A three-dimensional through-the-wall radar adaptive beam life detection method and system

Abstract

The application provides a three-dimensional through-wall radar adaptive beam life detection method and system, and relates to the field of data processing. In the method, a long-time sequence three-dimensional radar data cube is obtained, and a life candidate area is determined. The candidate area is input into a variational time sequence autoencoder to extract a life main rhythm. According to the life main rhythm, a short-time window multi-path structure is input into a neural ordinary differential equation to generate a reparameterization mapping curve, and time reparameterization is performed to form a life phase-locked time axis. Then, a multi-path phase slice is constructed according to the life phase, and a corresponding graph structure is established. A topological preserving graph neural network is used to identify stable life trajectory clusters. Finally, based on a meta-learning model, the spatial direction, propagation path and time modulation characteristics of the life trajectory cluster are adaptively adjusted, so that the adaptive beam and sparse inversion converge in the life phase-locked field and output the through-wall life detection result. By implementing the scheme, the precision of life micro-motion feature extraction can be improved in the through-wall life detection scene.

Description

Technical Field

[0001] This application relates to the technical field of data processing, specifically to a three-dimensional through-wall radar adaptive beam life detection method and system. Background Technology

[0002] In scenarios involving penetrating walls for life detection, millimeter-wave, microwave, or ultra-wideband radar signals are significantly affected by the spatial non-uniformity of the wall material when penetrating concrete walls, brick walls, lightweight partitions, and their internal structures. The interior of walls typically contains irregular dielectric textures, small cavities, damp areas, and locally reinforced structures such as metal ribs. These structures create numerous localized reverberation paths during electromagnetic wave propagation, causing the echo signal to exhibit strong fluctuations in the time delay, energy, and phase domains.

[0003] Because these localized reverberations are highly sensitive, their propagation characteristics can experience nonlinear time drift due to minute changes in the wall material, causing the multipath arrival sequence in adjacent time windows to lose its stable order. When radar systems extract thoracic micromotions of living organisms based on time-domain sparse weighting or compressed sensing inversion, existing technologies typically rely on the temporal coherence of multi-period echoes to distinguish between life-related paths and reverberation paths. However, in the presence of nonlinear reverberation drift, the internal ordering of multipath clusters changes randomly over time, causing the weak periodic modulation characteristics caused by the living organism to be masked by reverberation disturbances. This makes it difficult for the living signal to maintain coherence during inversion, making it difficult for the adaptive beam to be stably focused on the direction of the living organism, resulting in a significant decrease in the accuracy of extracting life micromotion features.

[0004] Therefore, there is an urgent need for a three-dimensional through-wall radar adaptive beam life detection method and system. Summary of the Invention

[0005] This application provides a three-dimensional through-wall radar adaptive beam life detection method and system, which can improve the accuracy of life micro-motion feature extraction in through-wall life detection scenarios.

[0006] A first aspect of this application provides a three-dimensional through-wall radar adaptive beam life detection method, the method comprising: acquiring a long-time-series three-dimensional radar data cube covering a wall region, and determining life candidate regions in the long-time-series three-dimensional radar data cube; inputting the long-time-series three-dimensional radar data cube corresponding to the life candidate regions into a variational temporal autoencoder to obtain the life master rhythm; inputting the multipath structure corresponding to each short time window into a neural network ordinary differential equation according to the life master rhythm to generate a reparameter mapping curve; performing time reparameterization processing on the multipath structure based on the reparameter mapping curve to obtain a life phase-locked time axis; constructing multipath phase slices for multiple life phase positions based on the life phase-locked time axis to construct a graph structure according to each multipath phase slice, and inputting the graph structure into a topology-preserving graph neural network to identify life trajectory clusters that maintain connectivity consistency in adjacent phase slices; and performing adaptive adjustment on the spatial direction, propagation path, and temporal modulation features corresponding to the life trajectory clusters based on a meta-learning model, so that the spatial weighting structure and sparse inversion weights of the adaptive beam converge around the life trajectory clusters in the life phase-locked field to output through-wall life detection results.

[0007] A second aspect of this application provides a three-dimensional through-wall radar adaptive beam life detection system. The system includes an acquisition module and a processing module. The acquisition module is used to acquire a long-time-series three-dimensional radar data cube covering a wall region and determine life candidate regions within the long-time-series three-dimensional radar data cube. The processing module is used to input the long-time-series three-dimensional radar data cube corresponding to the life candidate regions into a variational temporal autoencoder to obtain the main life rhythm. The processing module is further used to input the multipath structure corresponding to each short time window into a neural network constant differential equation based on the main life rhythm to generate a reparameter mapping curve. The processing module is also used to perform... The multipath structure undergoes time reparameterization processing to obtain a life phase-locked time axis. The processing module is further configured to construct multipath phase slices for multiple life phase positions based on the life phase-locked time axis, to construct a graph structure based on each multipath phase slice, and input the graph structure into a topology-preserving graph neural network to identify life trajectory clusters that maintain connectivity consistency in adjacent phase slices. The processing module is further configured to perform adaptive adjustments on the spatial direction, propagation path, and temporal modulation features corresponding to the life trajectory clusters based on a meta-learning model, so that the spatial weighting structure and sparse inversion weights of the adaptive beam converge around the life trajectory clusters in the life phase-locked field to output the through-wall life detection results.

[0008] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, and both the user interface and the network interface are used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method described above.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed, perform the method described above.

[0010] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: By extracting life candidate regions from a long-term three-dimensional radar data cube, the system establishes a focusing mechanism for life-related spatial regions in the initial stage, reducing computational costs in irrelevant regions during subsequent processing and improving the signal-to-noise ratio of life signals in the early processing stages. Subsequently, the life candidate regions are input into a variational temporal autoencoder to extract the life master rhythm, enabling the system to obtain a unified temporal reference throughout the entire detection cycle. This fundamentally avoids the fragmentation of periodic characteristics caused by multipath structure drift over time, forming a life-based phase-locking foundation that can dominate the entire process. After obtaining the life master rhythm, a neural frequent differential equation is introduced to generate a reparameterized mapping curve. Through life-driven temporal reparameterization, the multipath structure is migrated from the physical time axis to the life phase-locking time axis, causing multipath disturbances to detach from the life rhythm and become asynchronous noise, thereby achieving structural suppression of reverberation drift in the time dimension.

[0011] By constructing multipath phase slices and establishing a graph structure based on the life phase-locked time axis, and identifying cross-phase connected life trajectory clusters through a topology-preserving graph neural network, life-related paths are topologically enhanced within the graph structure, while reverberant paths, lacking phase coherence, are automatically separated, forming a life path aggregation mechanism under dual constraints in the time and spatial domains. A meta-learning model adaptively adjusts the spatial direction, propagation path, and temporal modulation characteristics of life trajectory clusters, enabling the adaptive beam and sparse inversion parameters to automatically adjust according to the current wall environment and life distribution. This ensures the beam prioritizes alignment with the direction corresponding to life trajectory clusters in three-dimensional space, and the sparse inversion prioritizes preserving the propagation trajectories corresponding to life paths, achieving adaptive and intelligent beam control and inversion optimization. This results in output through-wall life detection results that simultaneously possess convergence and consistency in three-dimensional space and life temporal rhythms. Therefore, it facilitates improved accuracy in extracting micro-motion features of life in through-wall life detection scenarios. Attached Figure Description

[0012] Figure 1A flowchart illustrating a three-dimensional through-wall radar adaptive beam life detection method provided in this application embodiment; Figure 2 A schematic diagram of a three-dimensional through-wall radar adaptive beam life detection system provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0013] Explanation of reference numerals in the attached figures: 21. Acquisition module; 22. Processing module; 31. Processor; 32. Communication bus; 33. User interface; 34. Network interface; 35. Memory. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0015] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0016] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0017] To address the aforementioned technical problems, this application provides a three-dimensional through-wall radar adaptive beam life detection method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating a three-dimensional through-wall radar adaptive beam life detection method provided in an embodiment of this application. The method is applied to a server and includes steps S110 to S160, as follows:

[0018] S110. Obtain a long-term three-dimensional radar data cube covering the wall area, and determine the life candidate area in the long-term three-dimensional radar data cube.

[0019] Specifically, a server refers to a computing device that undertakes the tasks of receiving, storing, processing, and analyzing data from through-wall radar. It can be a rack-mounted server, an industrial computer, or a high-configuration edge computing node. The server receives echo data continuously collected and uploaded by the through-wall radar through wired or wireless connections to the radar host, array front-end, or data acquisition terminal. It forms data records in the storage unit and runs life detection algorithms in the processing unit. The server not only handles simple data forwarding but also undertakes the complete algorithm chain, including data preprocessing, feature extraction, model inference, and result output, making it the computing core of the entire through-wall life detection system.

