A sleeping person recognition method based on a multi-time-scale breathing fluctuation model

Abstract

The present invention discloses a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model, which belongs to the technical field of personnel identification; the steps are as follows: Step 1: Setting a millimeter-wave radar as an acquisition device, using the millimeter-wave radar to collect signals inside the room to obtain a sampled signal; Step 2: Preprocessing the sampled signal to perform body motion detection and judgment of the resting stable state; Step 3: Extracting multi-dimensional features of the short-term respiratory fluctuation model from the resting stable state signal, specifically including chest and abdomen positioning, respiratory information extraction and correction, dynamic segmentation features, and chest and abdomen fluctuation model feature extraction; Step 4: Establishing a long-term respiratory fluctuation feature sequence and recognition network, and providing preliminary results for the person to be identified based on the stored personnel information; Step 5: Making a decision again on the preliminary result to obtain the final result. The present invention adopts the above method to achieve continuous monitoring in sleeping scenes, which can work all day and all weather.

Description

Technical Field

[0001] The present invention relates to the technical field of personnel information recognition, and in particular to a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model. Background Art

[0002] Traditional identity recognition methods have numerous limitations regarding privacy and continuous monitoring in home sleep environments. Visual cameras offer an intuitive recognition method, directly capturing a person's face or other features. However, these methods present significant privacy concerns and are highly dependent on ambient lighting conditions. Performance degrades significantly at night or in low-light environments, and recognition accuracy is significantly reduced when the face or other key areas are obscured. While thermal imaging cameras can operate in complete darkness and reduce the risk of privacy breaches by producing heat maps rather than clear facial images, this technology is costly, typically has low resolution, and is affected by temperature fluctuations. Biometric sensors (such as fingerprints and iris scans) are renowned for their high accuracy. The uniqueness of biometric features ensures very high recognition accuracy and fast response times. However, many biometric recognition technologies require direct contact between the user and the device, which can raise hygiene concerns. Skin damage or eye diseases can also affect recognition performance. Voiceprint recognition, with its convenience and natural interaction, allows authentication through voice without the need for additional hardware. However, background noise can significantly impact recognition accuracy, and there is a risk of voice imitation. Changes in health status, such as a cold or other condition that alters the voice, can cause recognition failure. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model, which can overcome the limitations of traditional technologies and use millimeter-wave radar as an identification tool, and is particularly suitable for home sleeping environments. Its non-contact monitoring characteristics ensure that users can continuously monitor their vital signs without being disturbed, while avoiding the risk of privacy leakage. The technology can work stably under various lighting conditions, and can penetrate thin obstacles, effectively capture subtle movements such as breathing and heartbeats, and has strong environmental adaptability. The frequency band used by millimeter-wave radar is less susceptible to interference from other radio signals, ensuring the stability and reliability of the data. Dynamic adaptability enables millimeter-wave radar to maintain data consistency when users frequently change postures.

[0004] To achieve the above objectives, the present invention provides a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model, comprising the following steps:

[0005] Step 1: Set up a millimeter-wave radar as a collection device, use the millimeter-wave radar to collect signals inside the room, obtain an echo signal, and set the echo signal as the sampling signal;

[0006] Step 2: Preprocess the sampled signal to detect body motion and determine the resting state;

[0007] Step 3: Perform multi-dimensional feature extraction of the short-term respiratory fluctuation model on the resting steady state signal, including chest and abdomen positioning, respiratory information extraction and correction, dynamic segmentation features, and chest and abdomen fluctuation model feature extraction;

[0008] Step 4: Establish a long-term respiratory fluctuation feature sequence and recognition network, and provide preliminary results for the person to be identified based on the stored personnel information;

[0009] Step 5: Make another decision based on the preliminary results to get the final result.

[0010] Preferably, the specific process in step 2 is as follows:

[0011] Step 21: Preprocess the sampled signal in step 1 to obtain a discrete echo signal; the target objects in the discrete echo signal include users and interfering devices; the discrete echo signal includes discrete signals at multiple sampling moments; based on the discrete signals at the sampling moments in each sliding window, determine the data to be used corresponding to each sliding window;

[0012] Step 22: For each different sliding window, based on the to-be-used data corresponding to the current sliding window and the sliding window before the current sliding window, determine the body movement index corresponding to the monitoring moment in the current sliding window; and perform gross body movement detection;

[0013] Step 23: Accumulate the body movement state in the long time scale. If there is no large body movement in the time period, then the time period is judged to be a resting and stable state.

