Method, device and computer equipment for enhancing sparse time-frequency of micro-motion signals

Abstract

The present application relates to a method, apparatus and computer device for enhancing sparse time-frequency of micro-motion signals. The method uses a generative network to generate a time-frequency distribution prior from random noise, designs the random term in the Langevin dynamics differential equation, and uses the stochastic gradient optimization method to accurately sample the posterior probability distribution of the parameters of the time-frequency distribution prior distribution generator through Markov chain Monte Carlo. The least mean square error criterion is used to obtain a good reconstruction effect on the sparse time-frequency distribution in the sparse time-frequency observation model. Experimental results show that compared with traditional time-frequency representations, this method has better time-frequency representation resolution under complex observation conditions, and improves the representation accuracy of target micro-Doppler features in complex observation scenarios.

Description

Technical Field

[0001] The present application relates to the technical field of radar target micro-motion signal characterization, and particularly to a sparse time-frequency enhancement method, device, and computer device for micro-motion signals. Background Art

[0002] As a powerful tool for non-stationary signal processing, time-frequency characterization can depict the time-varying Doppler modulation characteristics of micro-motion signals and is widely used in the research of radar target micro-motion characteristics. In actual observation scenarios, affected by complex observation conditions such as radar system transmit power limitations, target maneuvering flight, and electromagnetic interference in a strong confrontation environment, the signal-to-noise ratio of the actually received radar echo is often low, and some echo data is unavailable and missing. Under the above complex observation conditions, the performance of traditional time-frequency characterization methods such as the short-time Fourier transform significantly degrades, the time-frequency characterization of the target will be severely defocused and contain a large number of false points, resulting in reduced resolution and difficulty in extracting micro-motion features.

[0003] Current solutions are divided into methods based on ADMM, Bayesian optimization, and generative adversarial networks. The method based on ADMM loses detailed features while improving the focus of the time-frequency diagram; the method based on Bayesian optimization requires assuming a prior distribution for the time-frequency diagram and has high requirements for prior modeling, and existing models cannot fully represent it; while the performance of the method based on generative adversarial networks highly depends on the effectiveness of the pre-trained model in simulating the time-frequency diagram, and the gradient descent is limited during the training process.

[0004] Traditional time-frequency characterization has low resolution, degrades in performance under complex observation conditions, and existing improved algorithms have problems such as being unable to balance focus and detailed features, relying on prior modeling, and relying on pre-trained models. Summary of the Invention

[0005] Based on this, in view of the problems that traditional time-frequency characterization has low resolution, degrades in performance under complex observation conditions, existing improved algorithms cannot balance focus and detailed features, rely on prior modeling, and generative adversarial networks rely on pre-trained models, it is necessary to provide a sparse time-frequency enhancement method, device, and computer device for micro-motion signals that can improve the resolution of the time-frequency characterization of defective radar echoes.

[0006] A sparse time-frequency enhancement method for micro-motion signals, the method comprising:

[0007] Obtain radar echo data of a micro-motion target, and perform joint time-frequency analysis on the radar echo data to obtain a time-frequency characterization of the micro-motion signal.

[0008] Construct a sparse observation model of the time-frequency characterization of the radar echo data according to the time-frequency characterization of the micro-motion signal.

[0009] According to the sparse observation model of time-frequency representation and the time-frequency representation of micro-motion signals, the generative convolutional neural network is used to reparameterize the time-frequency representation of micro-motion signals to obtain the prior of the accurate time-frequency distribution;

[0010] The adaptive stochastic gradient Langevin dynamics method is adopted, and the posterior probability distribution of the parameters of the generative convolutional neural network is accurately sampled through Markov chain Monte Carlo to obtain the posterior probability distribution.

[0011] According to the prior of the accurate time-frequency distribution and the posterior probability distribution, the structured sparse optimization model of time-frequency representation is determined.

[0012] According to the minimum mean square error criterion, the time-frequency distribution under sparse observation in the structured sparse optimization model of time-frequency representation is reconstructed to obtain the enhanced two-dimensional time-frequency representation of micro-motion targets.

[0013] A sparse time-frequency enhancement device for micro-motion signals, the device includes:

[0014] The micro-motion signal video analysis module is used to obtain the radar echo data of the micro-motion target and perform joint time-frequency analysis on the radar echo data to obtain the time-frequency representation of the micro-motion signal.

[0015] The time-frequency representation sparse observation model construction module is used to construct the time-frequency representation sparse observation model of the radar echo data according to the time-frequency representation of the micro-motion signal.

[0016] The time-frequency distribution prior determination module is used to reparameterize the time-frequency representation of the micro-motion signal by using the generative convolutional neural network according to the time-frequency representation sparse observation model and the time-frequency representation of the micro-motion signal to obtain the prior of the accurate time-frequency distribution;

[0017] The posterior distribution determination module of the generative convolutional neural network parameters is used to accurately sample the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo by adopting the adaptive stochastic gradient Langevin dynamics method to obtain the posterior probability distribution.

[0018] The structured sparse optimization model determination module of time-frequency representation is used to determine the structured sparse optimization model of time-frequency representation according to the prior of the accurate time-frequency distribution and the posterior probability distribution.

[0019] The sparse time-frequency representation reconstruction module of micro-motion signals is used to reconstruct the time-frequency distribution under sparse observation in the structured sparse optimization model of time-frequency representation according to the minimum mean square error criterion to obtain the enhanced two-dimensional time-frequency representation of micro-motion targets.

[0020] A computer device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of any of the above-mentioned sparse time-frequency enhancement methods for micro-motion signals are implemented.

