Doppler through-the-wall radar localization method based on neural network time-frequency analysis

Abstract

The application discloses a Doppler through-wall radar positioning method based on neural network time-frequency analysis, comprising the following steps: constructing a multi-target motion echo signal after demodulation of a Doppler through-wall radar; performing time-frequency analysis processing on the signal through Winger-Ville distribution to obtain corresponding time-frequency distribution and construct a fusion data set; constructing a multi-input neural network model, training the model by using the fusion data set, updating the model parameters, and constructing an optimal time-frequency distribution prediction model; acquiring real-time collected echo signals, performing time-frequency analysis processing through the distribution to obtain corresponding time-frequency distribution, and constructing a real-time fusion data set; obtaining a predicted time-frequency distribution; extracting the multi-target instantaneous frequency of the predicted time-frequency distribution, estimating the distance and angle information of the target through a Doppler through-wall radar positioning algorithm, and completing the positioning of the multi-target; and the method can eliminate the interference of time-frequency distribution cross terms, improve positioning accuracy, and improve time-frequency analysis effect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of moving target positioning, and particularly relates to a Doppler through-wall radar positioning method based on neural network time-frequency analysis. BACKGROUND

[0002] In recent years, Doppler through-wall radar has been widely applied in the field of target positioning, mainly in civil and military fields, including indoor positioning, disaster rescue, and anti-terrorism reconnaissance. As one of the methods for non-contact target identification, radar technology has good long-distance target identification capability, and with the characteristics of high precision and strong anti-interference capability, it can realize tracking and positioning of targets. The Doppler through-wall radar transmits electromagnetic waves to contact moving targets through obstacles, and processes the echo signal at the receiving end, filters and demodulates the echo signal by using signal processing technology, and obtains the information of the moving target, so as to realize the synthesis of the moving track of the target.

[0003] Based on the traditional target positioning method, the echo signal of the moving target needs to be analyzed in time and frequency, and then the peak value extraction method is used to obtain the instantaneous frequency of the target in the time-frequency distribution, and the motion track of the target is fitted through the instantaneous frequency component; when the positioning of multiple targets is studied, the WVD time-frequency method can eliminate the "frequency ambiguity" phenomenon caused by windowed time-frequency transform, and the corresponding time-frequency distribution energy is concentrated; but WVD will produce cross-term interference, which seriously affects the effect of signal processing, and then the motion track of the target obtained by fitting appears obvious deviation.

[0004] At present, the cross-term in the time-frequency distribution is eliminated by the time-frequency analysis method, including PWVD and SPWVD, but these methods still have the following shortcomings: 1. When processing non-stationary signals, there are still a small amount of cross-terms; 2. In time-frequency analysis, when the signal is interfered by strong noise, the time-frequency analysis result may be affected; 3. Since the above time-frequency analysis method needs to adjust the appropriate smoothing parameter, the inappropriate parameter will cause the time-frequency analysis result to be excessively smoothed or sharp, and the large number of parameters makes the calculation complexity high; 4. Different parameters need to be selected for different signals to adapt, so the complexity of use is high.

[0005] In summary, the current Doppler radar positioning method based on time-frequency analysis is mostly affected by the cross-term in the time-frequency distribution, and at the same time, the positioning accuracy is not enough and the resolution is low. SUMMARY

[0006] The purpose of the present application is to provide a Doppler through-wall radar positioning method based on neural network time-frequency analysis, which can eliminate the interference of cross-terms in time-frequency distribution, improve positioning accuracy, and improve the effect of time-frequency analysis.

[0007] The Doppler through-wall radar localization method based on time-frequency analysis using neural networks provided by this invention includes the following steps:

[0008] S1. Construct the echo signal of multi-target motion after demodulation by Doppler through-wall radar;

[0009] S2. Perform time-frequency analysis on the echo signal constructed in step S1 using the Winger-Ville distribution to obtain the corresponding time-frequency distribution, and use the obtained time-frequency distribution and the constructed echo signal to construct a fused dataset;

[0010] S3. Construct a multi-input neural network model. Use the fusion dataset constructed in step S2 to train the constructed model, update the model parameters, and construct the optimal time-frequency distribution prediction model.

[0011] S4. Acquire the real-time acquired echo signal, perform time-frequency analysis processing through the Winger-Ville distribution to obtain the corresponding time-frequency distribution, and simultaneously construct a real-time fusion dataset using the real-time acquired echo signal and the corresponding time-frequency distribution;

[0012] S5. Using the real-time fusion dataset constructed in step S4, the predicted time-frequency distribution is obtained through the optimal time-frequency distribution prediction model constructed in step S3;

[0013] S6. Extract the instantaneous frequencies of the multi-targets obtained from the predicted time-frequency distribution in step S5, estimate the distance and angle information of the targets using the Doppler through-wall radar localization algorithm, and complete the localization of the multi-targets;

[0014] Step S1, which involves constructing the echo signal of multi-target motion after demodulation by Doppler through-wall radar, specifically includes:

[0015] Construct several random parameters and establish a fourth-order random instantaneous frequency function using the following formula:

[0016] IF i (t)=H i +2tJ i +3K i t 2 +4L i t 3

[0017] Among them, IF i (t) represents the instantaneous frequency of the i-th target, where the instantaneous frequency is a training parameter of the multi-input neural network model; H i J represents the first frequency offset of the i-th signal; i K represents the second frequency offset of the i-th signal; i L represents the third frequency shift of the i-th signal; idenotes the fourth frequency deviation of the ith signal; t denotes time;

[0018] The echo signal data simulating the multi-target motion is obtained, one-dimensional echo signal data is obtained, the received target echo signal x(t) is constructed, and the following formula is used to represent:

[0019]

[0020] Wherein, each signal includes N sub-signals; N represents a normal distribution; A i (t) represents the amplitude of the ith carrier signal; M represents a constant term of the signal;

[0021] Step S2: performing time-frequency analysis processing on the echo signal constructed in step S1 by Winger-Ville distribution to obtain a corresponding time-frequency distribution, and constructing a fusion data set using the obtained time-frequency distribution and the constructed echo signal, specifically including:

[0022] The time-frequency analysis processing is performed by Winger-Ville distribution, and the following formula is used to represent the processing process of Winger-Ville distribution:

[0023] x WVD (t,f)=∫x(t+τ / 2)·x * (t-τ / 2)·exp(-j2πfτ)dτ

[0024] Wherein, x WVD represents the time-frequency distribution obtained after time-frequency analysis processing by Winger-Ville distribution; x * (t) represents the conjugate of x(t); τ represents time delay; f represents instantaneous frequency;

[0025] The time-frequency distribution matrix obtained by the time-frequency analysis processing by Winger-Ville distribution is used to construct a fusion data set with the echo signal constructed in step S1;

[0026] Step S3: constructing a multi-input neural network model, training the constructed model using the fusion data set constructed in step S2, updating the model parameters, and constructing an optimal time-frequency distribution prediction model, specifically including:

[0027] (3-1) Constructing a multi-input neural network model:

[0028] The multi-input neural network model includes a multi-input coding structure, a cross-term suppression structure, and an output part.

[0029] 1) The multi-input coding structure includes an input signal, a two-dimensional convolution coding, and a one-dimensional complex convolution coding Convl.

[0030] The input includes a one-dimensional complex signal and a Winger-Ville distribution-time-frequency distribution (WVD-TFR), wherein the one-dimensional complex signal includes a signal real part and a signal imaginary part;

[0031] The two-dimensional convolution coding includes three two-dimensional convolution layers;

[0032] The one-dimensional complex convolution coding Convl includes two one-dimensional convolution layers;

[0033] The WVD-TFR of the input is processed by the two-dimensional convolution coding to obtain a two-dimensional convolution coded output matrix C;

[0034] The matrix A obtained by one-dimensional convolution of the real part signal and the matrix B obtained by one-dimensional convolution of the imaginary part signal are defined, and the output dimension is calculated using the following formula:

[0035] (A-B)+(A+B)i

[0036] The output dimension obtained by the above calculation is recombined to obtain a matrix D, which is then concatenated with the matrix C to obtain an output matrix E as the input of the cross-term suppression structure;

[0037] 2) In the cross-term suppression structure, 15 residual structure encoders and 15 residual structure decoders are included; each residual structure encoder includes a two-dimensional convolution layer with a size of (3, 3) and a Relu activation function; each residual structure decoder includes a two-dimensional deconvolution layer with a size of (3, 3) and a Relu activation function;

[0038] In the connection of the encoder and the decoder, the output of the first encoder is taken as the input of the second encoder and also as the input of the fifteenth decoder; the output of the second encoder is taken as the input of the third encoder and also as the input of the fourteenth decoder; the output of the third encoder is taken as the input of the fourth encoder and also as the input of the thirteenth decoder; the output of the fourth encoder is taken as the input of the fifth encoder and also as the input of the twelfth decoder; the output of the fifth encoder is taken as the input of the sixth encoder and also as the input of the eleventh decoder; the output of the sixth encoder is taken as the input of the seventh encoder and also as the input of the tenth decoder; the output of the seventh encoder is taken as the input of the eighth encoder and also as the input of the ninth decoder; the output of the eighth encoder is taken as the input of the ninth encoder; the output of the ninth encoder is taken as the input of the tenth encoder and also as the input of the seventh decoder; the output of the tenth encoder is taken as the input of the eleventh encoder and also as the input of the sixth decoder; the output of the eleventh encoder is taken as the input of the twelfth encoder and also as the input of the fifth decoder; the output of the twelfth encoder is taken as the input of the thirteenth encoder and also as the input of the fourth decoder; the output of the thirteenth encoder is taken as the input of the fourteenth encoder and also as the input of the third decoder; the output of the fourteenth encoder is taken as the input of the fifteenth encoder and also as the input of the second decoder; the output of the fifteenth encoder is taken as the input of the first encoder;

[0039] In step 1), the matrix E fused is not only taken as the input of the residual structure encoder, but also fused with the feature matrix F output by the residual structure decoder to be taken as the output of the cross term suppression structure together, to obtain the output matrix G of the cross term suppression structure;

[0040] 3) The output matrix G of the cross term suppression structure obtained in step 2) is taken through a two-dimensional convolution layer to obtain a feature matrix H; finally, through a two-dimensional convolution layer, the final optimized spectrum is obtained after dimension reduction; through the back propagation processing of the optimized spectrum, the optimization of the parameters in the network is realized;

[0041] (3-2) Training model:

[0042] The multi-input neural network model constructed in step (3-1) is trained by using the fusion data set constructed in step S1, and the validation set data is used to estimate the accuracy of the model;

[0043] Part of the data in the fusion data set constructed in step S1 is selected to form the validation set data;

[0044] (3-3) Constructing an optimal time-frequency distribution prediction model:

[0045] Repeat the above steps (3-2), and adjust the parameters of the model based on the previous training results until the set conditions are met. Stop the repeated training of the model and obtain the optimal time-frequency distribution prediction model.