[0020] Covering the wall area means that the radar's transmitted beam and imaging field of view correspond to the spatial extent of the wall area in the range direction, horizontal angular direction, and vertical angular direction. This ensures that each frame of acquired 3D radar data can fully describe the reflection intensity distribution of the wall and the space behind it. In the range direction, it is necessary to cover the distance between the radar and the wall surface, the propagation range through the wall thickness, and a certain depth behind the wall. In the angular direction, it is necessary to cover the horizontal and vertical extension range of the wall to ensure that there are no spatial omissions in the 3D radar data due to insufficient imaging field of view.

[0021] A long-term time-series 3D radar data cube represents the complete coverage of a wall and the space behind it by through-wall radar in three spatial dimensions: range, horizontal angle, and vertical angle. Multiple frames of 3D radar data are acquired over continuous or quasi-continuous time periods, stacking all the 3D radar data in chronological order to form a four-dimensional data set with both spatial 3D structure and temporal evolution structure. The wall region refers to the wall itself that radar waves can penetrate and generate effective echoes, as well as the spatial area within a certain depth range behind the wall. This includes the thickness range of the concrete wall and the indoor area behind it where living organisms can move, enabling the radar to capture scattered information from the wall structure and target objects within these spaces.

[0022] The life candidate region is the spatial basis for all subsequent life rhythm extraction, life phase-locked time axis construction and life trajectory cluster identification. Its role is to focus the processing on the spatial region most likely to contain micro-movements of life, and to avoid the algorithm being interfered with by wall reverberation or random noise in the whole 3D data, so that subsequent temporal modeling and spatial modeling are carried out in the region with life characteristics.

[0023] Furthermore, when acquiring long-term three-dimensional radar data cubes covering the wall area, a multi-band linear frequency modulated pulse is continuously transmitted through the through-wall radar array, and echo signals after refraction, reflection, and scattering by the wall are received in each transmission cycle. The echo signals are processed by range compression in the range dimension and by angular domain imaging and azimuth focusing in the angle dimension, thereby obtaining the three-dimensional radar data corresponding to the current time. During range compression, the echo signal received by each array element and the matched filtering result of the transmitted multi-band linear frequency modulated pulses can be used as the compressed output in the range dimension. Let the baseband echo received by a certain array element be the received signal sequence, and let the matched filter coefficient sequence be convolved with the received signal to obtain the range compression result. Then, based on the array geometry and the angle relationship of the incident wavefront, the range compression results of each array element are weighted and superimposed through beamforming weights, focusing the energy in different spatial directions into the corresponding angle and azimuth units, forming three-dimensional radar data indexed by range, horizontal angle, and vertical angle at the current time. As the through-wall radar array continuously transmits multi-band linear frequency modulated pulses at a fixed pulse repetition period and synchronously repeats the above processing, the multi-frame three-dimensional radar data obtained at different times are stacked in chronological order to form a long-time three-dimensional radar data cube indexed by range, angle, and time. Each voxel records the echo intensity or complex amplitude at a certain range unit, angle unit, and time sampling time. The range index corresponds to the discrete point of the round-trip propagation distance of the radar to the spatial voxel, the angle index corresponds to the discrete point of the spatial direction of the array beam scan, and the time index corresponds to the sampling time of the pulse transmission period.

[0024] When performing low-resolution spatial focusing processing on a long-time-series 3D radar data cube, smoothing aggregation is performed on adjacent range cells, adjacent angle cells, and adjacent height cells in the spatial dimension. This averages out the local texture scattering components of the wall and the high-frequency spatial fluctuations caused by fine structures on the spatial scale, resulting in a smoother 3D radar data field. Let the 3D data corresponding to a certain sampling time of the long-time-series 3D radar data cube be the spatial sample value, and let the spatial smoothing kernel be the spatial weight coefficient. A weighted summation is performed on the neighborhood of each spatial location. The sample values ​​in the neighborhood are multiplied by the corresponding spatial weight coefficient and then summed to obtain the spatially smoothed sample value. The spatial weight coefficient is used to control the contribution of different spatial cells in the neighborhood to the smoothing result. Non-negative weights that add up to one are selected so that the spatially smoothed sample values ​​can still maintain comparability at the energy scale. Subsequently, a large time window smoothing process is performed on the long-time series 3D radar data cube, covering multiple micro-motion cycles of life with the time window length. By selecting a time window containing several preceding and following time frames at the center of each time location, a weighted average is applied in the temporal direction to the spatial smoothing results within the window. This allows the energy of adjacent time frames to be accumulated in the temporal dimension, thereby suppressing rapid changes in instantaneous reverberation and random noise, and highlighting life-related scattering components with slow periodic fluctuations over long time scales. Under the simultaneous action of spatial and temporal smoothing, regions appear in the long-time series 3D radar data cube where energy changes slowly in the temporal dimension and forms connected weak energy fluctuations in the spatial dimension. These regions are regarded as weak spatial fluctuation regions in subsequent processing, characterized by weak amplitude periodic fluctuations in energy without drastic energy jumps within multiple time windows.

[0025] When performing energy field analysis based on spatially weak fluctuation regions, the long-term three-dimensional radar data cube, after spatial and temporal smoothing, is regarded as a three-dimensional energy field that varies with time. Using each spatial location as an index, the energy sequence of that location within the entire time range or multiple sliding time windows is analyzed to construct a three-dimensional distribution structure of energy changing with time. For the energy sequence of each spatial location, the average energy, energy fluctuation amplitude, and rhythmicity index of energy change within multiple time windows can be calculated. These temporal statistical characteristics are combined with the connectivity characteristics of the spatial neighborhood to form a criterion for describing whether the spatial location belongs to a spatially weak fluctuation region. Subsequently, through consistency screening of the three-dimensional distribution structure, regions that exhibit continuous weak rhythmic changes in time and maintain connectivity with adjacent voxels in space are found within multiple sliding time windows. These regions that repeatedly appear within multiple time windows and meet the dual conditions of weak temporal periodicity and spatial connectivity are marked as target regions. The target area physically corresponds to a spatial sub-region in the scene behind the wall where there is a more likely micro-movement of the thoracic circumference of a living organism. The target area is used as a life candidate region in the long-term three-dimensional radar data cube for the subsequent variational temporal autoencoder to extract the main rhythm of life and construct the spatial constraints of the life phase-locked time axis, so that the subsequent temporal modeling focuses on the region with life-related scattering characteristics rather than the global background region.

[0026] S120. Input the long-time three-dimensional radar data cube corresponding to the life candidate region into the variational time-series autoencoder to obtain the life master rhythm.

[0027] Specifically, the variational temporal autoencoder (VTE) is a deep generative model used to model the underlying patterns of change within long-term time-series data. Its core consists of an encoder structure, a latent space representation, and a decoder structure. The encoder structure receives the time-varying input sequence and compresses high-dimensional temporal radar data into a lower-dimensional latent space representation that can characterize the main change patterns. The latent space representation is not a single, fixed value, but a representation structure with certain statistical distribution characteristics. By introducing variational inference, the model learns the statistical features of the temporal evolution trajectory in the latent space, such as slowly changing periodic structures and long-term trends. The decoder structure reconstructs the main change patterns of the original time-series data in the time dimension based on the latent space representation, thus ensuring a correspondence between the latent space representation and the input time-series data. Compared to ordinary temporal autoencoders, the variational temporal autoencoder not only compresses and reconstructs the input sequence but also imposes constraints on the distribution of the latent space, making the latent representations of different time periods smoother and more continuous in time, making it more suitable for extracting slow, stable life rhythms such as breathing or heartbeat.

[0028] The dominant rhythm of life refers to the time-series pattern extracted from the long-term three-dimensional radar data cube corresponding to the life candidate region, which represents the overall rhythm of life activities. It manifests as a slow and approximately periodic change trajectory in the time dimension. This dominant rhythm of life is not directly equivalent to the echo intensity sequence of a specific voxel, but rather is the dominant rhythmic component extracted after joint modeling of multiple spatial locations and multiple time windows in the latent space by a variational temporal autoencoder. The dominant rhythm of life reflects the main temporal rhythm that all scattering paths related to the micro-movements of the human thoracic cavity within the life candidate region follow together, such as the large-scale slow oscillations formed by the respiratory cycle and possibly superimposed faster rhythmic components. For example, if there is a person standing still behind a wall, their ribcage will undergo a slight displacement with a period of about four seconds with each breath. The energy of multiple spatial voxels in the life candidate region will show a weak fluctuation of four-second period. After the variational temporal autoencoder performs joint compression and reconstruction of the time series of these voxels in the latent space, it will learn a stable four-second period rhythm in the latent trajectory. This rhythm is the life master rhythm, which can be used as a reference for the life phase-locked time axis in the subsequent time reparameterization process.