[0021] The above-mentioned micro-motion signal sparse time-frequency enhancement method, device and computer equipment. The method uses a generative network to generate a time-frequency distribution prior from random noise, designs the random term in the Langevin dynamics differential equation, and uses the stochastic gradient optimization method to accurately sample the posterior probability distribution of the parameters of the time-frequency distribution prior distribution generator through Markov chain Monte Carlo. The minimum mean square error criterion is used to obtain a good reconstruction effect on the sparse time-frequency distribution in the sparse time-frequency observation model. Experimental results show that compared with traditional time-frequency representations, this method has better time-frequency representation resolution under complex observation conditions, and improves the representation accuracy of target micro-Doppler features in complex observation scenarios. Description of the Drawings

[0022] Figure 1 It is a schematic flow chart of the micro-motion signal sparse time-frequency enhancement method in one embodiment;

[0023] Figure 2 It is a flow chart of the micro-motion signal sparse time-frequency enhancement method in another embodiment;

[0024] Figure 3 It is a schematic diagram of the time-frequency distribution of the simulated motion in another embodiment;

[0025] Figure 4 It is the experimental result comparison of the time-frequency distribution enhancement method of the simulated signal under different random sparsity rates in another embodiment, where Figure 4 (a), Figure 4 (b) and Figure 4 (c) are respectively schematic diagrams of the two-dimensional time-frequency representation of the target obtained by the short-time Fourier transform of the simulated target at sparsity rates of 70%, 50% and 30%; Figure 4 (d), Figure 4 (e) and Figure 4 (f) are respectively schematic diagrams of the two-dimensional time-frequency representation of the target obtained by this method for the simulated target at sparsity rates of 70%, 50% and 30%;

[0026] Figure 5 It is a schematic diagram of the PSNR change curve of the ASGLD process of the simulated signal under different random sparsity rates in another embodiment;

[0027] Figure 6 It is a schematic diagram of the physical experiment scenario and experimental results in another embodiment, where Figure 6 (a) is a schematic diagram of the TI AWR2243 millimeter-wave radar physical object used in the actual measurement, Figure 6 (b) is a schematic diagram of the data acquisition experiment scenario, Figure 6 (c) and Figure 6 (d) are respectively the slow time-range image and the time-frequency distribution map of the actual collected radar data;

[0028] Figure 7 Schematic diagram of the comparison experiment results of the measured signal time-frequency distribution enhancement method under different random sparsity rates in another embodiment, where Figure 7 (a), Figure 7 (b) and Figure 7 (c) are schematic diagrams of the target two-dimensional time-frequency representation obtained by short-time Fourier transform of the measured signal at sparsity rates of 70%, 50%, and 30% respectively, Figure 7 (d), Figure 7 (e) and Figure 7 (f) are schematic diagrams of the target two-dimensional time-frequency representation obtained by the method of this application for the measured signal at sparsity rates of 70%, 50%, and 30% respectively;

[0029] Figure 8 Graph of the change of PSNR during the ASGLD process of the measured signal under different random sparsity rates in another embodiment;

[0030] Figure 9 Structural diagram of a computer device in one embodiment. Specific implementation manners

[0031] In order to make the purpose, technical solutions and advantages of this application clearer, the following further elaborates on this application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0032] The micro-motion signal sparse time-frequency enhancement method proposed in this application is abbreviated as: ASGLD method.

[0033] Aiming at the problems of low resolution of traditional time-frequency representation, performance degradation under complex observation conditions, existing improved algorithms being unable to balance focus and detail features, relying on prior modeling, and generative adversarial networks relying on pre-trained models, etc., this application proposes a micro-motion signal sparse time-frequency enhancement method based on adaptive stochastic gradient Langevin dynamics that can improve the time-frequency representation resolution of defective radar echoes.

[0034] In one embodiment, as Figure 1 shown, a micro-motion signal sparse time-frequency enhancement method is provided, and the method includes the following steps:

[0035] Step 100: Obtain the radar echo data of the micro-motion target, and perform joint time-frequency analysis on the radar echo data to obtain the time-frequency representation of the micro-motion signal.

[0036] Specifically, the target micro-motion generates a modulation effect on the radar echo, which is mainly manifested as the micro-Doppler modulation of the echo. Therefore, the radar echo of the micro-motion target can be regarded as a multi-component non-stationary signal. Due to the Doppler spectrum broadening caused by micro-motion, the traditional Fourier transform-based spectral analysis method cannot accurately describe the micro-motion signal due to the lack of time and frequency localization. As a powerful tool for non-stationary signal processing, time-frequency analysis can describe the time-varying frequency components of a signal by characterizing the energy distribution of the signal on a two-dimensional time-frequency plane, and is widely used to analyze the micro-Doppler characteristics of targets.

[0037] Step 102: According to the time-frequency characterization of the micro-motion signal, construct a sparse observation model for the time-frequency characterization of radar echo data.

[0038] Specifically, based on the relationship between the radar echo of the micro-motion target and the time-frequency characterization to be solved, establish a sparse observation model for the time-frequency characterization of the received echo.

[0039] Step 104: According to the sparse observation model of the time-frequency characterization and the time-frequency characterization of the micro-motion signal, use a generative convolutional neural network to reparameterize the time-frequency characterization of the micro-motion signal to obtain a prior of the accurate time-frequency distribution;

[0040] Specifically, a commonly used method for calculating the sparse observation model of the time-frequency characterization of radar echo data is Monte Carlo sampling (MC). However, in this method, sampling is not directly performed on, but a generative convolutional neural network is used to reparameterize , and such a reparameterization method can obtain a prior of the accurate time-frequency distribution obtained by the convolutional neural network.

[0041] Step 106: Adopt the adaptive stochastic gradient Langevin dynamics method to accurately sample the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo to obtain the posterior probability distribution.

