A radar micro-motion feature extraction method based on deep learning

Abstract

The present application relates to the technical field of feature detection, in particular to a radar micro-motion feature extraction method based on deep learning, comprising the following steps: collecting the instantaneous amplitude, slope, time interval and phase change of the radar signal, normalizing into a time sequence feature group, constructing and fusing into a feature matrix, inputting into a time sequence convolution network model for multi-layer extraction and modeling, implementing multi-level discrimination and dynamic screening, and finally outputting a segment sequence based on attribution confidence probability through threshold judgment and multi-element joint screening.In the present application, four-element time sequence features are constructed by capturing instantaneous amplitude, slope, extreme point interval and phase change, revealing the internal correlation of target weak dynamics, automatically modeling the space-time correlation to mine nonlinear and non-stationary features, and implementing multi-level discrimination, dynamic screening and attribution confidence probability joint judgment on the features, so as to accurately lock the effective micro-motion signal segment under complex background interference, and significantly improve the robustness and accuracy of feature extraction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of feature detection, in particular to a radar micro-motion feature extraction method based on deep learning. BACKGROUND

[0002] The technical field of feature detection belongs to an important direction of signal and information processing, and its core matters include the identification and extraction of target features in complex backgrounds and multi-dimensional environments, covering the detection of target motion state, the capture of subtle structural changes, and the analysis of weak dynamic features in signals. The overall technical field systematically involves electromagnetic signal acquisition, feature pattern recognition, dynamic behavior representation, and other aspects, and is widely used in security monitoring, life detection, traffic control, and intelligent interaction scenarios. Among them, the traditional radar micro-motion feature extraction method refers to the way of processing the Doppler effect features caused by the local motion of the target in the radar echo signal. It mainly aims at how to separate the subtle motion information formed by the vibration of the target's limbs, the fluctuation of breathing, or the slight jitter of mechanical parts from the continuous echo. Common means usually use Fourier spectrum decomposition to analyze periodic micro-motion signals, short-time energy change-based detection to capture transient features, and time-frequency analysis to unfold the signal on the time axis and frequency axis to reveal the implicit micro-motion pattern, thereby forming a systematic extraction approach to micro-motion features.

[0003] The existing technology highly depends on Fourier spectrum decomposition and is mainly designed for periodic and stationary signals. When the target micro-motion presents nonlinear or transient characteristics, the spectral features will be dispersed, resulting in the loss of key information details and restricting the integrity of feature extraction. At the same time, in complex electromagnetic environments, detection methods based on short-time energy changes or single time-frequency patterns are easily disturbed by background clutter and noise signals. Due to the lack of confidence evaluation mechanism for feature sources, irrelevant disturbances are often misjudged as target micro-motion, causing a high false alarm rate. For example, in life detection applications, mechanical vibrations in the environment may be mistakenly identified as weak breathing signals, thereby seriously affecting the accuracy and reliability of the detection results. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art and to provide a radar micro-motion feature extraction method based on deep learning.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme: a radar micro-motion feature extraction method based on deep learning, comprising the following steps:

[0006] S1: Collecting radar signal continuous sampling data, obtaining instantaneous amplitude change, amplitude change slope, extreme point interval sampling time interval, and instantaneous phase change, combining them as a four-element time sequence feature group, calling a normalization processing function to normalize the four-element time sequence feature group, and generating a normalized time sequence feature block;

[0007] S2: Input the normalized temporal feature block into the feature matrix construction process, organize it into a temporal two-dimensional feature matrix, and fuse the instantaneous phase change to generate a fused feature matrix;

[0008] S3: Input the fused feature matrix into the temporal convolutional network model, perform multi-layer feature extraction, automatically model spatiotemporal correlation, obtain the discrimination probability distribution of each segment at the output end, extract the attribution confidence probability distribution, and perform multi-level discrimination for the discrimination feature distribution, perform dynamic filtering, and output confidence probability and feature parameters.

[0009] S4: Based on the attribution confidence probability distribution, perform probability threshold judgment on the signal segmentation attribution results according to the discriminant probability distribution, call the maximum relative fluctuation amplitude and root mean square parameter within the window, perform multivariate joint screening for data segments whose discriminant probability distribution is within the threshold interval, and output the screened segment sequence.

[0010] As a further aspect of the present invention, the normalized time series feature block includes an amplitude normalized sequence, a slope normalized sequence, an extreme value interval normalized sequence, and a phase change normalized sequence. The fused feature matrix is ​​specifically a two-dimensional time series feature matrix that fuses the phase change normalized sequence. The generation step of the attribution confidence probability distribution includes performing dynamic screening to determine dynamic screening feature parameters based on the segmented discrimination probability and the attribution confidence parameters. The screened segment sequence is specifically a signal segment sequence obtained after the multivariate joint screening.

[0011] As a further aspect of the present invention, the step of obtaining the normalized temporal feature block specifically comprises:

[0012] S101: Collects continuous sampling data of radar signals, calculates the instantaneous amplitude difference between adjacent sampling points and the time ratio to obtain the amplitude change slope, counts the number of samples between waveform inflection points to convert the time interval, and analyzes the phase information of the signal to extract the instantaneous phase change. Combines the instantaneous amplitude change, amplitude change slope, sampling time interval between extreme points and instantaneous phase change to establish the original multidimensional feature sequence.

[0013] S102: Align the four features in the original multidimensional feature sequence—the instantaneous amplitude change, the slope of the amplitude change, the sampling time interval between extreme points, and the instantaneous phase change—at the same time point, and arrange them in a preset order. Treat the four features at each time point as a unit to form a structured data set, and obtain a quaternary time series feature group.

[0014] S103: traversing all values of each feature dimension in the four-element time sequence feature group, determining the upper limit and lower limit of the values of multiple dimensions, performing scale scaling calculation on each original value under the dimension according to the difference between the upper limit and the lower limit of the values, and mapping to a preset value interval to generate a normalized time sequence feature block.

