Millimeter wave and ai cooperative sleep lamp sleep state perception system

Abstract

The application provides a millimeter wave and AI cooperative sleep lamp sleep state sensing system, and relates to the technical field of smart home control, and comprises an acquisition module used for emitting a millimeter wave signal to a bed body area through a millimeter wave radar and receiving a reflected echo to obtain an original echo signal; performing electromagnetic interference filtering on the original echo signal to eliminate electromagnetic interference generated by an LED lamp bead driving circuit to obtain a first filtered signal; and performing multipath reflection suppression on the first filtered signal to eliminate multipath reflection clutter caused by a lampshade material to obtain a second filtered signal. The application improves the stability and accuracy of sleep sensing and realizes intelligent cooperative control in a whole sleep cycle.

Description

Technical Field

[0001] This invention relates to the field of smart home control technology, and in particular to a sleep lamp sleep state sensing system that combines millimeter wave and AI. Background Technology

[0002] With the increasing popularity of smart homes and non-contact physiological monitoring technology, smart sleep lamps with integrated sleep sensing functions are gradually becoming the mainstream products of smart health terminals in the bedroom. Most existing smart sleep lamps attempt to use millimeter-wave radar to achieve non-contact sleep state monitoring. They can detect human breathing, body movement and other characteristics through radar echo signals and control the lighting without the need for wearable devices.

[0003] For example, a smart sleep aid lamp with built-in millimeter-wave radar integrates millimeter-wave radar, LED beads, metal heat dissipation structure, and plastic lampshade into the same lamp body. It can achieve basic human presence detection and simple sleep stage determination, and adjust the brightness and color temperature of the light accordingly. However, when the millimeter-wave radar is working, it is affected by electromagnetic interference generated by the LED bead driving circuit and multipath reflection noise caused by the lampshade material. At the same time, the metal heat dissipation structure of the lamp body will cause uneven spatial attenuation of the millimeter-wave signal, which will interfere with and mask the weak sleep characteristics such as human breathing and heartbeat in the echo signal, making it difficult to accurately extract effective physiological characteristics. Ultimately, this leads to large errors in sleep stage division and distortion of related monitoring data. Summary of the Invention

[0004] This invention provides a sleep state sensing system for a sleep lamp that integrates millimeter wave and AI, improving the stability and accuracy of sleep sensing and enabling intelligent collaborative control throughout the entire sleep cycle.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a sleep state sensing system for sleep lamps that integrates millimeter waves and AI includes:

[0007] The acquisition module is used to transmit millimeter-wave signals to the bed area via millimeter-wave radar and receive reflected echoes to obtain the original echo signal; to perform electromagnetic interference filtering on the original echo signal to eliminate electromagnetic interference generated by the LED lamp bead driving circuit to obtain the first filtered signal; and to perform multipath reflection suppression on the first filtered signal to eliminate multipath reflection clutter caused by the lamp cover material to obtain the second filtered signal.

[0008] The calculation module is used to perform metal shielding compensation on the second filtered signal to correct the signal attenuation caused by the metal heat dissipation structure of the lamp body and obtain the corrected echo signal. The metal shielding compensation includes collecting the spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertex of the inner wall of the lamp cover, and multiple auxiliary points on the surface of the lamp body. By minimizing the sum of squared distances from each point to the fitted sphere, a virtual reference sphere is fitted. The virtual reference sphere is then meshed to generate a two-dimensional spatial attenuation compensation matrix covering the bed area.

[0009] The calibration module is used to perform spatial point-by-point calibration on the calibrated echo signal using a two-dimensional spatial attenuation compensation matrix, and to extract weak feature signals reflecting the human sleep state from the calibrated echo signal.

[0010] The processing module is used to input the extracted weak feature signals into the pre-trained AI sleep state recognition algorithm, and output the current sleep stage, breathing rate, and number of times the user turns over. Based on the output sleep stage, it generates control commands to link the smart bedroom devices to execute corresponding sleep linkage strategies.

[0011] Furthermore, millimeter-wave signals are transmitted to the bed area via millimeter-wave radar, and reflected echoes are received to obtain the original echo signal; the original echo signal is then subjected to electromagnetic interference filtering to eliminate electromagnetic interference generated by the LED bead driving circuit, resulting in a first filtered signal, including:

[0012] Collect the operating current waveform of the LED bead driver circuit and use the operating current waveform as a reference interference signal;

[0013] The original echo signal and the reference interference signal are input into an adaptive filter. The adaptive filter dynamically updates the filter coefficients based on the least squares criterion to cancel out the interference components related to the reference interference signal from the original echo signal, thus obtaining the preliminary filtered signal.

