A method and system for analyzing the state of an AI-powered sleep lamp by capturing millimeter-wave breathing signals.

Abstract

This invention provides an AI-based sleep lamp state analysis method and system for capturing millimeter-wave respiratory signals, belonging to the field of data analysis technology. The method includes: extracting the respiratory time-domain signal from the pre-processed original radar echo signal of the target area to obtain the respiratory frequency value; determining the current sleep stage based on the respiratory time-domain signal and respiratory frequency value using a pre-trained AI sleep state classification model; and extracting the spatial phase difference matrix of the respiratory micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition to construct a first respiratory source point and a second respiratory source point. This invention captures respiratory signals using millimeter waves, combines an AI sleep state classification model with manifold operations to analyze sleep state and respiratory stability, constructs a control parameter set and achieves closed-loop calibration, and can adjust the sleep lamp's light and sound fields in a coordinated manner, improving the accuracy and intelligence of the sleep lamp's adaptive control.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to an AI sleep lamp status analysis method and system for capturing millimeter-wave breathing signals. Background Technology

[0002] Existing AI sleep lights typically integrate millimeter-wave radar antenna arrays to receive echo signals generated by human breathing movements and output adjustment commands for light brightness or sound field envelope based on signal processing results. In practical applications, such devices have limited utilization of the spatial phase difference matrix output by the antenna array, which may lead to a certain time deviation between the generation of adjustment commands and changes in the spatial distribution of breathing movements.

[0003] Specifically, the echo signals received by each antenna element in a millimeter-wave radar antenna array contain range displacement information of different regions on the surface of the human thoracic cavity. The echo phase difference values ​​between different antenna elements constitute a spatial phase difference matrix, the distribution pattern of which is related to the amplitude of motion and spatial symmetry of each point on the thoracic cavity surface. However, some existing sleep lamps, after acquiring the raw echo signals, typically only extract the amplitude of each antenna element or the phase average of a single range gate to calculate the overall respiratory rate, while rarely performing further processing on the relative changes between the elements in the spatial phase difference matrix.

[0004] For example, if, within a certain time period, the phase difference between the two left-hand elements in the antenna array increases by approximately 0.1 radians compared to the previous minute, while the phase difference between the two right-hand elements changes by less than 0.02 radians, gradient fluctuations appear in a local region of the spatial phase difference matrix. If the signal processing link of existing equipment directly uses this matrix as raw data storage or only uses it to calculate a single breathing frequency, the subsequently generated luminous flux or acoustic field envelope modulation coefficients will only change with the overall breathing frequency, without generating additional adjustment instructions due to local gradient fluctuations in the matrix.

[0005] In this scenario, there may be a time difference of several seconds to tens of seconds between the moment the device outputs light or sound field adjustments and the moment when local fluctuations in the spatial phase difference matrix occur. The adjustment command is only triggered after a measurable change in the overall breathing frequency, by which time the distribution pattern of the spatial phase difference matrix may have recovered or further evolved. Summary of the Invention

[0006] This invention provides an AI sleep lamp status analysis method and system for capturing millimeter-wave breathing signals, avoiding interference with user sleep caused by contact monitoring.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0008] In a first aspect, an AI sleep lamp state analysis method based on millimeter-wave respiratory signal capture is provided, the method comprising:

[0009] Step 1: Extract the respiratory time-domain signal from the preprocessed original radar echo signal of the target area, obtain the respiratory frequency value, and determine the current sleep stage based on the respiratory time-domain signal and respiratory frequency value using a pre-trained AI sleep state classification model; extract the spatial phase difference matrix of the respiratory micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition, and construct the first respiratory source point and the second respiratory source point.

[0010] Step 2: Map the first and second breathing source points to the Lie algebra manifold space to determine the first and second Lie algebra generators; obtain the breathing manifold structure tensor by performing Lie bracket operations on the first and second Lie algebra generators.

[0011] Step 3: Calculate the respiratory phase factor or respiratory symplectic geometric invariant based on the current sleep stage. Input the respiratory phase factor or respiratory symplectic geometric invariant along with the respiratory frequency value into the respiratory stability mapper to obtain the respiratory stability evaluation parameters.

[0012] Step 4: Calculate the current sleep stage, respiratory stability evaluation parameters, and the variance of the spatial distance between the first and second respiratory sources during continuous respiratory cycles to construct a set of sleep light control parameters.

[0013] Step 5: Generate light and sound driving commands based on the sleep light control parameter set to adjust the light and sound fields in a coordinated manner; collect light and sound feedback signals in real time to calibrate the output deviation of the actuator, and dynamically adjust the proportional adjustment coefficient in the sleep light control parameter set to form a closed-loop control.

[0014] Secondly, an AI sleep light state analysis system for capturing millimeter-wave respiratory signals includes:

[0015] The source point construction module is used to extract the respiratory time-domain signal from the pre-processed original radar echo signal of the target area, obtain the respiratory frequency value, and determine the current sleep stage based on the respiratory time-domain signal and respiratory frequency value through a pre-trained AI sleep state classification model; it also extracts the spatial phase difference matrix of the respiratory micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition, and constructs the first and second respiratory source points.

[0016] The computation module is used to map the first and second breathing source points to the Lie algebra manifold space, determine the first and second Lie algebra generators, and obtain the breathing manifold structure tensor by performing Lie bracket operations on the first and second Lie algebra generators.

[0017] The matching module is used to calculate the respiratory phase factor or respiratory symplectic geometric invariant based on the current sleep stage, and input the respiratory phase factor or respiratory symplectic geometric invariant and the respiratory frequency value into the respiratory stability mapper to obtain respiratory stability evaluation parameters.

[0018] The acquisition module is used to calculate the current sleep stage, respiratory stability evaluation parameters, and the variance of the spatial distance between the first and second respiratory sources in a continuous respiratory cycle, in order to construct a set of sleep light control parameters.

[0019] The adjustment module is used to generate light and sound driving commands based on the sleep light control parameter set to adjust the light and sound fields in a coordinated manner; to collect light and sound feedback signals in real time to calibrate the output deviation of the actuator, and to dynamically adjust the proportional adjustment coefficient in the sleep light control parameter set to form a closed-loop control.

[0020] The above-described solution of the present invention has at least the following beneficial effects:

[0021] By mining the spatial phase distribution and geometric structure information of respiratory micro-movements, a refined characterization of the user's respiratory state is achieved, effectively shortening the response delay between environmental regulation and the user's actual physiological needs, and improving the regulation matching degree. Introducing Lie algebraic manifold and symplectic geometric correlation operations quantifies the coupling relationship and motion trajectory characteristics of the spatial location of the respiratory source, enriching the evaluation dimensions of respiratory stability, and improving the accuracy and sensitivity of sleep state analysis. This effectively distinguishes between physiological rhythmic fluctuations and abnormal disturbances in breathing. Combining sleep state classification results with respiratory stability evaluation parameters, a multi-dimensional sleep lamp control parameter set is constructed to achieve adaptive continuous adjustment of the light and sound fields, replacing the discrete-stage trigger-based adjustment mode, improving the synchronicity of sleep environment regulation and user sleep comfort, and enhancing the sleep assistance effect of the intelligent sleep lamp. Real-time calibration of the actuator output deviation through light and sound feedback signals and dynamic adjustment of the proportional adjustment coefficient in the sleep lamp control parameter set form a complete closed-loop control, effectively reducing the impact of execution deviation and model drift on system performance. It uses non-contact millimeter-wave radar to capture breathing signals, avoiding interference with users' sleep caused by contact monitoring. At the same time, through refined status analysis and control logic, it optimizes the practicality and user experience of the smart sleep lamp. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an AI sleep light state analysis method for capturing millimeter-wave breathing signals, provided by an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of an AI sleep light status analysis system for capturing millimeter-wave breathing signals, provided by an embodiment of the present invention.

[0024] Figure 3This is a statistical diagram of the distribution of breathing source points in the Lie algebraic manifold space provided by an embodiment of the present invention.

[0025] Figure 4 This is a trend diagram showing the correlation between respiratory stability evaluation parameters and sleep stages provided in an embodiment of the present invention.

[0026] Figure 5 This is a schematic diagram of the spatial distance fluctuation variance of dual respiratory source points provided in an embodiment of the present invention. Detailed Implementation

[0027] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0028] like Figure 1 As shown, an embodiment of the present invention proposes an AI sleep light state analysis method based on millimeter-wave respiratory signal capture, the method comprising the following steps:

[0029] Step 1: Extract the respiratory time-domain signal from the preprocessed original radar echo signal of the target area, obtain the respiratory frequency value, and determine the current sleep stage based on the respiratory time-domain signal and respiratory frequency value using a pre-trained AI sleep state classification model; extract the spatial phase difference matrix of the respiratory micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition, and construct the first respiratory source point and the second respiratory source point.

[0030] Step 2: Map the first and second breathing source points to the Lie algebra manifold space to determine the first and second Lie algebra generators; obtain the breathing manifold structure tensor by performing Lie bracket operations on the first and second Lie algebra generators.

[0031] Step 3: Calculate the respiratory phase factor or respiratory symplectic geometric invariant based on the current sleep stage. Input the respiratory phase factor or respiratory symplectic geometric invariant along with the respiratory frequency value into the respiratory stability mapper to obtain the respiratory stability evaluation parameters.

[0032] Step 4: Calculate the current sleep stage, respiratory stability evaluation parameters, and the variance of the spatial distance between the first and second respiratory sources during continuous respiratory cycles to construct a set of sleep light control parameters.

[0033] Step 5: Generate light and sound driving commands based on the sleep light control parameter set to adjust the light and sound fields in a coordinated manner; collect light and sound feedback signals in real time to calibrate the output deviation of the actuator, and dynamically adjust the proportional adjustment coefficient in the sleep light control parameter set to form a closed-loop control.

[0034] In this embodiment of the invention, by mining the spatial phase distribution and geometric structure information of respiratory micro-movements, a refined characterization of the user's respiratory state is achieved, effectively shortening the response delay between environmental regulation and the user's actual physiological needs, and improving the regulation matching degree. By introducing Lie algebraic manifolds and symplectic geometric correlation operations, the coupling relationship and motion trajectory characteristics of the spatial location of the respiratory source are quantified, enriching the evaluation dimensions of respiratory stability and improving the accuracy and sensitivity of sleep state analysis. This effectively distinguishes between physiological rhythm fluctuations and abnormal disturbances in breathing. Combining sleep state classification results with respiratory stability evaluation parameters, a multi-dimensional sleep lamp control parameter set is constructed to achieve adaptive continuous adjustment of the light and sound fields, replacing the discrete-stage trigger-based adjustment mode, improving the synchronicity of sleep environment regulation and user sleep comfort, and enhancing the sleep assistance effect of the intelligent sleep lamp. Real-time calibration of the actuator output deviation through light and sound feedback signals and dynamic adjustment of the proportional adjustment coefficient in the sleep lamp control parameter set form a complete closed-loop control, effectively reducing the impact of execution deviation and model drift on system performance. It uses non-contact millimeter-wave radar to capture breathing signals, avoiding interference with users' sleep caused by contact monitoring. At the same time, through refined status analysis and control logic, it optimizes the practicality and user experience of the smart sleep lamp.