[0014] Preferably, a specific implementation of step 2 is as follows:

[0015] Step 21: Sample the signal Perform preprocessing, represents the millimeter wave antenna dimension index, represents the sample index, Represents the slow time index; perform fast Fourier transform on it to obtain the frequency domain signal; then use the moving average filter to suppress the clutter of the frequency domain signal to reduce background noise and unnecessary interference, and the filtered signal is generated by coherent accumulation matrix ,in, Represents the distance unit index;

[0016] Step 22: Right Calculate the body movement index per second using the double integration method , based on the obtained body movement index , perform gross body movement detection to determine whether there is gross body movement in the current state. The gross body movement judgment formula is as follows:

[0017] ;

[0018] For all times within the analysis range ,in is a sensitivity parameter, Indicates the average of the previous period When the time without body movement meets the set value, it is judged as a resting and stable state;

[0019] Step 23: If there is no gross body movement during the time period, the time period is determined to be a resting and stable state, and the information in the resting and stable state is further analyzed.

[0020] Preferably, the specific process of step 3 is as follows:

[0021] Step 31: Extract effective respiratory features from all short-time scale information in the resting steady state signal, and identify, locate, and correct the chest and abdomen area;

[0022] Step 32: Based on chest and abdomen positioning, extract phase information for all distance units within the chest and abdomen. All initially extracted respiratory phase information is corrected based on angle and distance position, converting the radial respiratory phase signal received by the radar into the actual vertical respiratory phase signal of the human body.

[0023] Step 33: A dynamic segmentation algorithm is used to extract dynamic segmentation features that reflect the geometric information of the respiratory signal. Finally, a dynamic segmentation feature sequence at the center of the chest and abdomen is selected as the current short-term respiratory signal sampling feature.

[0024] Step 34: An abdomen-back-chest fluctuation model is constructed and feature extraction is performed. The depth of inspiration and expiration of each distance unit in the chest and abdomen range is supplemented with the description of the current sleeping posture. Dynamic segmentation features and abdomen-back-chest fluctuation model features are used to form a multidimensional feature of the short-term respiratory fluctuation model to characterize the short-term posture information during sleep and its corresponding representative respiratory information.

[0025] Preferably, a specific implementation process of step 3 is as follows:

[0026] Step 31: For the resting state, The resting steady state signals filtered out ,right Extract short-term effective breathing features, and then identify and locate the chest and abdominal cavity positions, targeting signals in the resting state , using the minimum variance distortion-free response algorithm to calculate the distance angle spectrum Detect the target point by using the ordered statistical constant false alarm rate method to generate a point cloud spectrum ; Point cloud spectrum Perform analysis to calculate density spectra ;

[0027] Potential target objects are identified by determining the maximum peak in the density spectrum. If no peak is found, it means that there is no target. The potential target object is detected at the location of Perform a row-by-row search for the center to determine the upper, lower, left, and right boundaries. When the target object represents a human cluster, process the distance angle spectrum. To refine the position estimate, the process is as follows:

[0028] The first detected cluster The position is set as the initial reference point , for each subsequent cluster , calculate and current reference point When the deviation exceeds the predefined threshold, Set to new cluster location ,When the deviation does not exceed the predefined threshold, the original reference point is retained and the chest and abdomen range is determined according to the cluster position , complete the positioning of the chest and abdomen,

[0029] Step 32: Extract breathing information according to Select Respiratory phase information within the distance unit, in steps Seconds, window length Seconds use differential adaptive Kalman filter algorithm to extract this The distance unit represents the respiratory phase information;

[0030] Step 33: Dynamic segmentation feature extraction is performed on the respiratory information. To further extract dynamic segmentation features, the rise and fall of the transition period from 30% to 70% during inspiration and expiration are considered, and then the displacement and time area ratio of the attack are calculated;

