Life detection method and system based on radar signal processing

Abstract

The present invention provides a life detection method and system based on radar signal processing, which relates to the technical field of radar signal processing. The method and system include collecting radar echo signals in a target area, controlling a phased array radar transmitting antenna array to transmit an orthogonal frequency division multiplexed dual-frequency continuous wave radar signal, inputting the processed radar echo signal into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal; inputting a feature vector into a multi-task deep learning network with an integrated attention mechanism for feature fusion and target recognition to obtain a second processed signal; obtaining a target feature spectrum by calculating the time-frequency distribution of the second processed signal, judging whether there is a life form in the target area according to the time-varying characteristics of a respiratory frequency component and a heartbeat frequency component and the law of amplitude fluctuation; if it is determined that there is a life form in the target area, calculating the three-dimensional spatial coordinates of the life form through a multi-base station collaborative positioning algorithm, and outputting a detection result containing the three-dimensional spatial coordinates and the life form feature information.

Description

Technical Field

[0001] The present invention relates to radar signal processing technology, and in particular to a life detection method and system based on radar signal processing. Background Art

[0002] Single-frequency continuous wave radar is easily affected by environmental clutter and human interference, resulting in low detection accuracy and high false alarm rate. Although pulse radar has a certain anti-interference ability, its distance resolution and vital sign extraction accuracy are limited. Conventional signal processing methods used in existing solutions, such as simple filtering and spectrum analysis, are difficult to effectively suppress interference in complex environments and cannot meet actual application requirements.

[0003] Existing technologies mainly rely on the extraction of single features of breathing and heart rate, ignoring other important vital signs information, such as nonlinear features and time-frequency joint features. This results in low accuracy of vital sign recognition when the target is weak and the environment is complex, and it is difficult to effectively distinguish between living things and other moving targets.

[0004] To address the above issues, there is an urgent need for an improved method based on radar signal processing that can achieve high-precision detection and positioning of life forms in complex environments, effectively suppress environmental clutter and interference signals, and accurately extract weak life characteristics. Summary of the invention

[0005] The embodiments of the present invention provide a life detection method and system based on radar signal processing, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A life detection method based on radar signal processing is provided, comprising:

[0008] Collect radar echo signals in the target area, control the phased array radar transmitting antenna array to transmit orthogonal frequency division multiplexing dual-frequency continuous wave radar signals, the dual-frequency continuous wave radar signals include a first frequency signal and a second frequency signal, the first frequency signal and the second frequency signal are coded and modulated by Walsh orthogonal codes, the radar echo signals are received by a multi-channel radar receiving system, interference suppression and clutter cancellation processing is performed on the radar echo signals by using an adaptive space-time two-dimensional filter, and the processed radar echo signals are input into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal;

[0009] Performing digital sampling processing on the baseband signal to obtain a sampling signal, performing multi-stage static clutter elimination processing on the sampling signal using an adaptive notch filter group to obtain a first processed signal, inputting the first processed signal into a dual-tree complex wavelet transform module to perform multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extracting nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct a feature vector, wherein the nonlinear feature parameters include approximate entropy and sample entropy, and the time-frequency joint feature parameters include time-frequency ridge and instantaneous frequency, and inputting the feature vector into a multi-task deep learning network with an integrated attention mechanism to perform feature fusion and target recognition to obtain a second processed signal;

[0010] The target characteristic spectrum is obtained by calculating the time-frequency distribution of the second processed signal, and the target characteristic spectrum is denoised by using a morphological filter. At the same time, the target characteristic spectrum is selectively noise-suppressed by using an adaptive morphological filter. A multi-target segmentation algorithm based on a variational model is used to extract the respiratory frequency component and the heart rate frequency component in the target characteristic spectrum. It is judged whether there is a living body in the target area according to the time-varying characteristics of the respiratory frequency component and the heart rate frequency component and the amplitude fluctuation law. If it is determined that there is a living body in the target area, the three-dimensional spatial coordinates of the living body are calculated by a multi-base station collaborative positioning algorithm, and a detection result containing the three-dimensional spatial coordinates and the characteristic information of the living body is output.

[0011] In an optional embodiment,

[0012] Using an adaptive space-time two-dimensional filter to perform interference suppression and clutter cancellation processing on the radar echo signal, and inputting the processed radar echo signal into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal includes:

[0013] Collecting time-series sampling data of multiple spatial sampling channels of a radar receiving system, and constructing a space-time two-dimensional sampling data matrix with the time-series sampling data according to the spatial sampling channel sequence number and the time sampling sequence number, wherein the space-time two-dimensional sampling data matrix includes a target echo component, an interference component, and a clutter component;

[0014] Initializing the complex weight vector and correlation matrix of the adaptive space-time two-dimensional filter, performing an inner product operation on the complex weight vector and the space-time two-dimensional sampling matrix to obtain an initial filter output, and calculating a prediction error signal based on the obtained initial filter output and the expected signal;

[0015] Calculating a time-varying forgetting factor according to the prediction error signal, multiplying the product of the time-varying forgetting factor and the correlation matrix by the space-time two-dimensional sampling matrix to obtain a gain vector, wherein the gain vector is used to update the complex weight value vector;

[0016] The complex weight vector is updated by using the gain vector and the prediction error signal, and the space-time two-dimensional sampling matrix is ​​weighted filtered by using the updated complex weight vector to obtain a filtered signal, and the filtered signal is envelope detected to obtain an envelope signal, and a signal sample is selected on the envelope signal according to a preset time window length to construct a sliding time window, and the amplitude mean and amplitude standard deviation of the signal samples in the sliding time window are calculated;

[0017] The weighted sum of the amplitude mean and the amplitude standard deviation is used as a threshold value, and an amplitude decision is performed on the filtered signal according to the threshold value to obtain a decision signal, wherein the decision signal retains a signal amplitude and phase greater than the threshold value;

[0018] Generate a cosine waveform in-phase carrier and a sine waveform orthogonal carrier with the same frequency and phase as the radar transmission signal, and multiply the decision signal with the in-phase carrier and the orthogonal carrier to obtain a first product signal and a second product signal;

[0019] The first product signal and the second product signal are respectively subjected to low-pass filtering to remove high-frequency components to obtain an in-phase component and a quadrature component, and the in-phase component is used as a real part and the quadrature component is used as an imaginary part to obtain a final baseband signal.

[0020] In an optional embodiment,

[0021] The first product signal and the second product signal are respectively subjected to low-pass filtering to remove high-frequency components to obtain an in-phase component and an orthogonal component, and the in-phase component is used as a real part and the orthogonal component is used as an imaginary part to obtain a final baseband signal, comprising:

[0022] Multiplying the radar echo signal with the first product signal and the second product signal respectively to obtain an in-phase component and an orthogonal component;

[0023] Performing time-frequency analysis on the in-phase component and the orthogonal component, calculating the dynamic bandwidth parameter of the signal, adaptively adjusting the cutoff frequency of the low-pass filter according to the dynamic bandwidth parameter, and calculating the time-frequency concentration of the signal according to the dynamic bandwidth parameter, adaptively determining the shape factor of the Kaiser window function based on the time-frequency concentration, and generating a window function sequence using the shape factor;

[0024] The window function sequence is combined with the cutoff frequency to generate a low-pass filter coefficient, and the low-pass filter coefficient is used to filter the in-phase component and the orthogonal component respectively to obtain an in-phase baseband signal and an orthogonal baseband signal;

[0025] Calculating the signal-to-noise ratio of the in-phase baseband signal and the quadrature baseband signal, generating signal weights according to the signal-to-noise ratios, and multiplying the in-phase baseband signal and the quadrature baseband signal by the corresponding signal weights to obtain a weighted in-phase signal and a weighted quadrature signal;

[0026] The weighted in-phase signal is used as the real part and the weighted quadrature signal is used as the imaginary part to perform orthogonal combination to obtain a complex baseband output signal as the final baseband signal.

[0027] In an optional embodiment,

[0028] Inputting the first processed signal into a dual-tree complex wavelet transform module for multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extracting nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct feature vectors includes:

[0029] Constructing a dual-tree complex wavelet transform module including a symmetric orthogonal filter group, wherein the symmetric orthogonal filter group includes a first group of orthogonal filters and a second group of orthogonal filters, inputting the first processed signal into the symmetric orthogonal filter group for four-layer wavelet decomposition, and obtaining complex wavelet coefficients of different frequency bands based on output signals of the first group of orthogonal filters and the second group of orthogonal filters;

[0030] A reconstruction vector is constructed based on the complex wavelet coefficients, and the distance between the vectors of the reconstructed vector is calculated to obtain a distance parameter. The number of vectors that meet the distance parameter condition is counted according to a preset threshold to obtain a matching number. The conditional probability logarithm is calculated based on the matching number to obtain an approximate entropy feature; based on the matching number, the ratio of the template matching number in different dimensions is calculated to obtain a sample entropy feature;

[0031] Performing a continuous wavelet transform on the complex wavelet coefficients to obtain a time-frequency coefficient matrix, extracting the maximum value of the modulus value at each time point based on the time-frequency coefficient matrix to obtain a time-frequency ridge feature; performing a Hilbert transform on the first processed signal to obtain an analytical signal, and calculating the instantaneous frequency feature according to the phase information of the analytical signal;

[0032] The approximate entropy feature, the sample entropy feature, the time-frequency ridge feature and the instantaneous frequency feature are combined to construct a feature vector.

[0033] In an optional embodiment,

[0034] The feature vector is input into a multi-task deep learning network with an integrated attention mechanism for feature fusion and target recognition to obtain a second processed signal including:

[0035] Constructing a deep learning network including a feature extraction layer, a self-attention layer, and a multi-task output layer, inputting a feature vector into the feature extraction layer, which includes a plurality of cascaded residual blocks, each residual block including a first convolutional layer, a second convolutional layer, and a skip connection path, the feature vector is passed through the first convolutional layer to obtain a first transformation feature, the first transformation feature is passed through the second convolutional layer to obtain a second transformation feature, and the second transformation feature is added to the feature vector transmitted through the skip connection path to obtain a local feature;

[0036] The local features are input into a self-attention layer, the self-attention layer performs a linear transformation on the local features to obtain a query matrix, a key matrix and a value matrix, the query matrix is ​​multiplied by the transpose of the key matrix to obtain a correlation matrix, and the correlation matrix is ​​normalized to obtain an attention weight matrix;

[0037] Multiplying the attention weight matrix by the value matrix to obtain a first weighted feature, constructing multiple groups of self-attention layers in parallel to obtain multiple groups of first weighted features, and concatenating the multiple groups of first weighted features in the feature dimension to obtain a second weighted feature;

[0038] Constructing a multi-task output layer including a detection subnetwork and a classification subnetwork, wherein the detection subnetwork and the classification subnetwork are constructed in a parameter sharing manner, inputting the second weighted feature into the detection subnetwork and the classification subnetwork at the same time, performing target detection on the second weighted feature through the detection subnetwork to obtain a target position feature and a target confidence feature, and performing target classification on the second weighted feature through the classification subnetwork to obtain a target category feature;

[0039] Constructing a joint loss function, the joint loss function including a detection loss term and a classification loss term calculated according to target position features and target confidence features, and optimizing parameters of the detection subnetwork and the classification subnetwork based on the joint loss function;

[0040] The target position features and target confidence features output by the optimized detection subnetwork and the target category features output by the optimized classification subnetwork are combined to obtain a second processed signal.