[0014] The initial filtered signal is subjected to band-stop filtering to remove fixed-band interference caused by the switching frequency of the driving circuit and its harmonics, thus obtaining the first filtered signal.

[0015] Furthermore, multipath reflection suppression is applied to the first filtered signal to eliminate multipath reflection clutter caused by the lampshade material, resulting in a second filtered signal, including:

[0016] A range-dimensional fast Fourier transform is performed on the first filtered signal to obtain a range image sequence. Based on the geometric positional relationship between the millimeter-wave radar and the inner wall of the lampshade, the range cell corresponding to each reflection point on the inner wall of the lampshade in the range image is calculated, and the time-domain echo component in the range cell is extracted as the original estimate of the multipath reflection of the lampshade.

[0017] The original estimate of the multipath reflection of the lampshade is convolved with the pre-measured reflectance coefficient of the lampshade material to reconstruct the time-domain waveform of the multipath interference; the time-domain waveform of the multipath interference is subtracted from the first filtered signal to obtain the preliminary multipath suppression signal.

[0018] The initial multipath suppression signal is processed by spatial adaptive beamforming. By pointing the main lobe of the beam to the bed area and forming a null in the angular direction of the lamp cover, residual lamp cover directional clutter is filtered out to obtain the second filtered signal.

[0019] Furthermore, metal shielding compensation is applied to the second filtered signal to correct the signal attenuation caused by the metal heat dissipation structure of the lamp body, resulting in a corrected echo signal. Metal shielding compensation involves acquiring the spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertices of the inner wall of the lamp cover, and multiple auxiliary points on the lamp body surface. By minimizing the sum of squared distances from each point to the fitted sphere, a virtual reference sphere is fitted. The virtual reference sphere is then meshed to generate a two-dimensional spatial attenuation compensation matrix covering the bed area, including:

[0020] Metal shielding compensation is applied to the second filtered signal. The spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertex of the inner wall of the lamp cover, and multiple auxiliary points on the surface of the lamp body are collected by a three-dimensional scanning device to form an initial point set.

[0021] By setting the objective function of spherical fitting to minimize the sum of squared distances from each point to the fitted sphere, the initial point set is iteratively optimized to determine the coordinates of the center of the virtual reference sphere and the radius of the sphere, thus obtaining the virtual reference sphere.

[0022] The virtual reference sphere is divided into equally spaced latitude and longitude grids in the bed area. Each grid node corresponds to a spatial position in the bed area. Based on the relative geometric relationship between the spatial position and the millimeter-wave radar, the theoretical attenuation value caused by the metal heat dissipation structure is calculated. The theoretical attenuation value is used as the attenuation compensation coefficient of the grid node.

[0023] The attenuation compensation coefficients of each node are arranged in the row and column order of the grid to construct a two-dimensional spatial attenuation compensation matrix covering the bed area.

[0024] Furthermore, the virtual reference sphere is divided into equally spaced latitude and longitude grids across the bed area, with each grid node corresponding to a spatial location within the bed area. Based on the relative geometric relationship between the spatial location and the millimeter-wave radar, the theoretical attenuation value caused by the metal heat dissipation structure's obstruction is calculated. This theoretical attenuation value is then used as the attenuation compensation coefficient for the grid node, including:

[0025] A spherical coordinate system is established with the center of the virtual reference sphere as the origin. The latitude and longitude range of the bed area under the spherical coordinate system is determined. The latitude and longitude range is divided into equal intervals according to the preset spatial resolution to generate the latitude and longitude coordinates of each grid node.

[0026] The latitude and longitude coordinates of each grid node are converted into three-dimensional spatial coordinates to obtain the actual spatial position of the grid node in the bed area;

[0027] Based on the three-dimensional spatial coordinates of the millimeter-wave radar and the three-dimensional spatial coordinates of the grid nodes, a straight spatial path from the radar to the grid nodes is constructed.

[0028] Determine whether the spatial straight path intersects with the virtual reference sphere. If they intersect, calculate the path's crossing distance within the sphere and convert the crossing distance into a theoretical attenuation value based on the attenuation characteristics of electromagnetic waves in a metallic medium. If they do not intersect, set the theoretical attenuation value to one. Use the calculated theoretical attenuation value as the attenuation compensation coefficient for the mesh nodes.

