Non-orthogonal waveform MIMO radar imaging method based on deep learning

Abstract

The invention discloses a non-orthogonal waveform MIMO radar imaging method based on deep learning, and the method comprises the following steps: constructing an MIMO radar echo signal model, and obtaining a target echo signal in a whole imaging scene; separating signals of different waveforms in the target echo signal under the condition that the transmitted waveform satisfies an ideal orthogonal condition to obtain waveform vectors corresponding to the different waveforms, and performing azimuth compression to obtain an ideal target image of the MIMO radar; taking frequency domain echo signals at different receiving array elements in an MIMO radar echo signal model and the ideal target image as a training set, and training the constructed deep network model to obtain a trained deep network model; performing target imaging on the actual target echo signal by using the trained deep network model; according to the invention, the echo signal is processed through the trained deep network echo, the coupling interference between waveforms can be effectively suppressed, and the target high-resolution image can be efficiently acquired.

Description

Technical Field

[0001] The present invention relates to the field of MIMO radar imaging technology, and more specifically to a non-orthogonal waveform MIMO radar imaging method based on deep learning. Background Art

[0002] Compared with optical imaging, radar imaging has all-day and all-weather detection capabilities, can effectively detect the spatial morphological characteristics of target objects, and has a wide range of application needs in both military and civilian fields. Multiple-Input Multiple-Output (MIMO) radar imaging is a new radar imaging technology. By transmitting mutually orthogonal signals and sorting the signals at the receiving end, it can obtain observation channels and degrees of freedom far more than the actual number of receiving and transmitting array elements. Compared with traditional synthetic aperture radar (SAR) imaging and inverse synthetic aperture radar (ISAR) imaging, MIMO radar imaging has significant advantages such as high sampling data rate, no need for complex motion compensation, and easy to achieve three-dimensional imaging. In recent years, it has received widespread attention at home and abroad, and in-depth research has been carried out in imaging models, imaging methods, error correction, etc.

[0003] Transmitting orthogonal waveforms is the basis for leveraging the advantages of MIMO radar imaging technology. Ideal orthogonality between different transmit waveforms can be easily achieved through time diversity or frequency diversity, but it will reduce the data rate and spectrum utilization of MIMO radar imaging. Transmitting simultaneous and co-frequency orthogonal waveforms can give full play to the advantages of the MIMO radar system and is a current research hotspot in the design of MIMO radar orthogonal waveforms. However, existing studies have shown that the ideal orthogonal waveform requires that the spectra of different transmitted signals do not overlap with each other, and the currently common simultaneous and co-frequency waveforms do not meet the above requirements. This will result in mutual coupling noise and excessively high integrated sidelobes between different waveform components after matched filtering separation, which will seriously affect the imaging quality of MIMO radar.

[0004] To solve this problem, researchers have designed various MIMO waveforms from the perspective of waveform design, including multi-phase coded signals, discrete frequency coded signals, quasi-zero autocorrelation phase coded signals, positive and negative linear frequency modulation signals, OFDM chirp signals, and space-time coded waveforms. Although waveform optimization can suppress the influence of non-orthogonal waveform coupling to a certain extent, strictly speaking, completely orthogonal orthogonal waveforms of the same frequency do not exist under the matched filtering method. Therefore, waveform design can only alleviate but not completely solve this problem. In addition, some scholars have improved the imaging effect of non-orthogonal waveforms from the perspective of signal processing by designing receiving filters, introducing waveform polarization information, and using Clean technology, but the actual effect is limited. Sparse recovery or compressed sensing imaging is a new imaging method proposed in recent years that is different from the matched filter imaging theory. By constructing a linear mapping relationship between the MIMO radar echo signal and the target one-dimensional range image, two-dimensional image, and three-dimensional image, based on the sparsity of the target image, a sparse optimization algorithm is used to directly reconstruct a high-resolution, low-sidelobe target image, which can avoid the mutual coupling noise generated by the matched filtering of non-orthogonal waveforms. However, this non-orthogonal waveform MIMO radar imaging based on sparse recovery faces problems such as sensitivity to model errors, difficulty in selecting optimization algorithm hyperparameters, and sparse basis mismatch, resulting in unsatisfactory sparse imaging effects in practice. In addition, for large-scene, high-resolution, and high-dimensional imaging, the sparse optimization solution process has high computational complexity and requires a large amount of data sampling and storage, making it difficult to meet real-time, low-cost imaging requirements. Summary of the invention

[0005] The purpose of the present invention is to provide a non-orthogonal waveform MIMO radar imaging method based on deep learning. By processing the echo signal with a trained deep network model, the coupling interference between waveforms can be effectively suppressed, and a high-resolution image of the target can be efficiently acquired.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A non-orthogonal waveform MIMO radar imaging method based on deep learning, comprising the following steps:

[0008] Construct a MIMO radar echo signal model to obtain the target echo signal in the entire imaging scene;

[0009] Under the condition that the transmission waveform satisfies the ideal orthogonality, the signals of different waveforms in the target echo signal are separated to obtain waveform vectors corresponding to the different waveforms, and the waveform vectors are compressed in azimuth to obtain the ideal target image of the MIMO radar;

[0010] The frequency domain echo signals at different receiving array elements in the MIMO radar echo signal model and the ideal target image are used as training sets to train the constructed deep network model to obtain a trained deep network model;

[0011] The trained deep network model is used to perform target imaging on the actual target echo signal.