[0035] In a preferred embodiment of the present invention, step 1 includes:

[0036] Step 100a: Extract the breathing time-domain signal from the preprocessed original radar echo signal of the target area to obtain the breathing frequency value. Obtaining the breathing frequency value includes: performing a fast Fourier transform on the breathing time-domain signal to obtain a breathing frequency spectrum, and extracting the frequency corresponding to the energy peak from the breathing frequency spectrum as the breathing frequency value. Specifically, this includes:

[0037] The raw radar echo signals collected by millimeter-wave radar (using the 24GHz or 60GHz band, suitable for bedroom sleep monitoring scenarios, with a detection range of 0.5m to 5m) undergo systematic preprocessing. The preprocessing operations are performed sequentially as follows: clutter suppression, DC component removal, and bandpass filtering. The filtering frequency band is strictly set to 0.1Hz to 0.8Hz, which accurately covers the respiratory rate range of various sleep states (0.2Hz to 0.33Hz for adults when awake, slightly lower during light and deep sleep, both within 0.1Hz to 0.8Hz). Within 0.8Hz, the effective signal component corresponding to the micro-movements of human breathing is retained. The amplitude range of this effective signal component is limited to 5% to 30% of the total amplitude of the radar echo. Signals with an amplitude ratio below 5% are identified as environmental noise, and signals with an amplitude ratio above 30% are identified as interference signals such as large-amplitude limb shaking and equipment vibration. This limitation can effectively filter out heartbeat interference (heartbeat signal frequency 1Hz to 2Hz, which is outside the bandpass filtering range), large-amplitude limb shaking interference (amplitude ratio usually exceeds 30%), and environmental electromagnetic interference signals (frequency mostly above 1Hz). After preprocessing, the target area containing the human chest is locked within the radar detection range (the area with a stable signal strength and consistent with the chest size, located 0.5m to 2m away from the radar). From the radar echo signal corresponding to the target area, the signal component that exhibits periodic strength changes with respiratory movement is extracted. Specifically, this component is the signal component whose radar echo signal amplitude periodically increases and decreases as the human chest rises and falls with inhalation and exhalation. During inhalation, the chest expands, increasing the distance from the radar and decreasing the echo signal amplitude; during exhalation, the chest contracts, decreasing the distance from the radar and increasing the echo signal amplitude. Its change period is completely consistent with the human respiratory cycle (the adult respiratory cycle is about 3s to 5s). This component is the respiratory time-domain signal.

[0038] A Fast Fourier Transform (FFT) is performed on the respiratory time-domain signal to transform it into the frequency domain, yielding the respiratory frequency spectrum. This facilitates the selection of the dominant frequency (i.e., respiratory rate) of the respiratory signal. The formula for the Discrete Fast Fourier Transform is as follows: ;

[0039] In the formula, These are discrete sampled values ​​of the respiratory time-domain signal, that is, discrete values ​​obtained by collecting continuous respiratory time-domain signals at fixed time intervals (usually set to 0.01s). For sampling point index; This represents the total number of sampling points; For frequency domain indexing; It is a complex value in the frequency domain, and its form is: ( For the real part, (The imaginary part) contains both the amplitude and phase information of the frequency components; The imaginary unit satisfies = 1; Pi It is a natural constant. The respiratory frequency domain spectrum characterizes the energy distribution of the respiratory signal at different frequency components. The larger the energy amplitude, the stronger the signal corresponding to that frequency component. The energy of the respiratory signal is mainly concentrated in the component corresponding to its dominant frequency. The formula for calculating the energy amplitude of a single frequency point is: ;

[0040] In the formula, For frequency domain complex values The real part, that is, the real part of the complex number in the frequency domain (corresponding to the above). value); For frequency domain complex values The imaginary part, that is, the imaginary part of the complex number in the frequency domain (corresponding to the above). value); The energy amplitude at a corresponding frequency point, measured in mV², represents the signal strength of that frequency component. A larger energy amplitude indicates that the frequency component is closer to the actual human respiratory rate. By iterating through the energy amplitudes corresponding to all frequency points in the respiratory frequency spectrum, the peak point with the largest energy amplitude is selected. The frequency corresponding to this peak point is the respiratory rate value. The respiratory rate value represents the number of times a person breathes per unit of time. The unit is Hz. When converting to respiratory rate per minute, the value needs to be multiplied by 60. To obtain the energy amplitude Frequency domain index corresponding to the maximum That is, by traversing =0 to Find all possible values ​​of 1 The maximum time corresponding Value, that The value is the frequency domain index corresponding to the respiratory rate.

[0041] Step 100b: Based on the respiratory time-domain signal and respiratory frequency value, determine the current sleep stage using a pre-trained AI sleep state classification model. This includes: inputting a feature vector composed of the respiratory frequency value, the peak-to-trough amplitude difference, zero-crossing rate, and signal envelope change rate of the respiratory time-domain signal into the first-level random forest classifier of the pre-trained AI sleep state classification model; performing preliminary classification on the feature vector to obtain a preliminary classification result indicating whether the current sleep stage belongs to the awake or non-awake stage, and a classification confidence vector composed of the classification confidence scores output by each decision tree; when the preliminary classification result indicates a non-awake stage, inputting the respiratory frequency value, the peak-to-trough amplitude difference, zero-crossing rate, signal envelope change rate, and the classification confidence vectors of each decision tree into the second-level gradient boosting tree classifier of the pre-trained AI sleep state classification model to classify the non-awake stage, obtaining a classification result indicating whether the current sleep stage is a light sleep or deep sleep stage. Specifically, this includes:

[0042] The pre-trained AI sleep state classification model adopts a two-stage cascaded structure. The first stage is a random forest classifier, which is an ensemble learning algorithm consisting of 50 CART decision trees in parallel. Each decision tree has a depth of 8 layers and 10 leaf nodes. The parallel operation of 50 decision trees can reduce the classification error of a single decision tree. The second stage is a gradient boosting tree classifier, which is also an ensemble learning algorithm, consisting of 100 iterative decision trees connected in series. Each decision tree has a depth of 6 layers and 8 leaf nodes. The first-level random forest classifier is used to distinguish between the waking and non-waking stages. During the waking stage, the human respiratory rhythm is unstable and the respiratory rate fluctuates greatly (0.3Hz to 0.5Hz). During the non-waking stage (including light sleep and deep sleep), the human respiratory rhythm is relatively stable and the respiratory rate fluctuates less. The second-level gradient boosting tree classifier is used to further distinguish between the light sleep stage and the deep sleep stage within the non-waking stage. During the light sleep stage, the respiratory rhythm fluctuates slightly, and the respiratory rate fluctuates within the range of 0.02Hz to 0.05Hz. During the deep sleep stage, the respiratory rhythm is the most stable and the fluctuation is minimal, with the respiratory rate fluctuating within the range of no more than 0.02Hz.

[0043] Millimeter-wave radar respiratory signal data were collected from test subjects of different ages and body types in three sleep states: awake, light sleep, and deep sleep. The data collection time for each subject in each sleep state was no less than 30 minutes. For each set of respiratory time-domain signals, four core characteristic parameters were calculated: respiratory rate (calculated according to the method in step 100a), peak-to-trough amplitude difference (the difference between the peak and trough values ​​within a single respiratory cycle in the respiratory time-domain signal, taken as the average of 10 consecutive respiratory cycles, in mV), zero-crossing rate (the number of times the respiratory time-domain signal crosses the zero line per minute, characterizing the degree of fluctuation in the respiratory rhythm; a higher zero-crossing rate indicates a more unstable respiratory rhythm), and signal envelope change rate. Specifically, the original respiratory time-domain signal was first identified as a continuous signal, then subjected to a Hilbert transform, and phase shift was achieved through convolution operations. The specific transformation formula is as follows: In the formula The variable to be integrated is the one whose integration range covers the entire time domain of the signal. It is a time variable. It is pi (π). Through this convolution operation, the original respiratory time-domain signal can be transformed. All frequency components are phase-shifted by -90° to obtain the Hilbert transform result. An analytical signal is constructed from the original respiratory time-domain signal and its Hilbert transform result. ,in The imaginary unit (satisfying) = 1) The analyzed signal can completely preserve the amplitude and phase information of the original respiratory signal, and its amplitude... The calculation formula is: The envelope of the respiratory signal encapsulates the peak change trajectory of the original respiratory time-domain signal. After extracting the envelope, the first derivative is calculated, and then the absolute value of the first derivative is taken. Finally, the average of all absolute values ​​of the first derivative over the entire respiratory signal acquisition period (no less than 30 minutes) is taken as the final value of the signal envelope change rate, in mV / s. The corresponding real sleep stages are labeled using a sleep monitoring device (polysomnography, with an accuracy of no less than 98% in determining sleep stages), forming labeled training samples. Each label corresponds one-to-one with a feature parameter, used by the model to learn the correlation between features and sleep stages.

[0044] The training process of the first-level random forest classifier is as follows: the respiratory rate, peak-to-valley amplitude difference, zero-crossing rate, and signal envelope change rate are used as input data to form a four-dimensional feature vector. The awake and non-awake stages are used as classification labels (light sleep and deep sleep are uniformly labeled as non-awake, forming a binary classification task). During the training process, a bagging sampling method is used to train the decision trees independently, that is, 70% of the samples are randomly selected from the training sample set (with replacement) and different training samples are assigned to each decision tree to avoid excessive correlation between decision trees and improve the generalization ability of the model. After each decision tree is trained, the remaining 30% of test samples are classified. A majority voting mechanism is used to determine the preliminary classification result. That is, each of the 50 decision trees outputs a classification result, and the result with the most votes is the preliminary classification result for that sample. For example, if 30 decision trees classify it as conscious and 20 classify it as unconscious, the preliminary classification result is conscious. At the same time, the classification confidence of each decision tree is recorded. The classification confidence refers to the degree of confidence that each decision tree has in its own classification result. The value ranges from 0 to 1. It is calculated as the proportion of the number of samples in the decision tree that are classified as that category to the total number of training samples of that decision tree. The higher the confidence, the more reliable the classification result.

[0045] The training process of the second-level gradient boosting tree classifier is as follows: Samples classified as non-awake in the first level (training is only performed on samples from non-awake stages to reduce interference from irrelevant samples and focus on distinguishing between light and deep sleep) are used as input data. The input data consists of a high-dimensional feature vector (54 dimensions in total) composed of respiratory rate, peak-to-trough amplitude difference, zero-crossing rate, signal envelope change rate, and the classification confidence scores of all first-level decision trees (50 confidence scores in total). Light sleep stage and deep sleep stage are used as classification labels (binary classification task). A forward step-by-step iterative training method is adopted, i.e., training each decision tree one by one. Each decision tree is corrected using a gradient descent algorithm to correct the classification error of the previous decision tree (a logarithmic loss function is used), gradually reducing the overall classification error. During training, the classification accuracy is calculated in real time until it converges to a preset threshold, which is set to 95%. Specifically, the convergence condition is that training stops when the classification accuracy is not lower than 95% for 10 consecutive iterations. Ten consecutive iterations avoid accuracy fluctuations caused by random factors, ensuring stable model convergence and preventing overfitting.

[0046] Millimeter-wave radar echo signals are acquired in real time. Following the method in step 100a, the real-time respiratory rate, peak-to-valley amplitude difference, zero-crossing rate, and signal envelope change rate corresponding to the respiratory time-domain signal are calculated. These four parameters are combined into a feature vector, which is then input into the trained first-level random forest classifier. The classifier performs decision tree-by-decision discrimination on the feature vector. Each decision tree, based on the parameter magnitude of the feature vector and its own trained decision rules (e.g., respiratory rate > 0.35 Hz, zero-crossing rate > 25 breaths / minute, indicating a conscious state), determines whether the sample belongs to a conscious or unconscious state through majority voting. The system obtains a preliminary classification result indicating whether the current sleep stage is conscious or unconscious. Simultaneously, it combines the classification confidence scores from each decision tree in order of tree number to form a classification confidence vector (50 elements in total), which is used as input to the subsequent second-level classifier. If the preliminary classification result indicates conscious sleep, the current sleep stage is directly determined to be conscious, without needing to proceed to the second-level classification. If the preliminary classification result indicates unconscious sleep, the respiratory rate, peak-to-trough amplitude difference, zero-crossing rate, signal envelope change rate, and classification confidence vector are input together with the second-level gradient boosting tree classifier. The classifier then uses the high-dimensional feature vector to determine the sleep stage. The classification rules obtained from training are used to determine whether a person is in light sleep (when the respiratory rate fluctuates between 0.02Hz and 0.05Hz, the zero-crossing rate is 12 to 18 breaths / minute, the signal envelope change rate is 0.08 to 0.15mV / s, the peak-to-trough amplitude difference is 1.2 to 2.5mV, and the mean classification confidence vector is between 0.7 and 0.85); when the respiratory rate fluctuates between ≤0.02Hz, the zero-crossing rate is ≤12 breaths / minute, the signal envelope change rate is ≤0.08mV / s, the peak-to-trough amplitude difference is ≤1.2mV, and the mean classification confidence vector is ≥0.85). The final judgment result determines whether the current sleep stage is light sleep or deep sleep, with a judgment accuracy of no less than 95%, meeting the actual needs of sleep monitoring. At the same time, the second-level gradient boosting tree classifier outputs the classification confidence of the current sleep stage. This confidence value takes a value of [0,1], which is obtained by the maximum probability that the classifier predicts the current input feature vector as light sleep or deep sleep. For example, if the probability of predicting light sleep is 0.92 and the probability of deep sleep is 0.08, then the classification confidence is 0.92, which is used to characterize the reliability of the current sleep stage judgment. The closer the confidence is to 1, the more reliable the judgment result.