[0042] Specifically, the classical method for sampling the posterior probability distribution of the parameters of the generative convolutional neural network is Markov chain Monte Carlo sampling (MCMC). In this method, by designing the random term in the Langevin dynamics differential equation and using the stochastic gradient optimization method, the posterior probability distribution of the parameters of the time-frequency distribution prior distribution generator is accurately sampled through Markov chain Monte Carlo.

[0043] This method has important engineering application value by mining the prior information of the micro-motion target in the time-frequency domain and introducing adaptive stochastic gradient Langevin dynamics for sparse time-frequency enhancement.

[0044] Step 108: Determine the structured sparse optimization model of the time-frequency representation according to the precise time-frequency distribution prior and posterior probability distribution.

[0045] Step 110: Reconstruct the time-frequency distribution under sparse observation in the structured sparse optimization model of the time-frequency representation according to the minimum mean square error criterion to obtain the enhanced two-dimensional time-frequency representation of the micro-motion target.

[0046] Specifically, set the experimental parameters, use this method for iterative training to obtain the distribution prior, and finally obtain the final enhanced two-dimensional time-frequency representation of the micro-motion target.

[0047] In the above micro-motion signal sparse time-frequency enhancement method, the method uses a generative network to generate the time-frequency distribution prior from random noise, designs the random term in the Langevin dynamics differential equation, and uses the stochastic gradient optimization method to accurately sample the posterior probability distribution of the parameters of the time-frequency distribution prior generator through Markov chain Monte Carlo. The minimum mean square error criterion is used to obtain a good reconstruction effect for the sparse time-frequency distribution in the sparse time-frequency observation model. Experimental results show that compared with traditional time-frequency representations, this method has better time-frequency representation resolution under complex observation conditions, and improves the representation accuracy of the target micro-Doppler characteristics in complex observation scenarios.

[0048] In one embodiment, step 100 includes: obtaining the radar echo data of the micro-motion target; performing short-time Fourier transform on the radar echo data to obtain the time-frequency representation of the micro-motion signal; sampling the time-frequency representation of the micro-motion signal in the time and frequency dimensions, and gridifying the two-dimensional time-frequency representation to obtain the discrete signal model corresponding to the short-time Fourier transform; the expression of the discrete signal model is:

[0049] ;

[0050] where represents the serial number of the time sampling point of the time-frequency representation of the micro-motion signal, N represents the number of signal sampling points, represents the serial number of the frequency sampling point of the time-frequency representation of the micro-motion signal, M represents the number of points of the discrete Fourier transform, represents that the serial number of the time sampling point is m and the serial number of the frequency sampling point is n for the discrete signal corresponding to the short-time Fourier transform, represents the l th signal of the discrete micro-motion signal, represents the window function after discretization. , is the echo time sampling interval, , is the echo frequency sampling interval.

[0051] Convert the discrete signal model corresponding to the short-time Fourier transform into a matrix form to obtain the two-dimensional time-frequency representation matrix of the micro-motion signal.

[0052] Specifically, the short-time Fourier transform (STFT) can be used to process micro-motion signals due to its advantages such as linear invertibility and no cross-terms. Different from the Fourier transform that processes the entire signal duration, STFT divides the signal into multiple continuous short-time segments using a time window and performs Fourier transform on each short-time segment to obtain the frequency distribution information of different segments at the corresponding time points. Therefore, the micro-motion signal The short-time Fourier transform expression of is

[0053] ;

[0054] where, represents the time-frequency representation of the micro-motion signal obtained by STFT, is a symmetric window function centered on , such as Hamming window, Gaussian window, etc., which is used to ensure the smoothness of the signal within the window.

[0055] According to the above short-time Fourier transform expression of the micro-motion signal , the short-time segment of the signal can be defined as:

[0056] ;

[0057] where, represents the short-time segment of the signal at t moment, is the length of the time window function .

[0058] Sample the time and frequency dimensions of the time-frequency representation in the short-time Fourier transform expression of the micro-motion signal respectively. After gridifying the two-dimensional time-frequency representation, the discrete signal model corresponding to the short-time Fourier transform as shown in the expression of the above discrete signal model can be obtained.

[0059] For the convenience of subsequent derivation, this application sets .

[0060] Furthermore, rewrite the expression of the above discrete signal model into a matrix form to obtain the two-dimensional time-frequency representation matrix of the micro-motion signal; the expression of this two-dimensional time-frequency representation matrix of the micro-motion signal is:

[0061] ;

[0062] where, and respectively represent the complete echo observation vector and the two-dimensional time-frequency representation matrix, N represents the number of sampling points of the complete echo, is the short-time Fourier transform matrix.

[0063] In one embodiment, step 102 includes: performing inverse short-time Fourier transform on the time-frequency representation of the micro-motion signal and then performing random sparse sampling to obtain the observed sparse sampling echo vector; performing short-time Fourier transform on the observed sparse sampling echo vector to obtain the time-frequency representation sparse observation model of the radar echo data; the expression of the time-frequency representation sparse observation model is:

[0064] ;

[0065] wherein, represents the sparse observation time-frequency representation of the radar echo data, represents the time-frequency distribution of the signal under the full sampling condition, represents the inverse Fourier transform matrix, represents the short-time random sampling vector matrix, represents the observation noise vector.

[0066] Specifically, in actual situations, affected by multi-functional radar systems and target maneuvering flights, etc., the observation data of a single target by the radar is usually incomplete. At the same time, in the actual confrontation environment, due to environmental influences such as electromagnetic interference and noise pollution, some echo data may be unavailable and missing. Therefore, only partial echo observations in the actual scenario can be used to obtain the time-frequency representation of the target. According to the above expression of the two-dimensional time-frequency representation matrix of the micro-motion signal, the expression of the time-frequency distribution observation model of the micro-motion signal calculated by short-time Fourier transform under random sparse sampling is:

[0067] ;

[0068] wherein, represents the observed sparse sampling echo vector, represents the time-frequency distribution of the signal under the full sampling condition, represents the sparse sampling matrix, which is a downsampling vector composed of 0 and 1, taking 1 at the echo sampling point and 0 at the missing point, represents the inverse short-time Fourier transform, represents the observation noise vector.