[0029] Furthermore, the corrected echo signal is spatially calibrated point-by-point using a two-dimensional spatial attenuation compensation matrix. Weak feature signals reflecting human sleep states are then extracted from the calibrated echo signal, including:

[0030] The corrected echo signal is subjected to distance-dimensional fast Fourier transform and angle-dimensional fast Fourier transform to generate a distance-angle spectrum. Each pixel in the distance-angle spectrum corresponds to a spatial location in the bed area.

[0031] The spatial position of each pixel in the distance-angle spectrum is mapped to the grid nodes of the two-dimensional spatial attenuation compensation matrix, and the attenuation compensation coefficient corresponding to the pixel position is calculated by bilinear interpolation.

[0032] The amplitude of the corresponding pixel of the corrected echo signal is compensated point by point by the attenuation compensation coefficient to obtain the amplitude-calibrated echo signal; the amplitude-calibrated echo signal is subjected to phase unwrapping processing to extract the phase change information caused by the rise and fall of the human chest cavity;

[0033] Bandpass filtering was applied to the phase change information to separate the respiratory frequency component and the heart rate component. Moving target detection was performed on multiple consecutive frames of distance-angle spectrum to extract the micro-Doppler features corresponding to turning over and body movement.

[0034] By fusing respiratory rate components, heart rate components, and micro-Doppler features corresponding to turning over and body movement, a weak feature signal reflecting the human sleep state is constructed.

[0035] Furthermore, the extracted weak feature signals are input into a pre-trained AI sleep state recognition algorithm, which outputs the current sleep stage and indicators such as breathing rate and number of times the user turns over, including:

[0036] Weak feature signals are segmented by a sliding window to obtain feature sequences corresponding to multiple time windows. Each feature sequence contains respiratory rate components, heart rate components, and micro-Doppler features.

[0037] The feature sequence of each time window is input into a pre-trained AI sleep stage classifier and an AI breathing and turning regressor, which are trained based on machine learning algorithms;

[0038] The AI ​​sleep stage classifier outputs the probability distribution of the sleep stage to which the current time window belongs, and the AI ​​breathing and turning regressor outputs the respiratory rate value and the number of turnings detected for the current time window.

[0039] The probability distribution of sleep stages in multiple consecutive time windows is weighted and smoothed, and the stage with the highest probability is taken as the final current sleep stage. Kalman filtering is applied to the respiratory rate and turning frequency values ​​in multiple consecutive time windows to remove random noise and obtain stable respiratory rate and turning frequency indices.

[0040] Furthermore, based on the output sleep stage, control commands are generated to trigger corresponding sleep-related strategies to be executed by smart bedroom devices, including:

[0041] The system acquires the current sleep stage, as well as the respiratory rate and number of times the user turns over. Based on the current sleep stage, it matches the corresponding device control parameters in a preset linkage strategy library. The linkage strategy library stores the mapping relationship between different sleep stages and light brightness, color temperature, sleep aid sound type and volume, and smart curtain opening degree.

[0042] If the current sleep stage is the sleep induction period, the first control command is generated to gradually dim the brightness of the sleep light according to a linear decreasing curve, switch the color temperature to a warm yellow temperature, and trigger the sleep aid speaker to play preset white noise.

[0043] If the current sleep stage is deep sleep maintenance, a second control command is generated to turn off the brightness of the sleep light or adjust it to the lowest brightness, reduce the volume of the sleep aid sound to the preset background volume, and detect the status of the smart curtains. If the curtains are not closed, a closing command is generated.

[0044] If the current sleep stage is light sleep and the number of times the sleeper turns over exceeds a preset threshold, a third control command is generated to fine-tune the sleep lamp to a low-light mode and adjust the frequency of the sleep aid sound to assist in the transition to sleep.

[0045] If the current sleep stage is a waking period and the respiratory rate index exceeds the wake-up threshold, a fourth control command is generated to gradually increase the brightness of the sleep lamp to daytime mode according to the preset wake-up time, while stopping the sleep aid sound.

[0046] In a second aspect, a computing device includes:

[0047] One or more processors;

[0048] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to execute the system.

[0049] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, performs the system.