[0012] Furthermore, the deep network model includes: an echo transformation module and a feature imaging module;

[0013] The echo transform module includes: a matched filter network and an azimuth compression network;

[0014] The matched filter network is used to complete the range dimension compression of the target echo signal by inputting a non-orthogonal transmission waveform set;

[0015] The azimuth compression network is used to achieve two-dimensional compression of the target echo signal through echo rearrangement and azimuth compression according to the structure of the MIMO radar echo signal model;

[0016] The feature imaging module is constructed using a convolutional neural network and is used to establish a skip connection between the encoding and decoding of the neural network model to retain the detail information of the image during the reconstruction of the target image.

[0017] Furthermore, the method for constructing the feature imaging module is specifically as follows:

[0018] Each convolution operation or deconvolution operation is connected to a batch normalization process and a ReLu activation function; the deep network model contains a total of 5 convolution layers, and the number of input / output channels of each layer is 2 / 32, 32 / 64, 64 / 64, 64 / 64, and 64 / 128 respectively; it contains 5 deconvolution layers, and the number of input / output channels of each layer is 128 / 64, 64 / 64, 64 / 64, 64 / 32, and 32 / 1 respectively; MaxPool and MaxUnpool represent the maximum pooling layer and the maximum unpooling layer respectively, the pooling or unpooling window size is 2×2, the window span is 2, and there are 4 pooling layers and 4 unpooling layers.

[0019] Furthermore, the MIMO radar echo signal model is constructed to obtain the target echo signal in the entire imaging scene, specifically:

[0020] The MIMO radar echo signal model includes: a transmitting array element with a spacing of Nd, a receiving array element with a spacing of d; a target with U scattering points; the qth scattering point Q on the target, and the scattering point Q and the mth transmitting array element T m The distance between m Q and the scattering point Q and the nth receiving array element R n The distance betweenn Q;

[0021] The expression of the echo signal received by the nth receiving array element is:

[0022]

[0023] In the above formula, τ q,mn =(T m Q+R n Q) / c, τ q,mn The meaning is that the electromagnetic wave is transmitted by the mth transmitting element T m To the qth scattering point Q, and then to the nth receiving array element R n The propagation delay; n = 0, 1, L, N-1, c is the speed of light, σ q represents the scattering coefficient of the scattering point Q, λ is the radar operating wavelength; when the target sight angle is known, the echo envelopes of different transceiver channels are aligned by pre-compensating the envelope, τ q,mn No longer changes with m and n, then τ in formula (1) q,mn Denoted as τ q ;

[0024] Discretize the range imaging area. Assume that the range imaging range is ΔR. Then the number of sampling points of the imaging area discretization unit is: K = 2ΔR·f s / c, where f s is the distance sampling frequency, using f s Transmit waveform Sampling is performed to obtain the transmitted waveform vector The length is L;

[0025] Assume that the scattering point Q is located in the kth distance unit, k = 0, 1, L, k-1, and the distance kernel matrix of the kth distance grid is:

[0026]

[0027] In the above formula, 0 k×L represents a k×L-dimensional zero matrix; I L×L Represents an L×L dimensional identity matrix; 0 (K-k)×L represents a (Kk)×L-dimensional zero matrix;

[0028] Then the vector form of the echo signal of the scattering point Q received by the nth receiving array element is expressed as:

[0029]

[0030] in: For Rn The target complex scattering coefficient vector corresponding to the k-th range unit received at , the dimension of the target complex scattering coefficient vector is the number of transmitting array elements M, and S represents the transmitting waveform matrix;

[0031] The target echo signal is the superposition of the echo signals of all scattering points Q on the target. Let T = [T0, T1, ..., T K-1 ]∈C (K+L)×LK , Then the nth receiving array element R n The expression of the target echo signal of the entire imaging scene received at is:

[0032] y n =Te S·θ n +ε (3)

[0033] in, For R n The corresponding target complex-valued range image at , ε represents the receiving noise.

[0034] Furthermore, the signals of different waveforms in the target echo signal are separated to obtain waveform vectors corresponding to different waveforms, and azimuth compression is performed to obtain the target image of the MIMO radar, specifically:

[0035] The separation result of the target echo signal after the mth matched filter is expressed as:

[0036]

[0037] In the above formula, y mn Represents the target echo signal and the mth transmitted waveform The corresponding waveform components, where:

[0038]

[0039] Substituting equation (5) into equation (4), we can obtain

[0040] y mn =D m ·θ n (6)

[0041] in, D m is the matched filter reconstruction matrix corresponding to the mth transmitted waveform;

[0042] Assume that the scattering point Q is located on the kth distance grid, and the scattering coefficient vector θ to be restored is n It is expressed as: in, Then formula (6) can be further expressed as:

[0043]

[0044] Among them, y mn The result for the kth distance unit is:

[0045]

[0046] In the above formula, represents the transmission waveform vector corresponding to the mth transmission waveform, The * in the upper right corner represents the conjugate operation of the vector;

[0047] When the mth transmission waveform m=m', When m≠m', get That is, the separation result only contains the target signal corresponding to the mth transmission waveform, and there is no component of other waveforms; the complex scattering coefficients obtained by different transceiver channels After integration, the ideal target image of MIMO radar is obtained through azimuth compression.