[0047] Step 100c: Perform range-dimensional Fast Fourier Transform on the breathing micro-motion signals received by each antenna element, extract the complex values ​​of each antenna element within the target range gate, calculate the conjugate product of the complex values ​​of each antenna element within the target range gate, and obtain the spatial phase difference between different pairs of antenna elements in the antenna array. Specifically, this includes:

[0048] A multi-element linear antenna array is used to receive respiratory micro-motion echo signals. This array consists of eight antenna elements arranged linearly and uniformly, with an element spacing of 10 mm and a total array length of 70 mm. The multi-element linear antenna array can simultaneously receive respiratory micro-motion echo signals from different directions, capturing the spatial distribution information of these signals. For the respiratory micro-motion signal received by each antenna element, a range-time Fast Fourier Transform is performed. First, it is established that the respiratory micro-motion echo signal received by each antenna element is a range-time two-dimensional signal, denoted as […]. ,in For distance variables, As a time variable, the signal contains echo components at different distances; the continuous range signal received by each antenna element. Perform discrete sampling, sampling interval Set to 0.01m. It is the number of sampling points for the discrete distance signal. ( Windowing is performed on the distance sampling index, using the Hanning window (window function expression is...). In the formula, the fixed coefficient It is 0.5. (where the value is the Hanning window function); perform a Fast Fourier Transform on the windowed discrete distance signal, the transform formula is: ;

[0049] In the formula For distance frequency domain index, It is a distance-frequency domain complex value. The imaginary unit, Pi; for the transformed frequency domain complex value Take the amplitude value to obtain the distance frequency domain amplitude spectrum. The horizontal axis of this amplitude spectrum is distance (indexed by the frequency domain). The mapping is obtained, and the mapping relationship is as follows: The vertical axis represents signal strength (in mV), converting the time-domain signal to the range-frequency domain for clearer differentiation of signals at different distances. Based on the actual distance between the human chest and the radar (set to 0.8m to 1.5m, suitable for a bedroom sleeping scenario), a target range gate is established. The target range gate refers to the distance interval corresponding to the distance from the human chest. The gate width is set to 0.1m (e.g., if the distance from the human chest to the radar is 1.0m, then the target range gate is 0.95m to 1.05m). This is used to filter out the micro-respiratory signals corresponding to the human chest and exclude interference signals from walls, furniture, and other distances. The complex values ​​corresponding to each antenna element within the target range gate are extracted. The form ( For the real part, The specific meaning and extraction method of the imaginary part are consistent with the frequency domain complex value in step 100a. It contains the amplitude and phase information of the breathing micro-motion signal and is the core basis for calculating the spatial phase difference. A total of 8 corresponding complex values ​​are obtained from the 8 antenna elements.

[0050] For any two different antenna elements, let the complex value corresponding to the first antenna element be... (Format is) The second line unit corresponds to the complex value as follows: (Format is) Its conjugate value is The conjugate value refers to the conjugate value of a complex number. The complex number obtained by inverting the imaginary part is, i.e. The conjugate operation can eliminate the common phase error of the received signals from two antenna elements. (Calculation) and The result of the conjugate product, i.e. The phase angle of the conjugate product is calculated, which is the spatial phase difference between the antenna element pairs. The larger the distance difference, the larger the phase difference. The spatial phase difference between antenna element pairs, expressed in radians (rad), ranges from - to It can be converted into an angle system to characterize the phase offset of two antenna elements receiving the same breathing micro-motion signal; following the above method, all antenna element combinations (8 antenna elements have a total of 28 different combinations) are traversed, and the spatial phase difference between each pair of antenna elements is calculated one by one, resulting in a total of 28 spatial phase difference values.

[0051] Step 101c: Construct a spatial phase difference matrix by arranging the spatial phase differences between all antenna element pairs according to the order of the antenna elements; substitute each spatial phase difference in the spatial phase difference matrix into the steering vector of the multi-signal classification algorithm to construct a spatial spectrum function, specifically including:

[0052] All 28 spatial phase differences are arranged according to the physical arrangement of the antenna array (i.e., the installation order of the antenna elements in the actual device, numbered from 1 to 8 from left to right), forming an 8×8 two-dimensional matrix. This matrix is ​​the spatial phase difference matrix. The rows and columns of the matrix correspond to the antenna element numbers. The element in the I-th row and J-th column of the matrix is ​​the spatial phase difference between the I-th antenna element and the J-th antenna element (when I≠J, it is the calculated phase difference value; when I=J, the phase difference is set to 0, because the phase difference of the same antenna element is 0). This matrix can centrally store the spatial phase difference information between all pairs of antenna elements, clearly presenting the phase correlation relationship of each antenna element. Each set of spatial phase difference values ​​in the spatial phase difference matrix is ​​substituted into the steering vector corresponding to the multi-signal classification algorithm. The steering vector characterizes the response characteristics of the antenna array to signals in different spatial directions and is related to the antenna position, signal frequency, and spatial angle. Its calculation formula is: ;

[0053] in The antenna array steering vector is an 8-dimensional column vector related to the azimuth and elevation angles, with dimensions matching the number of antenna elements. The azimuth angle refers to the angle between the target on the horizontal plane and the area directly in front of the radar. The pitch angle, The imaginary unit, Pi The antenna element spacing is 10mm. For millimeter wave wavelength, Using the transpose symbol and combining the antenna array manifold characteristics with the radial distance corresponding to the target range gate, a spatial spectrum function for spatial target localization is constructed, i.e., the spatial spectrum function formula is: ;

[0054] In the formula, This is the conjugate transpose of the guide vector. The sign for conjugate transpose; The noise subspace projection matrix is ​​constructed from the spatial phase difference matrix. By performing eigenvalue decomposition on the spatial phase difference matrix, eight eigenvalues ​​and their corresponding eight eigenvectors are obtained. The eigenvalues ​​are sorted in ascending order, and the eigenvalues ​​with smaller values ​​(corresponding to noise signals) and their corresponding eigenvectors constitute the noise subspace. The projection matrix is ​​then constructed using this noise subspace to suppress interference signals. The spatial spectrum function characterizes the response intensity of respiratory micromotion signals at different azimuth and elevation angles in space. A larger function value indicates a stronger respiratory micromotion signal at that location, and a higher probability of the presence of a respiratory source at that location.

[0055] Step 102c involves performing a two-dimensional peak search on the spatial spectral function in both azimuth and pitch dimensions. The first azimuth and first pitch angles corresponding to the maximum peak response are extracted. Combined with the radial distance corresponding to the target distance gate, the first respiratory source point is calculated in the spatial coordinate system. The second azimuth and second pitch angles corresponding to the second maximum peak response are extracted. Combined with the radial distance corresponding to the target distance gate, the second respiratory source point is calculated in the spatial coordinate system. The first respiratory source point corresponds to the spatial location point where the end-expiratory chest wall displacement is maximum, and the second respiratory source point corresponds to the spatial location point where the end-inspiratory chest wall displacement is maximum. Specifically, this includes:

[0056] In the spatial spectrum function, azimuth and elevation angles are used as independent variables to perform a two-dimensional global peak search, with the azimuth angle ranging from - / 2 to / 2, with a step size of 0.01 rad (approximately 0.57°); pitch angle from 0 to / 2, with a step size of 0.01 rad, ensures the search covers all possible respiratory source locations without missing any peak points. After traversing all azimuth and elevation angle combinations, find the location with the largest peak in the spatial spectrum function response, and record the corresponding first azimuth and first elevation angles. This location with the largest peak corresponds to the location with the strongest respiratory signal, i.e., the spatial location with the largest displacement of the chest cavity at the end of expiration, because the chest cavity contracts to its smallest size at the end of expiration, is closest to the radar, and has the strongest reflected radar echo signal, corresponding to the largest spatial spectrum function value. Combined with the radial distance corresponding to the target range gate. Based on the transformation relationship between spherical and rectangular coordinate systems, the corresponding spatial coordinate point is calculated in the spatial rectangular coordinate system. This coordinate point is the first breathing source point. The spherical coordinate system uses the radar as its origin and describes the target position using radial distance, azimuth, and elevation angles. The spatial rectangular coordinate system also uses the radar as its origin. The axis is the horizontal direction directly in front of the radar. The axis represents the horizontal and vertical directions of the radar (right side is positive). The axis is perpendicular to the ground (upward is positive), using , , Three coordinates describe the target position, and the two are converted using a fixed formula.

[0057] Continue searching for the location of the second largest response peak in the spatial spectrum function, using the same search method as for the largest peak. Record the second azimuth and second elevation angles corresponding to this location. This location with the second largest peak corresponds to the location of the second strongest respiratory signal, i.e., the spatial location with the largest chest cavity displacement at the end of inhalation. Because the chest cavity expands to its maximum at the end of inhalation, it is furthest from the radar, resulting in the second strongest reflected radar echo signal and the second largest corresponding spatial spectrum function value. Using the same coordinate calculation method, combined with radial distance, the second respiratory source point is calculated. This coordinate point corresponds to the spatial location with the largest chest cavity displacement at the end of inhalation. The formula for calculating the coordinates in the spatial rectangular coordinate system is: ;

[0058] The two final calculated positioning points can represent the spatial position of the thoracic cavity at the end of expiration and the end of inspiration.

[0059] This embodiment uses Fast Fourier Transform to extract the peak energy of the respiratory frequency spectrum to determine the respiratory frequency value. Compared with directly estimating the frequency through the time-domain waveform, it can effectively suppress time-domain noise interference and improve the accuracy and stability of respiratory frequency extraction. A two-level cascaded structure is used to construct an AI sleep state classification model. First, a random forest is used to coarsely classify the awake and non-awake stages. Then, combined with classification confidence, a gradient boosting tree is used to subdivide the light sleep and deep sleep stages. Compared with a single classification model, this effectively improves the accuracy of sleep stage identification and reduces the probability of misclassification. Based on the echo signal from the antenna array, the spatial phase difference is calculated and a phase difference matrix is ​​constructed. This fully utilizes the spatial perception capability of millimeter-wave radar to capture the spatial distribution characteristics of respiratory motion, avoiding the limitations of single-channel signal detection. Combining multi-signal classification algorithms and two-dimensional angle peak search to solve for the respiratory source location point, the maximum displacement position of the thoracic cavity at the end of expiration and inspiration can be located, achieving a quantitative representation of the spatial characteristics of respiratory motion and overcoming the deficiency of only focusing on the time-domain and frequency-domain characteristics of breathing while ignoring the spatial micro-motion distribution.

[0060] In a preferred embodiment of the present invention, step 2 includes:

[0061] Step 200: Represent the first and second respiratory source points as a first coordinate vector and a second coordinate vector in a spatial coordinate system, respectively. Map the first and second coordinate vectors to the Lie algebra manifold space. The tangent space of the first coordinate vector is represented as the first Lie algebra generator, and the tangent space of the second coordinate vector is represented as the second Lie algebra generator. Specifically, this includes:

[0062] The first respiratory source point (chest position at the end of expiration) and the second respiratory source point (chest position at the end of inspiration) calculated in step 102c are both located in a spatial rectangular coordinate system with the radar as the origin (x-axis: horizontal in front of the radar; y-axis: positive to the right of the radar horizontally; z-axis: positive perpendicular to the ground upwards), consistent with the coordinate system of step 102c. The two positioning points are represented as a first coordinate vector and a second coordinate vector, with three components corresponding to the coordinate values ​​of each axis, fully describing their spatial positions. The two coordinate vectors are mapped to the Lie algebra space corresponding to a special orthogonal group. This space can characterize the nonlinear transformation properties of spatial vectors and is suitable for the coupled calculation of respiratory source positioning points. The mapping uses the Lie algebra exponential mapping method to convert the Euclidean coordinate vector into a Lie algebra tangent space vector. The specific calculation formula is as follows: ,in The Lie algebra tangent space vector is the result of mapping coordinate vectors to the Lie algebra manifold space, preserving the spatial characteristics of coordinate vectors; It is a spatial coordinate vector, that is, the first coordinate vector or the second coordinate vector obtained by solving in step 102c; For the exponential mapping operator, In the formula Describes the basis vectors of generators of Lie algebras Perform antisymmetry operations to ensure that the exponential mapping operator conforms to the rules of Lie algebra operations; The angle between the respiratory source location point and the radar's front is defined as , ranging from 0 to π / 2. 0 and 1 are fixed coefficients in the operator to ensure the accuracy of the mapping operation. After mapping, the tangent space vector corresponding to the first coordinate vector is used as the first Lie algebra generator, and the tangent space vector corresponding to the second coordinate vector is used as the second Lie algebra generator. Both retain the spatial characteristics and motion attributes of the corresponding respiratory source location point.