[0015] As a further scheme of the application, the acquisition step of the fusion feature matrix is specifically:

[0016] S201: calling the normalized time sequence feature block, setting a time step as the basis for data regularization, stacking the continuous feature blocks in time sequence into rows, and arranging the feature quantities of the normalized instantaneous amplitude change, amplitude change slope and sampling time interval between extreme points in each feature block as columns for structured arrangement, and establishing a time sequence two-dimensional feature matrix;

[0017] S202: according to the row and column structure of the time sequence two-dimensional feature matrix, extracting the corresponding instantaneous phase change sequence from the normalized time sequence feature block, taking the instantaneous phase change sequence as the current feature dimension, and performing splicing calculation with the existing feature vector of each time step in the time sequence two-dimensional feature matrix to generate a fusion feature matrix.

[0018] As a further scheme of the application, the acquisition step of the confidence probability and the feature parameter is specifically:

[0019] S301: inputting the fusion feature matrix into a time sequence convolution network, performing feature extraction on the space-time correlation in the matrix through multiple layers of convolution kernels and activation functions, and performing normalization processing on the output values of the full connection layer at the end of the network to calculate the class attribution probability of each segment and obtain the discriminant probability distribution;

[0020] S302: based on the discriminant probability distribution, tracing back the contribution values of multiple layers of features in the network to the output result, quantifying the influence weight of each input feature parameter on the discriminant result, and combining the weight with the original probability value in the discriminant probability distribution to generate an attribution confidence probability distribution;

[0021] S303: for the attribution confidence probability distribution, setting multiple progressive discriminant confidence thresholds, comparing each probability value in the distribution with the multiple thresholds one by one in value size, screening the probability values exceeding the discriminant confidence threshold, and extracting the original feature parameters corresponding to the probability values to output the confidence probability and the feature parameter.

[0022] As a further scheme of the application, the acquisition step of the screening segment sequence is specifically:

[0023] S401: According to the attribution confidence probability distribution, the attribution result of each signal segment in the discriminant probability distribution is called, its value is compared with the upper and lower limits of the set probability threshold, and the data segment in the threshold interval is recorded to establish a candidate segment index set;

[0024] S402: For each data segment recorded in the candidate segment index set, all signal sampling points in the corresponding time window are extracted, the difference between the maximum value and the minimum value of the signal amplitude in the window is calculated, and the square mean value of all sampling point amplitudes is calculated and then the square root is taken to obtain the fluctuation characteristic measurement value;

[0025] S403: The maximum relative fluctuation amplitude and the fluctuation amplitude filtering reference of each segment in the candidate segment index set and the corresponding fluctuation characteristic measurement value are compared, and the root mean square parameter and the energy filtering reference are compared, the segments whose two parameters exceed the respective references are filtered, and the filtered segment sequence is output in time sequence.

[0026] As a further scheme of the present application, the acquisition step of the discriminant probability distribution is specifically:

[0027] The non-normalized logic value output by the full connection layer in the time sequence convolution network model for any signal segment is obtained, and based on the total number of preset micro-motion target categories, the non-normalized logic value is calculated by applying a Softmax function, and the value of each dimension in the calculation result is taken as the category belonging probability of the corresponding category, wherein the calculation formula of the category belonging probability is:

[0028] ;

[0029] Wherein, is the category belonging probability of the current signal segment being judged as the i-th category, is the non-normalized logic value output by the full connection layer corresponding to the i-th category, is the non-normalized logic value output by the full connection layer corresponding to the i-th category, is the non-normalized logic value output by the full connection layer corresponding to the i-th category, is the non-normalized logic value output by the full connection layer corresponding to the i-th category, is the non-normalized logic value output by the full connection layer corresponding to the i-th category, is the total number of preset micro-motion target categories;

[0030] The category belonging probabilities of all micro-motion target categories are combined to establish a discriminant probability distribution.

[0031] As a further scheme of the present application, the acquisition step of the attribution confidence probability distribution is specifically:

[0032] The influence weight of each input feature parameter is multiplied with the class attribution probability of each class in the discriminant probability distribution to obtain the weighted contribution of the input feature parameters.

[0033] The weighted contributions of all input feature parameters related to the same class attribution probability are summed up to obtain the attribution confidence score of the corresponding class.

[0034] The attribution confidence scores of all micro-motion target classes are normalized so that the sum of the attribution confidence scores of all classes is 1 to generate an attribution confidence probability distribution.

[0035] As a further scheme of the present application, the step of obtaining the fluctuation characteristic metric value specifically comprises:

[0036] For each data segment recorded in the candidate segment index set, all signal sampling points in the corresponding time window are obtained, the maximum signal amplitude and the minimum signal amplitude in the all signal sampling points are calculated, and the arithmetic mean value and the root mean square value of the signal amplitudes of the all signal sampling points are calculated.

[0037] According to a preset fluctuation balance coefficient, the difference between the maximum signal amplitude and the minimum signal amplitude and the root mean square value of the signal amplitudes are weighted and summed to obtain a fluctuation characteristic metric value, and the calculation formula is:

[0038] ;

[0039] wherein, is the fluctuation characteristic metric value, is the fluctuation balance coefficient, is the maximum signal amplitude, is the minimum signal amplitude, is the root mean square value of the signal amplitudes.

[0040] As a further scheme of the present application, the step of obtaining the screened fragment sequence further comprises:

[0041] The fluctuation characteristic metric values of all data segments in the candidate segment index set are calculated to form a fluctuation characteristic set, the statistical distribution of the fluctuation characteristic set is analyzed, and the numerical mean value and the numerical standard deviation of the maximum-minimum amplitude difference part and the root mean square part of the signal amplitudes in the fluctuation characteristic set are calculated.

[0042] The numerical mean value of the maximum-minimum amplitude difference part is multiplied by a preset first multiple factor, and the result of multiplying the numerical standard deviation of the maximum-minimum amplitude difference part by a preset second multiple factor is added to determine the fluctuation amplitude screening reference together.

[0043] The numerical mean value of the signal amplitude root mean square part is multiplied by a preset third multiple factor, and the result of multiplying the numerical standard deviation of the signal amplitude root mean square part by a preset fourth multiple factor is added to determine the energy screening reference together.