[0031] According to the geometric characteristics of the respiratory waveforms in both exhalation and inspiration states, the inspiratory area, expiratory area, inhalation / expiration speed, inhalation / expiration depth, as well as the respiratory area ratio, respiratory speed ratio, respiratory depth ratio and respiratory rate, average exhalation / inhalation start-up period, average respiratory start-up period ratio, complexity of the signal before and after full lung capacity, packaging density, and linear envelope error were extracted as features to optimize the dynamic segmentation algorithm and obtain the dynamic segmentation feature sequence. All extracted dynamic segmentation sequence information features are corrected based on angle and distance position, and the geometric features of the respiratory signal received by the radar in the radial direction are converted into the geometric information of the human respiratory signal in the actual vertical direction. For the breathing area and breathing speed, a calibration formula including the radar placement height, radar placement angle, signal distance unit position, and signal angle unit position is used. In order to calibrate the breathing area and breathing speed more accurately, the distance unit position of the signal must also be considered. and angular unit position , the specific formula is as follows:

[0032] ;

[0033] in: is the range unit position of the signal, is the altitude of the radar, is the angular position of the target relative to the radar, is the radar installation angle, and the actual dynamic segmentation feature sequence after calibration is obtained choose The dynamic segmentation feature sequence is used as the representative respiratory signal sampling feature of the current window ;

[0034] Step 34: Construct a chest and abdomen undulation model by The dynamic segmentation feature analysis of the respiratory waveform of each distance unit is performed to determine the exhalation depth of different distance units within the entire chest and abdomen at that moment, providing a simple description of the current sleeping posture, as follows:

[0035] ;

[0036] Representatives in A sequence of inspiratory depths over distance units, where Corresponding to the reference point The inhalation depth of each distance unit is Derived by averaging the depths obtained from all respiratory waveforms within a single short time window after applying a dynamic segmentation algorithm;

[0037] A short chest and abdomen rise and fall sequence and the dynamic segmentation feature sequence of the current state sampling position , characterizes the short-term sleeping posture information and the corresponding representative breathing information, and obtains the multidimensional features of the short-term breathing fluctuation model in a long period of time .

[0038] Preferably, the process of step 4 is as follows:

[0039] Step 41: constructing a time series feature sequence, arranging the multidimensional features of the short-term respiratory fluctuation model previously obtained under long-term conditions in chronological order to construct a multidimensional feature time series feature matrix of the long-term respiratory fluctuation model;

[0040] Step 42: training a neural network to identify a multi-dimensional feature time series feature matrix of a long-term respiratory fluctuation model;

[0041] Step 43: Use a neural network to process the multi-dimensional feature time series feature matrix of the long-term respiratory fluctuation model and output a preliminary result. The preliminary result indicates the degree of matching between the currently detected data content and each monitored object.

[0042] Preferably, an implementation method of step 4 is as follows:

[0043] Step 41: First, construct a time series feature sequence based on the multidimensional features of the previous short-term breathing fluctuation model , construct a multi-dimensional feature time series feature matrix of the long-term respiratory fluctuation model; integrate the short-term features in multiple time periods to form a continuous time series data set; for each monitored object, extract a series of short-term respiratory fluctuation features from its long-term monitoring data, including inhalation depth, exhalation depth, respiratory rate and respiratory cycle; arrange the above features in chronological order to construct a multi-dimensional time series feature matrix , the construction formula is as follows:

[0044] ;

[0045] here, Indicates at a point in time Next The eigenvalues ​​of the dimensions, is the dimension of the eigenvector, is the number of time slices;

[0046] Step 42: Use the long short-term memory network LSTM to process the multi-dimensional time series feature matrix To train the LSTM network, the dataset must be divided into a training set and a test set. The training set contains a large number of time-series feature matrix samples of long-term breathing fluctuations, covering breathing patterns in various sleeping positions. By learning from the training set, the LSTM network can automatically adjust its internal parameters to minimize prediction errors and gradually improve its ability to identify individual identities. The test set is used to verify the model's generalization ability and actual performance, ensuring that the model can maintain high recognition accuracy on unseen data. During training, the LSTM network calculates outputs through forward propagation and uses the backpropagation algorithm to update weights. This process is iterated continuously until the model converges or reaches the preset stopping condition.

[0047] Step 43: Use the LSTM model trained in step 42, which is based on the input long-term respiratory fluctuation feature time series feature matrix Generate preliminary results.

[0048] Preferably, in step 5, the specific process is as follows:

[0049] Step 51: Conduct a comprehensive analysis of the preliminary results of step 43 and use statistical methods to process them:

[0050] Step 52: After processing in step 51, the final result is obtained.