[0041] In an optional embodiment,

[0042] The target characteristic spectrum is obtained by calculating the time-frequency distribution of the second processed signal, the target characteristic spectrum is subjected to noise reduction processing by using a morphological filter, and the target characteristic spectrum is subjected to selective noise suppression by using an adaptive morphological filter, and the respiratory frequency component and the heartbeat frequency component in the target characteristic spectrum are extracted by using a multi-target segmentation algorithm based on a variational model, and judging whether there is a living body in the target area according to the time-varying characteristics and amplitude fluctuation law of the respiratory frequency component and the heartbeat frequency component includes:

[0043] Performing short-time Fourier transform on the second processed signal to obtain a time-frequency distribution diagram, performing frame windowing processing on the second processed signal using a Hanning window function, performing fast Fourier transform on the windowed signal to obtain a spectrum sequence, and splicing the spectrum sequence in the time dimension to obtain a target characteristic spectrum;

[0044] Constructing a double-layer morphological filter structure, wherein the first layer of the double-layer morphological filter structure includes a basic morphological filter, wherein the basic morphological filter adopts a disk-shaped structure element, and the disk-shaped structure element is used to perform an opening operation on the target feature spectrum to remove the peak noise higher than the background to obtain a first intermediate signal, and a closing operation is performed on the first intermediate signal to fill the valley noise lower than the background to obtain a first filtered signal;

[0045] Calculating the mean, variance and skewness of each time-frequency point neighborhood in the first filtered signal to obtain local statistical characteristics, calculating an adaptive threshold based on the local statistical characteristics, and comparing the adaptive threshold with a preset decision criterion to obtain noise area identification information;

[0046] Dynamically adjusting the shape and size parameters of the structural elements of the adaptive morphological filter according to the noise area identification information, selectively filtering the time-frequency points identified as the noise area to obtain a second filtered signal, and keeping the time-frequency points identified as the target feature area unchanged;

[0047] An energy functional is constructed, wherein a data item of the energy functional represents the degree of matching between the second filtered signal and the respiratory and heartbeat feature model, a regularization item introduces a spatial continuity constraint, and an alternating direction multiplier method is used to optimize and solve the energy functional to obtain a frequency component segmentation result;

[0048] Performing region marking and connected domain analysis on the frequency component segmentation results, extracting the time-varying trajectories of the respiratory frequency component and the heartbeat frequency component, and calculating the periodicity index, amplitude fluctuation law and energy distribution characteristics of the time-varying trajectories;

[0049] The periodic index, amplitude fluctuation law and energy distribution characteristics are input into a pre-trained physiological feature matching model to obtain a feature matching score, and the signal-to-noise ratio of the respiratory frequency component and the heart rate frequency component is calculated. When the feature matching score and the signal-to-noise ratio simultaneously meet the preset threshold conditions, it is determined that there is a living body in the target area.

[0050] In an optional embodiment,

[0051] If it is determined that there is a life form in the target area, the three-dimensional spatial coordinates of the life form are calculated by a multi-base station collaborative positioning algorithm, and a detection result including the three-dimensional spatial coordinates and the life form feature information is outputted, including:

[0052] Acquire receiving signals of multiple radar base stations and perform bandpass filtering to obtain filtered receiving signals, perform envelope detection on the filtered receiving signals to obtain envelope signals, obtain the peak moment of the envelope signal by peak detection, and determine the peak moment as the arrival time;

[0053] Performing phase analysis on the filtered received signal to obtain a phase spectrum, calculating the phase spectrum difference between adjacent radar base stations to obtain a phase difference, using the phase difference as an arrival time difference, constructing a hyperbola equation based on the arrival time difference, and performing spatial filtering on the filtered received signal to obtain a filtered direction signal, calculating the covariance matrix of the filtered direction signal, obtaining an arrival angle by eigendecomposing the covariance matrix, and constructing an azimuth angle equation based on the arrival angle;

[0054] The hyperbolic equation and the azimuth equation are combined to obtain a positioning equation, a ratio of signal power to noise power in the filtered received signal is calculated to obtain a signal-to-noise ratio, and the signal-to-noise ratio is normalized to obtain a weight coefficient;

[0055] Performing Taylor expansion on the positioning equation to obtain a linear equation, multiplying the weight coefficient by the coefficient of the linear equation to obtain a weighted linear equation, iteratively calculating the weighted linear equation to obtain a position increment, stopping the iteration when the position increment is less than a position threshold, and outputting the three-dimensional space coordinates;

[0056] Calculating the geometric precision factor of the three-dimensional space coordinates, and when the geometric precision factor is greater than a second preset threshold, calculating the positioning errors of each base station combination to obtain an error matrix, and selecting the base station combination with the smallest positioning error as the optimal base station combination based on the error matrix;

[0057] The three-dimensional space coordinates are optimized using the optimal base station combination to obtain optimized three-dimensional space coordinates, statistical parameters of respiratory frequency components and heart rate frequency components in a continuous time window are extracted, and the optimized three-dimensional space coordinates are combined with the statistical parameters to obtain a detection result containing three-dimensional space coordinates and vital feature information.

[0058] According to a second aspect of the embodiments of the present invention,

[0059] Provided is a life detection system based on radar signal processing, comprising:

[0060] The first unit is used to collect radar echo signals of the target area, control the phased array radar transmitting antenna array to transmit a dual-frequency continuous wave radar signal of orthogonal frequency division multiplexing, wherein the dual-frequency continuous wave radar signal includes a first frequency signal and a second frequency signal, the first frequency signal and the second frequency signal are coded and modulated by Walsh orthogonal codes, the radar echo signal is received by a multi-channel radar receiving system, interference suppression and clutter cancellation processing is performed on the radar echo signal by using an adaptive space-time two-dimensional filter, and the processed radar echo signal is input into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal;

[0061] The second unit is used to perform digital sampling processing on the baseband signal to obtain a sampling signal, perform multi-stage static clutter elimination processing on the sampling signal using an adaptive notch filter group to obtain a first processed signal, input the first processed signal into a dual-tree complex wavelet transform module to perform multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extract nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct a feature vector, the nonlinear feature parameters include approximate entropy and sample entropy, the time-frequency joint feature parameters include time-frequency ridge and instantaneous frequency, and input the feature vector into a multi-task deep learning network with an integrated attention mechanism to perform feature fusion and target recognition to obtain a second processed signal;

[0062] The third unit is used to obtain a target characteristic spectrum by calculating the time-frequency distribution of the second processed signal, perform noise reduction processing on the target characteristic spectrum by using a morphological filter, and selectively suppress noise on the target characteristic spectrum by using an adaptive morphological filter, and extract the respiratory frequency component and the heart rate component in the target characteristic spectrum by using a multi-target segmentation algorithm based on a variational model, and judge whether there is a living body in the target area according to the time-varying characteristics and amplitude fluctuation law of the respiratory frequency component and the heart rate component; if it is determined that there is a living body in the target area, calculate the three-dimensional spatial coordinates of the living body by using a multi-base station collaborative positioning algorithm, and output a detection result including the three-dimensional spatial coordinates and the characteristic information of the living body.

[0063] According to a third aspect of the embodiments of the present invention,

[0064] An electronic device is provided, comprising:

[0065] processor;

[0066] a memory for storing processor-executable instructions;

[0067] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0068] A fourth aspect of the embodiments of the present invention is:

[0069] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0070] In this embodiment, orthogonal frequency division multiplexing dual-frequency continuous wave radar signals are combined with Walsh orthogonal code modulation, and adaptive space-time two-dimensional filters and adaptive notch filter groups are used to perform multi-level clutter and interference suppression, which effectively improves the signal-to-noise ratio of radar echo signals, thereby improving the reliability and accuracy of life detection in complex environments. Dual-tree complex wavelet transform is used for multi-scale time-frequency analysis to extract nonlinear feature parameters such as approximate entropy, sample entropy, time-frequency ridge, instantaneous frequency, and time-frequency joint feature parameters, and combined with a multi-task deep learning network with an integrated attention mechanism for feature fusion and target recognition, which can more comprehensively and finely characterize the micro-motion characteristics of life, thereby significantly improving the accuracy of life recognition. The respiratory frequency and heart rate components are accurately extracted by a multi-target segmentation algorithm based on a variational model, and combined with a multi-base station collaborative positioning algorithm, the accurate positioning of life can be achieved, and the characteristic information of life can be obtained at the same time, providing key information for subsequent rescue and life status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic diagram of a flow chart of a life detection method based on radar signal processing according to an embodiment of the present invention;

[0072] Figure 2 The figure is a schematic diagram of the structure of a life detection system based on radar signal processing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0075] Figure 1 FIG. 1 is a flow chart of a life detection method based on radar signal processing according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0076] S101. Collect radar echo signals in the target area, control the phased array radar transmitting antenna array to transmit orthogonal frequency division multiplexing dual-frequency continuous wave radar signals, the dual-frequency continuous wave radar signals include a first frequency signal and a second frequency signal, the first frequency signal and the second frequency signal are coded and modulated by Walsh orthogonal codes, the radar echo signals are received by a multi-channel radar receiving system, interference suppression and clutter cancellation processing is performed on the radar echo signals by using an adaptive space-time two-dimensional filter, and the processed radar echo signals are input into an orthogonal demodulation module for synchronous coherent demodulation to obtain baseband signals;

[0077] S102. Perform digital sampling processing on the baseband signal to obtain a sampling signal, use an adaptive notch filter group to perform multi-stage static clutter elimination processing on the sampling signal to obtain a first processed signal, input the first processed signal into a dual-tree complex wavelet transform module to perform multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extract nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct a feature vector, the nonlinear feature parameters include approximate entropy and sample entropy, the time-frequency joint feature parameters include time-frequency ridge and instantaneous frequency, and input the feature vector into a multi-task deep learning network with an integrated attention mechanism to perform feature fusion and target recognition to obtain a second processed signal;

[0078] S103. Obtain the target characteristic spectrum by calculating the time-frequency distribution of the second processed signal, perform noise reduction processing on the target characteristic spectrum by using a morphological filter, and selectively suppress the noise on the target characteristic spectrum by using an adaptive morphological filter, and extract the respiratory frequency component and the heart rate component in the target characteristic spectrum by using a multi-target segmentation algorithm based on a variational model, and judge whether there is a living body in the target area according to the time-varying characteristics of the respiratory frequency component and the heart rate component and the amplitude fluctuation law. If it is determined that there is a living body in the target area, calculate the three-dimensional spatial coordinates of the living body by using a multi-base station collaborative positioning algorithm, and output the detection result containing the three-dimensional spatial coordinates and the characteristic information of the living body.