[0044] Compared with the prior art, the advantages and positive effects of the present application are that:

[0045] In the present application, by synchronously capturing the instantaneous amplitude change, the amplitude change slope, the sampling time interval between extreme points and the instantaneous phase change, a more comprehensive four-element time sequence feature is constructed from the radar original signal, which deeply reveals the multiple internal correlations of the weak dynamic of the target. This process automatically models the space-time correlation, deeply mines and characterizes the nonlinear and non-stationary features that cannot be captured by traditional spectral analysis. At the same time, multi-level discrimination and dynamic screening are implemented for the identified feature distribution, combined with threshold judgment and multi-element joint screening of the attribution confidence probability distribution, to achieve accurate locking of effective micro-motion signals in complex background interference, and significantly improve the robustness and accuracy of micro-motion feature extraction. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The present application is a radar micro-motion feature extraction method based on deep learning, and the total flow chart is shown in the figure.

[0047] Figure 2 The present application is a normalized time sequence feature block generation flow chart.

[0048] Figure 3 The present application is a fusion feature matrix generation flow chart.

[0049] Figure 4 The present application is an attribution confidence probability distribution generation and screening flow chart.

[0050] Figure 5 The present application is a screening fragment sequence generation flow chart. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme realized by software will be described in detail below in combination with the system architecture diagram and the embodiment. It should be understood that the specific embodiments described herein are only used to explain the technical scheme of the present application, and do not constitute a limitation on the scope of protection.

[0052] In the description of the present application, the system architecture relationship or data processing flow indicated by the terms "hierarchy", "module", "interface", "data flow", "client", "server" and the like are defined based on the corresponding architecture diagram or flowchart of the embodiment. This way of expression is only used to clearly explain the logical relationship of each element in the technical solution, and is not limited to the physical deployment form. The "multiple" contains two or more technical units, including but not limited to multiple data nodes, processing threads, service instances or functional components, and other scalable elements. The specific number is determined according to the actual business scenario.

[0053] Please refer to Figure 1 and Figure 2 The present application provides a technical solution: a radar micro-motion feature extraction method based on deep learning, comprising the following steps:

[0054] S1: Obtain the instantaneous amplitude change, amplitude change slope, extreme point interval sampling time interval and instantaneous phase change by collecting radar signal continuous sampling data, and combine them as a four-element time sequence feature group. Call the normalization processing function to normalize the four-element time sequence feature group to generate a normalized time sequence feature block;

[0055] The normalized time sequence feature block includes amplitude normalization sequence, slope normalization sequence, extreme interval normalization sequence and phase change normalization sequence. The acquisition steps of the normalized time sequence feature block are as follows:

[0056] S101: Collect radar signal continuous sampling data, calculate the instantaneous amplitude difference and time ratio between adjacent sampling points to obtain the amplitude change slope, count the number of sampling points between waveform turning points to convert the time interval, and analyze the phase information of the signal to extract the instantaneous phase change. Combine the instantaneous amplitude change, amplitude change slope, extreme point interval sampling time interval and instantaneous phase change to establish an original multi-dimensional feature sequence;

[0057] S102: Align the four feature quantities of instantaneous amplitude change, amplitude change slope, extreme point interval sampling time interval and instantaneous phase change at the same time point in the original multi-dimensional feature sequence, and arrange them in a predetermined order. Take the four feature quantities at each time point as a unit to form a structured data set, and obtain a four-element time sequence feature group;

[0058] S103: Traverse all values of each feature dimension in the four-element time sequence feature group, determine the upper limit and lower limit of the values of multiple dimensions, and perform scale scaling calculation on each original value under the dimension according to the difference between the upper limit and lower limit of the values, and map it to the preset value interval to generate a normalized time sequence feature block.

[0059] Take a continuous wave radar working at K band (24GHz) as an example, which is used to monitor whether there is a micro unmanned aerial vehicle activity in a certain airspace. The radar system continuously samples the echo signal at a sampling frequency of 50kHz, obtaining in-phase (I) and quadrature (Q) signal data.

[0060] S101: At time t0, the radar system starts collecting data, continuously collecting for 20ms, obtaining 1000 continuous complex sampling points. For a sampling point at time t1=1.2ms, its I / Q value is (0.85V, 0.62V), then the instantaneous amplitude A1 is calculated by . Its instantaneous phase is calculated by . The next sampling point, at t2=1.22ms, its I / Q value is (0.88V, 0.65V), its instantaneous amplitude , instantaneous phase . The instantaneous amplitude change is the amplitude value sequence of each sampling point, for example, at t2, the instantaneous amplitude is 1.094V. The amplitude change slope is calculated by , that is . The instantaneous phase change is . To obtain the sampling time interval between extreme points, the system processes the amplitude sequence of the 1000 points collected, detects the local maximum and minimum points of the amplitude signal. For example, a local maximum of 1.12V is detected at the 58th sampling point (t=1.16ms), and a local minimum of 0.95V is detected at the 69th sampling point (t=1.38ms), then the number of sampling points between the two extreme points is 11, and the corresponding time interval is . Combining the instantaneous amplitude, amplitude change slope, sampling time interval between extreme points, and instantaneous phase change of each sampling time point, a length of 1000 original multi-dimensional feature sequence is established.

[0061] ​​S102: Align the original multi-dimensional feature sequence established in S101. At each sampling time, for example, t2=1.22 ms, there is an instantaneous amplitude value (1.094V), an amplitude change slope value (2100V / s), an instantaneous phase change amount (0.006 rad). For the extreme value point interval sampling time interval, the value remains unchanged between two consecutive extreme value points. For example, between the 58th point and the 68th point, the corresponding extreme value point interval sampling time interval value of all sampling points is 0.22 ms. According to the preset order of [instantaneous amplitude, amplitude change slope, extreme value point interval sampling time interval, instantaneous phase change], the four feature quantities are arranged. At t2=1.22 ms, the unit formed is [1.094, 2100, 0.22, 0.006]. Perform this operation on all data of 1000 sampling points, and finally form a 1000x4 structured data set, obtaining a four-element time sequence feature group.