[0046] The set conditions include determining whether the time-frequency resolution of the augmented time-frequency spectrogram output by the multi-input neural network model is within the set range. If the time-frequency resolution is within the set range, the model training is stopped and the corresponding model is determined as the optimal time-frequency distribution prediction model. If the time-frequency resolution is not within the set range, the network model is trained until the time-frequency resolution is within the set range.

[0047] Step S6 involves extracting the instantaneous frequencies of multiple targets from the predicted time-frequency distribution obtained in step S5, estimating the target's distance and angle information using a Doppler through-wall radar localization algorithm, and completing the localization of multiple targets. Specifically, this includes:

[0048] (6-1) Extracting the instantaneous frequencies of multiple targets in the predicted time-frequency distribution:

[0049] The peak detection method is used to extract the instantaneous frequencies of multiple targets in the predicted time-frequency distribution;

[0050] The peak detection method is expressed by the following formula:

[0051]

[0052] in, The x represents the instantaneous frequency curve obtained in the i-th iteration; WVD (t,f) represents the time-frequency distribution obtained after time-frequency analysis of the Winger-Ville distribution;

[0053] (6-2) Doppler through-wall radar localization algorithm:

[0054] The Doppler through-wall radar localization algorithm includes a range estimation algorithm and an angle-of-arrival (AOA) estimation algorithm. The range estimation algorithm estimates the distance between the target and the transmitter in real time based on the target's instantaneous frequency. The AOA estimation algorithm estimates the angle between the target and the transmitter in real time based on the target's instantaneous frequency.

[0055] (1) Distance estimation algorithm:

[0056] via transmitter T x and receiver R x1 Obtain the distance between the target and the radar;

[0057] Let f1 and f2 represent the two carrier frequencies of the transmitted signal from the transmitter, respectively; the transmitted signal of the Doppler through-wall radar is represented by the following formula:

[0058]

[0059] wherein T x (t) denotes that the transmitted signal comprises two carrier signals; φ1 denotes the initial phase of the first carrier signal, and φ2 denotes the initial phase of the second carrier signal;

[0060] The echo signal received by the radar receiver is processed to obtain the following formula:

[0061]

[0062]

[0063] wherein R x1 (t) denotes the echo signal received by the receiver R x1 ; R1 denotes the distance between the target and the transmitter; R2 denotes the distance between the target and the receiver R x1 ; R x2 (t) denotes the echo signal received by the receiver R x2 ; R3 denotes the distance between the target and the receiver R x2 ; and c denotes the speed of electromagnetic waves;

[0064] By approximating R1≈R2, the phases of the two receivers are expressed by the following formula:

[0065]

[0066]

[0067] wherein φ1 denotes the phase of the receiver R x1 ; and φ2 denotes the phase of the receiver R x2 ;

[0068] The distance formula of the target motion is expressed by the following formula:

[0069]

[0070] wherein R denotes the distance of the target; denotes the instantaneous frequency of the echo signal at the carrier frequency f1 of the receiver R x1 ; denotes the instantaneous frequency of the echo signal at the carrier frequency f1 of the receiver R x2 ;

[0071] (2) Angle of arrival estimation algorithm:

[0072] The transmitter T x and the receivers R x1 , Rx2 the angle of arrival between the target and the radar;

[0073] Let d represent the distance between the receiver R x1 and the target; x2

[0074] Let the phase difference of the echo signal between the receiver R x1 and the target be represented by the following formula: x2

[0075]

[0076] wherein θ represents the angle of arrival of the target; λ1 represents the wavelength corresponding to the echo signal with the carrier frequency f1;

[0077] The angle of arrival is calculated by the following formula:

[0078]

[0079] wherein, represents the instantaneous frequency of the echo signal with the carrier frequency f1; x2

[0080] (6-3) Positioning multiple targets:

[0081] The distance and the angle of arrival of the target calculated by step (6-2) are represented by the following formula in the two-dimensional Cartesian coordinate system:

[0082] x = R sin θ

[0083] y = R cos θ

[0084] wherein x represents the horizontal coordinate of the target movement; y represents the vertical coordinate of the target movement.

[0085] The Doppler through-wall radar positioning method based on neural network time-frequency analysis provided by the present application realizes the common input of the echo signal and the time-frequency distribution information obtained after the Wigner-Ville time-frequency analysis processing by constructing a multi-input neural network model; the model composed of the multi-input coding structure network and the cross-term suppression structure network is used to predict the time-frequency spectrum and eliminate the cross-term interference problem; the positioning accuracy of the method is improved, and the time-frequency analysis effect is improved. BRIEF DESCRIPTION OF DRAWINGS

[0086] Figure 1 The figure is a method flowchart of the method of the present application.