[0029] Furthermore, when dividing the long-term 3D radar data cube corresponding to the life candidate region into time segments containing multiple frames of continuous 3D radar data, and performing amplitude normalization and noise floor suppression processing on each time segment, a fixed length of time segment is first selected within the spatial range of the life candidate region, so that each time segment contains multiple frames of continuous 3D radar data, with each frame corresponding to the same range cell, horizontal angle cell, and vertical angle cell in space. For each time segment, without disrupting the spatial indexing relationship, the echo amplitude samples of all voxels in that time segment are collected into a set of time samples. The statistical mean and fluctuation range within that time segment are calculated, thereby performing amplitude normalization processing to bring the amplitude distribution of different time segments into a comparable range. Based on this, the noise floor of each time segment is estimated according to the long-term background noise level, and random noise components below the noise floor threshold are compressed or suppressed by setting a noise floor threshold, thereby retaining stable scattering components above the noise floor. The original amplitude signal can be normalized into a normalized amplitude signal by performing mean and standard deviation statistics on the amplitude signal at the same spatial location within each time segment, as follows:

[0030] in, This indicates that the distance index in the current time segment is Angle index is The height index is Time index is The original echo amplitude sample; This represents the statistical mean of the amplitudes of all samples in the current time segment; This represents the statistical standard deviation of the amplitude of all samples in the current time segment; To prevent extremely small positive numbers with a denominator of zero; These are normalized amplitude samples. Amplitude normalization ensures that different time segments have similar numerical distributions when input to the variational temporal autoencoder, while noise reduction is achieved through comparison. By using a preset noise floor threshold, samples that are significantly below the noise floor are compressed to a level close to zero, making it easier for subsequent models to focus on the effective scattering components related to micro-movements in life.

[0031] In the process of inputting the processed time segments into the encoder structure of the variational temporal autoencoder and generating the latent spatial distribution through multi-layer temporal convolution, gated recurrent units, and temporal attention structures, each time segment is spatially rearranged or dimensionality-reduced in a predetermined manner to form a high-dimensional time series in the temporal dimension, which is then fed into the encoder structure of the variational temporal autoencoder. The encoder structure first extracts the change patterns in the local temporal neighborhood on the time axis through multi-layer temporal convolution, capturing the difference between the slowly changing energy fluctuations and the rapidly changing reverberation perturbations in the life candidate region. Then, the gated recurrent unit performs long-span state modeling on the convolutional time feature sequence, enabling the encoder to remember the energy evolution trend within a multi-period range. The gating parameters in the gated recurrent unit determine the relative weights of the current moment features and historical moment features in the state update. On this basis, the temporal attention structure weights the feature importance of each time position, so that the slow periodic segments related to the life rhythm receive higher weights in the feature aggregation process, thereby strengthening the life rhythm component in the latent space.

[0032] The encoder structure ultimately outputs the latent spatial distribution parameters corresponding to each time location, including the mean and perturbation parameters in each dimension of the latent space. Regularization constraints are applied to these parameters through a variational inference mechanism, ensuring that the latent spatial distribution of different time segments both closely matches the observed temporal characteristics and maintains smoothness and continuity in the temporal dimension. Latent spatial sampling can be represented as:

[0033] in, Indicates that at time index The latent space sample vector at the location; Indicates that at time index The latent space mean parameter vector output by the encoder, where each dimension corresponds to the average state in a certain direction of the latent space; Indicates that at time index The latent space perturbation parameter vector output by the encoder, each dimension of which corresponds to the random fluctuation amplitude in a certain direction of the latent space; This represents a random vector that follows a zero-mean, unit-variance distribution and is used to introduce random perturbations into the latent space. This indicates an element-wise multiplication operation. The sampled data obtained using this sampling method... Along the time index The arrangement forms a potential trajectory, which reflects the main changing trend of the temporal structure within the candidate region of life, and avoids overfitting of the instantaneous reverberation through the regularization of the mean parameter and the perturbation parameter.

[0034] In the process of inputting the latent trajectory into the decoder structure of a variational temporal autoencoder to generate a reconstructed sequence reflecting slow periodic fluctuations, and performing rhythmic trajectory extraction and weak periodic enhancement processing based on multiple reconstructed sequences, the latent trajectory is first input as a time-series feature into the decoder structure. The decoder structure then maps the latent trajectory back to the reconstructed sequence at the original time resolution through a deconvolution layer or an inverse time mapping module, so that the latent vector corresponding to each time position is restored to a time sample simulating the slow fluctuations of radar scattering energy within the life candidate region. During training, the decoder structure minimizes the difference between the reconstructed sequence and the normalized time segment, ensuring that the slow periodic component in the latent trajectory is consistent with the real micro-motion rhythms of life in the life candidate region. For multiple spatial locations or multiple aggregation channels within the life candidate region, multiple reconstructed sequences can be obtained. By performing weighted aggregation on these reconstructed sequences in the spatial dimension, one or more candidate rhythmic signals representing the overall energy evolution of the life candidate region are constructed. The reconstructed sequences are averaged as follows:

[0035] in, Indicates that at time index The aggregated rhythm signal at the location; Indicates the first The reconstructed sequence is at time index Sample values ​​at; This indicates the number of reconstructed sequences participating in the aggregation. After obtaining the aggregated rhythm signal, a rhythm trajectory extraction method is used to search for trajectories with a significant weak periodic structure over a long period in the potential trajectories and reconstructed sequences. For example, by analyzing the stability of the peak spacing of the autocorrelation function, the degree of phase alignment during the period, and the stability of the period amplitude, the trajectory with the strongest phase continuity and periodic consistency in multiple time segments is selected as a candidate life rhythm trajectory. Subsequently, weak periodic enhancement processing is applied to perform period alignment and period superposition on the candidate life rhythm trajectory. Samples with the same phase points in multiple periods are superimposed and averaged to suppress non-periodic noise and amplify periodic components. This results in the final life master rhythm exhibiting a stable periodic structure and continuous phase evolution characteristics throughout the entire time range, thus providing a unified time reference for subsequent time reparameterization and life phase-locked time axis construction.

[0036] S130. Based on the main life rhythm, input the multipath structure corresponding to each short time window into the neural constant differential equation to generate the reparameter mapping curve.

[0037] Specifically, the reparameter mapping curve is a time-mapping trajectory obtained by the neuronormal differential equation under the constraint of a given life master rhythm. It describes the correspondence between the original physical time and the new life phase-locked time. For each short time window, the reparameter mapping curve maps the physical time sampling points within that short time window to a new time axis. This new time axis uses the life master rhythm as a reference, making the life-related micro-motion signals exhibit a more regular structure that is closer to the ideal period on this time axis. Numerically, the reparameter mapping curve is a function curve that monotonically changes with physical time, providing a new time coordinate value for each physical time point. By reparameterizing the time coordinates of the multipath structure, the multipath states can be rearranged in the time structure, causing the parts modulated by the life master rhythm to be clustered, while the parts affected by changes in wall texture and local reverberation appear as disturbances that are not aligned with the life rhythm on the new time axis.

[0038] Furthermore, when representing the multipath structure within each short time window as a sequence of multipath state vectors evolving with physical time through time delay axis normalization and energy normalization, the time delay distribution and energy distribution of the multipath are first jointly encoded within the current short time window, mapping the original multipath time delay index and multipath energy sample to a unified normalized feature space. Let the current short time window contain the... The multipath delay sample set at each physical time sampling point is: The corresponding multipath energy sample set is ,in This indicates the number of valid multipath paths detected at that sampling point in time. Indicates the first The original delay index of the multipath path, Indicates the first The original amplitude or energy of each multipath path. To construct a nonlinearly compressed and globally consistent normalized coordinate system on the delay axis, the global minimum of all multipath delays is first obtained within the current short time window. and global maximum value Then, fractional power compression and shift normalization are performed on the delay of each multipath path to map it to a normalized delay. :

[0039] in, Indicates the first At the sampling point of time, the first Standardized delays corresponding to multiple path numbers; The nonlinear compression exponent is greater than zero, controlling for the difference in distribution between long and short delays in the normalized space. This can enhance short-latency resolution. This can enhance the ability to resolve long delays; To avoid extremely small positive numbers with a denominator of zero, their values ​​are much smaller than... Regarding energy normalization, to avoid a single strong scattering path dominating subsequent feature representation, soft normalization weights are introduced to weight the multipath energy within the same sampling point at the same time, thus normalizing the original energy. Mapped to weights Based on this, normalized energy is constructed. :

[0040] in, Indicates the first At the sampling point of time, the first The soft normalized weights corresponding to the multipath paths, all The summation at the same sampling points at the same time is equal to one; The temperature coefficient is used to control the effect of energy differences on the weight distribution. When the value is large, the weight of high-energy paths increases significantly. When the weights are relatively small, the weights of each path tend to be averaged. This represents the mean of all multipath energy samples within the current short time window; This represents the standard deviation of all multipath energy samples within the current short time window; To avoid extremely small positive numbers with a standard deviation of zero, the standardized time delay vector, normalized energy vector, and corresponding spatial orientation codes, such as distance index codes and angle index codes, are then concatenated into a time index. High-dimensional multipath state vector at the location Then, a unified representation of the multipath state vector sequence is obtained through a whitening operation based on overall statistical characteristics:

[0041] in, Indicates the time sampling point The normalized multipath state vector at the location; For sampling points at time The original multipath state vector is obtained by splicing together standardized time delay, normalized energy, and spatial direction encoding. This represents the average vector of the multipath state vectors obtained statistically across the current short time window or multiple short time windows. The covariance matrix representing the multipath state vector; This represents the inverse square root matrix of the covariance matrix, used to weaken the correlation between various feature dimensions and bring them closer to an independent and identically distributed state. Through the above processing, the multipath structure within the current short time window is uniformly encoded as an index that varies with physical time. An evolutionary set of normalized multipath state vector sequences each It also includes the representation of time delay distribution, energy distribution, and spatial structure in a unified feature space.