[0063] Step 201 involves performing a combined multiplication operation on the first Lie algebra generator and the second Lie algebra generator to obtain a first intermediate tensor; performing a combined multiplication operation on the second Lie algebra generator and the first Lie algebra generator to obtain a second intermediate tensor; and subtracting the first intermediate tensor from the second intermediate tensor to obtain a commutator tensor. Specifically, this includes:

[0064] For the generators of the first Lie algebra (3D tangent space vector, ) and the generators of the second Lie algebra (3D tangent space vector, Perform tensor multiplication operations. It is the transpose symbol, calculated using the Lie algebra tensor outer product rule, with the following formula: ,in This is the first intermediate tensor, which is the intermediate result of the coupling operation between two Lie algebra generators, containing information about the coupling positions of the two generators. ; They are the generators of the first Lie algebra. The components in the x, y, and z directions, They are generators of the second Lie algebra. Components in the x, y, and z directions.

[0065] By swapping the order of operations on the two generators and using the same outer product rule, the second intermediate tensor is calculated using the following formula: ,in The second intermediate tensor has the same dimension as the first intermediate tensor and is a coupled intermediate result after exchanging the order of generator operations. Calculation result It is a 3×3 dimensional tensor, and Dimensions are consistent.

[0066] Subtracting the two intermediate tensors point by point yields the commutator tensor, i.e. ,in The commutator tensor is the core tensor characterizing the difference in nonlinear correlation between two Lie algebra generators. The commutator tensor is composed of the difference of the tensor outer product of the two generators. The magnitude of each element in the tensor corresponds to the coupling strength of the two respiratory source localization points in the corresponding spatial directions (x, y, z axes). The larger the value, the tighter the spatial correlation between the two localization points in that direction. The positive and negative values ​​correspond to the coupling direction (positive coupling indicates that the positional change trend is consistent, and negative coupling indicates that the change trend is opposite). Through the numerical characteristics of each element, the degree and direction of spatial correlation between the two respiratory source localization points can be quantified, thereby reflecting their spatial coupling relationship.

[0067] Step 202 involves performing index scaling operations on the commutator tensor under the action of the metric tensor of the Lie algebra manifold space to obtain the nonlinear coupling strength tensor between the first and second breathing source points, which serves as the breathing manifold structure tensor. Specifically, this includes:

[0068] Determine the metric tensor of the Lie algebraic manifold space This is a 3×3 symmetric tensor, suitable for nonlinear coupled operations. The commutator tensor obtained in step 201 is subjected to index raising / lowering operations under the influence of the metric tensor, i.e. ,in It is the nonlinear coupling strength tensor, i.e., the final breathing manifold structure tensor; This is the inverse matrix of the metric tensor, used to implement the invertibility of index raising and lowering operations. After the operation is completed, the result is... This is the nonlinear coupling strength tensor between the first and second respiratory source points, which is directly used as the respiratory manifold structure tensor. This tensor fully preserves the spatial coupling relationship and nonlinear characteristics of the two location points. The magnitude of each element in the tensor corresponds to the coupling strength of the two respiratory source location points in the corresponding spatial direction. The larger the element value, the closer the spatial connection between the two location points in that direction. The positive and negative values ​​of the element values ​​correspond to the direction of the coupling relationship (positive values ​​indicate positive coupling, that is, the position change trend of the two location points in that direction is consistent, and negative values ​​indicate negative coupling, that is, the position change trend is opposite). By the magnitude and sign of each element in the tensor, the degree and direction of spatial connection between the two respiratory source location points (chest position at the end of expiration and end of inspiration) can be quantified, thereby characterizing the laws of chest wall respiratory movement (subtle changes in respiratory amplitude and rhythm).

[0069] This embodiment effectively solves the problem that the nonlinear correlation of the respiratory source location points cannot be accurately described in Euclidean space by mapping the respiratory source location points to the Lie algebra manifold space and using Lie algebra generators to represent the spatial features of the location points, thereby improving the accuracy of the representation of the spatial positional relationship of the respiratory source. By calculating the Yizi tensor and the nonlinear coupling strength tensor, the nonlinear coupling relationship between the first respiratory source point and the second respiratory source point can be captured, and the motion characteristics of the thoracic cavity during breathing can be fully preserved.

[0070] In a preferred embodiment of the present invention, step 3 includes:

[0071] Step 300a: When the current sleep stage is light sleep, using the spatial phase difference matrix as the basic data of the manifold to be processed, and using the respiratory manifold structure tensor as the Riemannian metric, the respiratory manifold structure tensor is assigned to the manifold structure where the spatial phase difference matrix resides, thus constructing the Riemannian manifold space. Specifically, this includes:

[0072] Obtain the respiratory manifold structure tensor obtained in step 202, and the spatial phase difference matrix constructed in step 101c, while confirming that the current sleep stage determined in step 100b is the light sleep stage. Use the spatial phase difference matrix obtained in step 101c as the basic data for the manifold to be processed. This spatial phase difference matrix is ​​an 8×8 two-dimensional matrix, with each row and column corresponding to one of the eight antenna element numbers. The element in the I-th row and J-th column represents the spatial phase difference between the I-th and J-th antenna elements, completely containing the phase correlation information of the respiratory micro-motion signals received by all antenna elements. The respiratory manifold structure tensor obtained in step 202 is used as a Riemannian metric. This respiratory manifold structure tensor is a 3×3-dimensional nonlinear coupling strength tensor, which fully preserves the spatial coupling relationship between the first respiratory source point (the position of the thoracic cavity at the end of expiration) and the second respiratory source point (the position of the thoracic cavity at the end of inspiration). The nonlinear characteristics of respiratory motion are characterized by the magnitude and sign of each element of the tensor. The magnitude of the element corresponds to the coupling strength of the two respiratory source points in the corresponding spatial direction. The larger the value, the more significant the nonlinear correlation of respiratory motion in that direction. The positive and negative values ​​correspond to the coupling direction; positive coupling indicates that the change trend of the thoracic cavity position at the end of expiration and inspiration is consistent, while negative coupling indicates that the change trend is opposite, thus fully characterizing the nonlinear characteristics of respiratory motion. The respiratory manifold structure tensor is then assigned to the manifold structure containing the spatial phase difference matrix. Specifically, this is achieved by coupling the respiratory manifold structure tensor with the spatial phase difference matrix through matrix tensor product operations, giving the manifold containing the spatial phase difference matrix Riemannian geometric properties, thus constructing the Riemannian manifold space. The specific calculation formula is as follows: ,in To construct the complete Riemannian manifold space, Given an 8×8 spatial phase difference matrix, the resulting Riemannian manifold space has a dimension of 24×24, which preserves both the phase correlation information of the spatial phase difference matrix and the nonlinear metric properties of the breathing manifold structure tensor.

[0073] Step 301a involves performing eigenvalue decomposition on the respiratory manifold structure tensor and extracting the eigenvector direction corresponding to the largest eigenvalue as the partitioning reference direction. In the Riemannian manifold space, the spatial phase difference matrix is ​​used as the tensor to be translated. The tensor to be translated is then moved parallel to the partitioning reference direction, and the tensor obtained after translation is used as the respiratory phase connection matrix. Specifically, this includes:

[0074] The respiratory manifold structure tensor obtained in step 202 Eigenvalue decomposition is performed using the standard matrix eigenvalue solution method, i.e. ,in The first tensor of the respiratory manifold structure eigenvalues ​​( =1,2,3), Eigenvalues The corresponding eigenvectors have the same dimension as the respiratory manifold structure tensor, being 3-dimensional column vectors. After eigenvalue decomposition, the eigenvalue with the largest value is extracted, and the direction of the eigenvector corresponding to this largest eigenvalue is used as the partitioning reference direction. This direction reflects the main nonlinear coupling characteristics of the respiratory manifold structure tensor. In the Riemannian manifold space constructed in step 300a, the spatial phase difference matrix obtained in step 101c is... As the tensor to be translated, it is moved parallel to the aforementioned partitioning reference direction (the direction of the eigenvector corresponding to the largest eigenvalue). The parallel translation operation follows the parallel translation rules of Riemannian manifold space, and the specific calculation formula is as follows: ,in This is the tensor obtained after translation (i.e., the breathing phase connection matrix). This is the translation step size, with a value of 0.01. For the tensor to be translated in the bisection reference direction The covariant derivative on the x-axis is used to characterize the rate of change of the tensor to be translated along that direction. After the translation operation is completed, the result is...

[0075] This is the respiratory phase connection matrix, which retains the phase correlation information of the spatial phase difference matrix and incorporates the nonlinear characteristics of the respiratory manifold structure tensor. It reflects the phase change law of respiratory micro-motion signals during light sleep by the continuity, periodicity and fluctuation amplitude of the phase value changes of each element in the matrix. The phase value of the element corresponding to the antenna element pair in the matrix shows a stable periodic fluctuation with the respiratory cycle, and the fluctuation amplitude matches the amplitude of respiratory micro-motions during light sleep. The continuity of phase change corresponds to the coherence of respiratory motion, and the uniformity of phase fluctuation corresponds to the relative stability of breathing during light sleep, thus reflecting the phase change law of respiratory micro-motion signals during light sleep.

[0076] Step 302a involves performing eigenvalue decomposition on the respiratory phase communication matrix, extracting the first principal eigenvalue with the largest modulus and the second principal eigenvalue with the second largest modulus, calculating the ratio of the first principal eigenvalue to the second principal eigenvalue, and using this ratio as the respiratory phase factor. Specifically, this includes:

[0077] The respiratory phase connection matrix obtained in step 301a is subjected to eigenvalue decomposition, using the same method as the eigenvalue decomposition method for the respiratory manifold structure tensor in step 301a. After eigenvalue decomposition, the magnitude of each eigenvalue is calculated. The magnitude characterizes the intensity of the eigenvalue; a larger magnitude indicates a more significant and stronger phase feature, while a smaller magnitude indicates a weaker and less intense phase feature. From all eigenvalue magnitudes, the eigenvalue with the largest magnitude is extracted and designated as the first principal eigenvalue, and the eigenvalue with the second largest magnitude is extracted and designated as the second principal eigenvalue. The ratio of the first principal eigenvalue to the second principal eigenvalue is calculated to obtain the respiratory phase factor. This ratio quantifies the prominence of the principal features of the respiratory phase connection matrix, thereby reflecting the stability of breathing during light sleep. The larger the ratio, the more prominent the respiratory phase feature corresponding to the first principal eigenvalue, the more concentrated the phase distribution of the respiratory phase connection matrix, and the more stable and regular the phase change of the respiratory micromotion signal during light sleep, resulting in stronger respiratory stability. The smaller the ratio, the smaller the difference between the first and second principal eigenvalues, the more dispersed the respiratory phase feature, and the more fluctuating and irregular the phase change of the respiratory micromotion signal during light sleep, resulting in weaker respiratory stability.