[0048] Furthermore, the frequency domain echo signals at different receiving array elements in the MIMO radar echo signal model and the ideal target image are used as training sets; wherein the method for acquiring the frequency domain echo signals includes:

[0049] The echo signal received by the nth receiving array element is converted into a frequency domain form to obtain a frequency domain echo signal, which is specifically:

[0050]

[0051] In the above formula, For the transmission waveform The frequency domain form of e q,m,n (f) is the scattered field generated by the qth scattering point in the mth transmission-nth reception channel;

[0052] Among them, e q,m,n The expression of (f) is:

[0053]

[0054] In formula (10), f c is the center frequency, φ q,m,n is the azimuth angle between the qth scattering center and the equivalent transmitting and receiving array element corresponding to the mth transmitting and nth receiving channel, ΔR q,m,n is the sum of the wave paths from the qth scattering point to the mth transmitting array element and the nth receiving array element, is the parameter associated with the qth scattering center: q is the scattering coefficient; α qis a parameter describing the relationship between the scatterer frequency and geometry; L q , Represents the length and direction angle of the distribution scattering center; γ q To describe the relationship between the local scattering center and direction.

[0055] Furthermore, the deep network model is trained to obtain a trained deep network model, including:

[0056] Using the training set, the deep network model is trained alternately by iteratively performing forward propagation and error back propagation, and the mean square error between the ideal target image and the target image generated by the deep network model is used as the loss function to continuously optimize and adjust the parameters of the deep network model to obtain a trained deep network model.

[0057] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0058] The present invention proposes a non-orthogonal waveform MIMO radar imaging method based on deep learning, which obtains the target echo signal in the entire imaging scene; performs waveform separation on the target echo signal in the entire imaging scene to obtain waveform vectors corresponding to different transmitting waveforms, and performs azimuth compression to obtain an ideal target image of the MIMO radar; uses the frequency domain echo signals at different receiving array elements in the MIMO radar echo signal model and the ideal target image as training sets to train a deep network model; uses the trained deep network model to perform target imaging on the echo signals at different receiving array elements, and through waveform separation and azimuth compression, it is helpful to improve the imaging resolution of the MIMO radar, thereby more finely depicting the target details; uses the deep learning model to process the echo signal, which can maintain a high detection performance even in a low signal-to-noise ratio or complex environment, can better identify and separate the target signal, and can effectively suppress the coupling interference between waveforms, and efficiently obtain a high-resolution image of the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0060] The non-orthogonal waveform MIMO radar imaging method based on deep learning of the present invention is further described below in conjunction with the accompanying drawings;

[0061] Figure 1 It is a schematic diagram of the overall process of the non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention;

[0062] Figure 2 It is a non-orthogonal waveform MIMO radar imaging framework diagram based on a deep network in the non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention;

[0063] Figure 3 It is a network structure diagram of a deep network model in a non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention;

[0064] Figure 4 It is a waveform diagram of a multi-phase coded signal transmitted in a non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention; wherein, Figure 4 (a) is the autocorrelation, Figure 4 (b) is cross-correlated with the sum signal;

[0065] Figure 5 It is a schematic diagram of an ideal target image in the non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention;

[0066] Figure 6 It is the imaging result of the point target test data comparison method in the non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention; wherein, Figure 6 (a) is the imaging result obtained by the matched filtering method. Figure 6 (b) is the imaging result obtained by 1D-SR method. Figure 6 (c) is the imaging structure obtained by MB-SL0 method, Figure 6 (d) is the imaging result obtained by CNN method;

[0067] Figure 7 It is the MIMO-LI-Net imaging result of point target test data in the non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention;

[0068] Figure 8 It is a target CAD model in a non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention.

[0069] Fig. 9 It is the imaging result of the electromagnetic calculation test data comparison method in the non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention; wherein, Fig. 9 (a) is the imaging result obtained by the matched filtering method. Fig. 9 (b) is the imaging result obtained by 1D-SR method. Fig. 9 (c) is the imaging result obtained by MB-SL0 method, Fig. 9 (d) is the imaging result obtained by the CNN method.

[0070] Fig.10 It is the MIMO-LI-Net imaging result of the electromagnetic calculation test data in the non-orthogonal waveform MIMO radar imaging method based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0071] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0072] In order to better understand the purpose, structure and function of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0073] The present invention introduces deep learning technology into the field of MIMO radar imaging and proposes a new MIMO radar imaging method suitable for non-orthogonal waveforms. The method only needs to transmit common non-orthogonal signal waveforms and post-process the echoes through a trained deep network, thereby effectively suppressing the coupling interference between waveforms and efficiently acquiring high-resolution images of the target. The present invention also verifies through experimental results that training the proposed imaging network on a training set can obtain good imaging effects on point scattering target data and electromagnetic calculation data.

[0074] like Figure 1 Shown and Figure 2 As shown, the present invention provides a non-orthogonal waveform MIMO radar imaging method based on deep learning, comprising the following steps:

[0075] Construct a MIMO radar echo signal model to obtain the target echo signal in the entire imaging scene;

[0076] Under the condition that the transmission waveform satisfies the ideal orthogonality, the signals of different waveforms in the target echo signal are separated to obtain waveform vectors corresponding to the different waveforms, and the waveform vectors are compressed in azimuth to obtain the ideal target image of the MIMO radar;

[0077] The frequency domain echo signals at different receiving array elements in the MIMO radar echo signal model and the ideal target image are used as training sets to train the constructed deep network model to obtain a trained deep network model;

[0078] The trained deep network model is used to perform target imaging on the actual target echo signal.