[0042] When inputting the multipath state vector sequence and the phase of the corresponding short-time window's main life rhythm into a time vector field model parameterized by a deep neural network, and generating a continuously changing, monotonically increasing, and derivative-smooth reparameter mapping curve in the neural network's ordinary differential equations through integration using the time vector field model, the first step is to define a set of discrete-time sampling points on the physical time axis of the current short-time window. ,in The start time of the short time window. The end time of the short time window. This represents the number of time sampling points within this short time window. Based on the globally estimated master biological rhythms, the phase trajectory of the master biological rhythms on the physical time axis is restricted to the current short time window interval, which is defined as each time sampling point. Calculate the corresponding life master rhythm phase to form a phase sequence ,in Normalized to an interval within a period or Inside, used to characterize at time sampling points At what phase position is the life rhythm present? For each time sampling point, the normalized multipath state vector is... Phase with the main life rhythm The input vector of the cascaded time vector field model And through deep neural network functions Extract intermediate feature representations to control the rate of time reparameterization:

[0043] in, Indicates the time sampling point The intermediate feature vector of the time vector field extracted by the deep neural network; Indicates a set of parameters A controlled deep neural network model is proposed. This model maps input features through multiple fully connected layers, normalization layers, and nonlinear activation layers, enabling the output intermediate feature vector to reflect the temporal reparameterization trend under the combined influence of multipath structure and the main life rhythm phase at different time sampling points. To ensure that the reparameter mapping curve remains strictly monotonically increasing on the physical time axis and has smoothness in its derivative, a restricted activation function and a bias term are introduced to construct an instantaneous reparameter velocity function. This is used as the right-hand side function of the divine ordinary differential equation:

[0044] in, Indicates the time sampling point The instantaneous rate of change of the repeated parameter over time; The positive lower bound bias term is used to ensure that the instantaneous rate of change is greater than zero at any point in time, thereby ensuring that the reparameter mapping curve is strictly monotonically increasing. It is a non-negative scaling factor used to control the dynamic range of the instantaneous rate of change between different time sampling points; The activation function is a monotonically increasing function, such as the hyperbolic tangent or sigmoid function, used to restrict the results of linear combinations to a finite interval; This is the weight vector, used to linearly weight each dimension of the intermediate feature vector; For bias scalars. The instantaneous rate of change... Treat as reparameter time For physical time The derivative of the equation can be used to express the evolution of the repeated parameters in the form of a regular differential equation as follows:

[0045] in, Indicates physical time The time value of the repetitive parameter at a given time point changes continuously with time within a short time window, starting at the initial time. Given an initial re-parameter time value By applying the above ordinary differential equation in the interval Performing numerical integration on the sampled time points yields the results. The re-parameter time values ​​at the location are as follows: in, Let be the integration variable, representing the continuous time during the integration process; In time The instantaneous rate of change is output by the time vector field model. By selecting a numerical integrator with adaptive step-size control capability, such as a high-order Runge-Kutta method or a linear multistep method, the set of repetitive parameter time values ​​corresponding to all time sampling points within the entire short time window physical time interval can be obtained while ensuring numerical stability. This results in a reparameter mapping curve that changes continuously with time within the current short time window, is strictly monotonically increasing in value, and remains smooth in its derivative. This reparameter mapping curve provides a precise and differentiable time coordinate transformation basis for the subsequent rearrangement of the multipath state vectors on the life-locked loop time axis.

[0046] S140. Based on the reparameter mapping curve, perform time reparameterization processing on the multipath structure to obtain the life phase-locked time axis.

[0047] Specifically, the life-locked phase-time axis refers to a unified time reference axis obtained by further aligning the reparameterized time with the main life rhythm after time reparameterization. Phase-locking means strictly aligning the phase structure of the time axis with the phase structure of the main life rhythm, so that the same life rhythm phase, such as the peak of inspiration and the end of expiration in a respiratory cycle, corresponds to the same or similar position on the life-locked phase-time axis in different cycles and different short time windows. The life-locked phase-time axis no longer simply represents "how many milliseconds have passed in absolute time," but rather "which cycle and phase position of the life rhythm currently in," such as the rising edge of the tenth respiratory cycle or the falling segment of the fifteenth respiratory cycle. By uniformly projecting the multipath structure onto the life-locked phase-time axis after time reparameterization, cross-cycle and cross-time-window life phase slices can be constructed on this time axis. This allows life-related multipath trajectories to coincide in phase when superimposed in different cycles, while reverberant trajectories exhibit inconsistent distributions in different cycles due to the lack of a stable phase reference, making them easier to separate and suppress in subsequent trajectory clustering and graph structure analysis. For example, if the respiratory cycle corresponding to the main life rhythm is about four seconds, the life phase-locked time axis is equivalent to straightening the different four-second segments in phase and superimposing them on the same cycle template, so that each inspiratory peak falls in the same phase position, rather than being scattered in different positions on the original physical time axis.

[0048] Furthermore, when mapping the physical time sampling points corresponding to each short time window to the reparameter time values ​​on the initial phase-locked time axis via the reparameter mapping curve, and performing time coordinate migration on the multipath state vector sequence based on the reparameter time values, firstly, the set of physical time sampling points within the current short time window is denoted as a set of discrete time points. Each physical time point is converted into its corresponding reparameter time value via the reparameter mapping curve obtained in the previous stage, thus numerically switching from the physical time axis to the initial phase-locked time axis. The physical time corresponding to the k-th physical time sampling point within the current short time window can be denoted as... The output of the reparameter mapping curve at this time point is denoted as Then the discrete form can be written as:

[0049] in, This represents the reparameter time value corresponding to the k-th physical time sampling point on the initial phase-locked time axis; This represents the initial reparameter time offset corresponding to the start time of the current short time window, which can be determined based on the end reparameter time value of the previous short time window or global constraints. It represents the instantaneous repetitive parameter velocity obtained from the time vector field model and the constant differential equation at the r-th time step. Its value is a non-negative real number and is used to control the growth rate of the repetitive parameter time at that time step. This represents the weighting coefficients used for numerical integration, such as the time step weights in trapezoidal integration or higher-order integration methods, whose values ​​are determined by the selected integration method. Through the above discrete accumulation, a set of re-parameter time values ​​corresponding to each physical time point can be obtained within the current short time window. Subsequently, the multipath state vector sequence within the current short time window is rearranged from its physical time index to the target multipath state vector sequence indexed by the re-parameter time, that is, the original sequence marked as being in time... The multipath state vector at that point is replaced with a time position marked on the initial phase-locked time axis. This forms a target multipath state vector sequence defined on the initial phase-locked time axis, ensuring that the multipath evolution trajectory within the same short time window maintains a monotonic temporal order in the new time coordinate system and has a basis for alignment with the main life rhythm.