[0062] S103: Traverse the 1000x4 four-element time sequence feature group obtained in S102. First, determine the numerical upper and lower limits of each feature dimension. For example, after processing 1000 sampling points, it is found that the instantaneous amplitude dimension has a numerical range of [0.05V, 1.8V], a numerical upper limit of 1.8V, a numerical lower limit of 0.05V, and a difference of 1.75V. The range of the amplitude change slope dimension is [-5000V / s, 5200V / s], and the difference is 10200V / s. The range of the extreme value point interval sampling time interval is [0.1ms, 0.8ms], and the difference is 0.7ms. The range of the instantaneous phase change is [-0.1 rad, 0.1 rad], and the difference is 0.2 rad. Then, according to these differences, perform scale scaling calculation on each original value to map it to the preset numerical interval. Take the feature unit [1.094, 2100, 0.22, 0.006] at t2=1.22 ms as an example, and the normalization calculation process is as follows:

[0063] Amplitude normalized value: .

[0064] Slope normalized value: .

[0065] Extreme interval normalized value: .

[0066] Phase change normalized value: .

[0067] After normalization, the feature unit at t2 becomes [0.5966, 0.6961, 0.1714, 0.5300]. The same calculation is performed on the four-element time sequence feature group of all 1000 time points, and finally a 1000x4 normalized time sequence feature block is generated, which is composed of amplitude normalization sequence, slope normalization sequence, extreme interval normalization sequence and phase change normalization sequence.

[0068] Please refer to Figure 1 and Figure 3 , S2: input the normalized time sequence feature block into the feature matrix construction process, arrange it into a time sequence two-dimensional feature matrix, and fuse the instantaneous phase change to generate a fusion feature matrix;

[0069] The fusion feature matrix is specifically a two-dimensional time sequence feature matrix that has fused the phase change normalization sequence.

[0070] The acquisition step of the fusion feature matrix is specifically:

[0071] S201: call the normalized time sequence feature block, set the time step as the basis for data regularization, stack the continuous feature blocks in time sequence into rows, and arrange the normalized instantaneous amplitude change, amplitude change slope and extreme point interval sampling time interval in each feature block as columns, structure arrangement, establish a time sequence two-dimensional feature matrix;

[0072] S202: according to the row and column structure of the time sequence two-dimensional feature matrix, extract the corresponding instantaneous phase change sequence from the normalized time sequence feature block, take the instantaneous phase change sequence as the current feature dimension, and perform splicing calculation with the existing feature vector of each time step in the time sequence two-dimensional feature matrix, to generate a fusion feature matrix.

[0073] S201: Call the 1000x4 normalized timing feature block generated in S103, and set the time step to 100. This means that 100 consecutive sampling point data are taken as a processing unit. From the first time point of the feature block, the feature block data of the consecutive 100 time points are stacked in time order to form a 100x4 matrix. In this matrix, each row represents a time point, and each column corresponds to a feature dimension. Specifically, the normalized instantaneous amplitude change in each time point is taken as the first column, the amplitude change slope is taken as the second column, and the sampling time interval between extreme points is taken as the third column. At this time, the phase change sequence is not included temporarily, thereby establishing a 100x3 two-dimensional timing feature matrix. Taking the normalized data of the first to the 100th sampling point as an example, the dimension of the first two-dimensional timing feature matrix constructed is 100x3. Subsequently, the time window slides backward, and the data of the 101st to the 200th sampling point is processed to construct the second 100x3 matrix, and so on, until all 1000 data points are processed, and 10 two-dimensional timing feature matrices are generated.

[0074] S202: According to the row and column structure of the 10 100x3 two-dimensional timing feature matrices established in S201, the corresponding instantaneous phase change sequence is extracted from the normalized timing feature block. For the first 100x3 matrix (corresponding to the 1st to 100th sampling point), the first 100 phase change normalized values are extracted from the fourth column of the normalized timing feature block to form a 100x1 column vector. This phase change sequence is taken as a new feature dimension, and is concatenated with the 100x3 two-dimensional timing feature matrix. The specific operation is to concatenate the 100x1 column vector to the right side of the 100x3 matrix to form a 100x4 fusion feature matrix. For example, if the normalized feature of the first time point is [0.5966, 0.6961, 0.1714, 0.5300], the first row of the matrix constructed in S201 is [0.5966, 0.6961, 0.1714], and now the corresponding phase change normalized value 0.5300 is concatenated back, and the first row is restored to [0.5966, 0.6961, 0.1714, 0.5300]. This operation is performed on all 100 rows of data to finally generate a fusion feature matrix. This concatenation operation is performed on all 10 two-dimensional timing feature matrices generated in S201, and finally 10 100x4 fusion feature matrices are obtained, each of which is taken as an independent signal segment and input into the subsequent network model.

[0075] See Figure 1 and Figure 4S3: Input the fused feature matrix into the temporal convolutional network model, perform multi-layer feature extraction, automatically model spatiotemporal correlation, obtain the discrimination probability distribution of each segment at the output end, extract the attribution confidence probability distribution, and perform multi-level discrimination for the discrimination feature distribution, perform dynamic filtering, and output confidence probability and feature parameters.

[0076] The steps for generating the attribution confidence probability distribution include determining dynamic selection feature parameters based on the piecewise discriminant probability and attribution confidence parameters. The specific steps for obtaining the confidence probability and feature parameters are as follows:

[0077] S301: Input the fused feature matrix into the temporal convolutional network, extract features from the spatiotemporal correlation within the matrix through multiple convolutional kernels and activation functions, normalize the output values ​​of the fully connected layer at the end of the network, calculate the class assignment probability of each segment, and obtain the discrimination probability distribution.

[0078] S302: Based on the discriminant probability distribution, the contribution values ​​of multi-layer features in the reverse network to the output result are traced back, the influence weight of each input feature parameter on the discriminant result is quantified, and the weights are combined with the original probability values ​​in the discriminant probability distribution to generate the attribution confidence probability distribution.

[0079] S303: For the attribution confidence probability distribution, set multiple progressive discrimination confidence thresholds, compare each probability value in the distribution with the multi-level thresholds one by one, filter out probability values ​​that exceed the discrimination confidence thresholds, extract the original feature parameters corresponding to the probability values, and output the confidence probability and feature parameters.