[0051] Preferably, a specific implementation method of step 5 is as follows:

[0052] Step 51: For the preliminary results, a secondary judgment mechanism based on all long-term preliminary results within the extended period is used. The preliminary results of multiple time periods are comprehensively analyzed by majority voting to finally determine the person identification result within the extended period. The execution process is as follows:

[0053] Statistical frequency: For each preliminary result of a time period, record the corresponding candidate identity and count the number of times each identity appears in the entire ultra-long time period;

[0054] Determine the majority identity: Based on the statistical results, the identity with the most occurrences is selected as the final identity recognition result. If multiple identities have the same number of votes, further judgment is made based on pre-set rules;

[0055] Step 52: The result output by the majority voting method is used as the final result.

[0056] Therefore, the present invention adopts the above-mentioned method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model, which has the following advantages:

[0057] (1) In the present invention, millimeter wave radar is used as a detection tool, which has the advantages of being non-contact and non-interference; not infringing on personal privacy; being insensitive to light, dust, smoke and temperature; and being able to detect objects around the clock.

[0058] (2) In the present invention, we mainly focus on continuous person identification in sleeping posture scenarios to meet the personalized needs of long-term continuous sleep monitoring, and consider the multi-time scale life biological information extraction under the sleeping posture of the person, which is conducive to improving the recognition accuracy.

[0059] (3) In the present invention, it is of great significance for solving application scenarios such as sleep stages of multiple people and recognition of vital signs such as breathing and heart rate.

[0060] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model according to the present invention;

[0062] Figure 2 Schematic diagram of dynamic segmentation in an embodiment of a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model of the present invention;

[0063] Figure 3 Schematic diagram of the abdomen-back-chest fluctuation model constructed in an embodiment of a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. The specific model specifications need to be selected and determined based on the actual specifications of the device, etc. The specific selection calculation method adopts the existing technology in this field, so it will not be described in detail.

[0065] Example

[0066] like Figure 1 As shown, the present invention provides a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model, comprising the following steps:

[0067] Step 1: Set up a millimeter-wave radar as a collection device. In this embodiment, a 2-transmitter 4-receiver millimeter-wave radar is used to collect signals inside the room, obtain echo signals, and set the echo signals as sampling signals. The specific formula is as follows:

[0068] ;

[0069] In the above formula, represents the intermediate frequency signal, represents the noise signal, represents the millimeter wave antenna dimension index, represents the sample index of the fast time dimension, represents the slow time index;

[0070] Step 2: Preprocess the signal collected in step 1 to detect body movement and determine the resting state;

[0071] Step 21: Preprocessing is as follows: First, the sampled signal Perform a Fast Fourier Transform along the fast time dimension, which is the sample index , and obtain the frequency domain signal; then use the moving average filter to suppress the clutter of the frequency domain signal, thereby reducing background noise and unnecessary interference, and the filtered signal is generated by coherent accumulation matrix ,in represents the millimeter wave antenna dimension index, Represents the distance unit index, represents the slow time index;

[0072] Step 22: Right Calculate the body movement index per second using the double integration method , based on the obtained body movement index , perform gross body movement detection to determine whether there is gross body movement in the current state. The gross body movement judgment formula is as follows:

[0073] ;

[0074] For all times within the analysis range ,in is a sensitivity parameter, Indicates the average of the previous period ;

[0075] Step 23: If there is no gross body movement during the time period, the time period is determined to be a resting and stable state.

[0076] Step 3: Perform multi-dimensional feature extraction of the short-term respiratory fluctuation model, including chest and abdomen positioning, respiratory information extraction and correction, dynamic segmentation features, and chest and abdomen fluctuation model feature extraction. The specific process is as follows:

[0077] Step 31: For the resting state in step 23, The resting steady state signals filtered out ,right Extract short-term effective breathing features, and then identify and locate the chest and abdominal cavity positions, targeting signals in the resting state , using the minimum variance distortion-free response algorithm to calculate the distance angle spectrum , and detect the target point by the ordered statistical constant false alarm rate method to generate a point cloud , for point cloud spectrum Perform analysis to calculate density spectra , by determining the maximum peak in the density spectrum to identify potential target objects, if no peak is found, it means that there is no target object; when the maximum peak The potential target object is detected at the location of Perform row-by-row search for the center to determine the upper, lower, left, and right boundaries of the cluster. When the target object represents a human cluster, process the distance angle spectrum. to refine the position estimate.