[0050] A time overlap region is set between adjacent short time windows. After completing their respective time reparameterization, smooth stitching is performed based on the statistical consistency of the reparameter time values ​​and corresponding target multipath state vectors within the overlap region. First, a shared physical time interval is reserved for adjacent short time windows during global time partitioning. Data within this physical time interval is simultaneously attributed to the end of the previous short time window and the beginning of the next short time window, ensuring a certain length of overlap between the two short time windows. After reparameterizing the two short time windows based on the reparameter mapping curve, two sets of reparameter time values ​​and target multipath state vectors defined on the initial phase-locked time axis can be obtained in the overlap region, denoted as the reparameter time series of the previous short time window and the reparameter time series of the next short time window. To form a continuous and smooth intermediate phase-locked time axis globally, these two sets of time and state data in the overlap region need to be fused. The fused reparameter time at a certain aligned index position in the overlap region can be denoted as:

[0051] in, This represents the fusion reparameter time value at the k-th fusion position within the overlap region, used to construct the time coordinates on the intermediate phase-locked time axis; This indicates the reparameterized time value obtained at this position after reparameterization in the previous short time window; This indicates the reparameterized time value obtained after reparameterization in the next shorter time window at that position; This represents the weighting coefficients used to achieve smooth stitching, and their values ​​are located in the interval [range missing]. Furthermore, the first half of the overlapping region is biased towards the earlier short time window, and the second half towards the later short time window. For example, a linear or cosine function can be used to smoothly transition from 1 to 0 within the overlapping region. The corresponding target multipath state vector is also fused using similar weights, ensuring a continuous transition of energy distribution and state characteristics between the two short time windows within the overlapping region. By performing this smooth stitching based on the consistency of reparameter time values ​​and multipath state statistics within the overlapping regions of all adjacent short time windows, all initial phase-locked time axis segments can be stitched together over the global time range into an intermediate phase-locked time axis that is continuous and smooth in both time and state throughout the entire observation period.

[0052] When constructing a target time grid with fixed intervals on the intermediate phase-locked time axis and resampling the mapped target multipath state vector using interpolation, a unified time resolution is first selected on the already obtained intermediate phase-locked time axis. The initial and final resampling parameter time values ​​of the intermediate phase-locked time axis are then divided into a series of equally spaced target resampling parameter time points, forming a globally unified target time grid. Let the continuously defined resampling parameter time variables on the intermediate phase-locked time axis be denoted as... The nth repetitive parameter time point in the target time grid is denoted as So, on the intermediate phase-locked time axis, and The nearest actual reparameter time points and their corresponding target multipath state vectors can be used to construct the interpolation resampling results. The resampled multipath state vector at the nth target time grid point can be represented as:

[0053] in, Indicates the target time grid point on the intermediate phase-locked time axis. The resampled target multipath state vector at the location; This indicates the actual sampling time point on the intermediate phase-locked loop time axis. The target multipath state vector at the location; This represents the interpolation kernel function used to measure the target time grid points. Compared with the actual sampling time point The distance relationship between them is determined and interpolation weights are assigned, the values ​​of which vary with... The value decreases as the value increases, and various forms such as linear interpolation kernels, cubic spline interpolation kernels, or Gaussian kernels can be selected, and the value should be included in all interpolation operations. The condition that the weights sum to one is satisfied. By performing the above interpolation resampling on all target time grid points along the entire intermediate phase-locked time axis, a multipath state vector sequence with a uniform sampling step size in the reparameter time domain can be obtained. Since the intermediate phase-locked time axis has already redistributed the time structure in conjunction with the main life rhythm, the life-related paths on this time axis exhibit stable alignment of the rhythm structure across multiple cycles. After interpolation resampling, the evolution trajectory in each cycle presents a slow change with approximate periodicity. However, the reverberation path, not constrained by the main life rhythm, has difficulty maintaining consistency in phase position and amplitude patterns across different cycles. After resampling, it exhibits irregular drift and unstable energy distribution on the target time grid. Finally, this uniformly sampled intermediate phase-locked time axis and its corresponding multipath state vector sequence are used as the life phase-locked time axis. In the subsequent multipath phase slice construction and life trajectory cluster identification steps, it serves as the time coordinate reference, used to stably extract life-related trajectories in the time dimension and significantly reduce the interference of the reverberation path.

[0054] S150. Based on the life phase-locked time axis, construct multipath phase slices for multiple life phase positions, construct a graph structure according to each multipath phase slice, and input the graph structure into a topology-preserving graph neural network to identify life trajectory clusters that maintain connectivity consistency in adjacent phase slices.

[0055] Specifically, vital phase positions are a series of discrete or continuous phase intervals obtained by dividing the main vital rhythms based on the vital phase-locked time axis. These intervals are used to characterize the specific phase state of vital activity within a standard rhythmic cycle. Multipath phase slicing is a time-segmented data representation method built upon the vital phase-locked time axis and vital phase positions. It is the result of slicing the multipath structure according to vital phase rather than physical time. On the vital phase-locked time axis, for each vital phase position, all multipath state vectors from multiple cycles and time windows within the corresponding time interval are collected to form a multipath set at that phase position. This set is called the multipath phase slice for the corresponding vital phase position. The function of multipath phase slicing is to aggregate multipath structures across cycles at the same vital phase position, statistically strengthening recurring vital-related paths at that phase position, while weakening reverberant paths that randomly appear in different cycles due to their unstable distribution at the same phase position. For example, near the inspiratory peak of the respiratory cycle, the multipath phase slice focuses on all multipath states occurring before and after this moment, thereby capturing the multipath response pattern at the moment of maximum thoracic expansion.

[0056] A life trajectory cluster refers to a set of multipath paths with high consistency in spatial direction, propagation distance, and life phase evolution, identified through clustering or classification methods in the node embedding space of the topology-preserving graph neural network output. Physically, this set corresponds to the main propagation path from radar to the living organism and back to radar, along with its stable variants. The multipath paths in the life trajectory cluster recur at multiple life phase positions and in multiple cycles, spatially distributed along similar angles and distances, and temporally advancing with the regular phase progression of the main life rhythm. Therefore, in the graph structure, it appears as a continuously extending connected component spanning multiple multipath phase slices. In the embedding space of the topology-preserving graph neural network, the life trajectory cluster corresponds to a tightly clustered cloud of nodes, clearly separated from the dispersed nodes representing reverberation paths.

[0057] Furthermore, when performing phase normalization on the life phase-locked time axis, the recalculated time values ​​on the intermediate phase-locked time axis are first mapped to a unified life phase coordinate range, ensuring that life rhythms in different periods and time windows all fall within the same phase reference frame. Let the recalculated time values ​​on the life phase-locked time axis be... The life master rhythm periodic function obtained from global or local estimation is: By unfolding the life phase-locked timeline along the main life rhythm cycle, the corresponding normalized life phases are obtained. It can be obtained through the instantaneous frequency function of the life rhythm. The phase mapping relationship is constructed by integrating over the phase-locked time axis, as follows:

[0058] in, Indicates the time of reparameterization The corresponding normalized life phase is mapped to an interval. Inside; Indicates the initial reparameter time of the phase-locked loop time axis; Indicates the time of reparameterization The instantaneous frequency function of the main life rhythm can be obtained by taking the time derivative of the main life rhythm curve or by estimating it based on time-frequency analysis. This represents a modulo-1 operation, used to constrain the phase value within a standard period. Through the above integration, arbitrary multiparameter times on the phase-locked time axis are mapped to a unified phase coordinate system. Subsequently, a complete lifecycle is divided into [number] periods based on the normalized lifecycle phase interval. Each of the five life phase position intervals has a width of [missing information]. , No. Each life phase position interval can be represented as ,in From 0 to The integers ensure that any key parameter time point on the life phase-locked time axis can be accurately mapped to a certain life phase position range, thus laying the foundation for subsequent slicing of multipath states in the phase coordinate system.

[0059] When classifying the target multipath state vector sequence according to its life phase position interval and constructing multipath phase slices and graph structures, firstly, for each re-parameter time sample on the life phase-locked time axis, its life phase value is calculated using the phase mapping relationship obtained in the previous stage. Then, based on the life phase position interval to which this phase value belongs, the corresponding target multipath state vector is classified into the corresponding multipath phase slice. Let the target multipath state vector sequence be... The corresponding reparameter time is Then the normalized life phase is , No. The life phase position index of each target multipath state vector It can be written as:

[0060] in, For the first The index of the life phase position interval corresponding to the multipath state vector of each target takes a value from 0 to... Integers; This indicates the floor function. According to... The value of divides all target multipath state vectors into Groups, each group forming a multipath phase slice. For the first... All target multipath state vectors in a multipath phase slice Construct a graph structure , where the set of nodes The edge set is composed of the target multipath state vectors in this phase slice. This is used to characterize the geometric proximity, energy distribution similarity, and phase evolution smoothness relationships between multipath paths. A comprehensive similarity metric function can be defined. To decide whether to be at the node With nodes The edges are established and their weights are set as follows:

[0061] in, Represents a node With nodes Overall similarity between them; and Representing nodes respectively and nodes The spatial position vector corresponding to the target multipath state vector is used to characterize the geometric coordinates of the multipath path in three-dimensional space; This represents the Euclidean distance between the two spatial location vectors; and Representing nodes respectively and nodes The corresponding multipath energy characteristics can be normalized energy or energy statistics. Represents a node With nodes The phase shift during repeated observations across cycles is used to characterize its rhythmic consistency on the life phase-locked timeline. , , These are non-negative weighting coefficients used to adjust the impact of spatial distance, energy difference, and phase smoothness on overall similarity. According to... Whether the similarity exceeds a preset threshold determines whether to include it. Add to edge set and will As edge weights, a graph structure representation that reflects both geometric proximity and energy patterns and phase evolution smoothness is obtained within the multipath phase slice.