[0050] The above-described solution of the present invention has at least the following beneficial effects:

[0051] By performing electromagnetic interference filtering and multipath reflection suppression on the original echo signal, and by using a metal shielding compensation method that collects the spatial coordinates of relevant points on the lamp body, fits a virtual reference sphere, and generates a two-dimensional spatial attenuation compensation matrix, combined with a pre-trained AI sleep state recognition algorithm and sleep linkage strategy, this method overcomes the technical problems of existing sleep lamps where millimeter-wave radar is susceptible to electromagnetic interference from LED lamp bead driving circuits, multipath reflection noise from the lampshade, and signal attenuation caused by the lamp body's metal heat dissipation structure. These problems lead to inaccurate extraction of weak sleep features, large errors in sleep stage segmentation, distorted monitoring data, and insufficient intelligent linkage. As a result, this method achieves accurate extraction of weak human sleep features, accurate output of sleep stages, breathing rate, number of turns, and other indicators, realizing high-precision and high-stability sleep state perception, and linking with smart bedroom devices to complete adaptive and collaborative control throughout the entire sleep cycle. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of a sleep state sensing system for a sleep lamp that integrates millimeter wave and AI, provided by an embodiment of the present invention.

[0053] Figure 2 This is a schematic diagram illustrating the process of a sleep state sensing system for a millimeter wave and AI-integrated sleep lamp, provided by an embodiment of the present invention. The system performs point-by-point spatial calibration on the corrected echo signal using a two-dimensional spatial attenuation compensation matrix and extracts weak feature signals reflecting the human sleep state from the calibrated echo signal.

[0054] Figure 3 This is a statistical diagram illustrating the signal-to-noise ratio of signals at different filtering stages.

[0055] Figure 4This is a statistical diagram illustrating the accuracy of different sleep stages.

[0056] Figure 5 This is a diagram showing the comparison of sleep quality indicators before and after the system implementation. Detailed Implementation

[0057] 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, embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey its scope to those skilled in the art.

[0058] like Figure 1 As shown, an embodiment of the present invention proposes a sleep state sensing system for a sleep lamp that integrates millimeter wave and AI, comprising:

[0059] The acquisition module is used to transmit millimeter-wave signals to the bed area via millimeter-wave radar and receive reflected echoes to obtain the original echo signal; to perform electromagnetic interference filtering on the original echo signal to eliminate electromagnetic interference generated by the LED lamp bead driving circuit to obtain the first filtered signal; and to perform multipath reflection suppression on the first filtered signal to eliminate multipath reflection clutter caused by the lamp cover material to obtain the second filtered signal.

[0060] The calculation module is used to perform metal shielding compensation on the second filtered signal to correct the signal attenuation caused by the metal heat dissipation structure of the lamp body and obtain the corrected echo signal. The metal shielding compensation includes collecting the spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertex of the inner wall of the lamp cover, and multiple auxiliary points on the surface of the lamp body. By minimizing the sum of squared distances from each point to the fitted sphere, a virtual reference sphere is fitted. The virtual reference sphere is then meshed to generate a two-dimensional spatial attenuation compensation matrix covering the bed area.

[0061] The calibration module is used to perform spatial point-by-point calibration on the calibrated echo signal using a two-dimensional spatial attenuation compensation matrix, and to extract weak feature signals reflecting the human sleep state from the calibrated echo signal.

[0062] The processing module is used to input the extracted weak feature signals into the pre-trained AI sleep state recognition algorithm, and output the current sleep stage, breathing rate, and number of times the user turns over. Based on the output sleep stage, it generates control commands to link the smart bedroom devices to execute corresponding sleep linkage strategies.

[0063] In this embodiment of the invention, electromagnetic interference filtering and multipath reflection suppression are sequentially applied to the original echo signal. Spatial coordinates of key points on the lamp body are acquired, and a virtual reference sphere is fitted by minimizing the sum of squared distances from each point to the fitted sphere. The virtual reference sphere is then meshed to generate a two-dimensional spatial attenuation compensation matrix covering the bed area to achieve metal shielding compensation. Based on this matrix, the signal is spatially calibrated point-by-point, and weak feature signals reflecting the human sleep state are extracted. These extracted weak feature signals are then input into a pre-trained AI sleep state recognition algorithm to identify the sleep state and generate control commands to link the layered collaboration of smart bedroom devices. This technological approach overcomes the technical problems of existing sleep lamp integrated millimeter-wave sensing solutions, which are susceptible to electromagnetic interference from LED lamp bead driving circuits, multipath reflection noise from the lampshade, and signal attenuation from the metal heat dissipation structure of the lamp body. These problems lead to inaccurate extraction of weak sleep features, large errors in judging sleep stages and physiological indicators, and a lack of adaptability to smart device linkage. This approach achieves deep purification and precise calibration of millimeter-wave echo signals, stably and completely extracts weak sleep features such as human breathing and body movement, accurately outputs core indicators such as sleep stages, breathing frequency, and number of turns, and enables adaptive and collaborative control between bedroom smart devices and real-time sleep status.