[0079] The deep network model includes: an echo transformation module and a feature imaging module;

[0080] The echo transform module includes: a matched filter network and an azimuth compression network;

[0081] The matched filter network is used to complete the range dimension compression of the target echo signal by inputting a non-orthogonal transmission waveform set;

[0082] The azimuth compression network is used to complete two-dimensional compression of the target echo signal through echo rearrangement and azimuth compression according to the structure of the MIMO radar echo signal model;

[0083] The feature imaging module is constructed using a convolutional neural network, which is used to establish a skip connection between the encoding and decoding of the neural network model to achieve the retention of image detail information during the target image reconstruction process. The MIMO radar echo signal model is constructed to obtain the target echo signal in the entire imaging scene, specifically:

[0084] The MIMO radar echo signal model includes: a transmitting array element with a spacing of Nd, a receiving array element with a spacing of d; a target with U scattering points; the qth scattering point Q on the target, and the scattering point Q and the mth transmitting array element T m The distance between m Q and the scattering point Q and the nth receiving array element R n The distance between n Q;

[0085] The expression of the echo signal received by the nth receiving array element is:

[0086]

[0087] In the above formula, τ q,mn =(T m Q+R n Q) / c, n = 0, 1, L, N-1, c is the speed of light, σ q represents the scattering coefficient of the scattering point Q, λ is the radar operating wavelength; when the target sight angle is known, the echo envelopes of different transceiver channels are aligned by pre-compensating the envelope, τ q,mn No longer changes with m and n, then τ in formula (1) q,mn Denoted as τ q ;

[0088] Discretize the range imaging area. Assume that the range imaging range is ΔR. Then the number of sampling points of the imaging area discretization unit is: K = 2ΔR·f s / c, where f s is the distance sampling frequency, using f s Transmit waveform Sampling is performed to obtain the transmitted waveform vector The length is L;

[0089] Assume that the scattering point Q is located in the kth distance unit, k = 0, 1, L, k-1, and the distance kernel matrix of the kth distance grid is:

[0090]

[0091] In the above formula, 0 k×L represents a k×L-dimensional zero matrix; I L×L Represents an L×L dimensional identity matrix; 0 (K-k)×L Represents a (Kk)×L-dimensional zero matrix.

[0092] Then the vector form of the echo signal of the scattering point Q received by the nth receiving array element is expressed as:

[0093]

[0094] in: For R n The target complex scattering coefficient vector corresponding to the k-th range unit received at , the dimension of the target complex scattering coefficient vector is the number of transmitting array elements M, and S represents the transmitting waveform matrix;

[0095] The target echo signal is the superposition of the echo signals of all scattering points Q on the target. Let T = [T0, T1, ..., T K-1 ]∈C (K+L)×LK , Then the nth receiving array element R n The expression of the target echo signal of the entire imaging scene received at is:

[0096] y n =Te S·θ n +ε (3)

[0097] in, For R n The corresponding target complex-valued range image at , ε represents the receiving noise.

[0098] The signals of different waveforms in the target echo signal are separated to obtain waveform vectors corresponding to different waveforms, and azimuth compression is performed to obtain the target image of the MIMO radar, specifically:

[0099] The separation result of the target echo signal after the mth matched filter is expressed as:

[0100]

[0101] In the above formula, y mn Represents the target echo signal and the mth transmitted waveform The corresponding waveform components, where:

[0102]

[0103] Substituting equation (5) into equation (4), we can obtain

[0104] y mn =D m ·θ n (6)

[0105] in, D m is the matched filter reconstruction matrix corresponding to the mth transmitted waveform;

[0106] Assume that the scattering point Q is located on the kth distance grid, and the scattering coefficient vector θ to be restored is n It is expressed as: in, Then formula (6) can be further expressed as:

[0107]

[0108] Among them, y mn The result for the kth distance unit is:

[0109]

[0110] In the above formula, represents the transmission waveform vector corresponding to the mth transmission waveform, The * in the upper right corner represents the conjugate operation of the vector;

[0111] When the mth transmission waveform m=m', When m≠m', get That is, the separation result only contains the target signal corresponding to the mth transmission waveform, and there is no component of other waveforms; the complex scattering coefficients obtained by different transceiver channels After integration, the ideal target image of MIMO radar is obtained through azimuth compression.

[0112] It should be noted that: Because here To be separated from Multiply (0 to M-1, a total of M), so use To represent one of the.

[0113] The frequency domain echo signals at different receiving array elements in the MIMO radar echo signal model and the ideal target image are used as training sets; wherein the method for obtaining the frequency domain echo signals includes:

[0114] The echo signal received by the nth receiving array element is converted into a frequency domain form to obtain a frequency domain echo signal, which is specifically:

[0115]

[0116] In the above formula, For the transmission waveform The frequency domain form of e q,m,n (f) is the scattered field generated by the qth scattering point in the mth transmission-nth reception channel;

[0117] Among them, e q,m,n The expression of (f) is:

[0118]

[0119] In formula (10), f c is the center frequency, φ q,m,n is the azimuth angle between the qth scattering center and the equivalent transmitting and receiving array element corresponding to the mth transmitting and nth receiving channel, ΔR q,m,n is the sum of the wave paths from the qth scattering point to the mth transmitting array element and the nth receiving array element, is the parameter associated with the qth scattering center: q is the scattering coefficient; α q is a parameter describing the relationship between the scatterer frequency and geometry; L q , Represents the length and direction angle of the distribution scattering center; γ q To describe the relationship between the local scattering center and direction.