[0079] Among them, phased array radar is a radar system that controls the direction of the beam by adjusting the phase difference of each unit in the antenna array. Unlike traditional mechanically rotating antennas, phased array radars can quickly adjust the direction of the beam electronically to achieve rapid scanning of targets and multi-target tracking. Here, the phased array radar is used as a transmitting antenna array to transmit orthogonal frequency division multiplexing dual-frequency continuous wave radar signals, providing high-precision and fast-response detection capabilities.

[0080] Orthogonal frequency division multiplexing (OFDM) is a modulation technology that divides the frequency band into multiple narrowband subcarriers and transmits them orthogonally. This method can effectively reduce frequency band interference and enhance signal anti-noise performance. In this solution, OFDM technology enables dual-frequency continuous wave radar signals to be transmitted in parallel through different frequency channels, effectively improving signal transmission efficiency and data capacity.

[0081] The dual-frequency continuous wave radar signal is a waveform that is continuously transmitted at two different frequencies. The first frequency signal and the second frequency signal are independently encoded through modulation, and the two are guaranteed to be orthogonal to each other.

[0082] Walsh orthogonal code is a special coding sequence used in coded modulation, which has the characteristic of being mutually orthogonal, that is, there is no mutual interference between different codes in the time domain. Encoding the first frequency signal and the second frequency signal through Walsh orthogonal code can avoid interference caused by frequency overlap and improve the separation and recognition accuracy of signals.

[0083] Adaptive space-time two-dimensional filter is a filter that combines space and time information to dynamically suppress interference and eliminate clutter. Space-time two-dimensional filter can adaptively adjust filter parameters according to the different characteristics of target and noise, suppress noise in specific frequency bands, make target features clearer, and make signals purer.

[0084] The orthogonal demodulation module is used to receive radar echo signals and complete the coherent demodulation of the signal by synchronous comparison with the local reference signal. Coherent demodulation can restore high-frequency echo signals to low-frequency baseband signals, making it easier to extract target information. In this scheme, orthogonal demodulation ensures data synchronization and effectively retains the amplitude and phase information of the signal. Dual-tree complex wavelet transform is a signal processing technology based on wavelet decomposition, which realizes multi-scale and multi-directional decomposition of signals through a dual-tree structure. Dual-tree complex wavelet transform has good translation invariance and directional selectivity. In this scheme, the module performs multi-scale time-frequency analysis on the first processed signal, generates wavelet coefficients in different frequency bands, and provides detailed frequency information for subsequent feature extraction.

[0085] The time-frequency ridge refers to the high energy distribution path of the signal along the frequency direction on the time-frequency graph, which can reflect the dominant frequency component of the signal. The instantaneous frequency is the frequency distribution of the signal changing over time. The time-frequency ridge and instantaneous frequency, as joint time-frequency characteristic parameters, can help better capture the periodic characteristics of life and distinguish the frequency components of breathing and heartbeat.

[0086] The variational model is an image segmentation method based on the variational principle. It constructs an energy functional model to optimally segment images or signals. In this solution, the variational model is used for multi-target segmentation, accurately extracting the respiratory frequency and heart rate components, and effectively distinguishing the key characteristic signals of living organisms.

[0087] The multi-base station collaborative positioning algorithm is a method to achieve high-precision positioning through the cooperation of multi-base station signals. By integrating the signal information of multiple base stations, the three-dimensional coordinate position of the target can be more accurately calculated. In life detection, this algorithm can be used to determine the specific position of life in space and improve detection accuracy and coverage.

[0088] For example, first, a phased array radar transmitting antenna array is used to transmit a carefully designed detection signal. The signal adopts a dual-frequency continuous wave radar system using orthogonal frequency division multiplexing, and contains two signals of different frequencies - a first frequency signal and a second frequency signal. In order to distinguish the two signals and improve the anti-interference capability, Walsh orthogonal codes are used to encode and modulate them respectively. For example, the first line of code in the Walsh code sequence can be used to modulate the first frequency signal, and the second line of code can be used to modulate the second frequency signal. The radar echo signal is received by a multi-channel radar receiving system.

[0089] Next, the received radar echo signal is preprocessed to remove interference and clutter. An adaptive space-time two-dimensional filter is used for interference suppression and clutter cancellation. The filter can adaptively adjust the filter parameters according to the changes in the environment and effectively suppress various interferences and clutter. For example, the minimum mean square error (LMS) algorithm or the recursive least squares (RLS) algorithm can be used to update the filter weights. The filtered radar echo signal is sent to the orthogonal demodulation module for synchronous coherent demodulation to extract the baseband signal.

[0090] Then, the demodulated baseband signal is digitally sampled to obtain a discrete sampling signal. In order to further eliminate the influence of static clutter, an adaptive notch filter group is used to perform multi-stage processing on the sampled signal. The notch filter group can suppress the clutter component in a targeted manner according to the frequency characteristics of the static clutter. For example, multiple notches can be set according to the frequency distribution of the clutter, and each notch filter is responsible for suppressing clutter within a specific frequency range. After being processed by the notch filter group, a first processed signal is obtained.

[0091] The first processed signal is input into the dual-tree complex wavelet transform module for multi-scale time-frequency analysis. The dual-tree complex wavelet transform can effectively extract the time-frequency characteristics of the signal and has good directional selectivity. By analyzing the wavelet coefficients of different frequency bands, nonlinear characteristic parameters and time-frequency joint characteristic parameters can be extracted. Nonlinear characteristic parameters include approximate entropy and sample entropy, which are used to describe the complexity and regularity of the signal. For example, the approximate entropy and sample entropy of the wavelet coefficients can be calculated to quantify the nonlinear characteristics of the signal. The time-frequency joint characteristic parameters include time-frequency ridge and instantaneous frequency, which are used to describe the change law of the frequency of the signal over time. For example, the time-frequency ridge and instantaneous frequency of the wavelet coefficients can be extracted to characterize the time-frequency characteristics of the signal. The extracted characteristic parameters are combined into a feature vector.

[0092] The constructed feature vector is input into a multi-task deep learning network with integrated attention mechanism. The network can perform feature fusion and target recognition at the same time. The attention mechanism can make the network pay more attention to important features and improve the accuracy of target recognition. Multi-task learning can enable the network to learn other related tasks, such as feature extraction of living organisms, while learning target recognition, thereby improving the generalization ability of the network. The output of the network is the second processed signal.

[0093] The target characteristic spectrum is obtained by calculating the time-frequency distribution of the second processed signal. In order to remove the influence of noise, a morphological filter is used to perform noise reduction on the target characteristic spectrum. The morphological filter can effectively remove noise in the image while retaining the edge information of the image. For example, open and close operations can be used to remove noise. At the same time, an adaptive morphological filter is used to selectively suppress noise on the target characteristic spectrum. The adaptive morphological filter can adaptively adjust the filter parameters according to the characteristics of the noise, so as to better suppress the noise. A multi-target segmentation algorithm based on a variational model is used to extract the respiratory frequency component and the heartbeat frequency component in the target characteristic spectrum. The variational model can effectively segment different targets in the image and extract the characteristics of the target. For example, a variational model based on a level set can be used to extract the respiratory frequency component and the heartbeat frequency component.

[0094] According to the time-varying characteristics of the extracted respiratory frequency component and heart rate component and the amplitude fluctuation law, it is judged whether there is a life in the target area. For example, it is assumed that the respiratory frequency is between 0.2Hz and 0.5Hz, and the heart rate is between 0.8Hz and 1.5Hz. If the frequency components that meet these conditions are detected in the target characteristic spectrum, and their amplitudes show periodic fluctuations, it can be judged that there is a life in the target area. If it is determined that there is a life in the target area, the three-dimensional spatial coordinates of the life are calculated by a multi-base station collaborative positioning algorithm. The multi-base station collaborative positioning algorithm can use the observation data of multiple base stations to improve the positioning accuracy. For example, a positioning algorithm based on the time difference of arrival (TDOA) can be used. Finally, the detection results containing the three-dimensional spatial coordinates of the life and the characteristic information of the life, such as respiratory frequency, heart rate, etc., are output. For example, the detection result can be expressed as: (x, y, z, respiratory frequency, heart rate), where (x, y, z) represents the three-dimensional spatial coordinates of the life.

[0095] In this embodiment, by adopting advanced signal processing technologies, such as adaptive filtering, dual-tree complex wavelet transform and deep learning, interference and noise can be effectively suppressed, weak characteristic signals of life bodies can be extracted, thereby improving the sensitivity and accuracy of detection. Multiple life bodies can be detected and located at the same time, and the three-dimensional spatial coordinates and characteristic information of the life bodies, such as breathing rate and heart rate, can be provided. Adaptive technologies, such as adaptive filtering and adaptive morphological filtering, can be used to adaptively adjust parameters according to changes in the environment, thereby enhancing the robustness and adaptability of the system, enabling it to work stably in complex environments.

[0096] In an optional implementation, using an adaptive space-time two-dimensional filter to perform interference suppression and clutter cancellation processing on the radar echo signal, and inputting the processed radar echo signal into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal includes:

[0097] Collecting time-series sampling data of multiple spatial sampling channels of a radar receiving system, and constructing a space-time two-dimensional sampling data matrix with the time-series sampling data according to the spatial sampling channel sequence number and the time sampling sequence number, wherein the space-time two-dimensional sampling data matrix includes a target echo component, an interference component, and a clutter component;

[0098] Initializing the complex weight vector and correlation matrix of the adaptive space-time two-dimensional filter, performing an inner product operation on the complex weight vector and the space-time two-dimensional sampling matrix to obtain an initial filter output, and calculating a prediction error signal based on the obtained initial filter output and the expected signal;

[0099] Calculating a time-varying forgetting factor according to the prediction error signal, multiplying the product of the time-varying forgetting factor and the correlation matrix by the space-time two-dimensional sampling matrix to obtain a gain vector, wherein the gain vector is used to update the complex weight value vector;

[0100] The complex weight vector is updated by using the gain vector and the prediction error signal, and the space-time two-dimensional sampling matrix is ​​weighted filtered by using the updated complex weight vector to obtain a filtered signal, and the filtered signal is envelope detected to obtain an envelope signal, and a signal sample is selected on the envelope signal according to a preset time window length to construct a sliding time window, and the amplitude mean and amplitude standard deviation of the signal samples in the sliding time window are calculated;

[0101] The weighted sum of the amplitude mean and the amplitude standard deviation is used as a threshold value, and an amplitude decision is performed on the filtered signal according to the threshold value to obtain a decision signal, wherein the decision signal retains a signal amplitude and phase greater than the threshold value;

[0102] Generate a cosine waveform in-phase carrier and a sine waveform orthogonal carrier with the same frequency and phase as the radar transmission signal, and multiply the decision signal with the in-phase carrier and the orthogonal carrier to obtain a first product signal and a second product signal;

[0103] The first product signal and the second product signal are respectively subjected to low-pass filtering to remove high-frequency components to obtain an in-phase component and a quadrature component, and the in-phase component is used as a real part and the quadrature component is used as an imaginary part to obtain a final baseband signal.

