method 0.7138 0.7599 The present invention 0.8509 0.8844

[0094] The calculation formulas of the parameters in Table 1 are as follows:

[0095]

[0096] where f out is Figure 5 the separation result image in, f label is the label image, μ(·) is the image mean, σ 2 (·) is the image variance, conv(f out , f label ) represents the covariance of f out and f label , and C1 and C2 are stable constants, C1 = 0.01, C2 = 0.03.

[0097] As can be seen from Table 1, compared with the existing CEMD method, the present invention achieves a higher structural similarity in the separation results of both the rigid body component and the micro-motion component, indicating that its separation curve is more complete and the separation accuracy is higher.

[0098] According to the calculation of the running time of the actual program, the calculation time of the present invention and the existing CEMD method on a single image is obtained, and the separation efficiency of the present invention and the existing CEMD method is compared. The results are shown in Table 2:

[0099] Table 2 Comparison of separation efficiency between the existing CEMD method and the present invention

[0100] Calculation time for a single image (s) Existing technology CEMD method 0.48 The present invention 0.012

[0101] As can be seen from Table 2, the present invention has a significant improvement in the calculation time of a single image compared with the existing CEMD method, and can greatly improve the separation efficiency.

[0102] The simulation experiment shows that the micro-motion target echo separation method based on the complex-valued mask learning network proposed by the present invention can achieve higher micro-motion target separation accuracy and separation efficiency.

[0103] It should be noted that the step numbers in the specification and claims of the present invention are only for a clear description of the embodiments of the present invention for easy understanding, and their sequence numbers are not limited.

Claims

1. A method for separating spatial micro-motion target echoes based on a complex-valued mask learning network, characterized in that, including (1) Generate a data set and divide it into a training set and a test set; (2) Construct a complex-valued mask learning network including an encoder, a separation mask learner, and a decoder: The encoder is used to extract the features of the input one-dimensional echo signal; The separation mask learner includes a pre-module, a middle module, and a post-module, and is used to generate a signal separation mask; The decoder is used to process the separated signal and generate a final output; (3) Iteratively train the complex-valued mask learning network using the training set; (4) Input the test set into the trained network to obtain the rigid body component and the micro-motion component after echo separation.

2. The method according to claim 1, characterized in that The implementation of generating the data set and dividing it into a training set and a test set in (1) includes: In the MATLAB simulation software, use the geometric characteristics of the micro-motion signal to construct a motion model of the micro-motion target, and set initial parameters such as the signal carrier frequency and the pulse repetition frequency. After all the settings are completed, calculate the Doppler frequency of the target micro-motion to obtain one-dimensional simulation echo data s(t) with different rotation speeds and rotation radii. Each echo data uses the corresponding echo signal component as a label to form a data set; Randomly divide the data set into a training set and a test set.

3. The method according to claim 1, wherein The encoder in (2) selects a one-dimensional convolutional module with a convolution kernel size of 3×3, an output channel number of 128, and a stride of 1. It uses a convolutional structure to map the original signal to a high-dimensional space for feature extraction.

4. The method according to claim 1, wherein The structures and functions of the modules that make up the separation mask learner in (2) are as follows: The pre-module is composed of a normalization layer and a convolutional layer connected. The convolutional layer has a convolution kernel size of 3×3, a channel number of 64, and a stride of 1, and is used to adjust the representation form of the encoded features; The middle module includes: eight convolutional layers, eight PReLU activation layers, eight normalization layers, and four depthwise separable convolutional layers. Their connection relationship is: The first convolutional layer, the first PReLU activation layer, the first normalization layer, the first depthwise separable convolutional layer, the second PReLU activation layer, the second normalization layer, the second convolutional layer, the third convolutional layer, the third PReLU activation layer, the third normalization layer, the second depthwise separable convolutional layer, the fourth PReLU activation layer, the fourth normalization layer, the fourth convolutional layer, the fifth convolutional layer, the fifth PReLU activation layer, the fifth normalization layer, the third depthwise separable convolutional layer, the sixth PReLU activation layer, the sixth normalization layer, the sixth convolutional layer, the seventh convolutional layer, the seventh PReLU activation layer, the seventh normalization layer, the fourth depthwise separable convolutional layer, the eighth PReLU activation layer, the eighth normalization layer, and the eighth convolutional layer are connected in sequence, which is used to lightweight the network, reduce the computational amount, and enable the network to learn information in a longer time range; The post-module is composed of a PReLU activation layer, a convolutional layer, and a Sigmoid activation layer connected. The convolutional layer has a convolution kernel size of 3×3, a channel number of 1, and a stride of 1, and is used to map the features output by the previous layer to between (0, 1) to generate a mask for weighting the encoded features and controlling the retention or suppression of different information.

5. The method according to claim 4, wherein: The convolution kernels of the first convolution layer, the third convolution layer, the fifth convolution layer, and the seventh convolution layer all have a size of 1×1, the number of channels is 256, and the stride is 1; The convolution kernels of the second convolution layer, the fourth convolution layer, the sixth convolution layer, and the eighth convolution layer all have a size of 1×1, the number of channels is 128, and the stride is 1; The convolution kernel of the first depthwise separable convolution layer has a size of 3×3, the number of channels is 256, the stride is 1, and the dilation coefficient is 1; The convolution kernel of the second depthwise separable convolution layer has a size of 3×3, the number of channels is 256, the stride is 1, and the dilation coefficient is 2; The convolution kernel of the third depthwise separable convolution layer has a size of 3×3, the number of channels is 256, the stride is 1, and the dilation coefficient is 4; The convolution kernel of the fourth depthwise separable convolution layer has a size of 3×3, the number of channels is 256, the stride is 1, and the dilation coefficient is 8.

6. The method according to claim 1, characterized in that, The decoder in (2) selects a one-dimensional convolution module with a convolution kernel size of 3×3, an output channel number of 128, and a stride of 1. It restores the masked encoded features to the original dimension through feature mapping to obtain the separated signal.

7. The method according to claim 1, characterized in that In (3), the complex-valued mask learning network is iteratively trained using the training set, and its implementation includes the following: (3a) Define and initialize the entire network parameter set θ of the complex-valued mask learning network. Take an echo signal sample s from the training set and input it into the complex-valued mask learning network for forward propagation to output the micro-motion component and the rigid body component of the echo signal sample s; (3b) Set the loss function of the network Among them, MSE is the mean square error. The network calculates the mean square errors of the micro-motion component and the rigid body component respectively based on the micro-motion component and the rigid body component obtained in (3a) and the label values in the training set, and takes the mean of the two mean square errors as the loss function of the network; (3c) Obtain the loss function by backpropagation Gradient grad of the network parameter set θ: (3d) According to the gradient grad, use the Adam optimizer to update the network parameter set θ to obtain the currently updated parameter set: θ' = θ - η × grad, where η is the learning rate; (3e) Repeat steps (3a) to (3d) until the loss function converges to obtain the trained complex-valued mask learning network.

8. A spatial micro-motion target echo separation system based on a complex-valued mask learning network, characterized in that, It includes a dataset generation module for generating a dataset containing corresponding echo signal components through a micro-motion signal model and randomly dividing it into a training set and a test set; A network construction module for constructing a complex-valued mask learning network including an encoder, a separation masker, and a decoder to separate the echo signal in the dataset into a micro-motion signal component and a rigid body signal component; A network training module for training the complex-valued mask learning network using the training set; A network testing module for inputting the test set into the trained complex-valued mask learning network to obtain the echo separation result.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the spatial micro-motion target echo separation method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions for causing the computer to execute the spatial micro-motion target echo separation method according to any one of claims 1 to 7.

Images