[0080] The specific steps for obtaining the probability distribution are as follows:

[0081] Obtain the unnormalized logical values ​​of the fully connected layers in the temporal convolutional network model for any signal segment. Based on a preset total number of micro-motion target categories, apply the Softmax function to calculate the unnormalized logical values. Use the value of each dimension in the calculation result as the probability of belonging to the corresponding category. The formula for calculating the category belonging probability is as follows:

[0082] ;

[0083] in, The current signal segment is judged as the first. The probability of class classification. The corresponding output of the fully connected layer Unnormalized logical values ​​of a class The corresponding output of the fully connected layer Unnormalized logical values ​​of a class The total number of preset micro-motion target categories;

[0084] By combining the category assignment probabilities of all micro-movement target categories, a discriminative probability distribution is established. The specific steps for obtaining the attribution confidence probability distribution are as follows:

[0085] The system calls the discriminative probability distribution and the influence weights of multiple input feature parameters calculated through the backtracking network. For the class assignment probability of each category in the discriminative probability distribution, it multiplies it one by one with the influence weights of all input feature parameters corresponding to the class assignment probability to obtain the weighted contribution of multiple input feature parameters.

[0086] The weighted contributions of all input feature parameters related to the probability of belonging to the same category are summed, and the sum is used as the attribution confidence score of the corresponding category.

[0087] The attribution confidence scores of all micro-motion target categories are normalized so that the sum of the attribution confidence scores of all categories is 1, thus generating an attribution confidence probability distribution.

[0088] S301: Input the 10 100×4 fused feature matrices generated in S202 one by one into a pre-trained Temporal Convolutional Network (TCN) model. Taking the first 100×4 fused feature matrix as an example, this matrix first passes through the first convolutional layer of the network, which contains 64 convolutional kernels of size 1×3, to initially extract the temporal correlation of the input matrix. After the convolution operation, the ReLU activation function is passed to increase the non-linear expressive power of the network. Subsequently, the feature data flows through multiple residual modules containing dilated convolutions to further expand the receptive field and capture long-term dependencies. At the end of the network, after a global average pooling layer, a fully connected layer is connected. This fully connected layer outputs a 3-dimensional vector, corresponding to three preset micro-movement target categories: [drone, bird, background clutter]. For the first input signal segment, the unnormalized logits output by the fully connected layer are z=[z1,z2,z3]=[4.2,1.5,-0.8]. Here, z1 corresponds to "drone", z2 to "birds", and z3 to "background clutter". Then, the Softmax function is applied to normalize the logistic vector, and the probability of belonging to each category is calculated.

[0089] The formula for calculating the probability of category affiliation is: ;

[0090] This formula is used to transform the unnormalized logistic value vector output by a fully connected layer into a probability distribution, where each logistic value is processed by an exponential function and normalized by the sum of the exponentialized logistic values ​​of all categories. This indicates that the current signal segment is classified as the first... The probability of class classification is the final output probability value. the corresponding first class of the fully connected layer output. an unnormalized logical value of the class, whose numerical value represents the unnormalized score of the model for the class. the corresponding first class of the fully connected layer output. an unnormalized logical value of the class. C is the total number of preset micro-motion target classes, and in this embodiment, C = 3.

[0091] The calculation is performed on the logical value z = [4.2, 1.5, -0.8] output above:

[0092] The denominator calculation is:

[0093] The class probability calculation is:

[0094]

[0095]

[0096] The attribution probabilities of the three classes are combined to obtain the discriminant probability distribution of the signal segment as [0.9312, 0.0626, 0.0063]. This calculation is performed on all 10 signal segments to obtain the respective discriminant probability distributions.

[0097] S302: Based on the discriminant probability distribution [0.9312, 0.0626, 0.0063] obtained from the first signal segment, the contribution of each layer feature in the network to the output result is traced back. This process uses a variant method of gradient-weighted class activation mapping (Grad-CAM) to calculate the output probability of the “drone” class (the class with the highest probability) relative to each element (a total of 100 x 4 = 400 input feature parameters) in the input fusion feature matrix. These gradient values quantify the influence weight of each input feature parameter on the final result of being discriminated as “drone”. For example, the feature in the 10th row and the 2nd column of the input matrix (the slope normalized value at t = 0.2 ms) has a contribution of 0.0002 to the output probability of the “drone” class. ​​​​The influence weight of the feature in the 4th column of the 80th row (phase change normalized value at t = 1.6 ms) is 0.15. The original probability value in the discriminant probability distribution is combined with these influence weights. For the "drone" category, multiply its category attribution probability 0.9312 with the influence weights of all 400 input feature parameters on this category one by one to obtain 400 weighted contribution degrees. For example, the weighted contribution degrees of the above two features are 0.9312*0.08 = 0.0745 and 0.9312*0.15 = 0.1397, respectively. Then, the weighted contribution degrees of all 400 input feature parameters related to the "drone" category are aggregated and summed up to obtain the attribution confidence score of this category, for example, the sum is 15.8. The same operation is performed on "bird" and "background clutter" to obtain their attribution confidence scores, for example, 2.1 and 0.5, respectively. Finally, normalize the three attribution confidence scores [15.8, 2.1, 0.5] so that their sum is 1. The normalized sum is 15.8 + 2.1 + 0.5 = 18.4, and the attribution confidence probability distribution is [15.8 / 18.4, 2.1 / 18.4, 0.5 / 18.4] \approx [0.8587, 0.1141, 0.0272].

[0098] S303: For the attribution confidence probability distribution [0.8587, 0.1141, 0.0272] generated in S302, three progressive discriminant confidence thresholds are set: T1 = 0.70, T2 = 0.85, and T3 = 0.95. These thresholds are determined based on testing on a validation dataset containing 5000 labeled radar signal segments. The testing goal is to obtain different recall levels by adjusting the thresholds while ensuring that the precision is not less than 98%. Set the first discriminant confidence threshold T1 = 0.70, the second discriminant confidence threshold T2 = 0.85, and the third discriminant confidence threshold T3 = 0.95. Compare each probability value in the distribution with the three thresholds one by one in terms of numerical size. The attribution confidence probability of the "drone" category of this segment is 0.8587, which is greater than T1 = 0.70 and T2 = 0.85, but not greater than T3 = 0.95. Therefore, this probability value is screened out. Then, the original feature parameters corresponding to this probability value are extracted, i.e., all the data in the first 100x4 fusion feature matrix input into the network. Finally, the confidence probability (0.8587) of this segment and its corresponding original feature parameter set are output.