[0104] Exemplarily, first, the time-series sampling data is collected through multiple spatial sampling channels of the radar receiving system. Assume that there are 4 spatial sampling channels, and each channel collects data at 100 time sampling points. These data are arranged according to the spatial sampling channel sequence number and the time sampling sequence number to construct a 4-row 100-column space-time two-dimensional sampling data matrix. The matrix contains target echo components, interference components, and clutter components. For example, the matrix element (1, 1) represents the data of the first time sampling point of the first spatial sampling channel, and the data is a complex number, such as 1+2j, which represents the amplitude and phase of the signal.

[0105] Next, initialize the complex weight vector and correlation matrix of the adaptive space-time two-dimensional filter. The complex weight vector is a column vector whose dimension is the same as the number of rows of the space-time two-dimensional sampling matrix, which is 4 rows and 1 column here. The initial value can be set to a full zero vector or a random vector, such as [0, 0, 0, 0]T. The correlation matrix is ​​a square matrix whose dimension is the same as the dimension of the complex weight vector, which is 4 rows and 4 columns here. The initial value can be set to the unit matrix multiplied by a small positive number, such as 0.01 multiplied by the unit matrix. Perform an inner product operation on the complex weight vector and the space-time two-dimensional sampling matrix to obtain the initial filter output. For example, perform an inner product operation on [0, 0, 0, 0]T and the first column [1+2j, 3+4j, 5+6j, 7+8j]T to obtain 0.

[0106] Then, the prediction error signal is calculated based on the obtained initial filter output and the expected signal. The expected signal is a known target echo signal, such as 1+2j. The prediction error signal is the difference between the expected signal and the initial filter output, here 1+2j- 0 = 1+2j.

[0107] The time-varying forgetting factor is calculated based on the prediction error signal. The time-varying forgetting factor is a value between 0 and 1, which is used to control the memory length of the filter. For example, the time-varying forgetting factor can be adjusted according to the amplitude of the prediction error signal. The larger the error, the smaller the forgetting factor, and vice versa. Assume that the calculated time-varying forgetting factor is 0.9. The gain vector is obtained by multiplying the product of the time-varying forgetting factor and the correlation matrix and the space-time two-dimensional sampling matrix. The gain vector is used to update the complex weight vector.

[0108] The gain vector and the prediction error signal are used to update the complex weight vector. The updated complex weight vector is used to perform weighted filtering on the space-time two-dimensional sampling matrix to obtain a filtered signal. The filtered signal is envelope-detected to obtain an envelope signal.

[0109] According to the preset time window length, for example, 10 sampling points, select signal samples from the envelope signal to construct a sliding window. Calculate the amplitude mean and amplitude standard deviation of the signal samples in the sliding window. For example, if the signal samples in the first sliding window are [0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4], the amplitude mean is 0.95 and the amplitude standard deviation is 0.3.

[0110] The weighted sum of the amplitude mean and the amplitude standard deviation is used as the threshold value. For example, the amplitude mean and the amplitude standard deviation are multiplied by weights 0.8 and 0.2 respectively, and then added to obtain the threshold value, which is 0.95 * 0.8 + 0.3 * 0.2 = 0.82. The amplitude of the filtered signal is determined according to the threshold value to obtain a decision signal. The decision signal retains the signal amplitude and phase greater than the threshold value.

[0111] Generate a cosine waveform in-phase carrier and a sine waveform quadrature carrier with the same frequency and phase as the radar transmission signal. Assuming that the radar transmission signal frequency is 10GHz, the in-phase carrier is cos(2π * 10^10 * t) and the quadrature carrier is sin(2π * 10^10 * t). Multiply the decision signal with the in-phase carrier and the quadrature carrier to obtain the first product signal and the second product signal.

[0112] The first product signal and the second product signal are respectively subjected to low-pass filtering to remove high-frequency components to obtain an in-phase component and a quadrature component. The in-phase component is used as a real part and the quadrature component is used as an imaginary part to obtain a final baseband signal.

[0113] In this embodiment, the interference and clutter components in the radar echo signal can be effectively filtered out by adaptive space-time two-dimensional filtering, and the signal-to-noise ratio and signal-to-clutter ratio can be improved. Through threshold judgment, noise and residual interference can be effectively removed, and the probability and accuracy of target detection can be improved. By multiplying and low-pass filtering with the carrier of the same frequency and phase, synchronous coherent demodulation can be achieved, and the baseband signal can be extracted, providing a basis for subsequent target recognition and parameter estimation.

[0114] In an optional implementation, low-pass filtering the first product signal and the second product signal to remove high-frequency components to obtain an in-phase component and a quadrature component, and combining the in-phase component as a real part and the quadrature component as an imaginary part to obtain a final baseband signal includes:

[0115] Multiplying the radar echo signal with the first product signal and the second product signal respectively to obtain an in-phase component and an orthogonal component;

[0116] Performing time-frequency analysis on the in-phase component and the orthogonal component, calculating the dynamic bandwidth parameter of the signal, adaptively adjusting the cutoff frequency of the low-pass filter according to the dynamic bandwidth parameter, and calculating the time-frequency concentration of the signal according to the dynamic bandwidth parameter, adaptively determining the shape factor of the Kaiser window function based on the time-frequency concentration, and generating a window function sequence using the shape factor;

[0117] The window function sequence is combined with the cutoff frequency to generate a low-pass filter coefficient, and the low-pass filter coefficient is used to filter the in-phase component and the orthogonal component respectively to obtain an in-phase baseband signal and an orthogonal baseband signal;

[0118] Calculating the signal-to-noise ratio of the in-phase baseband signal and the quadrature baseband signal, generating signal weights according to the signal-to-noise ratios, and multiplying the in-phase baseband signal and the quadrature baseband signal by the corresponding signal weights to obtain a weighted in-phase signal and a weighted quadrature signal;

[0119] The weighted in-phase signal is used as the real part and the weighted quadrature signal is used as the imaginary part to perform orthogonal combination to obtain a complex baseband output signal as the final baseband signal.

[0120] Exemplarily, first, a reference signal for mixing is generated. Assume that the transmitting signal of the radar system is a linear frequency modulation signal with a center frequency of 10 GHz, a bandwidth of 100 MHz, and a pulse width of 10 μs. Then two orthogonal local oscillator signals can be generated, for example, the first local oscillator signal is a cosine signal with a frequency of 10 GHz, and the second local oscillator signal is a sine signal with a frequency of 10 GHz. The transmitting signal is multiplied by the two local oscillator signals to obtain a first product signal and a second product signal.

[0121] Next, the radar echo signal is multiplied by the first product signal and the second product signal respectively to obtain the in-phase component and the quadrature component. For example, assuming that the echo signal is a reflected signal delayed by 1 μs and attenuated by 3 dB. By multiplying the first and second product signals, the in-phase and quadrature components can be obtained.

[0122] Then, the in-phase component and the orthogonal component are subjected to time-frequency analysis. For example, the short-time Fourier transform (STFT) can be used to analyze the time-frequency characteristics of the signal. Assume that the analysis results show that the dynamic bandwidth of the signal fluctuates between 0.8MHz and 1.2MHz. According to the dynamic bandwidth parameter, the cutoff frequency of the low-pass filter is adaptively adjusted. For example, the cutoff frequency can be set to the average value of the dynamic bandwidth, that is, 1MHz. At the same time, the time-frequency aggregation of the signal is calculated based on the dynamic bandwidth. Assume that the calculated time-frequency aggregation is 0.9. The shape factor of the Kaiser window function is adaptively determined based on the time-frequency aggregation. For example, the shape factor can be set to 2 according to the empirical formula. The window function sequence is generated using the shape factor.

[0123] The window function sequence is combined with the cutoff frequency to generate low-pass filter coefficients. For example, the window function method can be used to design an FIR filter. The generated low-pass filter coefficients are used to filter the in-phase component and the orthogonal component respectively to obtain the in-phase baseband signal and the orthogonal baseband signal.

[0124] Calculate the signal-to-noise ratio of the in-phase baseband signal and the quadrature baseband signal. For example, the signal-to-noise ratio can be calculated based on the ratio of signal power to noise power. Assume that the calculated signal-to-noise ratio of the in-phase baseband signal is 15 dB, and the signal-to-noise ratio of the quadrature baseband signal is 12 dB. Generate signal weights based on the signal-to-noise ratio. For example, the signal-to-noise ratio can be converted into a linear value and normalized to obtain the corresponding signal weights. Assume that the calculated weight of the in-phase baseband signal is 0.8, and the weight of the quadrature baseband signal is 0.6. Multiply the in-phase baseband signal and the quadrature baseband signal by the corresponding signal weights to obtain a weighted in-phase signal and a weighted quadrature signal.

[0125] Finally, the weighted in-phase signal is used as the real part and the weighted orthogonal signal is used as the imaginary part, and an orthogonal combination is performed to obtain a complex baseband output signal as the final baseband signal.

[0126] In this embodiment, by adaptively adjusting the cutoff frequency of the low-pass filter and the shape factor of the Kaiser window function, out-of-band noise and interference can be effectively suppressed, and the signal-to-noise ratio and anti-interference performance of the signal can be improved. Through time-frequency analysis and calculation of dynamic bandwidth parameters, the characteristic information of the signal can be more accurately extracted, thereby improving the accuracy of radar signal processing. The method can adaptively adjust the filter parameters and weighting coefficients according to the dynamic bandwidth and time-frequency concentration of the signal, thereby enhancing the adaptability of the radar system to different environments and targets.

[0127] In an optional implementation, inputting the first processed signal into a dual-tree complex wavelet transform module for multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, and extracting nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct a feature vector includes:

[0128] Constructing a dual-tree complex wavelet transform module including a symmetric orthogonal filter group, wherein the symmetric orthogonal filter group includes a first group of orthogonal filters and a second group of orthogonal filters, inputting the first processed signal into the symmetric orthogonal filter group for four-layer wavelet decomposition, and obtaining complex wavelet coefficients of different frequency bands based on output signals of the first group of orthogonal filters and the second group of orthogonal filters;

[0129] A reconstruction vector is constructed based on the complex wavelet coefficients, and the distance between the vectors of the reconstructed vector is calculated to obtain a distance parameter. The number of vectors that meet the distance parameter condition is counted according to a preset threshold to obtain a matching number. The conditional probability logarithm is calculated based on the matching number to obtain an approximate entropy feature; based on the matching number, the ratio of the template matching number in different dimensions is calculated to obtain a sample entropy feature;

[0130] Performing a continuous wavelet transform on the complex wavelet coefficients to obtain a time-frequency coefficient matrix, extracting the maximum value of the modulus value at each time point based on the time-frequency coefficient matrix to obtain a time-frequency ridge feature; performing a Hilbert transform on the first processed signal to obtain an analytical signal, and calculating the instantaneous frequency feature according to the phase information of the analytical signal;

[0131] The approximate entropy feature, the sample entropy feature, the time-frequency ridge feature and the instantaneous frequency feature are combined to construct a feature vector.