[0120] It should be noted that the present invention introduces deep learning technology to solve the problem of non-orthogonal waveform coupling suppression, and proposes a new framework for non-orthogonal waveform MIMO radar imaging based on deep network. The core is to realize the mapping of non-orthogonal waveform MIMO radar complex echoes to target high-resolution images through deep network, and the key is to design a network structure suitable for non-orthogonal waveform MIMO radar imaging. The network structure based on the deep expansion of the sparse reconstruction algorithm can effectively process complex echoes, but for non-orthogonal MIMO radar imaging, the dimension of the sparse reconstruction model is too high, and the network is difficult to implement; and the structures such as convolutional neural networks and recurrent neural networks widely used in the general deep learning field are not suitable for the data characteristics of non-orthogonal waveform MIMO radar echoes, and it is difficult to learn the complex mapping from complex echoes to target images. The proposed framework organically combines MIMO radar imaging theory with advanced deep networks, provides a basis for network construction and parameter initialization through imaging theory, and uses advanced deep network modules to improve the nonlinear fitting ability of the imaging network. The overall network structure is divided into two parts, namely "echo transformation module" and "feature imaging module". The role of the "echo transform module" is to perform artificial feature extraction, and to transform the complex echoes by referring to traditional MIMO radar imaging theories such as matched filtering, sparse recovery, and time-frequency analysis, so as to achieve the transformation from the complex echo domain of "implicitly expressed" information to the feature domain of "explicitly expressed" information. The role of the "feature imaging module" is to perform automatic feature extraction, and to build a network structure using advanced results in the field of deep learning to complete the nonlinear mapping from echo features to target images. The overall structure of the framework is as follows: Figure 2 As shown, it mainly includes three parts: training data generation based on MIMO radar imaging model, deep network forward propagation, and error back propagation based on reconstruction loss. Based on the above framework, the present invention specifically designs a non-orthogonal waveform MIMO radar imaging method based on deep learning. The proposed imaging method is introduced from three aspects: deep network structure, training data generation, and network training strategy.

[0121] Train the deep network model to obtain a trained deep network model, including:

[0122] Using the training set, the deep network model is trained alternately by iteratively performing forward propagation and error back propagation, and the mean square error between the ideal target image and the target image generated by the deep network model is used as the loss function to continuously optimize and adjust the parameters of the deep network model to obtain a trained deep network model.

[0123] The method for constructing the characteristic imaging module is specifically as follows:

[0124] Each convolution operation or deconvolution operation is connected to a batch normalization process and a ReLu activation function; the deep network model contains a total of 5 convolution layers, and the number of input / output channels of each layer is 2 / 32, 32 / 64, 64 / 64, 64 / 64, and 64 / 128 respectively; it contains 5 deconvolution layers, and the number of input / output channels of each layer is 128 / 64, 64 / 64, 64 / 64, 64 / 32, and 32 / 1 respectively; MaxPool and MaxUnpool represent the maximum pooling layer and the maximum unpooling layer respectively, the pooling or unpooling window size is 2×2, the window span is 2, and there are 4 pooling layers and 4 unpooling layers.

[0125] It should be noted that: in view of the specific characteristics of the non-orthogonal waveform MIMO radar imaging problem, the present invention constructs a MIMO radar learning imaging network (MIMO radar learning imaging network, MIMO-LI-Net), the network structure is as follows Figure 2 As shown. The echo conversion module is constructed according to the range-Doppler imaging algorithm, including a matched filter network and an azimuth compression network: the matched filter network completes the range dimension compression of the target echo by inputting a non-orthogonal transmission waveform set; the azimuth compression network completes the two-dimensional compression of the target echo by echo rearrangement, azimuth compression and other processes according to the MIMO array structure configuration. In a specific implementation, the present invention can improve the echo conversion efficiency with the help of Fast Fourier Transform (FFT) by designing a uniform equivalent transceiver array and frequency domain matched filtering.

[0126] Due to the coupling interference between non-orthogonal waveforms and possible cross-range unit migration, the above two-dimensional compression results cannot achieve ideal imaging effects. Nevertheless, the echo transform process still extracts effective target features from the MIMO radar complex echo, which can provide a good initial input for the subsequent feature imaging process. The feature imaging module is constructed using a convolutional neural network and can be generally divided into two parts: encoding and decoding. Figure 3As shown in the figure, Conv and ConvTrans represent convolutional layers and deconvolutional layers respectively, (3×3, 1) indicates that the size of the convolutional kernel or deconvolutional kernel is 3×3 and the stride is 1, and each convolutional operation or deconvolution operation is followed by a batch normalization (BN) process and a ReLu activation function. The network contains 5 convolutional layers, and the number of input / output channels of each layer is 2 / 32, 32 / 64, 64 / 64, 64 / 64, and 64 / 128 respectively; it contains 5 deconvolutional layers, and the number of input / output channels of each layer is 128 / 64, 64 / 64, 64 / 64, 64 / 32, and 32 / 1 respectively. MaxPool and MaxUnpool represent the maximum pooling layer and the maximum unpooling layer respectively, (2×2, 2) means that the pooling or unpooling window size is 2×2 and the window span is 2, the maximum pooling and maximum unpooling layers use the same position index, and the network contains 4 pooling layers and 4 unpooling layers. Unlike general natural image processing, the present invention sends the real and imaginary parts of the echo transformation results as two separate channels to the CNN network. Compared with the input single-channel modulus data, the input dual-channel data contains the complete information of the transformed echo, which is conducive to image feature extraction. In addition, the feature imaging module establishes a jump connection between the encoding and decoding parts of the network, which helps to retain the detailed information of the image during the image reconstruction process.