[0099] See Figure 1 and Figure 5S4: performing probability threshold judgment on the signal segment attribution result based on the discriminant probability distribution through the attribution confidence probability distribution, calling the maximum relative fluctuation amplitude and the root mean square parameter in the window, performing multivariate joint screening on the data segment whose discriminant probability distribution is in the threshold interval, and outputting a screening fragment sequence. The screening fragment sequence is specifically a signal segment sequence obtained after multivariate joint screening.

[0100] The acquisition step of the screening fragment sequence is specifically:

[0101] S401: According to the attribution confidence probability distribution, the attribution result of each signal segment in the discriminant probability distribution is called, its value is compared with the upper and lower limits of the set probability threshold, and the data segment whose attribution result is in the threshold interval is recorded to establish a candidate segment index set;

[0102] S402: For each data segment recorded in the candidate segment index set, all signal sampling points in the corresponding time window are extracted, the difference between the maximum value and the minimum value of the signal amplitude in the window is calculated, and the square mean of the amplitudes of all sampling points is calculated and then the square root is taken to obtain a fluctuation characteristic metric value;

[0103] S403: Compare the maximum relative fluctuation amplitude and the fluctuation amplitude screening reference of each segment in the candidate segment index set and the corresponding fluctuation characteristic metric value, and compare the root mean square parameter and the energy screening reference at the same time. The segments whose two parameters exceed the respective references are screened and arranged in chronological order, and a screening fragment sequence is output.

[0104] The acquisition step of the fluctuation characteristic metric value is specifically:

[0105] For each data segment recorded in the candidate segment index set, all signal sampling points in the corresponding time window are obtained, the maximum signal amplitude and the minimum signal amplitude in all signal sampling points are calculated, and the arithmetic mean of the signal amplitudes and the root mean square value of the signal amplitudes of all signal sampling points are calculated;

[0106] According to the preset fluctuation balance coefficient, the difference between the maximum signal amplitude and the minimum signal amplitude, and the root mean square value of the signal amplitudes are calculated by weighted summation to obtain a fluctuation characteristic metric value, and the calculation formula is:

[0107] ;

[0108] Wherein, is the fluctuation characteristic metric value, is the fluctuation balance coefficient, is the maximum signal amplitude, is the minimum signal amplitude, is the root mean square value of the signal amplitudes.

[0109] The obtaining step of the screening fragment sequence further includes:

[0110] The fluctuation feature metric values of all data segments in the candidate segment index set are calculated to form a fluctuation feature set, the statistical distribution of the fluctuation feature set is analyzed, and the numerical mean and numerical standard deviation of the maximum-minimum amplitude difference part and the signal amplitude root mean square part in the fluctuation feature set are calculated respectively;

[0111] The numerical mean of the maximum-minimum amplitude difference part is multiplied by a preset first multiple factor, and the result of the numerical standard deviation of the maximum-minimum amplitude difference part multiplied by a preset second multiple factor is added to determine the fluctuation amplitude screening reference together;

[0112] The numerical mean of the signal amplitude root mean square part is multiplied by a preset third multiple factor, and the result of the numerical standard deviation of the signal amplitude root mean square part multiplied by a preset fourth multiple factor is added to determine the energy screening reference together.

[0113] S401: According to the confidence probability output by S303, and the discriminant probability distribution generated by S301, the attribution result of the signal segment is executed. The probability threshold value judgment. The probability threshold value interval set here is [0.60, 0.90], the setting of this interval aims to identify those data segments with high model discriminant result uncertainty. The upper and lower limits of this interval are determined according to the test analysis of the validation data set. Assuming that among the 10 signal segments, the highest discriminant probability of the 3rd, 7th and 8th segments is 0.88, 0.75 and 0.62 respectively, these values all fall within the [0.60, 0.90] interval. Therefore, the indexes (3, 7, 8) of these three data segments are recorded to establish the candidate segment index set.

[0114] The determination of the upper and lower limits of the above interval is specifically through the test analysis of the validation data set containing 5000 labeled radar signal segments. The statistics show that the segments with discriminant probability less than 0.60 have a misjudgment rate of more than 85%, and the segments with discriminant probability higher than 0.90 have a discriminant accuracy of more than 99%, therefore, [0.60, 0.90] is selected as the threshold interval for identifying data segments with high uncertainty.

[0115] S402: For data segments 3, 7, and 8 recorded in the candidate segment index set, extract all original signal sampling points within their corresponding time windows. For example, segment 3 corresponds to points 201 to 300 in the original sampled data. Extract the original amplitude values ​​(unnormalized values) of these 100 sampling points. Calculate the maximum and minimum signal amplitude values ​​among these 100 points. Assume that for segment 3, the maximum amplitude value is 1.52V, the minimum is 0.41V, and the difference is 1.11V. Simultaneously, calculate the root mean square (RMS) value of the amplitudes of these 100 sampling points by averaging the squared amplitude values ​​of each point and then taking the square root. Assume the calculated RMS value is 0.98V. Combine the maximum and minimum amplitude difference of 1.11V with the RMS value of 0.98V to obtain the fluctuation characteristic measure of this segment.

[0116] The specific steps for obtaining the fluctuation characteristic metric are as follows: For data segment number 3 in the candidate segment index set, obtain the amplitude values ​​of 100 original signal sampling points within its corresponding time window (t=4.0ms to t=6.0ms). Calculate the maximum signal amplitude. =1.52V, minimum signal amplitude =0.41V. Calculate the root mean square value of the signal amplitude. =0.98V. Based on the preset volatility equilibrium coefficient. Perform a weighted summation calculation.