[0062] When inputting the constructed graph structure into a topology-preserving graph neural network and identifying life trajectory clusters that maintain connectivity consistency in adjacent phase slices, the graph structure corresponding to all phase slices is first... Together with the cross-slice edges between phases, they form a joint graph structure. The initial features of each target multipath state vector node are set to its multipath state vector or a low-dimensional embedding vector obtained through linear mapping. The topology-preserving graph neural network iteratively updates the node features on the joint graph structure through a multi-layer message passing and aggregation mechanism, so that the position of the node in the embedding space can reflect its topological relationship and cross-phase connectivity in the multipath network. Schematic, the first... Nodes in a layered network The eigenvectors are denoted as Then, the attention-weighted graph convolution operation can be represented as:

[0063] in, Indicates the first Layer nodes Update the feature vector; Represents a node The set of adjacent nodes includes adjacent nodes within the same phase slice and cross-phase adjacent nodes in adjacent phase slices; Indicates the first The trainable weight matrix of the layer is used to linearly transform the features of its neighbors; Represents a nonlinear activation function; Indicates the first Layer nodes Its neighboring nodes The attention weights can be calculated based on node-to-feature pairs, as follows: in, The scoring function used to calculate the similarity between node pairs can be in bilinear form or a multilayer perceptron. Normalization is applied to all neighbors of the same node to 1. After multiple iterations, the node features integrate both the geometric and energy relationships within the current phase slice and cross-phase connectivity information between adjacent phase slices. Subsequently, based on the frequency of a node's occurrence in adjacent phase slices and its connectivity along the phase axis, combined with the degree of clustering in the node embedding space, all nodes are clustered or grouped based on density. Node sets that appear more than a preset threshold in multiple adjacent phase slices and maintain connectivity consistency in the cross-phase topology are identified as life trajectory clusters. Node sets that appear only sporadically in a few phase slices, have poor connectivity in the topology, or are scattered in the embedding space are classified as reverberant trajectory clusters. In this way, the topology-preserving graph neural network separates life trajectory clusters from reverberant trajectory clusters in the node embedding space, thereby identifying life trajectory clusters with stable connectivity and strong rhythmic consistency in adjacent phase slices, providing a high-confidence set of life target paths for subsequent adaptive beamforming and sparse inversion.

[0064] S160. Based on the meta-learning model, the spatial direction, propagation path and temporal modulation features corresponding to the life trajectory cluster are adaptively adjusted so that the spatial weighting structure and sparse inversion weight of the adaptive beam converge around the life trajectory cluster in the life phase-locked field, so as to output the results of through-wall life detection.

[0065] Specifically, a meta-learning model is a high-level model structure that learns "how to learn" across various different but related tasks. Its goal is not to optimize a single set of parameters for a single, fixed scenario, but rather to summarize a set of parameter generation rules that can quickly adapt to new scenarios. In the scenario of through-wall life detection, different wall materials, different indoor layouts, and different personnel positions and postures constitute a large number of differently distributed detection tasks. By training on these tasks, the meta-learning model obtains a mapping method that generates adaptive beams and sparse inversion initial hyperparameters based on environmental features and life trajectory cluster features. This enables the system to quickly adjust its beam and inversion strategy based on the currently observed life phase-locked field characteristics when facing new walls and new scenarios, without having to start parameter tuning from scratch.

[0066] Spatial orientation refers to the incident and exit directions of the multipath paths corresponding to a cluster of life trajectories in three-dimensional space, described by azimuth, elevation, or other equivalent angular coordinates plus distance information. In the array antenna coordinate system, spatial orientation directly determines which direction the adaptive beam points to enhance or suppress energy. Therefore, spatial orientation information is the key bridge for transforming the determination of "which cluster of multipath paths belongs to the cluster of life trajectories" into "beam pointing and beam shape". For example, if the cluster of life trajectories is concentrated near a certain azimuth and elevation angle relative to the array, the adaptive beam needs to increase gain in that spatial direction while suppressing angular regions related to wall reflections, ceiling reflections, etc.

[0067] A propagation path refers to one or more physical paths that electromagnetic waves actually take from the radar transmitter to the living organism and back to the receiver. These include direct transmission paths (e.g., paths that pass through walls directly to the living organism and return), paths that undergo multiple reflections inside walls, paths that reflect off indoor surfaces, and multi-hop combination paths. Each propagation path can be described by parameters such as propagation distance, the type of medium traversed, and the number of reflections or refractions, and corresponds to one or more arrival times in the time delay domain. In a cluster of life trajectories, the propagation path is the basic element constituting the cluster; different paths superimpose to form the overall waveform profile of life-related echoes. When processing propagation path information, the meta-learning model needs to distinguish between "life-related paths that should be preserved and enhanced" and "reverberant paths that should be compressed or suppressed in sparse inversion."

[0068] Temporal modulation characteristics refer to the rhythmic features of multipath paths corresponding to a life trajectory cluster as they change over time along the life phase-locked time axis. These include phase fluctuations caused by micro-motions, slow amplitude fluctuations, and periodic peaks and troughs. For human organisms, breathing, heartbeat, and slight posture adjustments all leave periodic or quasi-periodic temporal modulation traces on the multipath paths. These temporal modulation characteristics manifest as trajectories with phase continuity and periodic consistency on the life phase-locked time axis. Temporal modulation characteristics are an important basis for distinguishing life trajectory clusters from static scatterers, because the echoes of static scatterers are almost constant in time, while the echoes related to living organisms exhibit dynamic changes with a specific rhythm over time.

[0069] A life-locked field (LLDF) refers to a comprehensive scene representation formed by reconstructing multipath structures within a life-locked time axis and three-dimensional spatial coordinates, using the dominant life rhythm as the time reference and life trajectory clusters as the spatial and path references. In a LLDF, each spatial location and propagation path is projected onto a time coordinate aligned with the life phase, making life-related paths appear as spatiotemporally continuous and phase-smooth trajectories in the field. Reverberation paths, however, exhibit discontinuities and irregular drift due to the lack of phase alignment. LLDF provides a unified reference framework for the joint optimization of adaptive beamforming and sparse inversion, allowing the convergence objectives of both to be defined around life trajectory clusters.

[0070] The results of through-wall life detection are the output of the entire processing chain, including the determination of whether a life form exists behind the wall, the estimation of the life form's position in three-dimensional space, and the estimation of the life form's corresponding time modulation parameters, such as respiratory rate and respiratory rhythm stability. In the above technical solution, the results of through-wall life detection are not a simple binary judgment of "presence or absence of life," but a comprehensive decision based on life phase-locked field, life trajectory cluster, adaptive beam output, and sparse inversion results. It considers the energy concentration in the spatial direction, the consistency of the propagation path, and whether the time modulation characteristics are consistent with a reasonable life rhythm, thereby improving the reliability and accuracy of life detection in scenarios with strong reverberation, complex wall structures, and extremely rich multipaths.

[0071] Furthermore, when aggregating the distribution of life trajectory clusters on the life phase-locked time axis with their corresponding three-dimensional spatial directions, propagation paths, and temporal modulation features to form a life phase-locked field feature representation, firstly, for the target multipath state vector within each life trajectory cluster, its occurrence time, phase interval coverage, and number of cross-cycle repetitions are statistically analyzed on the life phase-locked time axis. Simultaneously, combined with its corresponding three-dimensional spatial direction vector, propagation path number sequence, and temporal modulation feature sequence, the multipath paths within the same life trajectory cluster are weighted and aggregated to construct a high-dimensional feature vector that can characterize the overall behavior of the life trajectory cluster. The first... The aggregation feature of a cluster of life trajectories on the life phase-locked time axis is represented as follows:

[0072] in, Indicates the first The life phase-locked field feature representation vector corresponding to each life trajectory cluster; This represents the aggregated spatial direction feature obtained by weighting the spatial direction vectors of all multipaths within a life trajectory cluster according to energy weights, including azimuth and pitch components. This represents the path distance feature after weighted aggregation of the propagation distance vectors of all multipaths within a life trajectory cluster, used to characterize the effective distance distribution from life targets to the radar array; This represents the propagation topology feature obtained by statistically aggregating the path numbers or propagation topology codes of all paths within a life trajectory cluster. This refers to the time modulation feature vector extracted using the time modulation sequence on the life phase-locked time axis, for example, encoding the main frequency, harmonic energy distribution and phase stability of the life micro-motion signal over multiple cycles; This represents statistics on the stability of the trajectory cluster across different wall scenarios and observation periods, such as the number of times the trajectory cluster appears across periods and the connectivity statistics in adjacent life phase positions.