[0078] Step 303a: When the current sleep stage is deep sleep, using the respiratory manifold structure tensor as the symplectic potential function, rearrange the components of the respiratory manifold structure tensor according to the symplectic form of the symplectic geometric space to construct an even-dimensional respiratory phase space. In the respiratory phase space, perform exterior differentiation on the symplectic potential function to obtain the symplectic structure metric matrix, specifically including:

[0079] Confirming that the current sleep stage determined in step 100b is deep sleep, the respiratory manifold structure tensor obtained in step 202 is acquired. This respiratory manifold structure tensor is used as the symplectic potential function, which describes the basic geometric features of the symplectic geometric space and adapts to the stationarity characteristics of the respiratory signal in deep sleep. The components of the respiratory manifold structure tensor are rearranged according to the symplectic form of the symplectic geometric space, using the standard symplectic matrix form. (in The identity matrix is ​​arranged by expanding the three components of the breathing manifold structure tensor to an even dimension, adding zero components to make the total dimension 4 (to meet the even dimension requirement of symplectic geometric space), and rearranging them to obtain a 4×4 symplectic potential function matrix. The specific arrangement is as follows ,in For the respiratory manifold structure tensor, the first Line 1 The elements of the column. A 4-dimensional respiratory phase space is constructed through the above arrangement. This respiratory phase space is a symplectic geometric space, capable of characterizing the stability and periodicity of respiratory movements during deep sleep. Within this respiratory phase space, the symplectic potential function... The exterior differential operation follows the symplectic geometric exterior differential rule, and the specific calculation formula is as follows: ,in The result of the exterior differential operation. The symplectic potential function matrix No. Line 1 Column elements, The coordinate variables of the respiratory phase space ( =1,2,3,4). The derivative of the coordinate variable, The symbol for the outer product is... Represents the differential of coordinate variables , , The triple outer product and outer differential operation yield the symplectic structure metric matrix, a 4×4 symmetric matrix that describes the symplectic geometric properties of the respiratory phase space and reflects the stationary correlation characteristics of respiratory signals during deep sleep. The closer the values ​​of the diagonal elements of the symplectic structure metric matrix are and the smaller the fluctuations, the stronger the stationarity of the respiratory signals during deep sleep in the corresponding spatial dimension. The closer the values ​​of the off-diagonal elements are to 0, the more stable the correlation between respiratory signals in different spatial dimensions and the less obvious the abrupt changes. The uniformity of the overall element distribution of the matrix directly corresponds to the stationarity of the respiratory signals during deep sleep and the stability of the correlations in each dimension, thus accurately reflecting the stationary correlation characteristics of the respiratory signals during deep sleep.

[0080] Step 304a: Perform Singlasman manifold decomposition on the symplectic structure metric matrix, decomposing the symplectic structure metric matrix into multiple mutually symplectic orthogonal subspaces according to the symplectic orthogonal complement relation, with each subspace corresponding to a symplectic subspace cluster; for two adjacent symplectic subspace clusters, calculate the minimum symplectic inner product angle from any vector in one symplectic subspace cluster to the symplectic orthogonal complement of the other symplectic subspace cluster, and use the minimum symplectic inner product angle as the symplectic orthogonal complement angle between the two symplectic subspace clusters; arrange all the symplectic orthogonal complement angles between the symplectic subspace clusters in the topological order of the symplectic subspace clusters to form a sequence, and use the sequence as a breathing symplectic geometric invariant, specifically including:

[0081] The Singlasman manifold decomposition is performed on the symplectic structure metric matrix. The core of Singlasman manifold decomposition is to decompose the symplectic structure metric matrix into multiple mutually symplectic orthogonal subspaces according to the symplectic orthogonal complement relation. This is equivalent to eigenvalue decomposition of the symplectic structure metric matrix, using the same method as the eigenvalue decomposition of the respiratory manifold structure tensor in step 301a. After decomposition, four eigenvalues ​​and four corresponding eigenvectors are obtained. The eigenvectors are grouped according to the sign of the eigenvalues, with each group spanning a symplectic subspace. It is ensured that the symplectic subspaces of different groups satisfy the symplectic orthogonal complement relation (i.e., the symplectic inner product of any vector in one subspace with any vector in another subspace is 0). Each subspace corresponds to a symplectic subspace cluster. After decomposition, for any two adjacent symplectic subspace clusters, the minimum symplectic inner product angle from any vector in one symplectic subspace cluster to the symplectic orthogonal complement space of the other symplectic subspace cluster is calculated. ,in Let be any vector of the first symplectic subspace variety. Let be any vector in the symplectic orthogonal complement of the second symplectic subspace variety. It is a symplectic matrix. The sign for transpose; the least symplectic inner product angle. ,in The minimum symplectic inner product angle is the angle that characterizes the degree of symplectic orthogonal correlation between two adjacent symplectic subspace varieties. The closer the angle is to 90°, the stronger the symplectic orthogonality of the two adjacent symplectic subspaces, the higher the degree of symplectic orthogonality correlation, and the more stable the correlation of the respiratory signal features represented by the two subspaces; angle The greater the deviation from 90°, the weaker the symplectic orthogonality between two adjacent symplectic subspaces, the lower the degree of symplectic orthogonal correlation, and the less stable the correlation of respiratory signal features represented by the two subspaces. The calculated minimum symplectic inner product angle is used as the symplectic orthogonal complementary angle between two symplectic subspaces. Following the topological order of the symplectic subspaces (i.e., the order of eigenvalues ​​during decomposition), all the symplectic orthogonal complementary angles between symplectic subspaces are arranged sequentially to form a sequence. This sequence is the respiratory symplectic geometric invariant, capable of characterizing the stability and periodicity of respiratory movements during deep sleep, and possessing geometric invariance, unaffected by minor fluctuations in the respiratory signal.

[0082] Step 305a: When the current sleep stage is a waking stage, the respiratory phase factor is assigned a value of 1, and the respiratory symplectic geometric invariant is assigned a zero vector to represent the state of non-periodic respiratory geometric features. Specifically, this includes: confirming that the current sleep stage determined in step 100b is a waking stage. Since human respiration in the waking stage has no fixed periodicity, the amplitude and rhythm of respiratory movements fluctuate greatly, and there are no stable respiratory geometric features, the respiratory phase factor and respiratory symplectic geometric invariant are assigned values. The respiratory phase factor is assigned a value of 1 because there is no obvious principal feature phase in waking stage respiration, and a ratio of 1 can represent the state of no prominent principal feature. The respiratory symplectic geometric invariant is assigned a zero vector, which can represent the state of no stable symplectic geometric correlation and no periodic respiratory geometric features, matching the irregularity of respiration in the waking stage, and ensuring that the respiratory feature parameters of different sleep stages are represented consistently.

[0083] Step 300b: The respiratory phase factor or respiratory symplectic geometric invariant, along with the respiratory frequency value, is input into the respiratory stability mapper to obtain respiratory stability evaluation parameters. The respiratory stability mapper includes a criterion module, a logarithmic transformation module, a linear filtering module, and an exponential transformation module. The criterion module determines whether the current input is a marker of the awake stage. If so, a preset stability benchmark value is directly output. If not, the logarithmic transformation module performs a logarithmic domain transformation on the product of the norm of the respiratory phase factor or respiratory symplectic geometric invariant and the respiratory frequency value to obtain the transformed signal. The linear filtering module performs a weighted moving average processing on the transformed signal in the logarithmic domain to obtain the filtered result. The exponential transformation module performs exponential restoration on the filtered result to obtain a respiratory coherence coefficient with a value range normalized to the [0, 1] interval. The respiratory coherence coefficient is used as a respiratory stability evaluation parameter, specifically including:

[0084] Based on the current sleep stage, determine the parameters to be input to the respiratory stability mapper. In the light sleep stage, input the respiratory phase factor from step 302a and the respiratory frequency value from step 100a; in the deep sleep stage, input the respiratory symplectic geometric invariant from step 304a and the respiratory frequency value from step 100a; in the awake stage, input the respiratory phase factor assigned in step 305a (value 1, representing an undominant phase), the respiratory symplectic geometric invariant (all-zero vector, representing no stable symplectic geometric correlation), and the respiratory frequency value from step 100a, ensuring that the parameter format is consistent. The respiratory stability mapper is a dedicated signal processing mapper adapted to the needs of respiratory characteristics and quantitative evaluation at various stages. Its hardware relies on the signal processing unit of a millimeter-wave radar monitoring system (shared with step 100 to reduce complexity), while its software integrates a proprietary computational algorithm that seamlessly integrates with the preceding respiratory feature extraction logic. It can directly receive output parameters from the preceding steps. Its core function is to collaboratively fuse and standardize respiratory feature parameters and respiratory frequency values ​​at each stage, eliminating dimensional differences and outputting normalized respiratory stability evaluation parameters. Specifically, it uses a built-in algorithm to link the two parameters, complementarily preserving respiratory geometric stability (light sleep phase concentration, deep sleep symplectic orthogonal correlation) and rhythmic stability; respiratory phase factor... Retaining the dimensionless, symplectic geometric invariant of respiration pass (Minimum symplectic inner product angle is 90°) Normalization, The normalized symplectic geometric invariant of respiration, respiratory rate pass (Minimum respiratory rate) =12 breaths / minute, maximum respiratory rate =20 times / minute Normalized respiratory rate) is normalized to the dimensionless interval [0,1] to avoid calculation errors.

[0085] The mapper consists of four modules working collaboratively: criterion, logarithmic transformation, linear filtering, and exponential transformation. The criterion module receives input parameters and determines the awake stage (the respiratory symplectic geometric invariant is a zero vector and the respiratory phase factor is 1). If so, it outputs a preset stability benchmark value of 0.5 (representing moderate instability); otherwise, it passes the normalized parameters to the logarithmic transformation module. The logarithmic transformation module fuses the normalized parameters and transforms them to the logarithmic domain, linearizing the nonlinear signal. There are two calculation formulas; for the light sleep stage, the logarithmic transformation result = , (dimensionless, [0,+∞)); during deep sleep, the result of the logarithmic transformation = The linear filtering module performs weighted moving average filtering on the logarithmic domain signal, with a window length of 5 (adapting to a 3- to 5-second respiratory cycle), uniform weighting (weight 0.2), and the filtering result... ( For sampling points, For the current index, (Zero padding at 4 o'clock) filters out environmental noise and minute breathing fluctuations; the exponential transformation module exponentially restores the filtered result, formula... ( (This refers to the respiratory coherence coefficient), because the value of the logarithmic domain signal... After restoration Normalized to the [0,1] interval; finally, As a parameter for evaluating respiratory stability, the closer it is to 1, the more stable the breathing (deep sleep is closest to 1), and the closer it is to 0, the less stable the breathing (when awake, it is close to 0.5, and when there are large fluctuations, it is even closer to 0).

[0086] This embodiment designs differentiated methods for extracting and processing respiratory feature parameters for three different sleep stages: light sleep, deep sleep, and wakefulness. This approach aligns with the essential characteristics of respiratory movements in each stage, resolving the issue of inaccurate characterization of respiratory stability across different sleep stages. Through Riemannian manifold construction, eigenvalue decomposition, and symplectic geometric decomposition, combined with explicit calculation formulas, core parameters such as respiratory phase factors and symplectic geometric invariants are extracted, fully preserving the geometric features and stability information of respiration at different sleep stages. The respiratory stability mapper, through multi-module collaborative work, combines respiratory feature parameters with respiratory frequency to normalize the parameters for evaluating respiratory stability.

[0087] In a preferred embodiment of the present invention, step 4 includes:

[0088] Calculate the current sleep stage, respiratory stability evaluation parameters, and the variance of the spatial distance between the first and second respiratory sources over a continuous respiratory cycle to construct a sleep light control parameter set. This set includes a luminous flux modulation coefficient and a sound field envelope modulation coefficient. The luminous flux modulation coefficient is positively correlated with the respiratory coherence coefficient and negatively correlated with the variance of the variance. The sound field envelope modulation coefficient is positively correlated with the confidence level of the deep sleep stage within the current sleep stage. Specifically, it includes:

[0089] Three core basic parameters are obtained: First, the current sleep stage, determined by step 100b, specifically divided into light sleep, deep sleep, and wakefulness stages. Simultaneously, the confidence level of the deep sleep stage within the current sleep stage is obtained (value range [0,1], the closer the confidence level is to 1, the higher the probability that the current sleep stage is deep sleep; the closer it is to 0, the lower the probability that the current sleep stage is deep sleep). Second, the respiratory stability evaluation parameter, namely the respiratory coherence coefficient output by step 300b (value range [0,1], the closer the respiratory coherence coefficient is to 1, the more stable the breathing; the closer it is to 0, the less stable the breathing). Third, the spatial coordinates of the first and second respiratory source points. The first respiratory source point is the end-expiratory thoracic cavity positioning point obtained by step 102c, and the second respiratory source point is the end-inspiratory thoracic cavity positioning point obtained by step 102c. The variance of the spatial distance between the first and second respiratory sources is calculated over a continuous respiratory cycle. This variance characterizes the stability of respiratory amplitude; a smaller variance indicates a more stable respiratory amplitude, while a larger variance indicates more drastic fluctuations in respiratory amplitude. The specific calculation process is as follows: First, the spatial distance between the two respiratory sources is calculated within a single respiratory cycle; second, the average spatial distance over a continuous respiratory cycle is calculated; based on the spatial distance in a single respiratory cycle and the average spatial distance, the variance calculation formula is used to obtain the variance of the spatial distance over a continuous respiratory cycle. This variance directly reflects the fluctuation of respiratory amplitude.