[0087] Figure 2 The figure is a network structure diagram of the method of the present application. ​​​

[0088] Figure 3 Doppler through-the-wall radar structure diagram for the method of the present application.

[0089] Figure 4 Method result comparison diagram of the method of the present application and the result of STFT, Hough-linear method: Figure 4 (a) is an IF estimation diagram obtained using the STFT method; Figure 4 (b) is a trajectory fitting diagram obtained using the STFT method; Figure 4 (c) is an IF estimation diagram obtained using the Hough-linear method;

[0090] Figure 4 (d) is a trajectory fitting diagram obtained using the Hough-linear method; Figure 4 (e) is an IF estimation diagram obtained using the method of the present application; Figure 4 (f) is a trajectory fitting diagram obtained using the method of the present application. DETAILED DESCRIPTION

[0091] As Figure 1 shown is a method flow diagram of the method of the present application: the Doppler through-the-wall radar positioning method based on neural network time-frequency analysis provided by the present application includes the following steps:

[0092] S1. Constructing the echo signal of multi-target motion after demodulation of the Doppler through-the-wall radar; specifically including:

[0093] Constructing a plurality of random parameters, and establishing a 4th random instantaneous frequency function using the following formula:

[0094] IF i (t) = H i + 2tJ i + 3K i t 2 + 4L i t 3

[0095] Wherein, IF i (t) represents the instantaneous frequency of the i-th target, and the instantaneous frequency is a training parameter of the multi-input neural network model; H i represents the first frequency offset of the i-th signal; J i represents the second frequency offset of the i-th signal; K i represents the third frequency offset of the i-th signal; L i represents the fourth frequency offset of the i-th signal; t represents time;

[0096] In the method of the present application, H iUniform distribution in the range [-150, 150]; J i Uniform distribution in the range [-100, 100]; K i Uniform distribution in the range [-50, 50]; L i Uniform distribution in the range [-25, 25];

[0097] The echo signal data of the simulated multi-target motion is obtained, one-dimensional echo signal data is constructed, and the received target echo signal x(t) is expressed by the following formula:

[0098]

[0099] Each signal includes N sub-signals; N represents a normal distribution; A i (t) represents the amplitude of the i-th carrier signal; M represents a constant term of the signal;

[0100] In the method of the application, the echo signal data of the multi-target motion collected by the Doppler wall-penetrating radar module is simulated; N is a normal distribution of (1, 6); the value range of M is [-25, 25];

[0101] S2. Time-frequency analysis and processing of the echo signal constructed in step S1 by Winger-Ville distribution is performed to obtain the corresponding time-frequency distribution, and the obtained time-frequency distribution and the constructed echo signal are used to construct a fusion data set; specifically including:

[0102] The time-frequency analysis and processing is performed by Winger-Ville distribution, and the processing process of Winger-Ville distribution is expressed by the following formula:

[0103] x WVD (t,f)=∫x(t+τ / 2)·x * (t-τ / 2)·exp(-j2πfτ)dτ

[0104] Wherein, x WVD represents the time-frequency distribution obtained after time-frequency analysis and processing by Winger-Ville distribution; x * (t) represents the conjugate of x(t); τ represents the time delay; f represents the instantaneous frequency;

[0105] The time-frequency distribution matrix obtained by time-frequency analysis and processing by Winger-Ville distribution is used together with the echo signal constructed in step S1 to construct a fusion data set;

[0106] S3. Construct a multi-input neural network model, use the fusion data set constructed in step S2 to train the constructed model, update the model parameters at the same time, and construct an optimal time-frequency distribution prediction model; specifically including:

[0107] As Figure 2 The network structure of the method is shown in the figure:

[0108] (3-1) Construct a multi-input neural network model:

[0109] The multi-input neural network model includes a multi-input encoding structure, a cross-term suppression structure, and an output part.

[0110] 1) The multi-input encoding structure includes an input signal, a two-dimensional convolutional encoding, and a one-dimensional complex convolutional encoding Convl.

[0111] The input includes a one-dimensional complex signal and a Winger-Ville distribution-time-frequency distribution (WVD-TFR), wherein the one-dimensional complex signal includes a signal real part and a signal imaginary part.

[0112] The two-dimensional convolutional encoding includes three two-dimensional convolutional layers.

[0113] The one-dimensional complex convolutional encoding Convl includes two one-dimensional convolutional layers.

[0114] The WVD-TFR of the input is processed by the two-dimensional convolutional encoding to obtain a two-dimensional convolutional encoding output matrix C.

[0115] The matrix A obtained by one-dimensional convolution of the real part signal and the matrix B obtained by one-dimensional convolution of the imaginary part signal are used to calculate the output dimension using the following formula:

[0116] (A-B)+(A+B)i

[0117] The output dimension obtained by the above calculation is then processed by recombining the matrix D and the matrix C to obtain an output matrix E, which is used as the input of the cross-term suppression structure.

[0118] In the multi-input encoding structure of the method, the input complex signal is a one-dimensional complex signal with a signal dimension of (2, 256), including a real part signal with a dimension of (1, 256) and an imaginary part signal with a dimension of (1, 256).