[0073] The feature vectors of all life trajectory clusters are stacked according to their cluster indices and combined with prior environmental features, such as wall type encoding and noise environmental statistics. Pooling and concatenation operations are performed in the time and cluster dimensions to form a life phase-locked field feature representation matrix or a global feature vector. This life phase-locked field feature representation is then input into a meta-learning model, which learns the mapping rules from life phase-locked field features to meta-parameter vectors of adaptive beamspace weighted structure and sparse inversion path weighted structure in a high-dimensional feature space. This results in the output of meta-parameter vectors used to control the downstream solution process.

[0074] In the array covariance matrix estimation based on the meta-parameter vector, while applying enhancement weights to angular sectors aligned with the spatial direction of life trajectory clusters and suppression weights to angular sectors related to reverberant trajectory clusters, the meta-parameter vector output by the meta-learning model is first decomposed into a set of angular weight parameters for controlling beam space weighting and a set of path weight parameters for controlling path sparsity constraints during sparse inversion. In array covariance matrix estimation, a direction-related weight function is constructed for each candidate spatial direction. This weight function is applied to the array's received snapshot data, and the spatiotemporal covariance is calculated after weighting. Let the array at time index... The receiving vector at that location is For each spatial direction Constructing the directional weight function The weighted covariance matrix estimate can then be expressed as:

[0075] in, Indicates in the metaparameter vector Weighted array covariance matrix under control; This represents the number of time snapshots used to estimate the covariance matrix; This represents a predefined set of angular sectors in the spatial orientation domain; This represents the angle weighting function controlled by the meta-parameter vector. It assigns a larger weight to angle sectors that are consistent with or close to the spatial direction of the life trajectory cluster, and a smaller or suppressive weight to angle sectors that are related to the reverberant trajectory cluster. Indicates the direction of the array The guiding vector; This represents the conjugate transpose. By applying angular weights during the covariance matrix estimation stage, adaptive beamforming algorithms, such as minimum variance distortionless response beamforming, automatically enhance the signal components in the life trajectory cluster direction and suppress the reverberation direction during subsequent solution processes.

[0076] Meanwhile, in the sparse inversion process, the coefficient vector corresponding to the multipath propagation path is denoted as... The path-weighted sparse reconstruction objective function is constructed as follows: in, Represents the observation data vector; This represents the forward model matrix composed of the multipath propagation dictionary; This is a multipath propagation path coefficient vector, with each dimension... Corresponding to the Complex amplitude of each propagation path; To target the first under meta-parameter vector control The strength of the regularization penalty applied to each propagation path, when the path When it belongs to the path set corresponding to the life trajectory cluster Smaller values ​​result in weaker sparsity suppression of the path in the minimization objective, making it easier to preserve; when the path When it belongs to the set of reverberant trajectory cluster paths, The large value makes it easier to compress to near zero during iterative updates. By combining adaptive beam optimization and sparse inversion iteration with a finite number of steps, in the beam domain, the energy gradually contracts toward the spatial direction corresponding to the life trajectory cluster, and in the path domain, the multipath energy gradually concentrates on the propagation path corresponding to the life trajectory cluster, thus forming a beam output and inversion result that converges around the life trajectory cluster in the life phase-locked field.

[0077] When outputting the through-wall life detection results based on beam output and inversion results, the spatial energy distribution obtained after adaptive beam processing and the path coefficient distribution obtained after sparse inversion processing are jointly analyzed in the life phase-locked time axis and three-dimensional spatial coordinate system. Life presence determination and life parameter estimation are achieved by constructing life detection statistics in the spatial and temporal domains. A comprehensive detection statistic can be defined to measure the energy concentration and temporal rhythm stability of life trajectory clusters in the life phase-locked field, as follows:

[0078] in, This represents the comprehensive life detection statistics, used to provide a confidence measure for the results of penetrating-wall life detection; This represents the set of indexes for trajectory clusters identified as life trajectory clusters; Indicates the first The confidence weight of each life trajectory cluster can be determined by combining the clustering results of the topology-preserving graph neural network, the cross-period stability of the trajectory cluster, and the output confidence of the meta-learning model. Indicates clusters of life trajectories The corresponding propagation path index set; Indicating the first sparse inversion Reconstruction coefficients of the propagation path; This represents the normalized proportion of the life trajectory cluster in the path domain energy; Indicates the first The first life trajectory cluster on the life phase-locked timeline Phase at each phase position; This represents an ideal life phase reference obtained based on the estimation of the main life rhythm; This represents the number of discrete phase samples used to evaluate phase consistency across one or more cycles. Secondly, the first term, energy proportion, characterizes the dominance of the life trajectory cluster in the inversion result energy, while the second term, phase consistency, characterizes the alignment of the life trajectory cluster between the actual phase trajectory and the ideal main life rhythm. According to... By comparing with a preset threshold, the system can output a determination of whether a living organism exists. It can also estimate the three-dimensional position of the living organism based on the spatial direction and propagation distance corresponding to the life trajectory cluster, and estimate the vital signs parameters such as the breathing rate and breathing stability of the living organism based on the time modulation feature sequence. This results in a through-wall life detection result that includes determination of the existence of life, spatial positioning, and estimation of vital signs.

[0079] This application also provides a three-dimensional through-wall radar adaptive beam life detection system, referring to... Figure 2 , Figure 2This application provides a schematic diagram of a three-dimensional through-wall radar adaptive beam life detection system. The system is a server, comprising an acquisition module 21 and a processing module 22. The acquisition module 21 acquires a long-term three-dimensional radar data cube covering a wall region and identifies candidate life regions within the cube. The processing module 22 inputs the long-term three-dimensional radar data cube corresponding to the candidate life regions into a variational temporal autoencoder to obtain the dominant life rhythm. The processing module 22 is also used to input the multipath structure corresponding to each short time window into a neural network constant differential equation based on the dominant life rhythm to generate a reparameter mapping curve. The processing module 22 is further used to... Based on the reparameter mapping curve, time reparameterization processing is performed on the multipath structure to obtain the life phase-locked time axis; the processing module 22 is also used to construct multipath phase slices for multiple life phase positions based on the life phase-locked time axis, so as to construct a graph structure according to each multipath phase slice, and input the graph structure into the topology-preserving graph neural network to identify life trajectory clusters that maintain connectivity consistency in adjacent phase slices; the processing module 22 is also used to perform adaptive adjustment on the spatial direction, propagation path and temporal modulation features corresponding to the life trajectory clusters based on the meta-learning model, so that the spatial weighting structure and sparse inversion weight of the adaptive beam converge around the life trajectory clusters in the life phase-locked field to output the through-wall life detection results.

[0080] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0081] This application also provides an electronic device, with reference to... Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: at least one processor 31, at least one network interface 34, a user interface 33, a memory 35, and at least one communication bus 32.

[0082] The communication bus 32 is used to enable communication between these components.

[0083] The user interface 33 may include a display screen and a camera. Optionally, the user interface 33 may also include a standard wired interface and a wireless interface.

[0084] The network interface 34 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0085] The processor 31 may include one or more processing cores. The processor 31 connects to various parts of the server via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in the memory 35, and calling data stored in the memory 35 to perform various server functions and process data. Optionally, the processor 31 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 31 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 31 and may be implemented as a separate chip.

[0086] The memory 35 may include random access memory (RAM) or read-only memory. Optionally, the memory 35 may include a non-transitory computer-readable storage medium. The memory 35 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 35 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 35 may also be at least one storage device located remotely from the aforementioned processor 31. Figure 3 As shown, the memory 35, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a three-dimensional through-wall radar adaptive beam life detection method.

[0087] exist Figure 3In the electronic device shown, the user interface 33 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 31 can be used to call the application program of a three-dimensional through-wall radar adaptive beam life detection method stored in the memory 35. When executed by one or more processors, the electronic device executes one or more methods as described in the above embodiments.