[0064] In a preferred embodiment of the present invention, a millimeter-wave signal is transmitted to the bed area via a millimeter-wave radar, and the reflected echo is received to obtain the original echo signal; the original echo signal is then subjected to electromagnetic interference filtering to eliminate electromagnetic interference generated by the LED bead driving circuit, resulting in a first filtered signal, comprising:

[0065] The process involves acquiring the operating current waveform of the LED bead driver circuit and using it as a reference interference signal. Specifically, this includes: real-time acquisition of the operating current of the LED bead driver circuit; setting the signal sampling frequency of the millimeter-wave radar to 1000 Hz; the inherent switching frequency of the LED bead driver circuit to 100 Hz; and extending its electromagnetic interference harmonic components to 500 Hz. Through a current sampling unit connected in series in the driver circuit, the operating current waveform is continuously acquired for 1 second. The acquired discrete current data sequence completely characterizes the time-domain features of the electromagnetic interference generated by the driver circuit, and this operating current waveform is directly used as a reference interference signal.

[0066] The original echo signal and the reference interference signal are input into an adaptive filter. The adaptive filter dynamically updates the filter coefficients based on the least squares criterion to cancel out the interference components related to the reference interference signal from the original echo signal and obtain the preliminary filtered signal. Specifically, the original echo signal received by the millimeter-wave radar and the generated reference interference signal are synchronously input into an 8th-order adaptive filter, and all eight initial filter coefficients of the filter are set to 0.1.

[0067] The original echo signal has a value of x(n) at the current moment, and the reference interference signal has the following values ​​at the current moment and the previous 7 moments: , , , , , , , The eight adaptive filter coefficients are as follows: The interference cancellation signal output by the filter is y(n), the error signal is e(n), the iteration step size is μ=0.01, and the adaptive filter performs calculations according to the least squares criterion: = + + + + + + + ;

[0068] Calculate the error signal: ;

[0069] The filter coefficients are updated in real time with the goal of minimizing the sum of squares of the error signals. The update formula is as follows: ;

[0070] After each signal sampling is completed, the 8 sets of filter coefficients are dynamically updated according to the above formula. By continuously iterating and optimizing the filter coefficients, the interference components in the original echo signal that are highly correlated with the reference interference signal are gradually eliminated, and the preliminary filtered signal is output after the initial interference suppression is completed.

[0071] The initial filtered signal is subjected to band-stop filtering to remove fixed-band interference caused by the switching frequency of the driving circuit and its harmonics, resulting in the first filtered signal. Specifically, this includes: performing second-order infinite-length unit impulse response band-stop filtering on the initial filtered signal; setting the stopband range of the band-stop filter to 90 Hz to 510 Hz for the 100 Hz switching frequency of the LED lamp bead driving circuit and its 200 Hz, 300 Hz, 400 Hz, and 500 Hz harmonic components; controlling the passband signal attenuation value to within 1 dB; and setting the stopband signal attenuation value to above 40 dB. The filtering calculation process is as follows: first, the initial filtered signal is converted into a frequency domain signal through a fast Fourier transform; the signal amplitude in the 90 Hz to 510 Hz range in the frequency domain is multiplied by a suppression coefficient of 0.01, while the signal amplitude in the passband range remains unchanged; then, the processed frequency domain signal is converted back to a time domain signal through an inverse fast Fourier transform to remove the fixed-band interference generated by the driving circuit, ultimately obtaining a pure first filtered signal.