[0127] In addition to the MIMO radar target echo, network training also requires the corresponding output target image. The present invention uses the target imaging result under the ideal orthogonal waveform condition as the target image. The specific method is as follows: using the equivalent transmit-receive shared array echo to replace the MIMO array echo shown in formula (9), at this time, the echoes of different transmission waveforms are not aliased, and there is no coupling interference; using the back-projection algorithm to image the equivalent transmit-receive shared array echo, even if there is cross-range unit migration, the algorithm can still obtain a focused image; suppressing the side lobes through two-dimensional windowing, and using twice the array length and twice the signal bandwidth to improve the image resolution.

[0128] The present invention can drive the imaging network to automatically learn the mapping of non-orthogonal waveform echoes to ideal target images from a large number of training samples through effective network learning. Figure 3 In the MIMO-LI-Net shown, the echo transform module does not contain learnable parameters, and the network only needs to learn the CNN parameters in the feature imaging module. The present invention uses the mean square error between the reconstructed target image and the ideal target image as the loss function and uses the back propagation algorithm to train the network.

[0129] The present invention further provides a verification process:

[0130] (1) First, the proposed imaging network is trained by generating simulation data, and then two sets of test data are used to verify the effectiveness of the proposed imaging method. The simulation experiment uses a 4-transmit 20-receive MIMO linear array, with the transmitting and receiving array element spacings of 3m and 60m respectively, the radar center frequency of 10GHz, the transmitting waveform is a multi-phase coded signal of 450 code elements, and the signal bandwidth is 300MHz. The normalized autocorrelation and (and sum signal) cross-correlation waveforms of the four signals are shown in Figure 4 As shown. It can be seen that the autocorrelation sidelobes of the four signals are low but the cross-correlation sidelobes are high, and the waveforms do not meet the ideal orthogonality. The target is set to be 5km away from the center of the array, and the corresponding theoretical imaging resolutions of distance and azimuth are 0.5m and 0.625m respectively. The training samples are generated using the method in Section 3.2. The number of scatterers for each target is uniformly randomly generated in the range of 0 to 500, and are all within the range of 50m×50m. The scattering coefficient of the scatterer is generated by a standard complex Gaussian distribution. Each scatterer is set to a distributed scattering center or a local scattering center with equal probability, and the length of the distributed scattering center is a uniform random number in the range of 0 to 2m. The size of the ideal target image is set to 256×256. A total of 40,000 MIMO radar echo and target image training sample pairs are generated to form a training data set for Figure 3 The MIMO-LI-Net network shown in the figure is trained. The training platform is Intel Core i9-9900 CPU and NVIDIA GeForce RTX 2080Ti GPU. The Adam optimizer is used for training in the PyTorch framework, and the learning rate is set to 0.0001. The trained imaging network is verified by constructing test data.

[0131] (2) Test the simulation point target data

[0132] First, consider an aircraft target model consisting of ideal scattering points, and the imaging parameters are consistent with the training data. Using the MIMO radar echo corresponding to the target and the ideal target image, the latter is as follows Figure 5As shown. Four existing MIMO radar imaging methods are selected for comparison. (1) Matched filtering method: distance compression is performed based on formula (4), and the result is subjected to azimuth FFT to achieve azimuth compression. (2) 1D-SR method: a one-dimensional vector sparse recovery (One Dimensional Sparse Recovery, 1D-SR) model is constructed to directly solve the target two-dimensional image. It should be noted that since the dimension of the one-dimensional vector model is very high (under the experimental parameter conditions, the perception matrix dimension is 11000×8000), the use of a sparse solution algorithm involves matrix inversion operations, and the computational efficiency is too low. Therefore, the present invention adopts an orthogonal matching pursuit (Orthogonal Matching Pursuit, OMP) algorithm with higher computational efficiency to solve the sparse model, and the sparsity parameter is set to 400. (3) MB-SL0 method: the MB-SL0 algorithm is used for distance compression, and then azimuth compression is achieved through FFT. (4) CNN method: The CNN network is trained for MIMO radar imaging. The network input is the imaging result of the matched filtering method. The network structure is similar to MIMO-LI-Net, but does not contain jump links between network layers. The network training method is consistent with the training method of the proposed imaging network.