[0090] Based on the calculated current sleep stage, respiratory coherence coefficient, fluctuation variance, and deep sleep stage confidence level, a set of control parameters for the sleep lamp is constructed. This set includes a luminous flux modulation coefficient and a sound field envelope modulation coefficient, both of which are normalized to the [0,1] interval to suit the adjustment range of the sleep lamp. The specific construction process is as follows: The luminous flux modulation coefficient is used to adjust the intensity of the sleep lamp's luminous flux. It is positively correlated with the respiratory coherence coefficient (the more stable the breathing, the gentler the luminous flux, and the closer the modulation coefficient is to 1), and negatively correlated with the fluctuation variance (the greater the fluctuation in breathing amplitude, the lower the luminous flux needs to be, and the closer the modulation coefficient is to 0). Luminous flux modulation coefficient = The light flux modulation coefficient (range [0,1]) and the proportional adjustment coefficient (preset to 0.1, calibrated according to the sensitivity of the sleep lamp and the range of respiratory fluctuations to ensure that the change of the modulation coefficient matches the breathing state) are used. When the respiratory coherence coefficient = 1 and the fluctuation variance = 0 (breathing is most stable and the amplitude has no fluctuation), the light flux modulation coefficient = 1, and the light flux reaches the intensity most suitable for sleep. When the respiratory coherence coefficient = 0 and the fluctuation variance is the largest (breathing is most unstable and the amplitude fluctuates most violently), the light flux modulation coefficient = 0, and the light flux is reduced to the minimum to avoid disturbing sleep.

[0091] The sound field envelope modulation coefficient is used to adjust the envelope intensity of the sound field of the sleep lamp. It is positively correlated with the confidence level of the deep sleep stage in the current sleep stage (the higher the confidence level of deep sleep, the smoother the sound field envelope, the lower the intensity, and the closer the modulation coefficient is to 1). The sound field envelope modulation coefficient = deep sleep stage confidence level × sound field calibration coefficient, where the sound field envelope modulation coefficient (value range [0,1]) and the sound field calibration coefficient (preset to 1 to ensure that the modulation coefficient is linearly correlated with the deep sleep confidence level). When the deep sleep stage confidence level = 1 (determined to be a deep sleep stage), the sound field envelope modulation coefficient = 1, the sound field envelope is the smoothest and does not interfere with deep sleep. When the deep sleep stage confidence level = 0 (non-deep sleep stage, i.e., light sleep or wakefulness), the sound field envelope modulation coefficient = 0, the sound field envelope intensity is moderate, and it is suitable for the auditory needs of light sleep or wakefulness. By integrating the luminous flux modulation coefficient and the sound field envelope modulation coefficient, a set of control parameters for the sleep lamp is obtained. This set of parameters is directly output to the sleep lamp control module to achieve adaptive adjustment of the luminous flux and sound field envelope of the sleep lamp, ensuring that the adjustment logic is accurately matched with the human sleep state and breathing state.

[0092] This embodiment, based on existing sleep and breathing-related parameters, obtains luminous flux and sound field envelope modulation coefficients through explicit calculation formulas, avoiding the subjectivity and blindness of sleep lamp adjustment and improving the accuracy of sleep lamp control. The control parameters are deeply bound to the human sleep and breathing state; the luminous flux modulation coefficient is related to breathing stability and breathing amplitude fluctuations, and the sound field envelope modulation coefficient is related to deep sleep confidence, realizing adaptive adjustment of the sleep lamp to meet the needs of different sleep stages and different breathing states. Through quantified control parameters, the adjustment of the sleep lamp is made more scientific, providing a low-interference light and sound environment during deep sleep and adapting to the human body's state during light sleep and wakefulness, thus helping to improve sleep quality.

[0093] In a preferred embodiment of the present invention, step 5 includes:

[0094] Step 500: Extract the luminous flux modulation coefficient and sound field envelope modulation coefficient from the sleep lamp control parameter set, map the luminous flux modulation coefficient to a light drive command, and map the sound field envelope modulation coefficient to a sound drive command. Specifically, this includes: obtaining the sleep lamp control parameter set output in step 4, including the luminous flux modulation coefficient (value range [0,1]) and the sound field envelope modulation coefficient (value range [0,1]), ensuring that both coefficients are within the normalized range. The light drive command is used to control the luminous flux output of the sleep lamp light source module. A linear mapping algorithm is used to map the luminous flux modulation coefficient to the pulse width modulation (PWM) duty cycle of the light drive command, thereby controlling the drive current of the light source module and achieving precise adjustment of the luminous flux. The specific mapping formula is: PWM duty cycle of the light drive command = luminous flux modulation coefficient × maximum PWM duty cycle of the light source module, where the PWM duty cycle of the light drive command (value range [0,1]) and the maximum PWM duty cycle of the light source module (preset to 1, corresponding to the maximum luminous flux output of the light source module). After mapping, the PWM duty cycle is encapsulated into light drive instructions. The instruction format matches the communication protocol of the sleep lamp light source module, ensuring that the light source module can correctly parse and execute it.

[0095] The acoustic drive command controls the acoustic envelope output of the sleep light's acoustic module. It also employs a linear mapping algorithm to map the acoustic envelope modulation coefficient to the driving voltage amplitude of the acoustic module, achieving precise adjustment of the acoustic envelope intensity. The specific mapping formula is: driving voltage amplitude of the acoustic drive command = acoustic envelope modulation coefficient × maximum driving voltage amplitude of the acoustic module, where the maximum driving voltage amplitude of the acoustic module is preset to 5V, corresponding to the maximum acoustic envelope intensity of the acoustic module. After mapping, the driving voltage amplitude is encapsulated into an acoustic drive command. The command format matches the communication protocol of the sleep light's acoustic module, ensuring correct parsing and execution by the acoustic module. After mapping, both the light drive command and the acoustic drive command are obtained.

[0096] Step 501: The light-driven command and sound-driven command are sent to the light source module and sound field module of the sleep lamp, respectively, and the light field feedback signal and sound field feedback signal are collected. Specifically, this includes: through the control unit of the sleep lamp (linked with the signal processing unit of the millimeter-wave radar monitoring system), using serial communication, the light-driven command is sent to the light source module of the sleep lamp, and the sound-driven command is sent to the sound field module of the sleep lamp. The communication baud rate is preset to 9600bps to ensure the stability and real-time performance of command transmission and avoid command loss or delay. After transmission, the control unit waits for the execution confirmation signal from the module to confirm that the light source module and sound field module have successfully received and started executing the command. If no confirmation signal is received, the command is resent to ensure the reliability of command execution. After the light source module and sound field module execute the drive command, they respectively acquire light field feedback signals and sound field feedback signals in real time through the light field sensor and sound field sensor built into the sleep lamp. The light field feedback signal is the actual luminous flux value output by the current light source module, acquired in real time by the light field sensor at a preset sampling frequency of 10Hz to ensure real-time data acquisition. The sound field feedback signal is the actual sound field envelope amplitude output by the current sound field module, acquired in real time by the sound field sensor, also at a preset sampling frequency of 10Hz, consistent with the light field feedback signal acquisition frequency. After acquisition, the light field feedback signal and sound field feedback signal are transmitted to the sleep lamp control unit in real time.

[0097] Step 502: Compare the light field feedback signal with the preset target light field distribution to obtain the light field deviation; compare the sound field feedback signal with the preset target sound field envelope to obtain the sound field deviation; use the light field deviation and sound field deviation to calibrate the driving gain of the light source module and sound field module in real time through a feedback control algorithm; obtain the classification confidence of the current sleep stage; when the classification confidence is lower than a preset threshold, adjust the proportional adjustment coefficient in the sleep lamp control parameter set to form closed-loop control, specifically including:

[0098] Obtain the preset target light field distribution and the preset target sound field envelope, where the preset target light field distribution corresponds to the target luminous flux value (unit: The preset target sound field envelope corresponds to the target sound field envelope value (unit: V). The two target values ​​are preset according to different sleep stages (the target value for light flux during deep sleep is preset to 50). The target value for the sound field envelope is preset to 1V; the target value for luminous flux during light sleep is preset to 80lm and the target value for the sound field envelope is preset to 2V; the target value for luminous flux during wakefulness is preset to 120lm and the target value for the sound field envelope is preset to 3V, precisely matching the current sleep stage. The light field deviation and sound field deviation are calculated separately, where the light field deviation... =|Actual acquired optical field feedback signal - Target luminous flux|; Sound field deviation =|Actual acquired sound field feedback signal - Target value of sound field envelope|.

[0099] A proportional-integral-derivative (PID) feedback control algorithm is employed to calibrate the drive gain of the light source module and the sound field module in real time using the light field deviation and sound field deviation, ensuring that the actual output matches the preset target. The specific calibration process is as follows: for the light source module, the output of the PID feedback control algorithm is the light drive gain calibration value. ,Right now ,in This is the light field scaling factor (default is 0.5). This is the light field integral coefficient (default value is 0.1). This is the differential coefficient of the light field (default value is 0.05). The calibration time (in seconds) is defined by the integral term to eliminate static bias and the derivative term to accelerate the calibration response. After calibration, the optical drive gain calibration value is... The PWM duty cycle superimposed on the original light drive command yields the calibrated light drive command. ,in The PWM duty cycle (value range [0,1]) of the original light drive command obtained in step 500 is used to send the calibrated light drive command to the light source module to realize real-time calibration of luminous flux.

[0100] For the sound field module, the output of the PID feedback control algorithm is the sound drive gain calibration value. ,Right now ,in This is the sound field scaling factor (default is 0.4). This is the sound field integral coefficient (default is 0.08). This is the sound field differential coefficient (default is 0.04). After calibration, the sound drive gain calibration value will be... The voltage amplitude superimposed on the original acoustic drive command yields the calibrated acoustic drive command. ,in The calibrated acoustic driving command is sent to the sound field module to obtain the driving voltage amplitude of the original acoustic driving command mapped in step 500, thereby realizing real-time calibration of the sound field envelope.

[0101] Obtain the classification confidence score (range [0,1]) of the current sleep stage determined in step 100b. This classification confidence score characterizes the reliability of the current sleep stage determination. The closer the classification confidence score is to 1, the more reliable the determination result; the closer it is to 0, the less reliable the determination result. A preset classification confidence score threshold (preset to 0.7, calibrated according to the accuracy of sleep stage determination) is set. The current classification confidence score is compared with the threshold. When the current classification confidence score is greater than or equal to the threshold, it indicates that the current sleep stage determination is reliable, and there is no need to adjust the respiratory stability mapper parameters. The existing parameters are maintained to ensure the continuity of respiratory stability evaluation and sleep light control. When the current classification confidence score is less than the threshold, it indicates that the current sleep stage determination is unreliable, which will lead to a large deviation in the respiratory coherence coefficient output by the respiratory stability mapper, thus affecting the accuracy of the sleep light control parameter set. In this case, the sleep light control parameter set needs to be adjusted. The proportional adjustment coefficient in the sleep lamp control parameter set (the proportional adjustment coefficient in the luminous flux modulation coefficient calculation formula in step 4) is adjusted as follows: original proportional adjustment coefficient × (1 + preset classification confidence threshold - classification confidence of the current sleep stage). The original proportional adjustment coefficient (preset to 0.1) can improve the adaptability of the respiratory stability mapper to the discrimination bias of the sleep stage, ensure the accuracy of the respiratory coherence coefficient calculation, and thus optimize the sleep lamp control parameter set, forming a closed-loop control of monitoring, discrimination, control, feedback, and adjustment, ensuring the stability and accuracy of the entire system.