[0119] The one-dimensional complex convolutional layer Convl includes two one-dimensional convolutional layers; the one-dimensional convolutional layer has a size of (1, 31), an input channel of 1, and an output channel of 2048; the one-dimensional complex signal (1, 2, 256) is divided into a real part signal (1, 1, 256) and an imaginary part signal (1, 1, 256), which are respectively processed by the one-dimensional convolutional layer to obtain an output dimension of (2048, 1, 256).

[0120] The matrix obtained by one-dimensional convolution of the real part signal (1, 1, 256) is defined as A, and the matrix obtained by one-dimensional convolution of the imaginary part signal (1, 1, 256) is defined as B. The output dimension is calculated by using the following formula:

[0121] (A-B)+(A+B)i

[0122] The final output dimension is calculated to be (2048, 1, 256); and the obtained output vector is reorganized into a matrix D with a dimension of (8, 256, 256);

[0123] The matrix dimension obtained by WVD time-frequency analysis of the signal is (256, 256); the size of the two-dimensional convolution layer is (3, 3), the input channel is 1, and the output channel is 2; the dimension of the WVD distribution matrix with a dimension of (1, 1, 256) is (2, 256, 256) after two-dimensional convolution; the dimension is (4, 256, 256) after two-dimensional convolution; and the dimension of the matrix C obtained after two-dimensional convolution is (8, 256, 256);

[0124] The two matrices with a dimension of (8, 256, 256) obtained above are fused by concat to obtain a matrix E with a dimension of (16, 256, 256);

[0125] The matrix obtained by fusion is used as the input of the cross-term suppression structure;

[0126] 2) In the cross-term suppression structure, 15 residual structure encoders and 15 residual structure decoders are included; each residual structure encoder includes a two-dimensional convolution layer with a size of (3, 3) and a Relu activation function; and each residual structure decoder includes a two-dimensional deconvolution layer with a size of (3, 3) and a Relu activation function;

[0127] In the encoder-decoder connection, the output of the first encoder serves as the input of the second encoder and also as the input of the fifteenth decoder; the output of the second encoder serves as the input of the third encoder and also as the input of the fourteenth decoder; the output of the third encoder serves as the input of the fourth encoder and also as the input of the thirteenth decoder; the output of the fourth encoder serves as the input of the fifth encoder and also as the input of the twelfth decoder; the output of the fifth encoder serves as the input of the sixth encoder and also as the input of the eleventh decoder; the output of the sixth encoder serves as the input of the seventh encoder and also as the input of the tenth decoder; and the output of the seventh encoder serves as the input of the eighth encoder and also as the input of the ninth decoder. The output of the eighth encoder serves as the input of the ninth encoder; the output of the ninth encoder serves as the input of the tenth encoder and also as the input of the seventh decoder; the output of the tenth encoder serves as the input of the eleventh encoder and also as the input of the sixth decoder; the output of the eleventh encoder serves as the input of the twelfth encoder and also as the input of the fifth decoder; the output of the twelfth encoder serves as the input of the thirteenth encoder and also as the input of the fourth decoder; the output of the thirteenth encoder serves as the input of the fourteenth encoder and also as the input of the third decoder; the output of the fourteenth encoder serves as the input of the fifteenth encoder and also as the input of the second decoder; the output of the fifteenth encoder serves as the input of the first encoder.

[0128] In step 1), the fused matrix E is not only used as the input of the residual structure encoder, but also fused with the feature matrix F output by the residual structure decoder, and together they are used as the output of the cross-term suppression structure to obtain the output matrix G of the cross-term suppression structure.

[0129] In the method of this invention, the final output feature matrix has a dimension of (16,256,256);

[0130] 3) The output matrix G of the cross-term suppression structure obtained in step 2) is passed through a two-dimensional convolutional layer to obtain the feature matrix H; finally, it is passed through a two-dimensional convolutional layer to reduce the dimensionality and obtain the final optimized spectrum; the parameters in the network are optimized by backpropagation on the optimized spectrum.

[0131] In the method of this invention, the (32,256,256) matrix G obtained in step 2) is passed through a two-dimensional convolutional layer with dimension (3,3), with 32 input channels and 8 output channels, to obtain a matrix H with dimension (8,256,256); finally, it is passed through a two-dimensional convolutional layer with dimension (1,1), with 8 input channels and 1 output channel, to obtain an optimized spectrum of (1,256,256) after dimensionality reduction.

[0132] (3-2) Training the model:

[0133] The multi-input neural network model constructed in step (3-1) is trained using the fusion data set constructed in step S1, and a validation set data is used to estimate the accuracy of the model;

[0134] Part of the data in the fusion data set constructed in step S1 is selected to form a validation set data;

[0135] (3-3) Constructing an optimal time-frequency distribution prediction model:

[0136] Repeat the above step (3-2) based on the training results of the last time, adjust the parameters of the model until the set conditions are met, stop the repeated training of the model, and obtain the optimal time-frequency distribution prediction model;

[0137] The set conditions include judging whether the time-frequency resolution of the enhanced time-frequency spectrum output by the multi-input neural network model is within the set range. If the time-frequency resolution is within the set range, stop the model training, and determine the corresponding model as the optimal time-frequency distribution prediction model. If the time-frequency resolution is not within the set range, continue to train the network model until the time-frequency resolution is included in the set range;