[0069] Since the inverse short-time Fourier transform and the short-time Fourier transform are not completely reversible when the number of Fourier transform points is not equal to the short-time sliding window length, when the observation vector changes from the sparse signal echo to the time-frequency distribution under sparse observation, the observation model cannot be directly expressed as:

[0070] ;

[0071] It can only be expressed as:

[0072] ;

[0073] Among them, represents the inverse Fourier transform matrix, represents the short-time random sampling vector matrix. Let , and an equivalent observation model as shown in the expression of the above time-frequency representation sparse observation model is obtained.

[0074] According to the signal model in the formula, the time-frequency enhanced representation problem can be abstracted as a process of recovering the two-dimensional time-frequency representation from the incomplete echo observation data . However, since the observation matrix is an underdetermined matrix, the solution of the formula is an ill-posed problem and there are multiple groups of solutions. At the same time, affected by the observation noise, if is directly calculated, it will lead to serious errors in the reconstructed time-frequency representation.

[0075] In view of this, this method introduces prior information related to the time-frequency representation recovery and establishes an optimization framework to solve the recovery result. According to the equivalent scattering center theory, the scattering of the target in the microwave frequency band can be approximated as the superposition of strong scattering centers. Since the number of strong scatterers on the micro-motion target is usually limited, it is manifested as a few continuous curves (time-frequency ridge lines) on the time-frequency plane, and the rest are background regions. Therefore, the echo of the micro-motion target has sparsity on the time-frequency plane, and sparse prior can be introduced to reconstruct and enhance the two-dimensional time-frequency representation.

[0076] According to the minimum mean square error criterion, the time-frequency distribution under sparse observation is reconstructed and enhanced.

[0077] In one of the embodiments, step 104 includes: according to the time-frequency representation sparse observation model, transforming the time-frequency enhanced representation problem into a process of recovering the two-dimensional time-frequency representation from the incomplete echo observation data; according to the minimum mean square error criterion, reconstructing and enhancing the time-frequency distribution under sparse observation, and the expression of the process of reconstructing and enhancing the time-frequency distribution under sparse observation is:

[0078] ;

[0079] Among them, represents the reconstructed and enhanced two-dimensional time-frequency representation, represents the posterior probability distribution of the sparse observation time-frequency, represents the sparse observation time-frequency representation of the radar echo data, represents the signal time-frequency distribution under full sampling conditions.

[0080] When using a generative convolutional neural network to reparameterize the signal time-frequency distribution under full sampling, an accurate time-frequency distribution prior is obtained; the process of reconstructing and enhancing the time-frequency distribution under sparse observations is transformed into:

[0081] ;

[0082] Among them, is the generative convolutional neural network, is the posterior probability distribution of the parameters of the generative convolutional neural network, is a fixed input sampled from the uniform random distribution and represents the parameters of the generative convolutional neural network.

[0083] In one embodiment, the generative convolutional neural network in step 104 is a U-Net network; the U-Net network is composed of an autoencoder with five layers and 128 channels in each layer.

[0084] Specifically, for calculating , the commonly used method is Monte Carlo sampling (MC). However, in this method, sampling is not directly performed on , but a generative convolutional neural network is used to reparameterize , where

[0085] ;

[0086] So can be rewritten as:

[0087] ;

[0088] Such a reparameterization method can obtain an accurate time-frequency distribution prior obtained by the convolutional neural network. Next, how to efficiently sample to obtain the posterior probability distribution required for calculating is crucial.

[0089] In one embodiment, step 106 includes: constructing a loss function of the generative convolutional neural network; constructing a sampling sample set according to the loss function; where the expression of the sampling sample set is:

[0090] ;

[0091] Among them, is the sampling sample set, represents the network loss function, represents the noise variance, represents a set small threshold,​ Denote the parameters of the generative convolutional neural network.

[0092] In the sampling sample set, adopt the method of adaptive stochastic gradient Langevin dynamics, and perform exact sampling on the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo to obtain the posterior probability distribution of the parameters of the generative convolutional neural network; the expression of the posterior probability distribution of the parameters of the generative convolutional neural network is:

[0093] ;

[0094] where, Denote the parameters of the generative convolutional neural network The k +(1)th sample of Denote the parameters of the generative convolutional neural network The k th sample of Denote a constant, Denote the learning rate of the k-th iteration, Denote the hyperparameters set manually, , Denote the adaptive stochastic term.

[0095] In one of the embodiments, construct the loss function of the generative convolutional neural network, and the expression of the loss function of the generative convolutional neural network is:

[0096] ;

[0097] where, Denote the degradation model, and the degradation model is the time-frequency representation sparse observation model, Denote l 2-norm, Denote the generative convolutional neural network, Denote the fixed input sampled from the uniform random distribution (the initial input set in the experiment is , which will change during the training process), Denote the sparse observation time-frequency representation of the radar echo data.

[0098] Specifically, since has a very high dimension, to effectively approximate the integral in , the sampling should be concentrated on the samples that contribute enough to the integral calculation, that is, concentrated on the significant subset. Therefore, in this embodiment, a restricted sampler is designed, and its sample set is concentrated in the sample set shown in the expression of the above sampling sample set.

[0099] Introducel The 2-norm is used to measure the similarity between the recovered time-frequency representation and the sparse time-frequency representation. For a very small , the loss function is as shown in the expression of the above loss function.