[0078] For the detected clusters of people, the highest response value is determined and a correction formula is applied to further stabilize these position estimates by comparing with previously established reference points. The first detected cluster is The position is set as the initial reference point For each subsequent cluster , calculate and current reference point The deviation value of ; The formula for the deviation value is as follows:

[0079] ;

[0080] In the above formula, If the deviation exceeds the predefined threshold, Set to new cluster location Otherwise, keep the original reference point and determine the chest and abdomen range according to the cluster position , complete the positioning of the chest and abdomen;

[0081] Step 32: Based on the chest and abdomen position positioning, the respiratory phase information is extracted. Select Each respiratory phase information within the distance unit , the specific calculation process is as follows: with a step size Seconds, window length Seconds use differential adaptive Kalman filter algorithm to extract this The distance unit represents the respiratory phase information.

[0082] All the initially extracted respiratory phase information is corrected based on angle and distance position, and the respiratory phase signal in the radial direction received by the radar is converted into the human respiratory phase signal in the actual vertical direction. , the calibration formula is as follows:

[0083] ;

[0084] in: is the range unit position of the signal, is the altitude of the radar, is the angular position of the target relative to the radar, is the radar installation angle, Indicates the human respiratory phase signal in the actual vertical direction after calibration;

[0085] Step 33: Dynamic segmentation feature extraction is performed on the respiratory information, such as Figure 2 As shown, dynamic segmentation feature extraction is performed on respiratory information. The respiratory phase information of each distance unit is obtained. In order to further extract dynamic segmentation features, the respiratory signal feature extraction method in related sitting scenes is referred to, including traditional dynamic segmentation technology of respiratory signals and other geometric information extracted in other studies. The dynamic segmentation technology considers the rise and fall of the transition period from 30% to 70% during inspiration and expiration, and then calculates the displacement and time area ratio of the attack. Based on the geometric shape characteristics of the respiratory waveform in both exhalation and inspiration states, the inspiratory area is extracted. , expiratory area , suction speed , exhalation speed , Inhalation Depth , Exhalation Depth , and on this basis further proposed the respiratory area ratio , respiratory rate ratio , breathing depth ratio At the same time, the respiratory rate is also extracted , average exhalation start period , average suction start cycle , average breathing start cycle ratio etc. to optimize the characteristics of the dynamic segmentation algorithm. Get the dynamic segmentation feature sequence ,This type of feature has shown high effectiveness in identity recognition problems in sitting scenes by other researchers. The formula is as follows:

[0086] ;

[0087] Among them, for a complete breathing cycle including exhalation and inhalation, it can be seen that there is an exhalation start point and an inhalation start point, so it is assumed that there is a short time scale window A breathing cycle, set the exhalation start point and the inhalation start point They are expressed as follows:

[0088] ;

[0089] ;

[0090] get:

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] In the above formula, Indicates the corresponding exhalation start point time, Indicates the corresponding exhalation start point time;

[0104] Considering the positional sensitivity of breathing signals in sleeping postures, all initially extracted breathing geometry features were corrected based on angle and distance. This converts the radial geometry of the breathing signal received by the radar into the actual vertical geometry of the human body's breathing signal. For both the breathing area and breathing rate, a calibration formula was used that incorporates radar placement height, radar placement angle, signal range unit position, and signal angle unit position.

[0105] In order to calibrate the respiratory area and respiratory rate more accurately, the distance unit position of the signal should also be considered. and angular unit position , the specific formula is as follows:

[0106] ;

[0107] in: is the range unit location of the signal. is the altitude of the radar. is the angular position of the target relative to the radar. is the radar installation angle. Get the actual dynamic segmentation feature sequence after calibration Final choice The dynamic segmentation feature sequence is used as the representative respiratory signal sampling feature of the current window .

[0108] like Figure 3 As shown in the figure, an abdomen-back-chest fluctuation model is established because the respiratory signal not only has significant differences between individuals, but also shows high sensitivity to the person's sleeping posture and chest and abdominal occlusion. Specifically, different sleeping postures and chest and abdominal occlusions will lead to complex changes in the characteristics of the respiratory signal, so that different individuals may show similar dynamic segmentation features, although these features may originate from different sleeping postures. Therefore, relying solely on dynamic segmentation features for identity recognition is difficult to meet the accuracy requirements in complex sleeping scenes. However, given that human sleeping habits are relatively stable, for a specific target individual, their long-term sleeping posture type and state are relatively fixed. Based on this feature, supplementing the characterization of the individual's sleeping posture information before extracting the short-term respiratory signal can significantly enhance the effectiveness and reliability of the respiratory signal.