[0138] The set conditions include judging whether the time-frequency resolution of the enhanced time-frequency spectrum output by the multi-input neural network model is within the set range. If the time-frequency resolution is within the set range, stop the model training, and determine the corresponding model as the optimal time-frequency distribution prediction model. If the time-frequency resolution is not within the set range, continue to train the network model until the time-frequency resolution is included in the set range;

[0139] S4. Obtain the real-time collected echo signal, perform time-frequency analysis processing through Winger-Ville distribution to obtain the corresponding time-frequency distribution, and simultaneously use the real-time collected echo signal and the corresponding time-frequency distribution to construct a real-time fusion data set;

[0140] S5. Use the real-time fusion data set constructed in step S4 to obtain the predicted time-frequency distribution through the optimal time-frequency distribution prediction model constructed in step S3;

[0141] S6. Extract the multi-target instantaneous frequency of the predicted time-frequency distribution obtained in step S5, estimate the distance and angle information of the target through the Doppler through-wall radar positioning algorithm, and complete the positioning of the multi-target; Specifically including:

[0142] (6-1) Extracting the multi-target instantaneous frequency of the predicted time-frequency distribution:

[0143] The multi-target instantaneous frequency of the predicted time-frequency distribution is extracted by peak detection method;

[0144] The peak detection method is expressed by the following formula:

[0145]

[0146] wherein, represents the instantaneous frequency curve obtained by the i-th iteration; x WVD (t,f) represents the time-frequency distribution obtained after time-frequency analysis processing by Winger-Ville distribution;

[0147] (6-2) Doppler through-wall radar positioning algorithm:

[0148] The Doppler through-wall radar positioning algorithm includes a distance estimation algorithm and an angle of arrival estimation algorithm; the distance estimation algorithm estimates the distance between the target and the transmitter in real time according to the instantaneous frequency of the target; the angle of arrival estimation algorithm estimates the angle between the target and the transmitter in real time according to the instantaneous frequency of the target;

[0149] (1) Distance estimation algorithm:

[0150] The distance between the target and the radar is obtained by the transmitter T x and the receiver R x1 ;

[0151] f1 and f2 are used to represent two carrier frequencies of the transmitted signal of the transmitter; the transmitted signal of the Doppler through-wall radar is expressed by the following formula:

[0152]

[0153] wherein, T x (t) represents that the transmitted signal includes two carrier signals; φ1 represents the initial phase of the first carrier signal, and φ2 represents the initial phase of the second carrier signal;

[0154] The echo signal received by the radar receiver is processed to obtain the following formula:

[0155]

[0156]

[0157] wherein, R x1 (t) represents the echo signal received by the receiver R x1 ; R1 represents the distance between the target and the transmitter; R2 represents the distance between the target and the receiver R x1 ; R x2 (t) represents the echo signal received by the receiver R x2 ; R3 represents the distance between the target and the receiver R x2 ; and c represents the speed of electromagnetic wave;

[0158] By approximating R1≈R2, the phases of the two receivers are represented by the following equation:

[0159]

[0160]

[0161] wherein, denotes the phase of the receiver R x1 ; denotes the phase of the receiver R x2 ;

[0162] The distance formula of the target motion is represented by the following equation:

[0163]

[0164] wherein, R denotes the distance of the target; denotes the instantaneous frequency of the receiver R x1 under the echo signal with the carrier frequency f1; denotes the instantaneous frequency of the receiver R x2 under the echo signal with the carrier frequency f1;

[0165] (2) Angle of arrival estimation algorithm:

[0166] The angle of arrival between the target and the radar is obtained by the transmitter T x and the receivers R x1 , R x2 ;

[0167] The distance between the receiver R x1 and the receiver R x2 is represented by d;

[0168] The phase difference of the echo signal between the receiver R x1 and the receiver R x2 is represented by the following equation:

[0169]

[0170] wherein, θ denotes the angle of arrival of the target; λ1 denotes the wavelength corresponding to the echo signal with the carrier frequency f1;

[0171] The angle of arrival is calculated by the following equation:

[0172]

[0173]

[0174] wherein, denotes the phase of the receiver Rx2 The instantaneous frequency of the echo signal with a carrier frequency of f1;

[0175] (6-3) Locating multiple targets:

[0176] like Figure 3 The diagram shown is a schematic of the Doppler through-wall radar structure of the present invention:

[0177] The target's range and angle of arrival, calculated using step (6-2), are expressed in the two-dimensional Cartesian coordinate system using the following formulas:

[0178] x=Rsinθ

[0179] y = Rcosθ

[0180] Where x represents the horizontal coordinate of the target's motion; y represents the vertical coordinate of the target's motion;

[0181] like Figure 4 The diagram shown compares the results of the method of this invention with those of the STFT and Hough-linear methods: Figure 4 (a) is a schematic diagram of the IF estimate obtained using the STFT method; Figure 4 (b) is a schematic diagram of trajectory fitting obtained using the STFT method; Figure 4 (c) is a schematic diagram of the IF estimate obtained using the Hough-linear method; Figure 4 (d) is a schematic diagram of trajectory fitting obtained using the Hough-linear method; Figure 4 (e) is a schematic diagram of the IF estimation obtained using the method of the present invention; Figure 4 (f) is a schematic diagram of trajectory fitting obtained using the method of the present invention; through Figure 4 This demonstrates that the method of the present invention can achieve a more accurate positioning effect.