[0088] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0089] This application also provides a computer-readable storage medium storing instructions. When executed by one or more processors, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0090] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A three-dimensional through-wall radar adaptive beam life detection method, characterized in that, The method includes: Acquire a long-time three-dimensional radar data cube covering the wall area, and determine life candidate areas in the long-time three-dimensional radar data cube; The long-time three-dimensional radar data cube corresponding to the life candidate region is input into the variational time-series autoencoder to obtain the life master rhythm; Based on the aforementioned life master rhythm, the multipath structure corresponding to each short time window is input into the neural ordinary differential equation to generate a reparameter mapping curve; Based on the reparameter mapping curve, the multipath structure is subjected to time reparameterization processing to obtain the life phase-locked time axis; Based on the life phase-locked time axis, multipath phase slices are constructed for multiple life phase positions to build a graph structure according to each multipath phase slice. The graph structure is then input into a topology-preserving graph neural network to identify life trajectory clusters that maintain connectivity and consistency in adjacent phase slices. Based on the meta-learning model, the spatial direction, propagation path and temporal modulation features corresponding to the life trajectory cluster are adaptively adjusted so that the spatial weighting structure and sparse inversion weight of the adaptive beam converge around the life trajectory cluster in the life phase-locked field, so as to output the through-wall life detection results. The meta-learning model adaptively adjusts the spatial direction, propagation path, and temporal modulation features corresponding to the life trajectory cluster, so that the spatial weighting structure and sparse inversion weights of the adaptive beam converge around the life trajectory cluster in the life phase-locked field, thereby outputting the through-wall life detection results, specifically including: The distribution of the life trajectory clusters on the life phase-locked time axis is aggregated with the corresponding three-dimensional spatial direction, propagation path and time modulation features to form a life phase-locked field feature representation. The life phase-locked field feature representation is then input into the meta-learning model to generate a meta-parameter vector for controlling the adaptive beam spatial weighting structure and the sparse inversion path weighting structure. Based on the meta-parameter vector, enhancement weights are applied to the angular sectors aligned with the spatial direction of the life trajectory cluster in the array covariance matrix estimation, and suppression weights are applied to the angular sectors associated with the reverberant trajectory cluster. Simultaneously, during the sparse inversion process, the regularization penalty intensity is reduced for the propagation path corresponding to the life trajectory cluster, and the regularization penalty intensity is increased for the propagation path corresponding to the reverberant trajectory cluster. This results in the life trajectory cluster being enhanced in three-dimensional space after a finite number of adaptive beam optimization and sparse inversion iterations, forming a beam output and inversion result that converges around the life trajectory cluster in the life phase-locked field. Based on the beam output and inversion results, the results of the through-wall life detection are output.

2. The three-dimensional through-wall radar adaptive beam life detection method according to claim 1, characterized in that, The process of acquiring a long-term time-series three-dimensional radar data cube covering the wall area and determining life candidate regions within the long-term time-series three-dimensional radar data cube specifically includes: By continuously transmitting multi-band linear frequency modulated pulses through a through-wall radar array and receiving echo signals after refraction, reflection and scattering through the wall, range compression, angular imaging and azimuth focusing are performed on each transmission cycle, so that multiple frames of three-dimensional radar data are stacked in time order to form a long time-series three-dimensional radar data cube covering the wall area. Low-resolution spatial focusing processing is performed on the long-time-series three-dimensional radar data cube to smooth the local texture scattering components of the wall, and large time window smoothing processing is performed on the long-time-series three-dimensional radar data cube to accumulate the energy of adjacent time frames, so that the life-related scattering components with slow periodic fluctuations over a long time scale form a spatially weak fluctuation region with temporal consistency. Based on the weak spatial fluctuation region, energy field analysis is performed to obtain the three-dimensional distribution structure of energy changing over time. The consistency of the three-dimensional distribution structure is used to identify target regions with continuous weak rhythmic changes and stable spatial connectivity in multiple sliding time windows, and these target regions are selected as candidate regions for life.

3. The three-dimensional through-wall radar adaptive beam life detection method according to claim 1, characterized in that, The step of inputting the long-time three-dimensional radar data cube corresponding to the life candidate region into a variational temporal autoencoder to obtain the main life rhythm specifically includes: The long-time three-dimensional radar data cube corresponding to the life candidate region is divided into time segments containing multiple frames of continuous three-dimensional radar data in chronological order, and amplitude normalization and noise floor suppression are performed on each time segment to ensure that the input distribution remains consistent across different time segments. The processed time segment is input into the encoder structure of the variational temporal autoencoder. A latent spatial distribution is generated through multi-layer temporal convolution, gated recurrent units and temporal attention structures. The mean parameter and perturbation parameter of the latent spatial distribution are subject to regularization constraints through a variational inference mechanism to obtain the latent trajectory. The potential trajectory is input into the decoder structure of the variational temporal autoencoder to generate a reconstructed sequence that reflects the characteristics of slow periodic fluctuations. Based on multiple reconstructed sequences, rhythm trajectory extraction and weak periodic enhancement processing are performed to determine the trajectory with phase continuity and periodic consistency over a long period of time as the main life rhythm.

4. The three-dimensional through-wall radar adaptive beam life detection method according to claim 1, characterized in that, The step of inputting the multipath structure corresponding to each short time window into the neural ordinary differential equation based on the main life rhythm to generate a reparameter mapping curve specifically includes: The multipath structure within each short time window is represented as a sequence of multipath state vectors that evolves with physical time through time delay axis normalization and energy normalization. The multipath state vector sequence and the phase of the corresponding short-time window of the life main rhythm are input together into the time vector field model parameterized by the deep neural network. The time vector field model generates a reparameter mapping curve that changes continuously with time and remains monotonically increasing and has a smooth derivative in the neural ordinary differential equation by integral method.

5. The three-dimensional through-wall radar adaptive beam life detection method according to claim 4, characterized in that, The step of performing time reparameterization processing on the multipath structure based on the reparameter mapping curve to obtain the life phase-locked time axis specifically includes: The physical time sampling points corresponding to each short time window are mapped to the reparameter time values ​​on the initial phase-locked time axis through the reparameter mapping curve, and the multipath state vector sequence is time-coordinate-shifted according to the reparameter time values ​​to form a target multipath state vector sequence defined on the initial phase-locked time axis. A time overlap region is set between adjacent short time windows, and after completing their respective time reparameterization, smooth splicing is performed based on the statistical consistency of the reparameterized time values ​​in the overlap region and the corresponding target multipath state vector, so as to form a continuous and smooth intermediate phase-locked time axis in the global scope. A target time grid with fixed intervals is constructed on the intermediate phase-locked time axis, and the mapped target multipath state vector is resampled by interpolation. This makes the life-related paths exhibit a slow evolution with an approximate period in the intermediate phase-locked time axis, while the reverberation path exhibits irregular drift, thus obtaining a life phase-locked time axis for subsequent life trajectory cluster identification.

6. The three-dimensional through-wall radar adaptive beam life detection method according to claim 5, characterized in that, The process of constructing multipath phase slices for multiple life phase positions based on the life phase-locked time axis, building a graph structure based on each multipath phase slice, and inputting the graph structure into a topology-preserving graph neural network to identify life trajectory clusters that maintain connectivity and consistency in adjacent phase slices specifically includes: Phase normalization processing is performed on the life phase-locked time axis to map the reparameter time value to the period-normalized life phase coordinate interval, and multiple life phase position intervals are divided according to the life phase coordinate interval. The target multipath state vector sequence is classified according to the life phase position interval to form multiple multipath phase slices, and a graph structure containing target multipath state vector nodes and geometric proximity relationships, energy distribution similarity relationships and phase evolution smoothness relationships is constructed based on each multipath phase slice; The graph structure is input into the topology-preserving graph neural network, so that the target multipath state vectors that appear a preset number of times in adjacent phase slices and maintain spatial connectivity are aggregated into life trajectory clusters in the node embedding space, and the reverberant trajectory clusters that exhibit jumps or breaks are dispersed in the node embedding space, thereby identifying life trajectory clusters that maintain connectivity in adjacent phase slices.

7. A three-dimensional through-wall radar adaptive beam life detection system, characterized in that, The system is used to execute the three-dimensional through-wall radar adaptive beam life detection method as described in any one of claims 1 to 6, the system comprising an acquisition module and a processing module, wherein... The acquisition module is used to acquire a long-time three-dimensional radar data cube covering the wall area and to determine the life candidate area in the long-time three-dimensional radar data cube. The processing module is used to input the long-time three-dimensional radar data cube corresponding to the life candidate region into the variational time-series autoencoder to obtain the life master rhythm; The processing module is also used to input the multipath structure corresponding to each short time window into the neural ordinary differential equation according to the main life rhythm, so as to generate a reparameter mapping curve. The processing module is also used to perform time reparameterization processing on the multipath structure based on the reparameter mapping curve to obtain the life phase-locked time axis; The processing module is further configured to construct multipath phase slices for multiple life phase positions based on the life phase-locked time axis, to construct a graph structure according to each multipath phase slice, and to input the graph structure into a topology-preserving graph neural network to identify life trajectory clusters that maintain connectivity consistency in adjacent phase slices; The processing module is also used to perform adaptive adjustments on the spatial direction, propagation path and temporal modulation features corresponding to the life trajectory cluster based on the meta-learning model, so that the spatial weighting structure and sparse inversion weight of the adaptive beam converge around the life trajectory cluster in the life phase-locked field, so as to output the through-wall life detection results.

8. An electronic device, characterized in that, The electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1 to 6.

Images