[0109] To achieve this, a method was proposed that combines short-term sleeping posture information with short-term respiratory signal characteristics under long-term conditions. Through long-term analysis, using multi-range unit respiratory waveform data received by radar, information reflecting the fluctuation characteristics of the entire chest and abdomen is extracted. Human respiratory movement affects not only the chest and abdomen, but also the shoulder position. Due to individual differences in breathing habits and body structure, the fluctuation amplitude, or respiratory depth, of each area also varies. In the radar received signal, this difference manifests as a distinct difference in the respiratory depth of each range unit respiratory cycle waveform. For example, in the supine position, the entire chest and abdomen form a roughly double-arc structure around the center of the thoracic cavity and the center of the abdominal cavity. The specific depth depends on the breathing method. This structure also exhibits significant morphological changes in different sleeping postures and occlusion conditions.

[0110] Step 34: Construct a chest and abdomen undulation model by The dynamic segmentation feature analysis of the respiratory waveform of each distance unit is performed to determine the exhalation depth of different distance units within the entire chest and abdomen at that moment, providing a simple description of the current sleeping posture, as follows:

[0111] ;

[0112] Representatives in A sequence of inspiratory depths over distance units, where Corresponding to the first The inhalation depth at each distance unit.

[0113] For each distance unit, the inhalation depth It is determined by averaging the depth values ​​obtained from all respiratory waveforms within a single short time window after applying a dynamic segmentation algorithm. The specific calculation formula is as follows:

[0114] ;

[0115] here, Represents the number of respiratory waveforms within a short time window; Indicates that within the same short time window The first respiratory waveform is obtained by applying the dynamic segmentation algorithm In order to characterize the short-term posture information during sleep and its corresponding representative breathing information, two feature sequences are used: the short-term chest and abdominal fluctuation feature sequence at the current sampling position and dynamic segmentation feature sequence These features are then used to form a short-term respiratory fluctuation model The multidimensional features of are as follows:

[0116] ;

[0117] Step 4: Establish a long-term respiratory fluctuation feature sequence and recognition network to provide preliminary results for personal identification. Because in the long-term continuous monitoring of complex sleep scenes, compared with the direct recognition of short-term features, the advantage of long-term sequence feature accumulation is utilized, and a long short-term memory network (LSTM) is used for processing and analysis. This method is particularly suitable for capturing subtle changes in respiratory signals and complex patterns that evolve over time, thereby providing more accurate identity identification and health status assessment. The specific process is as follows:

[0118] Step 41: Construct a temporal feature sequence based on the multidimensional features of the short-term breathing fluctuation model previously obtained under long-term conditions ,constructing a multi-dimensional time series feature matrix for the long-term respiratory fluctuation model.,This process involves integrating short-term features from multiple time periods to form a continuous time series dataset.

[0119] For each monitored subject, a series of short-term respiratory fluctuation features are extracted from their long-term monitoring data, including inhalation depth, exhalation depth, respiratory rate, and respiratory cycle. Then, these features are arranged in chronological order to construct a multi-dimensional time series feature matrix. The matrix not only contains the eigenvalues ​​of different time periods, but also preserves the temporal relationship between them. Specifically, for a series of time segments on a long time scale , where each time segment corresponds to a specific short-term respiratory fluctuation cycle, and its corresponding multidimensional feature vector set can be expressed as Next, we arrange these feature vectors in chronological order and organize them into a two-dimensional matrix , where each row represents a time segment The eigenvector of , and the entire matrix reflects all the feature information from the beginning to the end of the time period. The two-dimensional time series feature matrix The formula is as follows:

[0120] ;

[0121] here, Indicates at a point in time Next The eigenvalues ​​of the dimensions, is the dimension of the eigenvector, is the number of time slices.