Claims

1. A Doppler through-wall radar localization method based on time-frequency analysis using neural networks, comprising the following steps: S1. Construct the echo signal of multi-target motion after demodulation by Doppler through-wall radar; S2. Perform time-frequency analysis on the echo signal constructed in step S1 using the Winger-Ville distribution to obtain the corresponding time-frequency distribution, and use the obtained time-frequency distribution and the constructed echo signal to construct a fused dataset; S3. Construct a multi-input neural network model. Using the fusion dataset constructed in step S2, train the constructed model and update its parameters to build the optimal time-frequency distribution prediction model. Specifically, this includes: (3-1) Constructing a multi-input neural network model: A multi-input neural network model includes a multi-input encoding structure, a cross-term suppression structure, and an output part; 1) Multi-input coding structures include input signals, two-dimensional convolutional coding, and one-dimensional complex convolutional coding (Convl). The input includes a one-dimensional complex signal and a Winger-Ville distribution-time-frequency distribution, wherein the one-dimensional complex signal includes a real part and an imaginary part of the signal; Two-dimensional convolutional coding consists of three two-dimensional convolutional layers; One-dimensional complex convolutional coding (Convl) consists of two one-dimensional convolutional layers; The input WVD-TFR is processed through two-dimensional convolutional encoding to obtain the output matrix of two-dimensional convolutional encoding. ; Define the matrix obtained by one-dimensional convolution of the real part of the signal as follows: The matrix obtained by one-dimensional convolution of the imaginary part signal is: The output dimension is calculated using the following formula: The output dimension obtained through the above calculations is then used to reorganize the resulting matrix. With matrix Perform Concat fusion processing to obtain the output matrix. , as input to the cross-term suppression structure; 2) The cross-term suppression structure includes 15 residual structure encoders and 15 residual structure decoders; each residual structure encoder includes a [size missing]. A two-dimensional convolutional layer, a ReLU activation function; each residual structure decoder includes a [missing information - likely a variable name] of size [missing information - likely a variable name] A two-dimensional deconvolutional layer with a ReLU activation function; In the encoder-decoder connection, the output of the first encoder serves as the input of the second encoder and also as the input of the fifteenth decoder; the output of the second encoder serves as the input of the third encoder and also as the input of the fourteenth decoder; the output of the third encoder serves as the input of the fourth encoder and also as the input of the thirteenth decoder; the output of the fourth encoder serves as the input of the fifth encoder and also as the input of the twelfth decoder; the output of the fifth encoder serves as the input of the sixth encoder and also as the input of the eleventh decoder; the output of the sixth encoder serves as the input of the seventh encoder and also as the input of the tenth decoder; and the output of the seventh encoder serves as the input of the eighth encoder and also as the input of the ninth decoder. The output of the eighth encoder serves as the input of the ninth encoder; the output of the ninth encoder serves as the input of the tenth encoder and also as the input of the seventh decoder; the output of the tenth encoder serves as the input of the eleventh encoder and also as the input of the sixth decoder; the output of the eleventh encoder serves as the input of the twelfth encoder and also as the input of the fifth decoder; the output of the twelfth encoder serves as the input of the thirteenth encoder and also as the input of the fourth decoder; the output of the thirteenth encoder serves as the input of the fourteenth encoder and also as the input of the third decoder; the output of the fourteenth encoder serves as the input of the fifteenth encoder and also as the input of the second decoder; the output of the fifteenth encoder serves as the input of the first encoder. Step 1) is achieved by fusing the matrix. It serves not only as input to the residual structure encoder, but also as part of the feature matrix output by the residual structure decoder. The components are combined and used together as the output of the cross-term suppression structure to obtain the output matrix of the cross-term suppression structure. ; 3) The output matrix of the cross-term suppression structure obtained in step 2). The feature matrix is ​​obtained through a two-dimensional convolutional layer. Finally, a two-dimensional convolutional layer is used to reduce the dimensionality and obtain the final optimized spectrum. By performing backpropagation on the optimized spectrum, the parameters in the network are optimized. (3-2) Training the model: The multi-input neural network model constructed in step (3-1) is trained using the fusion dataset constructed in step S1, and the accuracy of the model is estimated using the validation set data. Select a portion of the data from the fusion dataset constructed in step S1 to form the validation set data; Select a portion of the data from the fusion dataset constructed in step S1 to form the validation set data; (3-3) Constructing the optimal time-frequency distribution prediction model: Repeat the above steps (3-2), and adjust the parameters of the model based on the previous training results until the set conditions are met. Stop the repeated training of the model and obtain the optimal time-frequency distribution prediction model. The set conditions include determining whether the time-frequency resolution of the augmented time-frequency spectrogram output by the multi-input neural network model is within the set range. If the time-frequency resolution is within the set range, the model training is stopped and the corresponding model is determined as the optimal time-frequency distribution prediction model. If the time-frequency resolution is not within the set range, the network model is trained until the time-frequency resolution is within the set range. S4. Acquire the real-time acquired echo signal, perform time-frequency analysis processing through the Winger-Ville distribution to obtain the corresponding time-frequency distribution, and construct a real-time fusion dataset using the real-time acquired echo signal and the corresponding time-frequency distribution. S5. Using the real-time fusion dataset constructed in step S4, the predicted time-frequency distribution is obtained through the optimal time-frequency distribution prediction model constructed in step S3; S6. Extract the instantaneous frequencies of the multi-targets from the predicted time-frequency distribution obtained in step S5, estimate the distance and angle information of the targets using the Doppler through-wall radar localization algorithm, and complete the localization of the multi-targets.