[0102] In this embodiment, a linear mapping algorithm is used to convert the normalized modulation coefficients into executable driving commands for the module, ensuring the accuracy of luminous flux and sound field envelope adjustment and avoiding adjustment deviations. A feedback control mechanism is introduced, which calculates the deviation by collecting feedback signals and uses a PID algorithm to calibrate the driving gain in real time, solving the problem of execution deviations of the light source and sound field module, ensuring that the actual output is consistent with the preset target, and improving the stability of sleep light control. By triggering the adjustment of respiratory stability mapper parameters through sleep stage classification confidence, a closed-loop control is formed, which adapts to scenarios with sleep stage discrimination deviations and avoids control failures caused by discrimination errors.

[0103] like Figure 2 As shown, an AI sleep light state analysis system for capturing millimeter-wave breathing signals includes:

[0104] The source point construction module is used to extract the respiratory time-domain signal from the pre-processed original radar echo signal of the target area, obtain the respiratory frequency value, and determine the current sleep stage based on the respiratory time-domain signal and respiratory frequency value through a pre-trained AI sleep state classification model; it also extracts the spatial phase difference matrix of the respiratory micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition, and constructs the first and second respiratory source points.

[0105] The computation module is used to map the first and second breathing source points to the Lie algebra manifold space, determine the first and second Lie algebra generators, and obtain the breathing manifold structure tensor by performing Lie bracket operations on the first and second Lie algebra generators.

[0106] The matching module is used to calculate the respiratory phase factor or respiratory symplectic geometric invariant based on the current sleep stage, and input the respiratory phase factor or respiratory symplectic geometric invariant and the respiratory frequency value into the respiratory stability mapper to obtain respiratory stability evaluation parameters.

[0107] The acquisition module is used to calculate the current sleep stage, respiratory stability evaluation parameters, and the variance of the spatial distance between the first and second respiratory sources in a continuous respiratory cycle, in order to construct a set of sleep light control parameters.

[0108] The adjustment module is used to generate light and sound driving commands based on the sleep light control parameter set to adjust the light and sound fields in a coordinated manner; to collect light and sound feedback signals in real time to calibrate the output deviation of the actuator, and to dynamically adjust the proportional adjustment coefficient in the sleep light control parameter set to form a closed-loop control.

[0109] Experimental example:

[0110] This experiment recruited 30 healthy adult subjects for a 14-night sleep monitoring study. The experiment used a self-controlled pre- and post-control design, with the first 7 nights as the baseline period (using a traditional sleep lamp) and the last 7 nights as the experimental period (using the AI ​​sleep lamp system described in this application).

[0111] Experimental conditions: The monitoring environment was a standardized sleep laboratory with a room temperature of 22±1℃, relative humidity of 55%±5%, and ambient noise <30dB. The millimeter-wave radar system adopted a 77GHz FMCW architecture with a transmit power of 12dBm, an antenna array configuration of 4 transmit and 8 receive, and a sampling rate of 1kHz. The AI ​​sleep state classification model was based on a two-level cascaded architecture (first-level random forest + second-level gradient boosting tree), pre-trained on 100,000 labeled samples. The sleep light source used a full-spectrum LED array (color temperature adjustable from 2700K to 6500K), and the sound field output used a directional speaker array (frequency range 20Hz-20kHz).

[0112] Step 1, construct the first and second respiratory source points:

[0113] Specifically, the original radar echo signal was first bandpass filtered (passband 0.1-0.8Hz) to remove heartbeat signals and high-frequency noise interference. Then, an adaptive clutter suppression algorithm was used to eliminate static background reflections. A Fast Fourier Transform (FFT, 30s window, 50% overlap) was performed on the preprocessed signal, and the frequency corresponding to the energy peak was extracted from the frequency domain spectrum as the respiratory rate. Experimental data showed that the average respiratory rate of the subjects before falling asleep was 16.8±2.3 breaths / minute, which decreased to 14.2±1.8 breaths / minute after entering the light sleep stage, and further decreased to 12.5±1.5 breaths / minute during the deep sleep stage.

[0114] A four-dimensional feature vector is constructed by combining the respiratory rate value with the peak-to-trough amplitude difference, zero-crossing rate, and signal envelope change rate of the respiratory time-domain signal. This vector is then input into the first-level random forest classifier (200 base learners, maximum depth 15) of a pre-trained AI sleep state classification model. The first-level classifier outputs a binary classification result of awake / unawakened state and the classification confidence vector for each decision tree. When the initial classification result indicates an unawakened state, the aforementioned features and confidence vectors are input together into a second-level gradient boosting tree classifier (XGBoost, learning rate 0.05, maximum depth 8) to further distinguish between light sleep and deep sleep stages. As shown in Table 1, the classification performance of the proposed method outperforms traditional single-level classification methods in all four sleep stages. Figure 3 As shown, the classification performance of the method in this application is superior to that of traditional single-level classification methods in all four sleep stages.

[0115] Depend on Figure 3 It is evident that the two-level cascaded classification model of this application achieves a recognition accuracy of 96.8% in the awake stage, 94.2% in the light sleep stage, 91.5% in the deep sleep stage, and 93.7% in the REM stage. Compared with the traditional single-level random forest method (overall accuracy of 82.5%), the overall accuracy of the method in this application is improved to 94.2%, with the most significant improvement in the recognition ability of the deep sleep stage (from 76.3% to 91.5%). This is attributed to the specialized optimization of the second-level gradient boosting tree for fine classification of non-awake stages. Simultaneously, the spatial phase difference matrix (dimension 8×8) of respiratory micro-motion signals is extracted from the 8-channel signal of the antenna array, and the direction of arrival is estimated using the MUSIC algorithm to locate the two main respiratory motion source points: the chest and abdomen. The first respiratory source point corresponds to chest breathing (average position deviation ±2.3cm), and the second respiratory source point corresponds to abdominal breathing (average position deviation ±3.1cm). The dual-source point localization accuracy is 37.2% higher than the single-source point method, providing richer spatial information for subsequent manifold analysis.

[0116] Step 2, obtain the respiratory manifold structure tensor:

[0117] Specifically, the position coordinates of the two breathing source points in three-dimensional space are mapped to the Lie algebra space. The Cartesian coordinates are converted into a rotation matrix representation on the Lie group SO(3) through exponential mapping, and then Lie algebra generators are extracted through logarithmic mapping. The first Lie algebra generator represents the rotational motion characteristics of thoracic breathing, and the second Lie algebra generator represents the rotational motion characteristics of abdominal breathing.

[0118] Performing Lie bracket operations on two Lie algebra generators yields the breathing manifold structure tensor. This tensor is a 3×3 symmetric matrix with components... The autocorrelation strength characterizing thoracic breathing. The strength of the autocorrelation characterizing diaphragmatic breathing. = This characterizes the degree of coupling between thoracic and abdominal respiration. Experimental data show that the manifold structure tensor exhibits significant differences across different sleep stages, specifically during wakefulness. =2.12±0.45、 =1.68±0.38、 =0.52±0.15; Light sleep stage =1.45±0.32、 =1.23±0.28、 =0.38±0.12; Deep sleep stage =0.98±0.21、 =0.87±0.19、 =0.25±0.08.

[0119] Figure 4 The Lie algebra norm distributions of 30 subjects at two respiratory sources in a Lie algebra manifold space are presented. The average Lie algebra norm of the first respiratory source is 2.35 ± 0.42, and that of the second respiratory source is 1.87 ± 0.38. The norm difference between the two sources reflects the difference in the dominance of thoracoabdominal breathing in respiratory motion. Through mapping to the Lie algebra manifold space, subtle differences in respiratory motion that were previously difficult to distinguish in Euclidean space are significantly amplified, providing a highly discriminative feature representation for subsequent stability analysis.

[0120] Step 3, Obtain respiratory stability evaluation parameters:

[0121] Specifically, the respiratory phase factor is defined as the instantaneous phase angle of the respiratory time-domain signal in phase space, calculated through Hilbert transform. The respiratory symplectic geometric invariant is extracted from the symplectic form of the respiratory manifold based on symplectic geometric theory, representing the conserved quantities of the respiratory dynamics system. The respiratory stability mapper uses a feedforward neural network structure (3 nodes in the input layer, 64-32-16 nodes in the hidden layer, 1 node in the output layer, and ReLU activation function) to map the respiratory phase factor, symplectic geometric invariant, and respiratory frequency into respiratory stability evaluation parameters. The closer the respiratory stability evaluation parameter value is to 1, the more stable the breathing. Experimental data show that the respiratory stability evaluation parameter is 0.42±0.15 during wakefulness, 0.68±0.12 during light sleep, and 0.89±0.07 during deep sleep.

[0122] Step 4, construct the sleep light control parameter set:

[0123] The sleep light control parameter set contains five core parameters, namely... Encode the sleep stages (wakefulness = 3, light sleep = 1, deep sleep = 0, REM = 2). As a parameter for evaluating respiratory stability, The mean of the spatial distance between the two source points. The variance of the spatial distance between the two source points over 100 consecutive respiratory cycles is given. The comprehensive modulation coefficients include the light flux modulation coefficient and the sound field envelope modulation coefficient.

[0124] The spatial distance between the two source points reflects the degree of coordination of thoracic and diaphragmatic breathing. Experimental data shows that during the waking phase, the spatial distance between the two source points... =0.35±0.08m and the variance of fluctuation is large at =0.012, indicating a light sleep stage. =0.28±0.05m, fluctuation variance is 0.006, deep sleep stage =0.22±0.03m and the variance of fluctuation is minimal, i.e., 0.002. Variance of fluctuation is an important indicator for judging respiratory stability and a key threshold parameter for triggering photoacoustic modulation.

[0125] like Figure 5 The study demonstrated the fluctuations in the spatial distance between the two sources over 100 consecutive respiratory cycles. During the awake phase (cycles 1-30), the distance fluctuated dramatically and without a clear pattern. After entering the light sleep phase (cycles 31-65), the fluctuation amplitude decreased and became periodic. During the deep sleep phase (cycles 66-100), the distance stabilized at a low level with minimal fluctuations. This fluctuation characteristic provides a reliable basis for real-time control of the sleep lamp.

[0126] Step 5, Closed-loop control:

[0127] Specifically, the luminous flux modulation coefficient and sound field envelope modulation coefficient are extracted from the sleep light control parameter set. A linear mapping is used to generate drive commands. These commands are sent to the light source module and sound field module via serial port, and the luminous field feedback signal (actual luminous flux) and sound field feedback signal (actual sound field envelope) are acquired at a sampling rate of 10Hz. The luminous field deviation and sound field deviation are calculated. The target value is preset according to the deep sleep stage (50...). ,1V), light sleep stage (80 ,2V), awake phase (120) (3V). A PID controller is used to calibrate the drive gain, including the light drive gain, the post-calibration command, the sound drive gain, and the post-calibration command. The classification confidence of the current sleep stage is obtained. If it is lower than the threshold of 0.7, the proportional adjustment coefficient is adjusted to = the original proportional adjustment coefficient 0.1 × (1 + the preset classification confidence threshold 0.7 - the classification confidence of the current sleep stage) to adapt to the discrimination bias. Experimental data show that the steady-state error of the closed-loop calibration system is <2%, the response time is <300ms, and the overshoot is <8%, meeting the real-time control requirements.