[0122] In this way, we can effectively capture the changing trends of an individual's breathing patterns in different sleeping positions, providing rich input data for subsequent machine learning model training. In addition, this time series feature matrix can reflect an individual's long-term breathing habits and their changing patterns. For each monitored subject, a series of short-term breathing fluctuation features, including inhalation depth, exhalation depth, respiratory rate, and respiratory cycle, are extracted from their long-term monitoring data; these features are arranged in chronological order to construct a multi-dimensional time series feature matrix. ;

[0123] Step 42: Train the neural network to identify the multi-dimensional feature time series feature matrix of the long-term respiratory fluctuation model; first, the data set needs to be divided into a training set and a test set. The training set contains a large number of long-term respiratory fluctuation feature time series feature matrix samples, covering breathing patterns in various sleeping positions; through learning the training set, the LSTM network can automatically adjust internal parameters to minimize prediction errors and gradually improve the ability to identify individual identities. The test set is used to verify the generalization ability and actual performance of the model to ensure that the model can maintain a high recognition accuracy on unseen data. During the training process, the LSTM network calculates the output through forward propagation and uses the backpropagation algorithm to update the weights. This process is iterated continuously until the model converges or reaches the preset stopping condition. Finally, a fully trained LSTM model is obtained.

[0124] Step 43: Use the long short-term memory network LSTM to process the long-term breathing fluctuation feature time series feature matrix , based on the input long-term respiratory fluctuation feature time series feature matrix Generate preliminary results;

[0125] Step 5: Combine the preliminary results to make a decision on personnel identification and obtain the final result. The specific process is as follows:

[0126] Step 51: To further improve the accuracy and reliability of identity recognition, a secondary judgment mechanism based on all preliminary results within an extended period of time is adopted. The extended period of time here is set to more than 30 minutes. This mechanism uses majority voting to comprehensively analyze the preliminary results of multiple time periods and ultimately determines the person's identity recognition result within the extended period of time.

[0127] After completing the long-term rough judgment, a majority voting method is used to comprehensively analyze all the long-term preliminary results. The majority voting method is a simple and effective ensemble learning method. Its core idea is to count the occurrence frequency of each candidate identity and select the identity with the most votes as the final recognition result. The specific implementation steps are as follows:

[0128] Statistical frequency: For each preliminary result of a time period, record the corresponding candidate identity and count the number of times each identity appears in the entire extended time period. The extended time period here is set to be more than 30 minutes.

[0129] Determine the majority identity: Based on the statistical results, the identity with the most occurrences is selected as the final identity recognition result. If multiple identities have the same number of votes, further judgment is made based on pre-set rules (such as giving priority to the identity with the most common historical records).

[0130] Step 52: Following step 51, the final person identification result for the extended time period is obtained. This result is based on the preliminary results from all time periods and is integrated through majority voting, resulting in higher accuracy and reliability. Furthermore, the secondary decision mechanism effectively reduces the potential for misjudgments within a single time period, improving the robustness and stability of the entire system in complex sleep scenarios. The result output by the majority voting method is used as the final result.