[0117] The formula for calculating the volatility characteristic metric is: This formula is used to calculate a volatility characteristic metric, which is derived from the volatility equilibrium coefficient. Peak-to-peak ripple characteristics of the signal ( ) and the overall energy level of the signal ( ) are weighted and combined. Among them, This is the final calculated fluctuation characteristic metric. and These represent the maximum and minimum signal amplitudes within the time window, respectively, and their difference reflects the dynamic fluctuation range of the signal. This is the root mean square value of the signal amplitude, representing the average energy of the signal within this time window. Volatility equalization coefficient. Used to adjust for amplitude fluctuations and the proportion of energy in the final measurement. The value was determined experimentally. In a test set containing 1000 "drone" signal samples and 1000 "strong clutter" signal samples, the value was determined by traversing... The values ​​range from 0 to 1 (step size 0.05), and the two classes of samples are calculated for each value. The ratio of the inter-class dispersion to the intra-class dispersion of the values. When At this time, the ratio is maximum, indicating that the The ability to distinguish between "drone" micro-motion signals and "strong clutter" signals is the strongest. Therefore, in the embodiment, the "drone" micro-motion signals and the "strong clutter" signals are set as .

[0118] Table 1 gives the fluctuation characteristic parameters of each segment in the candidate segment index set.

[0119] Table 1 candidate segment fluctuation characteristic parameter table

[0120]

[0121] As shown in Table 1, the fluctuation characteristic parameters of the candidate segments are listed. The data of the third segment is substituted into the formula: ;

[0122] The result 1.0515 indicates the comprehensive fluctuation characteristics of the third segment.

[0123] S403: Jointly screen the candidate segment index set and the corresponding fluctuation characteristic metric value. First, the fluctuation amplitude screening reference and the energy screening reference need to be determined. The two references are dynamically calculated according to the statistical distribution of the fluctuation characteristic set of all data segments in the candidate segment index set. Analyze the statistical characteristics of the three candidate segments in Table 1: the maximum minimum amplitude difference part of the data is [1.11, 0.60, 0.75], the numerical average is 0.82V, and the numerical standard deviation is 0.262V. The root mean square of the signal amplitude part of the data is [0.98, 0.85, 0.55], the numerical average is 0.793V, and the numerical standard deviation is 0.221V. Set the first multiple factor to 1.0, the second multiple factor to 0.5, the third multiple factor to 1.0, and the fourth multiple factor to 0.5. The basis for setting these factors is that in historical data testing, using "mean + 0.5 times standard deviation" as the screening threshold can eliminate about 70% of false alarm segments while retaining more than 95% of real target segments. The fluctuation amplitude screening reference = . The energy screening reference = .

[0124] Next, the candidate segments are compared one by one:

[0125] For the third segment:

[0126] The maximum relative fluctuation amplitude (amplitude difference) is 1.11V, which is greater than the fluctuation amplitude screening reference 0.951V; the root mean square parameter is 0.98V, which is greater than the energy screening reference 0.9035V. Both parameters exceed their respective reference values, and the segment passes the screening.

[0127] For the seventh segment:

[0128] The amplitude difference is 0.60V, which is less than the reference 0.951V. The segment is rejected. For the 8th segment: the root mean square parameter is 0.55V, which is less than the reference 0.9035V. The segment is rejected. After screening, only the 3rd segment is retained. Arrange all the screened segments (only the 3rd segment in this example) in chronological order, and output the final screened segment sequence.

[0129] The above embodiments demonstrate the preferred implementation of the present application, and any equivalent adjustment of the technical solutions based on software engineering methods is within the protection scope, including but not limited to: implementing algorithm logic in different programming languages, service reconstruction of functional modules, adjustment of data interaction protocols, optimization of resource scheduling strategies, etc. Any implementation derived through reasonable modification of the data processing flow, service calling link or system architecture level without deviating from the technical core of the present application should be considered within the protection scope defined by the claims of the present application.

Claims

1. A radar micro-motion feature extraction method based on deep learning, characterized in that, Includes the following steps: S1: By collecting continuous sampling data of radar signals, instantaneous amplitude change, amplitude change slope, sampling time interval between extreme points and instantaneous phase change are obtained and combined as a four-element time series feature group. The normalization processing function is called to normalize the four-element time series feature group to generate a normalized time series feature block. S2: Input the normalized temporal feature block into the feature matrix construction process, organize it into a temporal two-dimensional feature matrix, and fuse the instantaneous phase change to generate a fused feature matrix; S3: Input the fused feature matrix into the temporal convolutional network model, perform multi-layer feature extraction, automatically model spatiotemporal correlation, obtain the discrimination probability distribution of each segment at the output end, extract the attribution confidence probability distribution, and perform multi-level discrimination for the discrimination feature distribution, perform dynamic filtering, and output confidence probability and feature parameters. S4: Based on the attribution confidence probability distribution, perform probability threshold judgment on the signal segmentation attribution results according to the discriminant probability distribution, call the maximum relative fluctuation amplitude and root mean square parameter within the window, perform multivariate joint screening for data segments whose discriminant probability distribution is within the threshold interval, and output the screened segment sequence.

2. The radar micro-motion feature extraction method based on deep learning according to claim 1, characterized in that, The normalized time series feature block includes an amplitude normalized sequence, a slope normalized sequence, an extreme value interval normalized sequence, and a phase change normalized sequence. The fused feature matrix is ​​specifically a two-dimensional time series feature matrix that fuses the phase change normalized sequence. The generation step of the attribution confidence probability distribution includes performing dynamic screening to determine dynamic screening feature parameters based on the segmented discrimination probability and the attribution confidence parameter. The screened segment sequence is specifically a signal segment sequence obtained after the multivariate joint screening.

3. The radar micro-motion feature extraction method based on deep learning according to claim 2, characterized in that, The specific steps for obtaining the normalized temporal feature block are as follows: S101: Collects continuous sampling data of radar signals, calculates the instantaneous amplitude difference between adjacent sampling points and the time ratio to obtain the amplitude change slope, counts the number of samples between waveform inflection points to convert the time interval, and analyzes the phase information of the signal to extract the instantaneous phase change. Combines the instantaneous amplitude change, amplitude change slope, sampling time interval between extreme points and instantaneous phase change to establish the original multidimensional feature sequence. S102: Align the four features in the original multidimensional feature sequence—the instantaneous amplitude change, the slope of the amplitude change, the sampling time interval between extreme points, and the instantaneous phase change—at the same time point, and arrange them in a preset order. Treat the four features at each time point as a unit to form a structured data set, and obtain a quaternary time series feature group. S103: Traverse all values ​​of each feature dimension in the four-dimensional time series feature group, determine the upper and lower limits of values ​​for multiple dimensions, and perform scaling calculations on each original value under the dimension based on the difference between the upper and lower limits, and map it to a preset value range to generate a normalized time series feature block.