[0100] A classic sampling method of this kind is Markov Chain Monte Carlo (MCMC). This method uses a sampler for the posterior distribution related to the weights of a deep neural network by applying Langevin dynamics. Langevin dynamics was originally used to describe the stochastic differential equation of the dynamics of a molecular system, and its expression is:

[0101] ;

[0102] where represents the network loss function. The discretized form of the above equation is the Stochastic Gradient Langevin Dynamics (SGLD) equation:

[0103] ;

[0104] where , obviously the above equation can be regarded as a stochastic gradient descent method with added Gaussian white noise.

[0105] To calculate efficiently enough, it is necessary to restrict the sampler so that the sampling is concentrated in the sampling sample set shown in the expression of the above sampling sample set. This method proposes an Adaptive Stochastic Gradient Langevin Dynamics (ASGLD) method, adding an adaptive random term to the Stochastic Gradient Langevin Dynamics equation:

[0106] ;

[0107] The discretized form expression of the Stochastic Gradient Langevin Dynamics equation after adding the adaptive random term is:

[0108] ;

[0109] where the random term is adaptive for the loss function .

[0110] When training a deep neural network through SGD, the value of the loss function continuously decreases as the number of iterations increases. For ASGLD, during the early iteration process when the loss function remains at a relatively large value, it is very close to classical SGD; when at this time, the noise level decreases, and the perturbation to SGD is smaller, while the next sample may have a smaller ; when at this time, the noise level increases, and the perturbation to SGD is larger, and the next sample may have a larger . In both cases, the iterative process shown by the discrete form expression of the stochastic gradient Langevin dynamics equation after adding the adaptive random term can pull the subsequent samples back to when the current sample is far from the feasible set . At the same time, it is necessary to select an appropriate constant such that the adaptive random term can be ignored when .

[0111] Given an untrained generative network, the network is trained through this method for a total of K iterations. Assume that after rounds of "warm-up" training, sampling of parameters close to the feasible set starts. Calculate the minimum mean square error estimate in during the subsequent number of training times.

[0112] In one embodiment, step 110 includes: reconstructing the time-frequency distribution under sparse observation in the structured sparse optimization model of the time-frequency representation according to the minimum mean square error criterion to obtain an enhanced two-dimensional time-frequency representation of the micro-motion target; the expression of the enhanced two-dimensional time-frequency representation of the micro-motion target is:

[0113] ;

[0114] wherein, represents the reconstructed and enhanced two-dimensional time-frequency representation, represents the generative convolutional neural network, represents the fixed input sampled from the uniform random distribution , represents the sparse observed time-frequency representation of the radar echo data, represents the parameters of the generative convolutional neural network, represents the number of warm-up training, K represents the number of training times, represents the posterior distribution of the parameters of the generative convolutional neural network, the generative convolutional neural network parameter the k th sample.

[0115] Specifically, for the time-frequency enhanced reconstruction of the micro-motion target echo under sparse sampling conditions, the degradation model in the expression of the loss function of the generative convolutional neural network is the time-frequency observation model in , and a U-Net composed of an autoencoder with five layers and 128 channels in each layer is used as the generative convolutional neural network to sample a fixed input from the uniform random distribution .

[0116] The expression of the enhanced two-dimensional time-frequency representation of the micro-motion target is the minimum mean square error estimation.

[0117] In this application, the learning rate is set to 0.003, the number of training times is 3000, and the number of "warm-up" training times is 1000. Experimental verification and analysis are carried out on the simulated data with a duration of 1.5 seconds and a signal-to-noise ratio of 20 dB and the measured data with a duration of 1.5 seconds under the conditions of 30%, 50%, and 70% random sparsity rates. The time-frequency distribution reconstructed by the adaptive stochastic gradient Langevin method is compared with the time-frequency distribution directly calculated by STFT.

[0118] The schematic diagram of the simulated target time-frequency distribution is as Figure 3 shown. For a radar with a working carrier frequency of and a pulse repetition frequency PRF = 200 Hz, irradiated at an average horizon angle of , with a frequency conical spin of and a precession angle of , the radar echo of the scattering model within 12 seconds is simulated. Assuming the bottom diameter of the object is 0.4 m and the height is 2 m. Two main scattering points located at (0, 0.2, -0.4) m and (0, -0.15, 0.1) m on the simulated point scattering model are selected to calculate their simulated radar echoes. The short-time Fourier transform of the echo with a 65-point Hamming window sliding window is performed to obtain the simulated echo time-frequency distribution as shown in Figure 3 .

[0119] The results of the sparse time-frequency reconstruction of the simulated signal under different sparsity rates are as shown in Figure 4 , where Figure 4 (a), Figure 4 (b) and Figure 4 (c) are the schematic diagrams of the two-dimensional time-frequency representation of the simulated target obtained by the short-time Fourier transform at sparsity rates of 70%, 50%, and 30% respectively; Figure 4 (d), Figure 4 (e) and Figure 4(f) Schematic diagrams of the two-dimensional time-frequency representations of the simulation targets obtained by the present method at sparsity rates of 70%, 50%, and 30% respectively. As the sparsity rate increases, the continuity of the time-frequency curve obtained by STFT is gradually disrupted, and the noise-like interference in the background becomes more serious. However, the time-frequency distributions reconstructed by the ASGLD method used in this application have good recovery effects on the details such as the width of the time-frequency curve in the frequency dimension, the gradual change of intensity, and the intensity of the time-frequency curve intersection points in the frequency dimension, and the time-frequency distributions reconstructed at different sparsity rates all have very high recovery quality.