2. The Doppler through-wall radar localization method based on neural network time-frequency analysis according to claim 1, characterized in that... Step S1, which involves constructing the echo signal of multi-target motion after demodulation by Doppler through-wall radar, specifically includes: Construct several random parameters and establish a fourth-order random instantaneous frequency function using the following formula: in, Indicates the first The instantaneous frequency of each target, where the instantaneous frequency is a training parameter of the multi-input neural network model; Indicates the first One frequency shift of a signal; Indicates the first The second frequency shift of the signal; Indicates the first The three frequency shifts of a signal; Indicates the first The fourth frequency deviation of the signal; Indicates time; Simulate echo signal data from multiple moving targets to obtain one-dimensional echo signal data, and construct the received target echo signal. It is expressed by the following formula: Each signal includes Individual signals; Represents a normal distribution; Indicates the first The amplitude of each carrier signal; The constant term representing the signal.

3. The Doppler through-wall radar localization method based on neural network time-frequency analysis according to claim 2, characterized in that... Step S2 involves performing time-frequency analysis on the echo signal constructed in step S1 using the Winger-Ville distribution to obtain the corresponding time-frequency distribution. A fused dataset is then constructed using the obtained time-frequency distribution and the constructed echo signal. Specifically, this includes: Time-frequency analysis was performed using the Winger-Ville distribution, and the processing procedure for the Winger-Ville distribution is expressed by the following formula: in, This represents the time-frequency distribution obtained after time-frequency analysis processing using the Winger-Ville distribution. express The conjugate; Indicates time delay; Indicates instantaneous frequency; A fused dataset is constructed using the time-frequency distribution matrix obtained through time-frequency analysis of the Winger-Ville distribution and the echo signal constructed in step S1.

4. The Doppler through-wall radar localization method based on neural network time-frequency analysis according to claim 3, characterized in that... Step S6 involves extracting the instantaneous frequencies of multiple targets from the predicted time-frequency distribution obtained in step S5, estimating the target's distance and angle information using a Doppler through-wall radar localization algorithm, and completing the localization of multiple targets. Specifically, this includes: (6-1) Extracting the instantaneous frequencies of multiple targets in the predicted time-frequency distribution: The peak detection method is used to extract the instantaneous frequencies of multiple targets in the predicted time-frequency distribution; The peak detection method is expressed by the following formula: in, Indicates the first The instantaneous frequency curve obtained from the next iteration; This represents the time-frequency distribution obtained after time-frequency analysis processing using the Winger-Ville distribution. (6-2) Doppler through-wall radar localization algorithm: The Doppler through-wall radar localization algorithm includes a range estimation algorithm and an angle-of-arrival (AOA) estimation algorithm. The range estimation algorithm estimates the distance between the target and the transmitter in real time based on the target's instantaneous frequency. The AOA estimation algorithm estimates the angle between the target and the transmitter in real time based on the target's instantaneous frequency. (6-3) Locating multiple targets: The target's distance and angle of arrival, calculated using step (6-2), are expressed in the two-dimensional Cartesian coordinate system using the following formulas: in, The horizontal coordinate representing the target's motion; The vertical coordinate represents the target's motion.

5. The Doppler through-wall radar localization method based on neural network time-frequency analysis according to claim 4, characterized in that... The distance estimation algorithms included are specifically: via transmitter and receiver Obtain the distance between the target and the radar; use , These represent the two carrier frequencies of the transmitted signal from the transmitter; the transmitted signal of the Doppler through-wall radar is expressed by the following formula: in, This indicates that the transmitted signal includes two carrier signals; Indicates the initial phase of the first carrier signal. This indicates the initial phase of the second carrier signal; After processing the echo signal received by the radar receiver, the following formula is obtained: in, Indicates receiver The received echo signal; Indicates the distance between the target and the transmitter; Indicates target and receiver The distance between them; Indicates receiver The received echo signal; Indicates target and receiver The distance between them; Indicates the speed of electromagnetic waves; By approximation The phase of the two receivers is expressed by the following formula: in, Indicates receiver The phase; Indicates receiver The phase; The following formula can be used to express the distance traveled by the target: in, Indicates the distance to the target; Indicates receiver At carrier frequency The instantaneous frequency of the echo signal; Indicates receiver At carrier frequency The instantaneous frequency of the echo signal.

6. The Doppler through-wall radar localization method based on time-frequency analysis using neural networks according to claim 5, characterized in that... The included angle of arrival estimation algorithms specifically include: via transmitter and receiver , Obtain the angle of arrival between the target and the radar; use Indicates receiver With receiver The distance between them; The receiver is represented by the following formula. With receiver Phase difference between the echo signals: in, Indicates the angle of arrival of the target; Indicates the carrier frequency is The wavelength corresponding to the echo signal; The angle of arrival is calculated using the following formula: in, Indicates receiver At carrier frequency The instantaneous frequency of the echo signal.

Images