4. The radar micro-motion feature extraction method based on deep learning according to claim 3, characterized in that, The specific steps for obtaining the fused feature matrix are as follows: S201: Call the normalized time series feature block, set the time step as the basis for data normalization, stack the continuous feature blocks in the order of time into rows, and use the normalized instantaneous amplitude change, amplitude change slope and sampling time interval between extreme points in each feature block as columns to perform structured arrangement and establish a time series two-dimensional feature matrix. S202: Based on the row and column structure of the time-series two-dimensional feature matrix, extract the corresponding instantaneous phase change sequence from the normalized time-series feature block, use the instantaneous phase change sequence as the current feature dimension, and simultaneously perform concatenation calculation with the existing feature vector of each time step in the time-series two-dimensional feature matrix to generate a fused feature matrix.

5. The radar micro-motion feature extraction method based on deep learning according to claim 4, characterized in that, The specific steps for obtaining the confidence probability and feature parameters are as follows: S301: Input the fused feature matrix into a temporal convolutional network, extract features from the spatiotemporal correlation within the matrix through multiple convolutional kernels and activation functions, normalize the output values ​​of the fully connected layer at the end of the network, calculate the class assignment probability of each segment, and obtain the discrimination probability distribution. S302: Based on the discriminant probability distribution, the contribution values ​​of multi-layer features in the reverse network to the output result are traced back, the influence weight of each input feature parameter on the discriminant result is quantified, and the weight is combined with the original probability value in the discriminant probability distribution to generate the attribution confidence probability distribution. S303: For the attribution confidence probability distribution, set multiple progressive discrimination confidence thresholds, compare each probability value in the distribution with the multi-level thresholds one by one, filter out probability values ​​that exceed the discrimination confidence thresholds, extract the original feature parameters corresponding to the probability values, and output the confidence probability and feature parameters.

6. The radar micro-motion feature extraction method based on deep learning according to claim 5, characterized in that, The specific steps for obtaining the selected fragment sequence are as follows: S401: Based on the attribution confidence probability distribution, call the attribution result of each signal segment in the discrimination probability distribution, compare its value with the upper and lower limits of the set probability threshold, and record the data segments whose attribution results are within the threshold range to establish a candidate segment index set; S402: For each data segment recorded in the candidate segment index set, extract all signal sampling points within the corresponding time window, calculate the difference between the maximum and minimum values ​​of the signal amplitude within the window, and take the square root of the square mean of the amplitudes of all sampling points to obtain the fluctuation characteristic measurement value. S403: Compare the maximum relative fluctuation amplitude of each segment in the candidate segment index set with the fluctuation amplitude screening benchmark, and compare the root mean square parameter with the energy screening benchmark. Screen segments whose two parameters exceed their respective benchmarks, and arrange them in chronological order to output the screening segment sequence.

7. The radar micro-motion feature extraction method based on deep learning according to claim 5, characterized in that, The specific steps for obtaining the discrimination probability distribution are as follows: Obtain the unnormalized logical value output by the fully connected layer in the temporal convolutional network model for any signal segment, and calculate the unnormalized logical value using the Softmax function based on the preset total number of micro-motion target categories. Use the value of each dimension in the calculation result as the probability of belonging to the corresponding category. The formula for calculating the category belonging probability is as follows: ； in, The current signal segment is judged as the first. The probability of class classification. The corresponding output of the fully connected layer Unnormalized logical values ​​of a class The corresponding output of the fully connected layer Unnormalized logical values ​​of a class The total number of preset micro-motion target categories; By combining the category attribution probabilities of all micro-movement target categories, a discrimination probability distribution is established.

8. The radar micro-motion feature extraction method based on deep learning according to claim 7, characterized in that, The specific steps for obtaining the attribution confidence probability distribution are as follows: The discriminative probability distribution and the influence weights of multiple input feature parameters calculated through the backtracking network are called. For the category belonging probability of each category in the discriminative probability distribution, it is multiplied one by one by the influence weights of all input feature parameters corresponding to the category belonging probability to obtain the weighted contribution of multiple input feature parameters. The weighted contributions of all input feature parameters related to the probability of belonging to the same category are summed, and the sum is used as the attribution confidence score of the corresponding category. The attribution confidence scores of all micro-motion target categories are normalized so that the sum of the attribution confidence scores of all categories is 1, thereby generating an attribution confidence probability distribution.

9. The radar micro-motion feature extraction method based on deep learning according to claim 6, characterized in that, The specific steps for obtaining the fluctuation characteristic metric are as follows: For each data segment recorded in the candidate segment index set, obtain all signal sampling points within its corresponding time window, calculate the maximum and minimum signal amplitudes among all signal sampling points, and calculate the arithmetic mean and root mean square value of the signal amplitudes among all signal sampling points. Based on a preset volatility equilibrium coefficient, a weighted summation is performed on the difference between the maximum and minimum signal amplitudes, and the root mean square value of the signal amplitude, to obtain a volatility characteristic metric. The calculation formula is as follows: ； in, This is a measure of fluctuation characteristics. The volatility equilibrium coefficient, For the maximum signal amplitude, Minimum signal amplitude, This is the root mean square value of the signal amplitude.

10. The radar micro-motion feature extraction method based on deep learning according to claim 9, characterized in that, The step of obtaining the filtered fragment sequence further includes: Calculate the fluctuation characteristic measurement values ​​of all data segments in the candidate segment index set to form a fluctuation characteristic set, analyze the statistical distribution of the fluctuation characteristic set, and calculate the mean and standard deviation of the maximum and minimum amplitude difference part and the root mean square part of the signal amplitude in the fluctuation characteristic set respectively. The mean of the maximum and minimum amplitude difference is multiplied by a preset first multiple factor, and the result of multiplying the standard deviation of the maximum and minimum amplitude difference by a preset second multiple factor is added to determine the fluctuation amplitude screening benchmark. The average value of the root mean square portion of the signal amplitude is multiplied by a preset third multiplier factor, and the result of multiplying the standard deviation of the root mean square portion of the signal amplitude by a preset fourth multiplier factor is added to determine the energy screening benchmark.

Images