[0132] Exemplarily, first, a first processing signal to be processed is obtained. This signal can be various types of signals, such as speech signals, electrocardiogram signals, earthquake signals, etc. Take a simulated electrocardiogram signal with a sampling rate of 1000 Hz and a duration of 1 second as an example. The signal contains multiple frequency components and nonlinear dynamic characteristics.

[0133] Next, a dual-tree complex wavelet transform module is constructed. The core of this module is two groups of orthogonal filters, which are called the first group of orthogonal filters and the second group of orthogonal filters. These two groups of filters need to be carefully designed to meet the requirements of symmetry and orthogonality. For example, the db4 wavelet can be selected as the first group of orthogonal filters, and the corresponding second group of orthogonal filters can be obtained by performing a certain transformation on the db4 wavelet to form a symmetric orthogonal filter group. The first processed signal is input into this symmetric orthogonal filter group for four-layer wavelet decomposition. Each layer of decomposition will decompose the signal into a high-frequency part and a low-frequency part. After four layers of decomposition, the complex wavelet coefficients of four high-frequency sub-bands and one low-frequency sub-band are obtained. Assuming that the high-frequency coefficient obtained by the first layer of decomposition is H1 and the low-frequency coefficient is L1, the second layer of decomposition decomposes L1 into H2 and L2, and so on, and finally obtains the complex wavelet coefficients of five frequency bands of H1, H2, H3, H4 and L4.

[0134] Then, nonlinear feature parameters and time-frequency joint feature parameters are extracted based on the obtained complex wavelet coefficients. First, the approximate entropy feature and the sample entropy feature are calculated. Select an embedding dimension m, for example, m=2, and a distance threshold r, for example, r=0.2*standard deviation. Construct a reconstructed vector and calculate the distance between the reconstructed vectors. Count the number of vectors whose distance is less than the threshold r, that is, the number of matches. Calculate the logarithm of the conditional probability based on the number of matches to obtain the approximate entropy feature. Similarly, calculate the ratio of the number of template matches in different dimensions based on the number of matches to obtain the sample entropy feature. Assuming that the embedding dimension m=2 and the threshold r=0.2*standard deviation, the approximate entropy value is calculated to be 0.5 and the sample entropy value is 0.8.

[0135] Next, the time-frequency ridge features and instantaneous frequency features are extracted. Perform continuous wavelet transform on the complex wavelet coefficients to obtain the time-frequency coefficient matrix. For each time point, find the position with the largest modulus value, and connect these positions to form the time-frequency ridge features. For example, at 0.1 seconds, the maximum modulus value appears at 50Hz, at 0.2 seconds, the maximum modulus value appears at 60Hz, and so on, to obtain the time-frequency ridge. Perform Hilbert transform on the first processed signal to obtain the analytical signal, and calculate the instantaneous frequency features based on the phase information of the analytical signal. For example, at 0.1 seconds, the calculated instantaneous frequency is 55Hz, at 0.2 seconds, the calculated instantaneous frequency is 62Hz, and so on, to obtain the instantaneous frequency curve.

[0136] Finally, the calculated approximate entropy features, sample entropy features, time-frequency ridge features, and instantaneous frequency features are combined into a feature vector. For example, the approximate entropy value of 0.5, the sample entropy value of 0.8, the time-frequency ridge, and the instantaneous frequency curve are combined together to form the final feature vector. This feature vector can be used for subsequent classification, recognition, and other tasks.

[0137] In this embodiment, by combining multiple methods such as dual-tree complex wavelet transform, approximate entropy, sample entropy, time-frequency ridge and instantaneous frequency, the nonlinear dynamic characteristics and time-frequency information of the signal can be captured more comprehensively, thereby improving the accuracy of feature extraction. Dual-tree complex wavelet transform has good time-frequency localization ability and noise resistance, and approximate entropy and sample entropy are insensitive to small changes in the signal. These characteristics make the extracted features more robust and less susceptible to noise and interference. This method can be applied to various types of signal processing tasks, such as fault diagnosis, biomedical signal analysis, image processing, etc., and has broad application prospects.

[0138] In an optional implementation, inputting the feature vector into a multi-task deep learning network with an integrated attention mechanism for feature fusion and target recognition to obtain a second processed signal includes:

[0139] Constructing a deep learning network including a feature extraction layer, a self-attention layer, and a multi-task output layer, inputting a feature vector into the feature extraction layer, which includes a plurality of cascaded residual blocks, each residual block including a first convolutional layer, a second convolutional layer, and a skip connection path, the feature vector is passed through the first convolutional layer to obtain a first transformation feature, the first transformation feature is passed through the second convolutional layer to obtain a second transformation feature, and the second transformation feature is added to the feature vector transmitted through the skip connection path to obtain a local feature;

[0140] The local features are input into a self-attention layer, the self-attention layer performs a linear transformation on the local features to obtain a query matrix, a key matrix and a value matrix, the query matrix is ​​multiplied by the transpose of the key matrix to obtain a correlation matrix, and the correlation matrix is ​​normalized to obtain an attention weight matrix;

[0141] Multiplying the attention weight matrix by the value matrix to obtain a first weighted feature, constructing multiple groups of self-attention layers in parallel to obtain multiple groups of first weighted features, and concatenating the multiple groups of first weighted features in the feature dimension to obtain a second weighted feature;

[0142] Constructing a multi-task output layer including a detection subnetwork and a classification subnetwork, wherein the detection subnetwork and the classification subnetwork are constructed in a parameter sharing manner, inputting the second weighted feature into the detection subnetwork and the classification subnetwork at the same time, performing target detection on the second weighted feature through the detection subnetwork to obtain a target position feature and a target confidence feature, and performing target classification on the second weighted feature through the classification subnetwork to obtain a target category feature;

[0143] Constructing a joint loss function, the joint loss function including a detection loss term and a classification loss term calculated according to target position features and target confidence features, and optimizing parameters of the detection subnetwork and the classification subnetwork based on the joint loss function;

[0144] The target position features and target confidence features output by the optimized detection subnetwork and the target category features output by the optimized classification subnetwork are combined to obtain a second processed signal.

[0145] For example, first, a deep learning network is constructed, which includes three main parts: feature extraction layer, self-attention layer and multi-task output layer.

[0146] The feature extraction layer is composed of multiple cascaded residual blocks. Each residual block contains two convolutional layers and a skip connection path. The input feature vector is first transformed by the first convolutional layer to obtain the first transformed feature. Then, the first transformed feature is transformed by the second convolutional layer to obtain the second transformed feature. Finally, the second transformed feature is added to the original feature vector directly transmitted through the skip connection path to obtain the output of the residual block, that is, the local feature. For example, the input feature vector dimension is 256, and the first transformed feature with a dimension of 256 is obtained by the first convolutional layer (for example: the convolution kernel size is 3x3, and the number of convolution kernels is 256). Then, the second transformed feature with a dimension of 256 is obtained by the second convolutional layer (for example: the convolution kernel size is 3x3, and the number of convolution kernels is 256). Finally, it is added to the original feature vector with a dimension of 256 to obtain a local feature with a dimension of 256. After multiple residual blocks are cascaded, more representative local features are finally obtained.

[0147] Next, the local features are input into the self-attention layer. The self-attention layer first performs a linear transformation on the local features to obtain three matrices: a query matrix, a key matrix, and a value matrix. For example, the local feature dimension is 256. After three different linear transformations (for example, the transformation matrix size is 256x256), the query matrix, key matrix, and value matrix with dimensions of 256 are obtained. Then, the query matrix is ​​multiplied by the transpose of the key matrix to obtain the correlation matrix. The correlation matrix is ​​then normalized, for example, using the Softmax function, to obtain the attention weight matrix. Finally, the attention weight matrix is ​​multiplied by the value matrix to obtain the first weighted feature. In order to enhance the expressive power of the network, multiple groups of self-attention layers are constructed in parallel, for example, 3 groups are constructed to obtain multiple groups of first weighted features. These multiple groups of first weighted features are spliced ​​in the feature dimension, for example, 3 groups of first weighted features with dimensions of 256 are spliced ​​into second weighted features with a dimension of 768.

[0148] Finally, the second weighted feature is input to the multi-task output layer. The multi-task output layer consists of a detection subnetwork and a classification subnetwork, and these two subnetworks are constructed in a parameter-sharing manner. The detection subnetwork is used to predict the location and confidence of the target, such as outputting the bounding box coordinates of the target and the probability of the target's existence. The classification subnetwork is used to predict the category of the target, such as outputting the probability that the target belongs to different categories. The two subnetworks receive the second weighted feature as input at the same time, and perform target detection and classification respectively. For example, after the second weighted feature with a dimension of 768 is input to the detection subnetwork and the classification subnetwork, the detection subnetwork outputs 4 coordinate values ​​and 1 confidence value, and the classification subnetwork outputs 10 category probability values ​​(assuming there are 10 categories).

[0149] In order to train this network, a joint loss function is constructed. This function contains two items: detection loss term and classification loss term. The detection loss term is calculated based on the predicted target position and confidence and the actual target position and confidence, for example, using mean square error to calculate the position loss and using cross entropy loss to calculate the confidence loss. The classification loss term is calculated based on the predicted target category and the actual category, for example, using cross entropy loss. Based on this joint loss function, the parameters of the detection subnetwork and the classification subnetwork are optimized, for example, using back propagation algorithm and stochastic gradient descent algorithm.

[0150] After training and optimization, the target position features and target confidence features output by the optimized detection subnetwork and the target category features output by the optimized classification subnetwork are combined to obtain the final second processed signal. For example, 4 coordinate values, 1 confidence value and 10 category probability values ​​are combined to form a vector containing 16 values ​​as the final second processed signal.

[0151] In this embodiment, by introducing residual blocks and self-attention mechanisms, the network can better extract and fuse key information from input features, thereby improving the expressiveness of features. Through the multi-task learning framework, target detection and classification tasks can share network parameters and feature representations, thereby achieving mutual promotion and improving overall performance. Through parameter sharing and joint optimization, the number of model parameters and computational complexity are reduced, thereby improving the efficiency of target recognition.