[0072] In this embodiment of the invention, the operating current waveform of the LED bead driving circuit is collected as a reference interference signal. The original echo signal and the reference interference signal are input into an adaptive filter with dynamically updated filter coefficients based on the least squares criterion to cancel out relevant interference components. Then, the preliminary filtered signal is subjected to band-stop filtering to filter out fixed-band interference composed of the switching frequency and harmonics of the driving circuit. Therefore, this invention overcomes the technical problem that electromagnetic interference generated by the LED bead driving circuit is easily mixed into the original echo signal, resulting in low purity and complex interference components in the original echo signal, which affects the accuracy of subsequent signal processing. Thus, it effectively eliminates various electromagnetic interferences brought by the LED bead driving circuit and obtains a pure and stable first filtered signal.

[0073] In a preferred embodiment of the present invention, multipath reflection suppression is applied to the first filtered signal to eliminate multipath reflection clutter caused by the lampshade material, resulting in a second filtered signal, including:

[0074] A range-dimensional Fast Fourier Transform (FFT) is performed on the first filtered signal to obtain a range image sequence. Based on the geometric positional relationship between the millimeter-wave radar and the inner wall of the lampshade, the range cells corresponding to each reflection point on the inner wall of the lampshade in the range image are calculated, and the time-domain echo components within the range cells are extracted as the original estimate of the lampshade's multipath reflection. Specifically, the signal bandwidth of the millimeter-wave radar is set to 1 GHz, the signal sampling frequency is set to 2 GHz, and a 128-point range-dimensional FFT is performed on the first filtered signal to convert the time-domain first filtered signal into a frequency-domain range image sequence. Each cell in the range image sequence corresponds to a fixed detection distance in space. The range resolution is determined by the radar signal bandwidth, and its calculation formula is as follows: , among which, among which The speed of light is 3 × 10⁻⁶. 8 meters per second The radar signal bandwidth is taken as 1 GHz; substituting the value into the formula, the range resolution can be obtained. =0.15 meters. Through preliminary structural calibration, the vertical distance between the millimeter-wave radar and the inner wall of the lampshade is 0.3 meters. The range cell number corresponding to each reflection point on the inner wall of the lampshade in the range image is calculated by dividing the actual distance by the range resolution: Range cell number = actual distance to the inner wall of the lampshade ÷ range resolution = 0.3 meters ÷ 0.15 meters = 2. In the second range cell of the range image sequence, all time-domain echo components corresponding to that cell are extracted. This component is the original estimate of the multipath reflection of the lampshade, completing the accurate positioning and preliminary extraction of lampshade clutter.

[0075] The original estimate of the multipath reflection of the lampshade is convolved with the pre-measured reflectance coefficient of the lampshade material to reconstruct the time-domain waveform of the multipath interference. The time-domain waveform of the multipath interference is subtracted from the first filtered signal to obtain the preliminary multipath suppression signal. Specifically, this involves: pre-measuring the reflectance coefficient of the plastic lampshade material used in this embodiment as 0.4 through experiments; convolving the obtained original estimate of the lampshade multipath reflection with this reflectance coefficient to reconstruct the complete time-domain waveform of the lampshade multipath interference. The convolution calculation formula is as follows: ,in, The reconstructed time-domain waveform of multipath interference. The original estimate of multipath reflection for the lampshade. The reflectance coefficient of the lampshade material is set to 0.4. After waveform reconstruction, time-domain clutter cancellation is performed. The initial multipath suppression signal is equal to the first filtered signal minus the multipath interference time-domain waveform. This operation directly eliminates the main interference components of the lampshade multipath reflection in the first filtered signal and outputs the initial multipath suppression signal.

[0076] The initial multipath suppression signal undergoes spatial adaptive beamforming processing. This involves pointing the main lobe of the beam towards the bed area and creating a null in the direction of the lampshade to filter out residual clutter in the lampshade direction, resulting in a second filtered signal. Specifically, this process uses a millimeter-wave radar with a 4-element uniform linear array structure to perform spatial adaptive beamforming processing on the initial multipath suppression signal. Through prior spatial calibration, the center angle of the radar beam corresponding to the bed area is 0 degrees, and the angle corresponding to the lampshade area is ±30 degrees. The null depth for beamforming is set to -40dB. During processing, the weighting coefficients of the antenna array are adjusted in real time to precisely point the main lobe of the beam towards the 0-degree bed area, ensuring that the effective echo signal generated by the human body passes through without attenuation. Simultaneously, a signal null with a depth of -40dB is created in the ±30-degree direction of the lampshade, completely suppressing residual clutter in the lampshade direction. After spatial adaptive beamforming processing, all residual clutter caused by the lampshade is filtered out, ultimately yielding a second filtered signal free from multipath reflection interference.