[0128] Through experimental verification of the above embodiments, the AI ​​sleep lamp state analysis method for capturing millimeter-wave respiratory signals described in this application has the following beneficial effects:

[0129] A two-level cascaded AI sleep state classification model accurately identifies four sleep stages: wakefulness, light sleep, deep sleep, and REM sleep, achieving an overall accuracy of 94.2%, a 14.2% improvement over traditional single-level classification methods. This provides a reliable state perception foundation for differentiated control of sleep lights. Through Lie algebraic manifold space mapping and Lie bracket operations, the spatial motion characteristics of the two respiratory sources are transformed into quantifiable manifold structure tensors, effectively characterizing the coupling relationship and dynamic evolution of thoracoabdominal breathing, providing a high-dimensional feature representation for respiratory stability analysis. By jointly calculating the respiratory phase factor, symplectic geometric invariants, and a respiratory stability mapper, continuous respiratory stability evaluation parameters are constructed, smoothly reflecting the gradual change process of sleep states, avoiding the jump problems of discrete classification, and improving the comfort of control. By integrating a multi-dimensional set of control parameters, including sleep stage, respiratory stability, distance between the two sources, and their fluctuation variance, intelligent linkage adjustment of the light and sound fields was achieved. The photoacoustic linkage response delay was as low as 285ms, and the closed-loop calibration convergence time was only 12.5 seconds, which is superior to the 620ms and 28.0 seconds of traditional methods. Through a closed-loop calibration mechanism with real-time feedback signals, the proportional adjustment coefficient was dynamically adjusted, effectively compensating for the output deviation of the actuator and ensuring the precise execution of photoacoustic control.

[0130] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for analyzing the state of an AI sleep lamp by capturing millimeter-wave breathing signals, characterized in that, The method includes: Step 1: Extract the respiratory time-domain signal from the preprocessed original radar echo signal of the target area, obtain the respiratory frequency value, and determine the current sleep stage based on the respiratory time-domain signal and respiratory frequency value using a pre-trained AI sleep state classification model; extract the spatial phase difference matrix of the respiratory micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition, and construct the first respiratory source point and the second respiratory source point. Step 2: Map the first and second breathing source points to the Lie algebra manifold space to determine the first and second Lie algebra generators; obtain the breathing manifold structure tensor by performing Lie bracket operations on the first and second Lie algebra generators. Step 3: Calculate the respiratory phase factor or respiratory symplectic geometric invariant based on the current sleep stage. Input the respiratory phase factor or respiratory symplectic geometric invariant along with the respiratory frequency value into the respiratory stability mapper to obtain the respiratory stability evaluation parameters. Step 4: Calculate the current sleep stage, respiratory stability evaluation parameters, and the variance of the spatial distance between the first and second respiratory sources during continuous respiratory cycles to construct a set of sleep light control parameters. Step 5: Generate light and sound driving commands based on the sleep light control parameter set to adjust the light and sound fields in a coordinated manner; collect light and sound feedback signals in real time to calibrate the output deviation of the actuator, and dynamically adjust the proportional adjustment coefficient in the sleep light control parameter set to form a closed-loop control.

2. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 1, characterized in that, The acquisition of the respiratory rate value includes: The respiratory time-domain signal is subjected to a fast Fourier transform to obtain the respiratory frequency domain spectrum, and the frequency corresponding to the energy peak value is extracted from the respiratory frequency domain spectrum as the respiratory frequency value.

3. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 2, characterized in that, The current sleep stage is determined using a pre-trained AI sleep state classification model, including: The respiratory rate value and the peak-to-valley amplitude difference, zero-crossing rate and signal envelope change rate of the respiratory time domain signal are used to form a feature vector, which is then input into the first-level random forest classifier of the pre-trained AI sleep state classification model. The feature vector is initially classified to obtain the preliminary classification result of whether the current sleep stage belongs to the awake stage or the non-awake stage, and the classification confidence vector composed of the classification confidence output by each decision tree. When the initial classification result is a non-awake stage, the respiratory rate value, the peak-to-valley amplitude difference of the respiratory time domain signal, the zero-crossing rate, the signal envelope change rate, and the classification confidence vector of each decision tree are jointly input into the second-level gradient boosting tree classifier of the pre-trained AI sleep state classification model to classify the non-awake stage and obtain the classification result of whether the current sleep stage is a light sleep stage or a deep sleep stage.

4. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 3, characterized in that, Extract the spatial phase difference matrix of the breathing micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition, and construct the first breathing source point and the second breathing source point, including: The range-dimensional fast Fourier transform is performed on the breathing micro-motion signals received by each antenna element to extract the complex values ​​of each antenna element in the target range gate. The conjugate product of the complex values ​​of each antenna element in the target range gate is calculated to obtain the spatial phase difference between different pairs of antenna elements in the antenna array. The spatial phase differences between all antenna element pairs are arranged in the order of the antenna elements to form a spatial phase difference matrix; each spatial phase difference in the spatial phase difference matrix is ​​substituted into the steering vector of the multi-signal classification algorithm to construct a spatial spectrum function. A two-dimensional peak search is performed on the spatial spectral function in the azimuth and pitch dimensions. The first azimuth and first pitch angles corresponding to the maximum peak response are extracted. Combined with the radial distance corresponding to the target distance gate, the first respiratory source point is calculated in the spatial coordinate system. The second azimuth and second pitch angles corresponding to the second maximum peak response are extracted. Combined with the radial distance corresponding to the target distance gate, the second respiratory source point is calculated in the spatial coordinate system. The first respiratory source point corresponds to the spatial position point corresponding to the maximum chest wall displacement at the end of expiration, and the second respiratory source point corresponds to the spatial position point corresponding to the maximum chest wall displacement at the end of inspiration.

5. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 4, characterized in that, Step 2 includes: The first respiratory source point and the second respiratory source point are respectively represented as the first coordinate vector and the second coordinate vector in the spatial coordinate system. The first coordinate vector and the second coordinate vector are mapped to the Lie algebra manifold space. The tangent space of the first coordinate vector is represented as the first Lie algebra generator, and the tangent space of the second coordinate vector is represented as the second Lie algebra generator. Perform a combined multiplication operation on the first Lie algebra generator and the second Lie algebra generator to obtain the first intermediate tensor. Perform a combined multiplication operation on the second Lie algebra generator and the first Lie algebra generator to obtain the second intermediate tensor. Subtract the first intermediate tensor from the second intermediate tensor to obtain the commutator tensor. The commutator tensor is subjected to index raising and lowering operations under the action of the metric tensor of the Lie algebra manifold space to obtain the nonlinear coupling strength tensor between the first and second breathing source points, which is used as the breathing manifold structure tensor.

6. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 5, characterized in that, Based on the current sleep stage, calculate the respiratory phase factor or the respiratory symplectic geometric invariant, including: When the current sleep stage is light sleep, the spatial phase difference matrix is ​​used as the basic data of the manifold to be processed, and the respiratory manifold structure tensor is used as the Riemannian metric. The respiratory manifold structure tensor is assigned to the manifold structure where the spatial phase difference matrix is ​​located, and the Riemannian manifold space is constructed. The respiratory manifold structure tensor is decomposed into eigenvalues, and the eigenvector direction corresponding to the largest eigenvalue is extracted as the partitioning reference direction. In the Riemannian manifold space, the spatial phase difference matrix is ​​used as the tensor to be translated. The tensor to be translated is moved in parallel along the partitioning reference direction, and the tensor obtained after translation is used as the respiratory phase connection matrix. The respiratory phase communication matrix is ​​decomposed into eigenvalues ​​to extract the first principal eigenvalue with the largest modulus and the second principal eigenvalue with the second largest modulus. The ratio of the first principal eigenvalue to the second principal eigenvalue is calculated and used as the respiratory phase factor. When the current sleep stage is deep sleep, the respiratory manifold structure tensor is used as the symplectic potential function. The components of the respiratory manifold structure tensor are rearranged according to the symplectic form of the symplectic geometric space to construct an even-dimensional respiratory phase space. In the respiratory phase space, the exterior differentiation operation is performed on the symplectic potential function to obtain the symplectic structure metric matrix. The symplectic structure metric matrix is ​​decomposed into Singlasman manifolds, and then into multiple mutually symplectic orthogonal subspaces according to the symplectic orthogonal complement relation. Each subspace corresponds to a symplectic subspace cluster. For two adjacent symplectic subspace clusters, the minimum symplectic inner product angle from any vector in one symplectic subspace cluster to the symplectic orthogonal complement of the other symplectic subspace cluster is calculated, and the minimum symplectic inner product angle is taken as the symplectic orthogonal complement angle between the two symplectic subspace clusters. The symplectic orthogonal complement angles between all symplectic subspace clusters are arranged in the topological order of the symplectic subspace clusters to form a sequence, and the sequence is taken as the breathing symplectic geometric invariant. When the current sleep stage is the waking stage, the respiratory phase factor is assigned a value of 1, and the respiratory symplectic geometric invariant is assigned a zero vector to represent the state of non-periodic respiratory geometric features.

7. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 6, characterized in that, The respiratory stability mapper includes a criterion module, a logarithmic transformation module, a linear filtering module, and an exponential transformation module. The criterion module determines whether the current input is a marker of the awake stage; if so, it directly outputs a preset stability benchmark value; if not, the logarithmic transformation module performs a logarithmic domain transformation on the product of the respiratory phase factor or the norm of the respiratory symplectic geometric invariant and the respiratory frequency value to obtain the transformed signal. The linear filtering module performs a weighted moving average process on the transformed signal in the logarithmic domain to obtain the filtered result. The exponential transformation module performs exponential restoration on the filtered result to obtain the respiratory coherence coefficient with a value range normalized to the [0, 1] interval, and uses the respiratory coherence coefficient as a parameter for evaluating respiratory stability.

8. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 7, characterized in that, The sleep lamp control parameter set includes a light flux modulation coefficient and a sound field envelope modulation coefficient. The light flux modulation coefficient is positively correlated with the respiratory coherence coefficient and negatively correlated with the fluctuation variance. The sound field envelope modulation coefficient is positively correlated with the confidence level of the deep sleep stage in the current sleep stage.

9. The AI ​​sleep light state analysis method based on millimeter-wave respiratory signal capture according to claim 8, characterized in that, Step 5 includes: The luminous flux modulation coefficient and the sound field envelope modulation coefficient are extracted from the control parameter set of the sleep light. The luminous flux modulation coefficient is mapped to the light driving command, and the sound field envelope modulation coefficient is mapped to the sound driving command. The light-driven command and the sound-driven command are sent to the light source module and the sound field module of the sleep lamp respectively, and the light field feedback signal and the sound field feedback signal are collected. The light field feedback signal is compared with the preset target light field distribution to obtain the light field deviation. The sound field feedback signal is compared with the preset target sound field envelope to obtain the sound field deviation. Using the light field deviation and sound field deviation, the driving gain of the light source module and the sound field module is calibrated in real time through a feedback control algorithm. The classification confidence of the current sleep stage is obtained. When the classification confidence is lower than the preset threshold, the parameters in the breathing stability mapper are adjusted to form a closed-loop control.

10. An AI sleep lamp state analysis system for capturing millimeter-wave respiratory signals, wherein the system is applied in the method as described in any one of claims 1 to 9, characterized in that, include: The source point construction module is used to extract the respiratory time-domain signal from the pre-processed original radar echo signal of the target area, obtain the respiratory frequency value, and determine the current sleep stage based on the respiratory time-domain signal and respiratory frequency value through a pre-trained AI sleep state classification model; it also extracts the spatial phase difference matrix of the respiratory micro-motion signal of the antenna array corresponding to the original radar echo signal acquisition, and constructs the first and second respiratory source points. The computation module is used to map the first and second breathing source points to the Lie algebra manifold space, determine the first and second Lie algebra generators, and obtain the breathing manifold structure tensor by performing Lie bracket operations on the first and second Lie algebra generators. The matching module is used to calculate the respiratory phase factor or respiratory symplectic geometric invariant based on the current sleep stage, and input the respiratory phase factor or respiratory symplectic geometric invariant and the respiratory frequency value into the respiratory stability mapper to obtain respiratory stability evaluation parameters. The acquisition module is used to calculate the current sleep stage, respiratory stability evaluation parameters, and the variance of the spatial distance between the first and second respiratory sources in a continuous respiratory cycle, in order to construct a set of sleep light control parameters. The adjustment module is used to generate light and sound driving commands based on the sleep light control parameter set to adjust the light field and sound field in a coordinated manner. Real-time acquisition of light and sound feedback signals is used to calibrate the output deviation of the actuator, and the proportional adjustment coefficient in the control parameter set of the sleep light is dynamically adjusted to form a closed-loop control.

Images