[0152] In an optional implementation, a target characteristic spectrum is obtained by calculating the time-frequency distribution of the second processed signal, a morphological filter is used to perform noise reduction processing on the target characteristic spectrum, and an adaptive morphological filter is used to selectively suppress noise on the target characteristic spectrum, and a multi-target segmentation algorithm based on a variational model is used to extract the respiratory frequency component and the heart rate frequency component in the target characteristic spectrum. Judging whether there is a living body in the target area according to the time-varying characteristics and amplitude fluctuation law of the respiratory frequency component and the heart rate frequency component includes:

[0153] Performing short-time Fourier transform on the second processed signal to obtain a time-frequency distribution diagram, performing frame windowing processing on the second processed signal using a Hanning window function, performing fast Fourier transform on the windowed signal to obtain a spectrum sequence, and splicing the spectrum sequence in the time dimension to obtain a target characteristic spectrum;

[0154] Constructing a double-layer morphological filter structure, wherein the first layer of the double-layer morphological filter structure includes a basic morphological filter, wherein the basic morphological filter adopts a disk-shaped structure element, and the disk-shaped structure element is used to perform an opening operation on the target feature spectrum to remove the peak noise higher than the background to obtain a first intermediate signal, and a closing operation is performed on the first intermediate signal to fill the valley noise lower than the background to obtain a first filtered signal;

[0155] Calculating the mean, variance and skewness of each time-frequency point neighborhood in the first filtered signal to obtain local statistical characteristics, calculating an adaptive threshold based on the local statistical characteristics, and comparing the adaptive threshold with a preset decision criterion to obtain noise area identification information;

[0156] Dynamically adjusting the shape and size parameters of the structural elements of the adaptive morphological filter according to the noise area identification information, selectively filtering the time-frequency points identified as the noise area to obtain a second filtered signal, and keeping the time-frequency points identified as the target feature area unchanged;

[0157] An energy functional is constructed, wherein a data item of the energy functional represents the degree of matching between the second filtered signal and the respiratory and heartbeat feature model, a regularization item introduces a spatial continuity constraint, and an alternating direction multiplier method is used to optimize and solve the energy functional to obtain a frequency component segmentation result;

[0158] Performing region marking and connected domain analysis on the frequency component segmentation results, extracting the time-varying trajectories of the respiratory frequency component and the heartbeat frequency component, and calculating the periodicity index, amplitude fluctuation law and energy distribution characteristics of the time-varying trajectories;

[0159] The periodic index, amplitude fluctuation law and energy distribution characteristics are input into a pre-trained physiological feature matching model to obtain a feature matching score, and the signal-to-noise ratio of the respiratory frequency component and the heart rate frequency component is calculated. When the feature matching score and the signal-to-noise ratio simultaneously meet the preset threshold conditions, it is determined that there is a living body in the target area.

[0160] Exemplarily, first, a short-time Fourier transform is performed on the preprocessed micro-motion signal to obtain a time-frequency distribution diagram of the signal. Specifically, the signal is framed and windowed using the Hanning window function, and then each frame of the windowed signal is fast Fourier transformed to obtain a spectrum sequence. Finally, the spectrum sequences of all frames are spliced ​​together in chronological order to form a target characteristic spectrum. For example, for a signal with a sampling rate of 100 Hz and a duration of 10 seconds, a Hanning window with a window length of 0.5 seconds (50 sampling points) and an overlap rate of 50% is used for framing, and 19 frames of data can be obtained. After fast Fourier transforming each frame of data, 19 spectrum sequences are obtained, and these spectrum sequences are spliced ​​together to obtain the target characteristic spectrum.

[0161] In order to remove the noise in the target feature spectrum, a two-layer morphological filter structure is used. The first layer uses a basic morphological filter and uses a disk-shaped structure element to perform an opening operation on the target feature spectrum to remove the peak noise above the background and obtain the first intermediate signal. Then, a closing operation is performed on the first intermediate signal to fill the valley noise below the background and obtain the first filtered signal. For example, a disk-shaped structure element with a radius of 3 can be used for opening and closing operations.

[0162] Next, the first filtered signal is subjected to adaptive morphological filtering. First, local statistical characteristics such as the mean, variance, and skewness of each time-frequency point neighborhood in the first filtered signal are calculated. Then, an adaptive threshold is calculated based on these local statistical characteristics. The calculated adaptive threshold is compared with a preset decision criterion to obtain noise area identification information. For example, it can be set that if the local variance of a certain time-frequency point is greater than a preset threshold, it is identified as a noise area.

[0163] According to the noise area identification information, the shape and size parameters of the structure element of the adaptive morphological filter are dynamically adjusted. The time-frequency points identified as the noise area are selectively filtered to obtain the second filtered signal. The time-frequency points identified as the target feature area remain unchanged. For example, for the time-frequency points identified as the noise area, a larger structure element can be used for filtering; while for the target feature area, it remains unchanged or a smaller structure element is used for filtering.

[0164] In order to extract the respiratory and heartbeat frequency components from the second filtered signal, a multi-objective segmentation algorithm based on a variational model is used. An energy functional is constructed, in which the data term characterizes the degree of matching between the second filtered signal and the respiratory and heartbeat feature model, and the regularization term introduces a spatial continuity constraint. The energy functional is optimized and solved using the alternating direction multiplier method to obtain the frequency component segmentation result. For example, the respiratory and heartbeat frequency ranges can be set to 0.2Hz-0.4Hz and 0.8Hz-1.5Hz, respectively.

[0165] The frequency component segmentation results are marked with regions and connected domains are analyzed to extract the time-varying trajectories of the respiratory frequency component and the heartbeat frequency component. Then, the periodicity index, amplitude fluctuation law, and energy distribution characteristics of the time-varying trajectory are calculated. For example, the instantaneous frequency, instantaneous amplitude, and signal energy of the respiratory and heartbeat frequencies can be calculated.

[0166] Finally, the extracted periodic indicators, amplitude fluctuation rules and energy distribution characteristics are input into the pre-trained physiological feature matching model to obtain the feature matching score. At the same time, the signal-to-noise ratio of the respiratory frequency component and the heartbeat frequency component is calculated. When the feature matching score and the signal-to-noise ratio meet the preset threshold conditions at the same time, it is determined that there is a living body in the target area. For example, the feature matching score threshold can be set to 0.8 and the signal-to-noise ratio threshold can be set to 5dB. Among them, the pre-trained physiological feature matching model is used in the life detection scheme to identify and confirm whether there are real life features in the signal. Its core function is to evaluate the similarity between the processed signal features and the typical physiological signal features of living bodies. The key to this model is to identify the small periodic change characteristics of breathing and heartbeat, so as to distinguish effective life signals from background noise.

[0167] Specifically, first, the physiological feature matching model requires a large amount of breathing and heartbeat signal data of living organisms for training. The data is usually obtained through real breathing and heartbeat signals collected in a variety of radar monitoring environments to cover different individual characteristics, frequency variation ranges and different environmental noise conditions. In order to improve the generalization of the model, it is also necessary to collect typical interference signals of non-living organisms as negative samples of the model to help the model learn to distinguish between life feature signals and non-life feature signals. Key signal features such as periodic indicators, amplitude fluctuation rules, and energy distribution are extracted from the collected signals. These features can reflect the physiological cycle and intensity changes of breathing and heartbeat. For example, the frequency of breathing signals is generally between 0.2 Hz and 0.5 Hz, while that of heartbeat signals is between 0.8 Hz and 1.5 Hz. These features are obtained by analyzing the time-frequency distribution and energy curve of the signal. According to the characteristics of the above features, models such as neural networks or support vector machines are often used to achieve classification. In particular, for the recognition of nonlinear and complex periodic features, multi-layer neural network structures are more suitable. Through layer-by-layer feature extraction, high-dimensional feature vectors reflecting the breathing and heartbeat characteristics of living organisms can be gradually abstracted. The sample data is input into the model for training by category. The model parameters are iteratively adjusted to minimize the classification error, so that the model can accurately distinguish between breathing, heartbeat signals and non-life signals. The cross-validation method is introduced during the training process to ensure the stability and robustness of the model on different samples. During the optimization process, a loss function with a high correlation with signal features is used to improve the model's sensitivity to periodicity and energy distribution features. After the model training is completed, a new set of data is used for verification to ensure that the model has a high recognition rate for breathing and heartbeat signals in the actual environment. At the same time, the model's ability to exclude various noise and clutter signals is verified so that the model can adapt to different scenarios and signal changes.

[0168] In this embodiment, through double-layer morphological filtering and adaptive noise suppression, the noise interference in the target feature spectrum is effectively removed, and the extraction accuracy of the breathing and heartbeat frequency components is improved. The multi-target segmentation algorithm based on the variational model can effectively extract the target frequency components from the complex time-frequency spectrum diagram, and has strong robustness to noise and interference. This method can realize the detection of vital signs without direct contact with the target, and has a wider range of application scenarios.

[0169] In an optional implementation, if it is determined that there is a life form in the target area, the three-dimensional spatial coordinates of the life form are calculated by a multi-base station collaborative positioning algorithm, and a detection result including the three-dimensional spatial coordinates and the life form characteristic information is outputted, including:

[0170] Acquire receiving signals of multiple radar base stations and perform bandpass filtering to obtain filtered receiving signals, perform envelope detection on the filtered receiving signals to obtain envelope signals, obtain the peak moment of the envelope signal by peak detection, and determine the peak moment as the arrival time;

[0171] Performing phase analysis on the filtered received signal to obtain a phase spectrum, calculating the phase spectrum difference between adjacent radar base stations to obtain a phase difference, using the phase difference as an arrival time difference, constructing a hyperbola equation based on the arrival time difference, and performing spatial filtering on the filtered received signal to obtain a filtered direction signal, calculating the covariance matrix of the filtered direction signal, obtaining an arrival angle by eigendecomposing the covariance matrix, and constructing an azimuth angle equation based on the arrival angle;

[0172] The hyperbolic equation and the azimuth equation are combined to obtain a positioning equation, a ratio of signal power to noise power in the filtered received signal is calculated to obtain a signal-to-noise ratio, and the signal-to-noise ratio is normalized to obtain a weight coefficient;

[0173] Performing Taylor expansion on the positioning equation to obtain a linear equation, multiplying the weight coefficient by the coefficient of the linear equation to obtain a weighted linear equation, iteratively calculating the weighted linear equation to obtain a position increment, stopping the iteration when the position increment is less than a position threshold, and outputting the three-dimensional space coordinates;

[0174] Calculating the geometric precision factor of the three-dimensional space coordinates, and when the geometric precision factor is greater than a second preset threshold, calculating the positioning errors of each base station combination to obtain an error matrix, and selecting the base station combination with the smallest positioning error as the optimal base station combination based on the error matrix;

[0175] The three-dimensional space coordinates are optimized using the optimal base station combination to obtain optimized three-dimensional space coordinates, statistical parameters of respiratory frequency components and heart rate frequency components in a continuous time window are extracted, and the optimized three-dimensional space coordinates are combined with the statistical parameters to obtain a detection result containing three-dimensional space coordinates and vital feature information.

[0176] Exemplarily, first, each base station receives the radar signal reflected from the target area. In order to remove interference noise, the received signal is bandpass filtered to retain only the signal component in the frequency band where the target life body reflection signal is located. Then, envelope detection is performed on the filtered signal to obtain an envelope signal, and the peak moment of the envelope signal is found through a peak detection algorithm. The peak moment is the time when the radar signal arrives at the base station.

[0177] The location of the target life form can be calculated using the arrival time difference of the signals received by multiple base stations. Specifically, the filtered received signal is subjected to phase analysis to obtain the phase spectrum of the received signal at each base station. Then, the phase spectrum difference of the received signals between adjacent base stations is calculated to obtain the phase difference, which is proportional to the arrival time difference. Using the arrival time difference, multiple hyperbolic equations can be constructed. For example, assuming that there are two base stations A and B, and the distance difference between them to the target life form is a constant, then the target life form is located on a hyperbola with A and B as the focus.