[0133] The imaging results of the four contrast methods are as follows Figure 6 As shown. Figure 6 (a) It can be seen that due to the non-ideal orthogonality of the transmitted waveform, the imaging results of the matched filtering method have serious mutual coupling noise. Figure 6 (b) is the imaging result of the 1D-SR method. It can be seen that the 1D-SR method can effectively suppress the mutual coupling noise, but due to the basis mismatch problem of the sparse recovery model, the scattering points of the sparsely reconstructed image have a certain position offset compared with the ideal target image; and the sparsity parameter of the OMP algorithm needs to be set manually. When the sparsity is small, the scattering points will be lost, and when the sparsity is large, the computational complexity will increase sharply. Figure 6 (c) It can be seen that the MB-SL0 method can effectively suppress the range coupling interference caused by waveform non-orthogonality, but the side lobes are relatively high in azimuth. In addition, due to the decoupling of range and azimuth, when there is cross-range unit migration at scattering points farther away in azimuth, the imaging results will be defocused. Figure 6 (d) is the imaging result of the CNN method. It can be seen that the CNN method can effectively suppress the image coupling interference caused by non-orthogonal waveforms. However, due to the lack of targeted design of the network structure, the network has a poor recovery effect on weak scattering points in the image. The imaging results of the MIMO-LI-Net trained by the present invention are shown in Figure 2. Figure 7 As shown in the figure, it can be seen that most of the scattering points on the target are well restored and the coupling interference between non-orthogonal waveforms is effectively suppressed.

[0134] In order to quantitatively analyze the above imaging results, the normalized mean square error (NMSE) and target background ratio (TBR) are used as evaluation indicators, and the threshold of TBR is set to 0.0001. The calculation results of the indicators of each method are shown in Table 1. It can be seen that the imaging indicators of the matched filtering method are the worst, while the invented method has achieved the best results in both NMSE and TBR indicators. At the same time, in order to compare the computational efficiency of various methods, Table 1 also gives the running time of different methods under the CPU. It can be seen that the matched filtering method has the highest computational efficiency; the 1D-SR method has the longest running time, reaching nearly 100s, and with the increase of the imaging model dimension and the increase of the sparsity of the OMP algorithm, the algorithm's demand for storage space and computing resources will increase sharply.

[0135] The running time of the MB-SL0 method, the CNN method and the method of this invention is relatively short, all of which are on the order of 0.1 s. Moreover, compared with the MB-SL0 method, the running efficiency of the CNN method and the method of this invention is less affected by the increase in the dimension of the imaging model.

[0136] Table 1 Performance comparison of different methods

[0137]

[0138] (3) Electromagnetic calculation data test

[0139] This section uses electromagnetic computing to simulate the target and further test the performance of the proposed imaging method. The CAD model of the target is shown in Figure 8 As shown in Figure 1, physical optics and moment method are used as electromagnetic scattering calculation methods to generate the target scattering field, and the MIMO radar echo of the target can be generated using equation (9). The imaging results of the target by each comparison method and the proposed imaging method are shown in Figure 1. Fig. 9 , Fig.10 As shown in Figure 2, it can be seen that for non-ideal point scattering model targets, the proposed method can also effectively suppress the mutual coupling noise of non-orthogonal waveforms while retaining the main scattering center of the target. Compared with the results of other comparative methods, the results of the proposed method have lower side lobes and fewer false points.

[0140] In summary, the present invention proposes a non-ideal orthogonal waveform MIMO radar imaging method based on deep learning. With the powerful nonlinear fitting ability of deep learning, a MIMO radar deep imaging network is constructed and trained on massive simulation training data to automatically learn the mapping relationship between non-orthogonal waveform MIMO radar echoes and ideal target images. The learned imaging network can effectively suppress the coupling interference between non-ideal orthogonal transmission waveforms and improve the imaging quality of MIMO radar. The effectiveness and timeliness of the proposed method are verified by experiments on point scattering model data and electromagnetic scattering calculation data.

[0141] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-orthogonal waveform MIMO radar imaging method based on deep learning, characterized in that: The following steps are involved: Construct a MIMO radar echo signal model to obtain the target echo signal in the entire imaging scene; Under the condition that the transmission waveform satisfies the ideal orthogonality, the signals of different waveforms in the target echo signal are separated to obtain waveform vectors corresponding to the different waveforms, and the waveform vectors are compressed in azimuth to obtain the ideal target image of the MIMO radar; The frequency domain echo signals at different receiving array elements in the MIMO radar echo signal model and the ideal target image are used as training sets to train the constructed deep network model to obtain a trained deep network model; The trained deep network model is used to perform target imaging on the actual target echo signal.

2. The non-orthogonal waveform MIMO radar imaging method based on deep learning according to claim 1, characterized in that: The deep network model includes: an echo transformation module and a feature imaging module; The echo transform module includes: a matched filter network and an azimuth compression network; The matched filter network is used to complete the range dimension compression of the target echo signal by inputting a non-orthogonal transmission waveform set; The azimuth compression network is used to achieve two-dimensional compression of the target echo signal through echo rearrangement and azimuth compression according to the structure of the MIMO radar echo signal model; The feature imaging module is constructed using a convolutional neural network and is used to establish a skip connection between the encoding and decoding of the neural network model to retain the detail information of the image during the reconstruction of the target image.

3. The non-orthogonal waveform MIMO radar imaging method based on deep learning according to claim 2, characterized in that: The method for constructing the characteristic imaging module is specifically as follows: Each convolution operation or deconvolution operation is connected to a batch normalization process and a ReLu activation function; the deep network model contains a total of 5 convolution layers, and the number of input / output channels of each layer is 2 / 32, 32 / 64, 64 / 64, 64 / 64, and 64 / 128 respectively; it contains 5 deconvolution layers, and the number of input / output channels of each layer is 128 / 64, 64 / 64, 64 / 64, 64 / 32, and 32 / 1 respectively; MaxPool and MaxUnpool represent the maximum pooling layer and the maximum unpooling layer respectively, the pooling or unpooling window size is 2×2, the window span is 2, and there are 4 pooling layers and 4 unpooling layers.