[0120] The curve of the peak signal-to-noise ratio (PSNR, abbreviated) of the reconstructed time-frequency map compared with the full-sampling time-frequency map during the process of reconstructing the sparse time-frequency distribution of the simulation signal by the ASGLD method changes with the number of iterations as shown in Figure 5 shown. It can be seen from Figure 5 that the time-frequency maps reconstructed from the simulation signal at the three sparsity rates all have high quality and are close to each other, with a PSNR close to 40 dB.

[0121] In the part of obtaining the measured data, in this embodiment, an AWR2243 multi-channel frequency-modulated continuous-wave (FMCW) radar device developed by Texas Instruments (TI) is used to collect experimental data. The working frequency band of this device covers 76 GHz to 81 GHz, is equipped with 12 transmitting channels and 16 receiving channels, and the maximum detection range can reach 350 meters. Its antenna system has a field of view angle of ±70°, and the angle resolution ability reaches 1.4°, which is suitable for various application scenarios such as target imaging, detection, and micro-Doppler effect analysis. By adjusting parameters such as the coherent processing interval, effective bandwidth, number of Chirp pulses per frame, and Chirp rise time of the radar in this application, the optimal configuration of the system working bandwidth, ranging accuracy, and velocity resolution can be achieved. Figure 6 (a) is a schematic diagram of the physical TI AWR2243 millimeter-wave radar used in the measurement. The system parameters of the AWR2243 millimeter-wave radar are set as shown in Table 1. Figure 6 (b) is the experimental scene of data collection. Figure 6 (c) and Figure 6 (d) are the slow-time range image and time-frequency distribution map of the actually collected radar data respectively. In the experimental setup, two cylindrical targets covered with aluminum foil are symmetrically arranged on both sides of the rotating platform, and the platform rotates at a preset angular velocity. At the same time, the reflected signals of the targets are obtained by using the millimeter-wave radar system. By changing variables such as the relative distance between the target and the radar, the angular velocity of the rotating platform, and the working bandwidth of the radar, the target echo information under different conditions can be obtained.

[0122] Table 1 AWR2243 Millimeter Wave Radar System Parameter Settings

[0123]

[0124] The experimental results of comparing the enhanced methods for the time-frequency distribution of measured signals at different random sparsity rates are as Figure 7 shown. Figure 7 (a), Figure 7 (b) and Figure 7 (c) are the two-dimensional time-frequency representations of the measured signals obtained by short-time Fourier transform at sparsity rates of 70%, 50%, and 30% respectively; Figure 7 (d), Figure 7 (e) and Figure 7 (f) are the two-dimensional time-frequency representations of the measured signals obtained by the proposed method at sparsity rates of 70%, 50%, and 30% respectively. For the time-frequency distribution obtained from the simulated signal, the intensity of one scattering point is weakened, and the width of the time-frequency curve becomes narrower due to over-focusing. The time-frequency reconstructions obtained by the ASGLD method highly restore the details of these time-frequency characteristics.

[0125] The PSNR variation curves of the ASGLD process for measured signals at different random sparsity rates are as Figure 8 shown. Figure 8 The PSNR variation curves of the comparison between the reconstructed time-frequency diagram and the full-sampling time-frequency diagram during the reconstruction of the sparse time-frequency distribution of the simulated signal by the ASGLD method in []. The overall PSNR is higher than that of the actual measurement signal, and the PSNR iteration curves at different sparsity rates are closer to the actual measurement data.

[0126] It should be understood that although Figure 1 each step in the flowchart of [] is shown in sequence according to the indication of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 at least a part of the steps in [] may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential either, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0127] In one embodiment, a micro-motion signal sparse time-frequency enhancement device is provided, including: a micro-motion signal video analysis module, a time-frequency representation sparse observation model construction module, a time-frequency distribution prior determination module, a generative convolutional neural network parameter posterior distribution determination module, a structured sparse optimization model determination module for time-frequency representation, and a micro-motion signal sparse time-frequency representation reconstruction module, where:

[0128] The micro-motion signal video analysis module is used to obtain radar echo data of a micro-motion target and perform joint time-frequency analysis on the radar echo data to obtain a time-frequency representation of the micro-motion signal.

[0129] The time-frequency representation sparse observation model construction module is used to construct a time-frequency representation sparse observation model of the radar echo data according to the time-frequency representation of the micro-motion signal.

[0130] The time-frequency distribution prior determination module is used to reparameterize the time-frequency representation of the micro-motion signal by using a generative convolutional neural network according to the time-frequency representation sparse observation model and the time-frequency representation of the micro-motion signal to obtain an accurate time-frequency distribution prior;

[0131] The generative convolutional neural network parameter posterior distribution determination module is used to accurately sample the posterior probability distribution of the generative convolutional neural network parameters by using the adaptive stochastic gradient Langevin dynamics method through Markov chain Monte Carlo to obtain the posterior probability distribution.

[0132] The structured sparse optimization model determination module for time-frequency representation is used to determine a structured sparse optimization model for time-frequency representation according to the accurate time-frequency distribution prior and the posterior probability distribution.

[0133] The micro-motion signal sparse time-frequency representation reconstruction module is used to reconstruct the time-frequency distribution under sparse observation in the structured sparse optimization model of time-frequency representation according to the minimum mean square error criterion to obtain an enhanced two-dimensional time-frequency representation of the micro-motion target.

[0134] In one of the embodiments, the micro-motion signal video analysis module is further used to obtain radar echo data of the micro-motion target; perform short-time Fourier transform on the radar echo data to obtain a time-frequency representation of the micro-motion signal; sample the time-frequency representation of the micro-motion signal in the time and frequency dimensions, grid the two-dimensional time-frequency representation to obtain a discrete signal model corresponding to the short-time Fourier transform as described in the above discrete signal model expression; and convert the discrete signal model corresponding to the short-time Fourier transform into a matrix form to obtain a two-dimensional time-frequency representation matrix of the micro-motion signal.