[0178] In addition to the arrival time difference, the arrival angle can also be used to further determine the location of the target life form. The filtered received signal is spatially filtered to obtain a filtered direction signal. Then, the covariance matrix of the filtered direction signal is calculated and the eigendecomposition is performed to obtain the arrival angle. The arrival angle can be used to construct an azimuth equation. For example, if the target life form is located due east of base station A, an azimuth equation can be constructed to describe this relationship.

[0179] By combining the hyperbola equation and the azimuth equation, we can get a positioning equation group. In order to improve the positioning accuracy, it is necessary to consider the quality of the received signal of each base station. The ratio of the signal power to the noise power of the received signal of each base station is calculated to get the signal-to-noise ratio. The signal-to-noise ratio is normalized to get the weight coefficient. The larger the weight coefficient, the better the quality of the received signal of the base station.

[0180] In order to solve the positioning equations, the positioning equations are Taylor expanded to obtain linear equations. The weight coefficients are multiplied by the coefficients of the linear equations to obtain weighted linear equations. The weighted linear equations are iteratively calculated to obtain position increments. When the position increment is less than the preset position threshold, the iteration is stopped and the three-dimensional spatial coordinates of the target organism are output.

[0181] In order to evaluate the positioning accuracy, the geometric precision factor of the three-dimensional space coordinates is calculated. If the geometric precision factor is greater than the preset second threshold, the positioning result needs to be further optimized. The positioning error of each base station combination is calculated to obtain an error matrix. The base station combination with the smallest positioning error is selected as the optimal base station combination. The three-dimensional space coordinates are optimized using the optimal base station combination to obtain the optimized three-dimensional space coordinates.

[0182] In addition to location information, vital signs information of the target life can also be extracted. Statistical parameters of the respiratory frequency component and the heart rate component in the continuous time window, such as mean value, variance, etc., are extracted. The optimized three-dimensional space coordinates are combined with the statistical parameters to obtain detection results containing three-dimensional space coordinates and life feature information.

[0183] For example, suppose there are three radar base stations A, B and C, and the peak times of the signals they receive are 1s, 1.2s and 1.1s respectively. The phase differences between adjacent base stations are π / 4 and π / 6 respectively. By calculation, the initial three-dimensional coordinates of the target life form are (10, 20, 30). The calculated geometric precision factor is 0.8, which is greater than the preset second threshold value of 0.5. The positioning errors of each base station combination are calculated, and it is found that the combined positioning error of base stations A and B is the smallest. The three-dimensional coordinates are optimized using the combination of base stations A and B, and the optimized three-dimensional coordinates are (11, 21, 31). The statistical parameters of the respiratory frequency component and the heart rate component are extracted, for example, the respiratory frequency is 16 times / minute and the heart rate is 70 times / minute. The final output detection results are: three-dimensional coordinates (11, 21, 31), respiratory frequency 16 times / minute, and heart rate 70 times / minute.

[0184] In this embodiment, through multi-base station collaborative positioning, weighted linear equation iterative calculation and optimal base station combination selection based on geometric precision factor, the accuracy of life body positioning can be effectively improved, positioning error can be reduced, and more accurate target location information can be obtained. This method can not only obtain the three-dimensional spatial coordinates of the life body, but also extract vital signs information such as breathing rate and heart rate, provide more comprehensive target information, and help to conduct a deeper analysis and judgment of the target state. Through signal-to-noise ratio weighting and optimal base station selection, the impact of environmental noise and base station failure on the positioning results can be effectively reduced, the robustness and reliability of the system can be enhanced, and it can work stably even in complex environments.

[0185] Figure 2 FIG. 1 is a schematic diagram of the structure of a life detection system based on radar signal processing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0186] The first unit is used to collect radar echo signals of the target area, control the phased array radar transmitting antenna array to transmit a dual-frequency continuous wave radar signal of orthogonal frequency division multiplexing, wherein the dual-frequency continuous wave radar signal includes a first frequency signal and a second frequency signal, the first frequency signal and the second frequency signal are coded and modulated by Walsh orthogonal codes, the radar echo signal is received by a multi-channel radar receiving system, interference suppression and clutter cancellation processing is performed on the radar echo signal by using an adaptive space-time two-dimensional filter, and the processed radar echo signal is input into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal;

[0187] The second unit is used to perform digital sampling processing on the baseband signal to obtain a sampling signal, perform multi-stage static clutter elimination processing on the sampling signal using an adaptive notch filter group to obtain a first processed signal, input the first processed signal into a dual-tree complex wavelet transform module to perform multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extract nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct a feature vector, the nonlinear feature parameters include approximate entropy and sample entropy, the time-frequency joint feature parameters include time-frequency ridge and instantaneous frequency, and input the feature vector into a multi-task deep learning network with an integrated attention mechanism to perform feature fusion and target recognition to obtain a second processed signal;

[0188] The third unit is used to obtain a target characteristic spectrum by calculating the time-frequency distribution of the second processed signal, perform noise reduction processing on the target characteristic spectrum by using a morphological filter, and selectively suppress noise on the target characteristic spectrum by using an adaptive morphological filter, and extract the respiratory frequency component and the heart rate component in the target characteristic spectrum by using a multi-target segmentation algorithm based on a variational model, and judge whether there is a living body in the target area according to the time-varying characteristics and amplitude fluctuation law of the respiratory frequency component and the heart rate component; if it is determined that there is a living body in the target area, calculate the three-dimensional spatial coordinates of the living body by using a multi-base station collaborative positioning algorithm, and output a detection result including the three-dimensional spatial coordinates and the characteristic information of the living body.

[0189] According to a third aspect of the embodiments of the present invention,

[0190] An electronic device is provided, comprising:

[0191] processor;

[0192] a memory for storing processor-executable instructions;

[0193] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0194] A fourth aspect of the embodiments of the present invention is:

[0195] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0196] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0197] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A life detection method based on radar signal processing, characterized in that: include: Collect radar echo signals in the target area, control the phased array radar transmitting antenna array to transmit orthogonal frequency division multiplexing dual-frequency continuous wave radar signals, the dual-frequency continuous wave radar signals include a first frequency signal and a second frequency signal, the first frequency signal and the second frequency signal are coded and modulated by Walsh orthogonal codes, the radar echo signals are received by a multi-channel radar receiving system, interference suppression and clutter cancellation processing is performed on the radar echo signals by using an adaptive space-time two-dimensional filter, and the processed radar echo signals are input into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal; Performing digital sampling processing on the baseband signal to obtain a sampling signal, performing multi-stage static clutter elimination processing on the sampling signal using an adaptive notch filter group to obtain a first processed signal, inputting the first processed signal into a dual-tree complex wavelet transform module to perform multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extracting nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct a feature vector, wherein the nonlinear feature parameters include approximate entropy and sample entropy, and the time-frequency joint feature parameters include time-frequency ridge and instantaneous frequency, and inputting the feature vector into a multi-task deep learning network with an integrated attention mechanism to perform feature fusion and target recognition to obtain a second processed signal; The target characteristic spectrum is obtained by calculating the time-frequency distribution of the second processed signal, and the target characteristic spectrum is denoised by using a morphological filter. At the same time, the target characteristic spectrum is selectively noise-suppressed by using an adaptive morphological filter. A multi-target segmentation algorithm based on a variational model is used to extract the respiratory frequency component and the heart rate frequency component in the target characteristic spectrum. It is judged whether there is a living body in the target area according to the time-varying characteristics of the respiratory frequency component and the heart rate frequency component and the amplitude fluctuation law. If it is determined that there is a living body in the target area, the three-dimensional spatial coordinates of the living body are calculated by a multi-base station collaborative positioning algorithm, and a detection result containing the three-dimensional spatial coordinates and the characteristic information of the living body is output.

2. The method according to claim 1, characterized in that: Using an adaptive space-time two-dimensional filter to perform interference suppression and clutter cancellation processing on the radar echo signal, and inputting the processed radar echo signal into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal includes: Collecting time-series sampling data of multiple spatial sampling channels of a radar receiving system, and constructing a space-time two-dimensional sampling data matrix with the time-series sampling data according to the spatial sampling channel sequence number and the time sampling sequence number, wherein the space-time two-dimensional sampling data matrix includes a target echo component, an interference component, and a clutter component; Initializing the complex weight vector and correlation matrix of the adaptive space-time two-dimensional filter, performing an inner product operation on the complex weight vector and the space-time two-dimensional sampling matrix to obtain an initial filter output, and calculating a prediction error signal based on the obtained initial filter output and the expected signal; Calculating a time-varying forgetting factor according to the prediction error signal, multiplying the product of the time-varying forgetting factor and the correlation matrix by the space-time two-dimensional sampling matrix to obtain a gain vector, wherein the gain vector is used to update the complex weight value vector; The complex weight vector is updated by using the gain vector and the prediction error signal, and the space-time two-dimensional sampling matrix is ​​weighted filtered by using the updated complex weight vector to obtain a filtered signal, and the filtered signal is envelope detected to obtain an envelope signal, and a signal sample is selected on the envelope signal according to a preset time window length to construct a sliding time window, and the amplitude mean and amplitude standard deviation of the signal samples in the sliding time window are calculated; The weighted sum of the amplitude mean and the amplitude standard deviation is used as a threshold value, and an amplitude decision is performed on the filtered signal according to the threshold value to obtain a decision signal, wherein the decision signal retains a signal amplitude and phase greater than the threshold value; Generate a cosine waveform in-phase carrier and a sine waveform orthogonal carrier with the same frequency and phase as the radar transmission signal, and multiply the decision signal with the in-phase carrier and the orthogonal carrier to obtain a first product signal and a second product signal; The first product signal and the second product signal are respectively subjected to low-pass filtering to remove high-frequency components to obtain an in-phase component and a quadrature component, and the in-phase component is used as a real part and the quadrature component is used as an imaginary part to obtain a final baseband signal.

3. The method according to claim 2, characterized in that The first product signal and the second product signal are respectively subjected to low-pass filtering to remove high-frequency components to obtain an in-phase component and an orthogonal component, and the in-phase component is used as a real part and the orthogonal component is used as an imaginary part to obtain a final baseband signal, comprising: Multiplying the radar echo signal with the first product signal and the second product signal respectively to obtain an in-phase component and an orthogonal component; Performing time-frequency analysis on the in-phase component and the orthogonal component, calculating the dynamic bandwidth parameter of the signal, adaptively adjusting the cutoff frequency of the low-pass filter according to the dynamic bandwidth parameter, and calculating the time-frequency concentration of the signal according to the dynamic bandwidth parameter, adaptively determining the shape factor of the Kaiser window function based on the time-frequency concentration, and generating a window function sequence using the shape factor; The window function sequence is combined with the cutoff frequency to generate a low-pass filter coefficient, and the low-pass filter coefficient is used to filter the in-phase component and the orthogonal component respectively to obtain an in-phase baseband signal and an orthogonal baseband signal; Calculating the signal-to-noise ratio of the in-phase baseband signal and the quadrature baseband signal, generating signal weights according to the signal-to-noise ratios, and multiplying the in-phase baseband signal and the quadrature baseband signal by the corresponding signal weights to obtain a weighted in-phase signal and a weighted quadrature signal; The weighted in-phase signal is used as the real part and the weighted quadrature signal is used as the imaginary part to perform orthogonal combination to obtain a complex baseband output signal as the final baseband signal.