4. The non-orthogonal waveform MIMO radar imaging method based on deep learning according to claim 1, characterized in that: The MIMO radar echo signal model is constructed to obtain the target echo signal in the entire imaging scene, specifically: The MIMO radar echo signal model includes: a transmitting array element with a spacing of Nd, a receiving array element with a spacing of d; a target with U scattering points; the qth scattering point Q on the target, and the scattering point Q and the mth transmitting array element T m The distance between m Q and the scattering point Q and the nth receiving array element R n The distance between n Q; The expression of the echo signal received by the nth receiving array element is: In the above formula, τ q,mn =(T m Q+R n Q) / c, τ q,mn The meaning is that the electromagnetic wave is transmitted by the mth transmitting element T m To the qth scattering point Q, and then to the nth receiving array element R n The propagation delay; n = 0, 1, L, N-1, c is the speed of light, σ q represents the scattering coefficient of the scattering point Q, λ is the radar operating wavelength; when the target sight angle is known, the echo envelopes of different transceiver channels are aligned by pre-compensating the envelope, τ q，mn No longer changes with m and n, then τ in formula (1) q，mn Denoted as τ q ; Discretize the range imaging area. Assume that the range imaging range is ΔR. Then the number of sampling points of the imaging area discretization unit is: K = 2ΔR·f s / c, where f s is the distance sampling frequency, using f s Transmit waveform Sampling is performed to obtain the transmitted waveform vector The length is L; Assume that the scattering point Q is located in the kth distance unit, k = 0, 1, L, k-1, and the distance kernel matrix of the kth distance grid is: In the above formula, 0 k×L represents a k×L-dimensional zero matrix; I L×L Represents an L×L dimensional identity matrix; 0 (K-k)×L represents a (Kk)×L-dimensional zero matrix; Then the vector form of the echo signal of the scattering point Q received by the nth receiving array element is expressed as: in: For R n The target complex scattering coefficient vector corresponding to the k-th range unit received at , the dimension of the target complex scattering coefficient vector is the number of transmitting array elements M, and S represents the transmitting waveform matrix; The target echo signal is the superposition of the echo signals of all scattering points Q on the target. Let T = [T0, T1, ..., T K-1 ]∈C (K+L)×LK , Then the nth receiving array element R n The expression of the target echo signal of the entire imaging scene received at is: and n =Te S·θ n +ε (3) in, For R n The corresponding target complex-valued range image at , ε represents the receiving noise.

5. The non-orthogonal waveform MIMO radar imaging method based on deep learning according to claim 1, characterized in that: The signals of different waveforms in the target echo signal are separated to obtain waveform vectors corresponding to different waveforms, and azimuth compression is performed to obtain the target image of the MIMO radar, specifically: The separation result of the target echo signal after the mth matched filter is expressed as: In the above formula, y mn Represents the target echo signal and the mth transmitted waveform The corresponding waveform components, where: Substituting equation (5) into equation (4), we can obtain y mn =D m ·i n (6) in, D m is the matched filter reconstruction matrix corresponding to the mth transmitted waveform; Assume that the scattering point Q is located on the kth distance grid, and the scattering coefficient vector θ to be restored is n It is expressed as: in, Then formula (6) can be further expressed as: Among them, y mn The result for the kth distance unit is: In the above formula, represents the transmission waveform vector corresponding to the mth transmission waveform, The * in the upper right corner represents the conjugate operation of the vector; When the mth transmission waveform m=m', When m≠m', get That is, the separation result only contains the target signal corresponding to the mth transmission waveform, and there is no component of other waveforms; the complex scattering coefficients obtained by different transceiver channels After integration, the ideal target image of MIMO radar is obtained through azimuth compression.

6. The non-orthogonal waveform MIMO radar imaging method based on deep learning according to claim 4, characterized in that: The frequency domain echo signals at different receiving array elements in the MIMO radar echo signal model and the ideal target image are used as training sets; wherein the method for obtaining the frequency domain echo signals includes: The echo signal received by the nth receiving array element is converted into a frequency domain form to obtain a frequency domain echo signal, which is specifically: In the above formula, For the transmission waveform The frequency domain form of e q，m，n (f) is the scattered field generated by the qth scattering point in the mth transmission-nth reception channel; Among them, e q，m，n The expression of (f) is: In formula (10), f c is the center frequency, φ q，m，n is the azimuth angle between the qth scattering center and the equivalent transmitting and receiving array element corresponding to the mth transmitting and nth receiving channel, ΔR q，m，n is the sum of the wave paths from the qth scattering point to the mth transmitting array element and the nth receiving array element, is the parameter associated with the qth scattering center: q is the scattering coefficient; α q is a parameter describing the relationship between the scatterer frequency and geometry; L q , Represents the length and direction angle of the distribution scattering center; γ g To describe the relationship between the local scattering center and direction.

7. The non-orthogonal waveform MIMO radar imaging method based on deep learning according to claim 6, characterized in that: Train the deep network model to obtain a trained deep network model, including: Using the training set, the deep network model is trained alternately by iteratively performing forward propagation and error back propagation, and the mean square error between the ideal target image and the target image generated by the deep network model is used as the loss function to continuously optimize and adjust the parameters of the deep network model to obtain a trained deep network model.

Images