[0135] In one embodiment, the time-frequency representation sparse observation model construction module is further configured to perform inverse short-time Fourier transform on the time-frequency representation of the micro-motion signal and then perform random sparse sampling to obtain the observed sparse sampling echo vector; perform short-time Fourier transform on the observed sparse sampling echo vector to obtain the time-frequency representation sparse observation model of the radar echo data as shown in the above time-frequency representation sparse observation model expression.

[0136] In one embodiment, the time-frequency distribution prior determination module is further configured to convert the time-frequency enhancement representation problem into a process of recovering the two-dimensional time-frequency representation from the incomplete echo observation data according to the time-frequency representation sparse observation model; according to the minimum mean square error criterion, use the above expression of the time-frequency distribution under sparse observation for the reconstruction enhancement process to reconstruct and enhance the time-frequency distribution under sparse observation.

[0137] Use a generative convolutional neural network to reparameterize the signal time-frequency distribution under full sampling , to obtain an accurate time-frequency distribution prior; where are the parameters of the generative convolutional neural network; convert the process of reconstructing and enhancing the time-frequency distribution under sparse observation into:

[0138] ;

[0139] where, is the generative convolutional neural network, is the posterior probability distribution of the parameters of the generative convolutional neural network, is a fixed input sampled from the uniform random distribution .

[0140] In one embodiment, the generative convolutional neural network in the time-frequency distribution prior determination module is a U-Net network; the U-Net network consists of an autoencoder with five layers and 128 channels in each layer.

[0141] In one embodiment, the generative convolutional neural network parameter posterior distribution determination module is further configured to construct a loss function of the generative convolutional neural network; construct a sampling sample set according to the loss function; where the sampling sample set is as shown in the above sampling sample set expression.

[0142] In the sampling sample set, use the adaptive stochastic gradient Langevin dynamics method to accurately sample the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo to obtain the posterior probability distribution of the parameters of the generative convolutional neural network as shown in the above posterior probability distribution expression of the parameters of the generative convolutional neural network.

[0143] In one embodiment, the generative convolutional neural network parameter posterior distribution determination module is further configured to construct the loss function of the generative convolutional neural network as shown in the expression of the loss function of the generative convolutional neural network described above.

[0144] In one embodiment, the micro-motion signal sparse time-frequency representation reconstruction module is further configured to reconstruct the time-frequency distribution under sparse observation in the structured sparse optimization model of the time-frequency representation according to the least mean square error criterion, so as to obtain the enhanced two-dimensional time-frequency representation of the micro-motion target as shown in the expression of the enhanced two-dimensional time-frequency representation of the micro-motion target described above.

[0145] For the specific limitations on the micro-motion signal sparse time-frequency enhancement device, reference can be made to the limitations on the micro-motion signal sparse time-frequency enhancement method in the foregoing text, which will not be elaborated here. Each module in the above micro-motion signal sparse time-frequency enhancement device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above respective modules.

[0146] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 9 shown. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a micro-motion signal sparse time-frequency enhancement method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0147] Those skilled in the art can understand that Figure 9 the structure shown in

[0148] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps in the above method embodiment are implemented.

[0149] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0150] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A sparse time-frequency enhancement method for micro-motion signals, characterized in that The method includes: Obtaining radar echo data of a micro-motion target, and performing joint time-frequency analysis on the radar echo data to obtain a time-frequency representation of the micro-motion signal; Constructing a sparse observation model of the time-frequency representation of the radar echo data according to the time-frequency representation of the micro-motion signal; Reparameterizing the time-frequency representation of the micro-motion signal by using a generative convolutional neural network according to the sparse observation model of the time-frequency representation and the time-frequency representation of the micro-motion signal to obtain a prior of the exact time-frequency distribution; Using the adaptive stochastic gradient Langevin dynamics method to accurately sample the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo to obtain the posterior probability distribution; Determining a structured sparse optimization model of the time-frequency representation according to the exact time-frequency distribution prior and the posterior probability distribution; Reconstructing the time-frequency distribution under sparse observation in the structured sparse optimization model of the time-frequency representation according to the minimum mean square error criterion to obtain an enhanced two-dimensional time-frequency representation of the micro-motion target.

2. The method for enhancing sparse time-frequency of micro-motion signals according to claim 1, wherein Obtaining radar echo data of a micro-motion target, and performing joint time-frequency analysis on the radar echo data to obtain a time-frequency representation of the micro-motion signal, including: Obtaining radar echo data of a micro-motion target; Performing short-time Fourier transform on the radar echo data to obtain a time-frequency representation of the micro-motion signal; Sampling the time-frequency representation of the micro-motion signal in the time and frequency dimensions, and gridifying the two-dimensional time-frequency representation to obtain a discrete signal model corresponding to the short-time Fourier transform as: ； Among them, represents the serial number of the time sampling point of the micro-motion signal time-frequency characterization, N represents the number of signal sampling points, represents the serial number of the frequency sampling point of the micro-motion signal time-frequency characterization, M represents the number of points of the discrete Fourier transform, represents that the serial number of the time sampling point is m and the serial number of the frequency sampling point is n the discrete signal corresponding to the short-time Fourier transform, represents the l th signal of the discrete micro-motion signal, represents the window function after discretization; Converting the discrete signal model corresponding to the short-time Fourier transform into a matrix form to obtain a two-dimensional time-frequency representation matrix of the micro-motion signal.