[0077] In this embodiment of the invention, the original estimate of multipath reflection of the lampshade is extracted by performing a range-dimensional fast Fourier transform on the first filtered signal and combining the geometric positional relationship between the millimeter-wave radar and the inner wall of the lampshade. This original estimate is then convolved with the reflection coefficient of the lampshade material to reconstruct the time-domain waveform of the multipath interference, which is then canceled out from the first filtered signal. The preliminary multipath suppression signal is then subjected to spatial adaptive beamforming processing, and the main lobe of the beam is pointed towards the bed area, forming a null in the lampshade direction. This overcomes the technical problem that multipath reflection clutter caused by the lampshade material has strong correlation and residual interference, and conventional suppression methods cannot eliminate such clutter, which easily masks the weak echo signal of the human body in the bed area. Thus, it achieves accurate reconstruction and deep suppression of multipath reflection clutter of the lampshade, filters out residual clutter in the lampshade direction, improves the signal-to-noise ratio of the echo signal, and completely preserves the effective echo of the human body in the bed area.

[0078] In a preferred embodiment of the present invention, the second filtered signal is compensated for by metal shielding to correct the signal attenuation caused by the metal heat dissipation structure of the lamp body, thereby obtaining the corrected echo signal. The metal shielding compensation includes acquiring the spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertices of the inner wall of the lamp cover, and multiple auxiliary points on the lamp body surface. A virtual reference sphere is obtained by minimizing the sum of squared distances from each point to the fitted sphere. The virtual reference sphere is then meshed to generate a two-dimensional spatial attenuation compensation matrix covering the bed area, including:

[0079] Metal shielding compensation is applied to the second filtered signal. Spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertices of the inner wall of the lampshade, and multiple auxiliary points on the lamp body surface are acquired using a 3D scanning device to form an initial point set. Specifically, before applying metal shielding compensation to the second filtered signal, the spatial coordinates of key structures of the lamp body are acquired using a 3D scanning device. The measurement accuracy of the 3D scanning device is set to 0.01 mm, enabling precise positioning of feature points on the lamp body surface. Three types of core feature points—the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, and the vertices of the inner wall of the lampshade—are acquired sequentially. Simultaneously, eight auxiliary detection points are evenly selected on the lamp body surface, resulting in the acquisition of a total of 11 feature points' 3D spatial coordinates. In this embodiment, the phase center of the millimeter-wave radar antenna is used as the origin of the 3D coordinate system. The 11 acquired feature points together constitute the initial point set for metal shielding compensation. All coordinate units are unified to meters, providing a realistic spatial data foundation for subsequent virtual reference spherical fitting.

[0080] By setting the objective function of spherical fitting to minimize the sum of squared distances from each point to the fitted sphere, the initial point set is iteratively optimized to determine the coordinates of the center and the radius of the virtual reference sphere, thus obtaining the virtual reference sphere. Specifically, this step uses minimizing the sum of squared distances from each spatial point in the initial point set to the fitted sphere as the optimization objective, and solves for the coordinates of the center and the radius of the virtual reference sphere through an iterative optimization algorithm. The standard equation of any sphere in space is: in,( , , Let be the coordinates of the center of the sphere to be solved. Let be the radius of the sphere to be solved, ( , , ) represents the three-dimensional spatial coordinates of any feature point in the initial point set.

[0081] The objective function for spherical fitting is defined as the sum of squared distances from all feature points to the fitted sphere. The expression for the objective function is: in, The index of the feature points in the initial point set, with a value range from 1 to 11. , , ) is the first The three-dimensional spatial coordinates of each feature point are used to minimize the objective function, with an iteration convergence threshold of 1×10⁻⁶. −6 The maximum number of iterations is 50. In each iteration, the distance from each feature point to the currently fitted sphere is recalculated, and the center coordinates and radius values ​​are continuously updated until the objective function value is less than the convergence threshold, at which point the iteration terminates. Through iterative optimization, the center coordinates of the virtual reference sphere are finally determined to be (0, 0, 0.05) meters, and the radius of the sphere is 0.12 meters, thus completing the fitting of the virtual reference sphere.