4. The method according to claim 1, characterized in that Inputting the first processed signal into a dual-tree complex wavelet transform module for multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extracting nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct feature vectors includes: Constructing a dual-tree complex wavelet transform module including a symmetric orthogonal filter group, wherein the symmetric orthogonal filter group includes a first group of orthogonal filters and a second group of orthogonal filters, inputting the first processed signal into the symmetric orthogonal filter group for four-layer wavelet decomposition, and obtaining complex wavelet coefficients of different frequency bands based on output signals of the first group of orthogonal filters and the second group of orthogonal filters; A reconstruction vector is constructed based on the complex wavelet coefficients, and the distance between the vectors of the reconstructed vector is calculated to obtain a distance parameter. The number of vectors that meet the distance parameter condition is counted according to a preset threshold to obtain a matching number. The conditional probability logarithm is calculated based on the matching number to obtain an approximate entropy feature; based on the matching number, the ratio of the template matching number in different dimensions is calculated to obtain a sample entropy feature; Performing a continuous wavelet transform on the complex wavelet coefficients to obtain a time-frequency coefficient matrix, extracting the maximum value of the modulus value at each time point based on the time-frequency coefficient matrix to obtain a time-frequency ridge feature; performing a Hilbert transform on the first processed signal to obtain an analytical signal, and calculating the instantaneous frequency feature according to the phase information of the analytical signal; The approximate entropy feature, the sample entropy feature, the time-frequency ridge feature and the instantaneous frequency feature are combined to construct a feature vector.

5. The method according to claim 1, characterized in that: The feature vector is input into a multi-task deep learning network with an integrated attention mechanism for feature fusion and target recognition to obtain a second processed signal including: Constructing a deep learning network including a feature extraction layer, a self-attention layer, and a multi-task output layer, inputting a feature vector into the feature extraction layer, which includes a plurality of cascaded residual blocks, each residual block including a first convolutional layer, a second convolutional layer, and a skip connection path, the feature vector is passed through the first convolutional layer to obtain a first transformation feature, the first transformation feature is passed through the second convolutional layer to obtain a second transformation feature, and the second transformation feature is added to the feature vector transmitted through the skip connection path to obtain a local feature; The local features are input into a self-attention layer, the self-attention layer performs a linear transformation on the local features to obtain a query matrix, a key matrix and a value matrix, the query matrix is ​​multiplied by the transpose of the key matrix to obtain a correlation matrix, and the correlation matrix is ​​normalized to obtain an attention weight matrix; Multiplying the attention weight matrix by the value matrix to obtain a first weighted feature, constructing multiple groups of self-attention layers in parallel to obtain multiple groups of first weighted features, and concatenating the multiple groups of first weighted features in the feature dimension to obtain a second weighted feature; Constructing a multi-task output layer including a detection subnetwork and a classification subnetwork, wherein the detection subnetwork and the classification subnetwork are constructed in a parameter sharing manner, inputting the second weighted feature into the detection subnetwork and the classification subnetwork at the same time, performing target detection on the second weighted feature through the detection subnetwork to obtain a target position feature and a target confidence feature, and performing target classification on the second weighted feature through the classification subnetwork to obtain a target category feature; Constructing a joint loss function, the joint loss function including a detection loss term and a classification loss term calculated according to target position features and target confidence features, and optimizing parameters of the detection subnetwork and the classification subnetwork based on the joint loss function; The target position features and target confidence features output by the optimized detection subnetwork and the target category features output by the optimized classification subnetwork are combined to obtain a second processed signal.

6. The method according to claim 1, characterized in that The target characteristic spectrum is obtained by calculating the time-frequency distribution of the second processed signal, the target characteristic spectrum is subjected to noise reduction processing by using a morphological filter, and the target characteristic spectrum is subjected to selective noise suppression by using an adaptive morphological filter, and the respiratory frequency component and the heartbeat frequency component in the target characteristic spectrum are extracted by using a multi-target segmentation algorithm based on a variational model, and judging whether there is a living body in the target area according to the time-varying characteristics and amplitude fluctuation law of the respiratory frequency component and the heartbeat frequency component includes: Performing short-time Fourier transform on the second processed signal to obtain a time-frequency distribution diagram, performing frame windowing processing on the second processed signal using a Hanning window function, performing fast Fourier transform on the windowed signal to obtain a spectrum sequence, and splicing the spectrum sequence in the time dimension to obtain a target characteristic spectrum; Constructing a double-layer morphological filter structure, wherein the first layer of the double-layer morphological filter structure includes a basic morphological filter, wherein the basic morphological filter adopts a disk-shaped structure element, and the disk-shaped structure element is used to perform an opening operation on the target feature spectrum to remove the peak noise higher than the background to obtain a first intermediate signal, and a closing operation is performed on the first intermediate signal to fill the valley noise lower than the background to obtain a first filtered signal; Calculating the mean, variance and skewness of each time-frequency point neighborhood in the first filtered signal to obtain local statistical characteristics, calculating an adaptive threshold based on the local statistical characteristics, and comparing the adaptive threshold with a preset decision criterion to obtain noise area identification information; Dynamically adjusting the shape and size parameters of the structural elements of the adaptive morphological filter according to the noise area identification information, selectively filtering the time-frequency points identified as the noise area to obtain a second filtered signal, and keeping the time-frequency points identified as the target feature area unchanged; An energy functional is constructed, wherein a data item of the energy functional represents the degree of matching between the second filtered signal and the respiratory and heartbeat feature model, a regularization item introduces a spatial continuity constraint, and an alternating direction multiplier method is used to optimize and solve the energy functional to obtain a frequency component segmentation result; Performing region marking and connected domain analysis on the frequency component segmentation results, extracting the time-varying trajectories of the respiratory frequency component and the heartbeat frequency component, and calculating the periodicity index, amplitude fluctuation law and energy distribution characteristics of the time-varying trajectories; The periodic index, amplitude fluctuation law and energy distribution characteristics are input into a pre-trained physiological feature matching model to obtain a feature matching score, and the signal-to-noise ratio of the respiratory frequency component and the heart rate frequency component is calculated. When the feature matching score and the signal-to-noise ratio simultaneously meet the preset threshold conditions, it is determined that there is a living body in the target area.

7. The method according to claim 1, characterized in that If it is determined that there is a life form in the target area, the three-dimensional spatial coordinates of the life form are calculated by a multi-base station collaborative positioning algorithm, and a detection result including the three-dimensional spatial coordinates and the life form feature information is outputted, including: Acquire receiving signals of multiple radar base stations and perform bandpass filtering to obtain filtered receiving signals, perform envelope detection on the filtered receiving signals to obtain envelope signals, obtain the peak moment of the envelope signal by peak detection, and determine the peak moment as the arrival time; Performing phase analysis on the filtered received signal to obtain a phase spectrum, calculating the phase spectrum difference between adjacent radar base stations to obtain a phase difference, using the phase difference as an arrival time difference, constructing a hyperbola equation based on the arrival time difference, and performing spatial filtering on the filtered received signal to obtain a filtered direction signal, calculating the covariance matrix of the filtered direction signal, obtaining an arrival angle by eigendecomposing the covariance matrix, and constructing an azimuth angle equation based on the arrival angle; The hyperbolic equation and the azimuth equation are combined to obtain a positioning equation, a ratio of signal power to noise power in the filtered received signal is calculated to obtain a signal-to-noise ratio, and the signal-to-noise ratio is normalized to obtain a weight coefficient; Performing Taylor expansion on the positioning equation to obtain a linear equation, multiplying the weight coefficient by the coefficient of the linear equation to obtain a weighted linear equation, iteratively calculating the weighted linear equation to obtain a position increment, stopping the iteration when the position increment is less than a position threshold, and outputting the three-dimensional space coordinates; Calculating the geometric precision factor of the three-dimensional space coordinates, and when the geometric precision factor is greater than a second preset threshold, calculating the positioning errors of each base station combination to obtain an error matrix, and selecting the base station combination with the smallest positioning error as the optimal base station combination based on the error matrix; The three-dimensional space coordinates are optimized using the optimal base station combination to obtain optimized three-dimensional space coordinates, statistical parameters of respiratory frequency components and heart rate frequency components in a continuous time window are extracted, and the optimized three-dimensional space coordinates are combined with the statistical parameters to obtain a detection result containing three-dimensional space coordinates and vital feature information.

8. A life detection system based on radar signal processing, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect radar echo signals of the target area, control the phased array radar transmitting antenna array to transmit a dual-frequency continuous wave radar signal of orthogonal frequency division multiplexing, wherein the dual-frequency continuous wave radar signal includes a first frequency signal and a second frequency signal, the first frequency signal and the second frequency signal are coded and modulated by Walsh orthogonal codes, the radar echo signal is received by a multi-channel radar receiving system, interference suppression and clutter cancellation processing is performed on the radar echo signal by using an adaptive space-time two-dimensional filter, and the processed radar echo signal is input into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal; The second unit is used to perform digital sampling processing on the baseband signal to obtain a sampling signal, perform multi-stage static clutter elimination processing on the sampling signal using an adaptive notch filter group to obtain a first processed signal, input the first processed signal into a dual-tree complex wavelet transform module to perform multi-scale time-frequency analysis to obtain wavelet coefficients of different frequency bands, extract nonlinear feature parameters and time-frequency joint feature parameters according to the wavelet coefficients of different frequency bands to construct a feature vector, the nonlinear feature parameters include approximate entropy and sample entropy, the time-frequency joint feature parameters include time-frequency ridge and instantaneous frequency, and input the feature vector into a multi-task deep learning network with an integrated attention mechanism to perform feature fusion and target recognition to obtain a second processed signal; The third unit is used to obtain a target characteristic spectrum by calculating the time-frequency distribution of the second processed signal, perform noise reduction processing on the target characteristic spectrum by using a morphological filter, and selectively suppress noise on the target characteristic spectrum by using an adaptive morphological filter, and extract the respiratory frequency component and the heart rate component in the target characteristic spectrum by using a multi-target segmentation algorithm based on a variational model, and judge whether there is a living body in the target area according to the time-varying characteristics and amplitude fluctuation law of the respiratory frequency component and the heart rate component; if it is determined that there is a living body in the target area, calculate the three-dimensional spatial coordinates of the living body by using a multi-base station collaborative positioning algorithm, and output a detection result including the three-dimensional spatial coordinates and the characteristic information of the living body.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images