3. The sparse time-frequency enhancement method for micro-motion signals according to claim 1, wherein Constructing a sparse observation model of the time-frequency representation of the radar echo data according to the time-frequency representation of the micro-motion signal, including: Performing inverse short-time Fourier transform on the time-frequency representation of the micro-motion signal and then performing random sparse sampling to obtain an observed sparse sampling echo vector; Performing short-time Fourier transform on the observed sparse sampling echo vector to obtain a sparse observation model of the time-frequency representation of the radar echo data as: ； Among them, represents the sparse observation time-frequency characterization of radar echo data, represents the signal time-frequency distribution under full sampling conditions, represents the inverse Fourier transform matrix, represents the short-time random sampling vector matrix, represents the observation noise vector.

4. The method for enhancing sparse time-frequency of micro-motion signals according to claim 1, wherein Reparameterizing the time-frequency representation of the micro-motion signal by using a generative convolutional neural network according to the sparse observation model of the time-frequency representation and the time-frequency representation of the micro-motion signal to obtain a prior of the exact time-frequency distribution, including: Converting the time-frequency enhancement representation problem into a process of recovering the two-dimensional time-frequency representation from incomplete echo observation data according to the sparse observation model of the time-frequency representation; According to the minimum mean square error criterion, the process of reconstructing and enhancing the time-frequency distribution under sparse observation is: ； Among them, represents the reconstructed and enhanced two-dimensional time-frequency representation, represents the posterior probability distribution of the sparse observation time-frequency, represents the sparse observation time-frequency representation of the radar echo data, represents the signal time-frequency distribution under the full sampling condition; Using a generative convolutional neural network to reparameterize the time-frequency distribution of the signal under full sampling conditions to obtain a prior of the exact time-frequency distribution; Converting the process of reconstructing and enhancing the time-frequency distribution under sparse observation into: ； Among them, is a generative convolutional neural network, is the posterior probability distribution of the parameters of the generative convolutional neural network, is a fixed input sampled from a uniform random distribution and represents the parameters of the generative convolutional neural network.

5. The micro-motion signal sparse time-frequency enhancement method according to claim 4, characterized in that The generative convolutional neural network is a U-Net network; the U-Net network consists of an autoencoder with five layers and 128 channels in each layer.

6. The sparse time-frequency enhancement method for micro-motion signals according to claim 1, wherein, Using the adaptive stochastic gradient Langevin dynamics method to accurately sample the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo to obtain the posterior probability distribution, including: Constructing a loss function of the generative convolutional neural network; Construct a sampling sample set according to the loss function; wherein the sampling sample set is as follows: ； Among them, is the sampling sample set, represents the network loss function, represents the noise variance, represents the set small threshold, represents the parameters of the generative convolutional neural network; In the sampling sample set, adopt the adaptive stochastic gradient Langevin dynamics method, and perform accurate sampling on the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo, and obtain the posterior probability distribution of the parameters of the generative convolutional neural network as follows: ； Among them, represents the parameters of the generative convolutional neural network of the k +(1)th sample, represents the parameters of the generative convolutional neural network of the k th sample, represents a constant, represents the learning rate of the k-th iteration, represents a hyperparameter set manually, , represents an adaptive random term.

7. The sparse time-frequency enhancement method for micro-motion signals according to claim 6, characterized in that Construct the loss function of the generative convolutional neural network as follows: ； Among them, represents a degradation model, and the degradation model is a time-frequency representation sparse observation model. represents norm represents a generative convolutional neural network. represents a fixed input sampled from a uniform random distribution and represents the sparse observation time-frequency representation of radar echo data.

8. The method for enhancing sparse time-frequency of micro-motion signals according to claim 1, wherein According to the least mean square error criterion, reconstruct the time-frequency distribution under sparse observation in the structured sparse optimization model of the time-frequency representation, and obtain the enhanced two-dimensional time-frequency representation of the micro-motion target as follows: ； Among them, represents the reconstructed and enhanced two-dimensional time-frequency representation, represents the generative convolutional neural network, represents sampling from a uniform random distribution to obtain a fixed input, represents the sparse observed time-frequency representation of radar echo data, represents the generative convolutional neural network parameters, represents the number of warm-up training, K represents the number of training times, represents the posterior distribution of the generative convolutional neural network parameters, The generative convolutional neural network parameters of the k th sample.

9. A fine motion signal sparse time-frequency enhancement device, characterized in that, The device includes: A micro-motion signal video analysis module, configured to acquire radar echo data of a micro-motion target, and perform joint time-frequency analysis on the radar echo data to obtain a time-frequency representation of the micro-motion signal; A time-frequency representation sparse observation model construction module, configured to construct a time-frequency representation sparse observation model of radar echo data according to the time-frequency representation of the micro-motion signal; A time-frequency distribution prior determination module, configured to reparameterize the time-frequency representation of the micro-motion signal by using a generative convolutional neural network according to the time-frequency representation sparse observation model and the time-frequency representation of the micro-motion signal, and obtain an accurate time-frequency distribution prior; A generative convolutional neural network parameter posterior distribution determination module, configured to perform accurate sampling on the posterior probability distribution of the parameters of the generative convolutional neural network through Markov chain Monte Carlo by using the adaptive stochastic gradient Langevin dynamics method, and obtain the posterior probability distribution; A structured sparse optimization model determination module of the time-frequency representation, configured to determine a structured sparse optimization model of the time-frequency representation according to the accurate time-frequency distribution prior and the posterior probability distribution; A micro-motion signal sparse time-frequency representation reconstruction module, configured to reconstruct the time-frequency distribution under sparse observation in the structured sparse optimization model of the time-frequency representation according to the least mean square error criterion, and obtain the enhanced two-dimensional time-frequency representation of the micro-motion target.

10. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the micro-motion signal sparse time-frequency enhancement method according to any one of claims 1 to 8 are implemented.

Images