[0131] Therefore, the present invention adopts a method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model, which can overcome the limitations of traditional technologies and use millimeter-wave radar as an identification tool, which is particularly suitable for home sleeping environments. Its non-contact monitoring characteristics ensure that users can continuously monitor their vital signs without being disturbed, while avoiding the risk of privacy leakage. This technology can work stably under various lighting conditions, and can penetrate thin obstacles, effectively capture subtle movements such as breathing and heartbeats, and has strong environmental adaptability. The frequency band used by millimeter-wave radar is less susceptible to interference from other radio signals, ensuring the stability and reliability of the data. Dynamic adaptability enables millimeter-wave radar to maintain data consistency when users frequently change postures.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model, characterized by: The following steps are involved: Step 1: Set up a millimeter-wave radar as a collection device, use the millimeter-wave radar to collect signals inside the room, obtain an echo signal, and set the echo signal as the sampling signal; Step 2: Preprocess the sampled signal to detect body motion and determine the resting state. The specific process is as follows: Step 21: Preprocess the sampled signal in step 1 to obtain a discrete echo signal; the target objects in the discrete echo signal include users and interfering devices; the discrete echo signal includes discrete signals at multiple sampling moments; based on the discrete signals at the sampling moments in each sliding window, determine the data to be used corresponding to each sliding window; Step 22: For each different sliding window, based on the to-be-used data corresponding to the current sliding window and the sliding window before the current sliding window, determine the body movement index corresponding to the monitoring moment in the current sliding window; and conduct gross body movement detection; Step 23: Accumulate the body movement state in the long time scale. If there is no large body movement in the time period, then the time period is judged to be a resting and stable state; Step 3: Perform multi-dimensional feature extraction of the short-term respiratory fluctuation model on the resting steady state signal, including chest and abdomen positioning, respiratory information extraction and correction, dynamic segmentation features, and abdomen-back-chest fluctuation model feature extraction. The specific process is as follows: Step 31: Extract effective respiratory features from all short-time scale information in the resting steady state signal, and identify, locate, and correct the chest and abdomen area; Step 32: Based on chest and abdomen positioning, extract phase information for all distance units within the chest and abdomen. All initially extracted respiratory phase information is corrected based on angle and distance position, converting the radial respiratory phase signal received by the radar into the actual vertical respiratory phase signal of the human body. Step 33: A dynamic segmentation algorithm is used to extract dynamic segmentation features that reflect the geometric information of the respiratory signal. Finally, a dynamic segmentation feature sequence at the center of the chest and abdomen is selected as the current short-term respiratory signal sampling feature. Step 34: Construct an abdomen-back-chest fluctuation model and perform feature extraction. The depth of inspiration and expiration for each distance unit in the chest and abdomen range supplements the description of the current sleeping posture. Dynamic segmentation features and abdomen-back-chest fluctuation model features are used to construct a multidimensional feature of the short-term respiratory fluctuation model, representing the short-term posture information during sleep and its corresponding representative respiratory information. Step 4: Establish a long-term respiratory fluctuation feature sequence and recognition network, and provide preliminary results for the person to be identified based on the stored personnel information; Step 5: Make another decision based on the preliminary results to get the final result.

2. The method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model according to claim 1, characterized in that: A specific implementation of step 2 is as follows: Step 21: Sample the signal Perform preprocessing, represents the millimeter wave antenna dimension index, represents the sample index, represents the slow time index; Perform fast Fourier transform on it to obtain frequency domain signal; then use moving average filter to suppress clutter on frequency domain signal to reduce background noise and unnecessary interference, and the filtered signal generates matrix through coherent accumulation. ,in, Represents the distance unit index; Step 22: Right Calculate the body movement index per second using the double integration method , based on the obtained body movement index , perform gross body movement detection to determine whether there is gross body movement in the current state. The gross body movement judgment formula is as follows: ； For all times within the analysis range ,in is a sensitivity parameter, Indicates the average of the previous period When the time without body movement meets the set value, it is judged as a resting and stable state; Step 23: If there is no gross body movement during the time period, the time period is determined to be a resting and stable state, and the information in the resting and stable state is further analyzed.

3. The method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model according to claim 1, characterized in that: The process of step 4 is as follows: Step 41: constructing a time series feature sequence, arranging the multidimensional features of the short-term respiratory fluctuation model previously obtained under long-term conditions in chronological order to construct a multidimensional feature time series feature matrix of the long-term respiratory fluctuation model; Step 42: training a neural network to identify a multi-dimensional feature time series feature matrix of a long-term respiratory fluctuation model; Step 43: Use a neural network to process the multi-dimensional feature time series feature matrix of the long-term respiratory fluctuation model and output a preliminary result. The preliminary result indicates the degree of matching between the currently detected data content and each monitored object.

4. The method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model according to claim 1, characterized in that: In step 5, the specific process is as follows: Step 51: Conduct a comprehensive analysis of the preliminary results of step 4 and use statistical methods to process them: Step 52: After processing in step 51, the final result is obtained.

5. The method for identifying people in sleeping positions based on a multi-time-scale respiratory fluctuation model according to claim 4, characterized in that: A specific implementation method of step 5 is as follows: Step 51: For the preliminary results, a secondary judgment mechanism based on all long-term preliminary results within the extended period is used. The preliminary results of multiple time periods are comprehensively analyzed by majority voting to finally determine the person identification result within the extended period. The execution process is as follows: Statistical frequency: For each preliminary result of a time period, record the corresponding candidate identity and count the number of times each identity appears in the entire ultra-long time period; Determine the majority identity: Based on the statistical results, the identity with the most occurrences is selected as the final identity recognition result. If multiple identities have the same number of votes, further judgment is made based on pre-set rules; Step 52: The result output by the majority voting method is used as the final result.

Images