[0082] The virtual reference sphere is divided into equally spaced latitude and longitude grids over the bed area. Each grid node corresponds to a spatial location within the bed area. Based on the relative geometric relationship between the spatial location and the millimeter-wave radar, the theoretical attenuation value caused by the metal heat dissipation structure's obstruction is calculated. This theoretical attenuation value is used as the attenuation compensation coefficient for the grid nodes. Specifically, this involves establishing a spherical coordinate system with the center of the virtual reference sphere as the origin, defining the range of the bed area within the spherical coordinate system as follows: azimuth from -60 degrees to +60 degrees, elevation from 0 degrees to 45 degrees, and equally spaced latitude and longitude ranges of the bed area with a spatial resolution of 5 degrees. A uniformly distributed grid of nodes is generated, each uniquely corresponding to an actual spatial location within the bed area. After converting the latitude and longitude coordinates of each grid node into three-dimensional spatial coordinates, a straight spatial path is constructed from the origin of the millimeter-wave radar coordinates to that grid node. It is then determined whether this straight path intersects with a virtual reference sphere. If the path intersects with the virtual reference sphere, the traversal distance within the sphere is calculated. Finally, the theoretical attenuation value is calculated based on the attenuation characteristics of millimeter waves in metallic media. In this embodiment, the attenuation coefficient of millimeter waves in the metal heat dissipation structure is set to 0.8 dB per centimeter. The formula for calculating the theoretical attenuation value is as follows: in, This is the theoretical attenuation value. The distance the path traverses within the virtual reference sphere. The attenuation coefficient for the metallic medium is used to convert the theoretical attenuation value into a linear domain attenuation compensation coefficient. If the path does not intersect with the virtual reference sphere, the attenuation compensation coefficient of the grid node is directly set to 1.0.

[0083] The attenuation compensation coefficients of each node are arranged in the row and column order of the grid to construct a two-dimensional spatial attenuation compensation matrix covering the bed area. Specifically, this involves: traversing all latitude and longitude grid nodes within the bed area, and arranging the attenuation compensation coefficients corresponding to each grid node in sequence according to the rule that the azimuth angle corresponds to the matrix row number and the elevation angle corresponds to the matrix column number. After partitioning, a total of 25 rows and 10 columns of grid nodes are obtained, corresponding to the generation of a 25-row and 10-column two-dimensional spatial attenuation compensation matrix. Each element in the matrix is ​​the attenuation compensation coefficient corresponding to the spatial position of the bed. After filling all the attenuation compensation coefficients into the corresponding positions in the matrix, the construction of the two-dimensional spatial attenuation compensation matrix covering the entire bed area is completed.

[0084] In this embodiment of the invention, an initial point set is formed by collecting the spatial coordinates of multiple key points on the lamp body. The virtual reference sphere is iteratively optimized with the goal of minimizing the sum of the squared distances from each point to the fitted sphere. The virtual reference sphere is divided into equally spaced latitude and longitude grids, and the corresponding theoretical attenuation values ​​are calculated. Then, a two-dimensional spatial attenuation compensation matrix covering the bed area is constructed to perform metal shielding compensation for the second filtered signal. Therefore, this method overcomes the technical problems of uneven spatial attenuation of millimeter-wave signals and distortion of echo signal amplitude caused by the metal heat dissipation structure of the lamp body, which cannot truly reflect the human body reflected signals. Thus, it achieves accurate correction of signal attenuation and realizes precise point-by-point spatial compensation of echo signals in the bed area, ensuring the authenticity and integrity of the signal.

[0085] In a preferred embodiment of the present invention, the virtual reference sphere is divided into equally spaced latitude and longitude grids in the bed area, with each grid node corresponding to a spatial location in the bed area. Based on the relative geometric relationship between the spatial location and the millimeter-wave radar, the theoretical attenuation value caused by the metal heat dissipation structure blocking is calculated, and the theoretical attenuation value is used as the attenuation compensation coefficient of the grid node, including:

[0086] A spherical coordinate system is established with the center of the virtual reference sphere as the origin. The latitude and longitude range of the bed area under the spherical coordinate system is determined. The latitude and longitude range is then divided into equal intervals according to a preset spatial resolution to generate the latitude and longitude coordinates of each grid node. Specifically, this includes: establishing a spherical coordinate system with the center of the virtual reference sphere as the origin. The azimuth angle of the spherical coordinate system corresponds to the longitude in latitude and longitude, with a value range of -180 degrees to +180 degrees. The pitch angle corresponds to the latitude in latitude and longitude, with a value range of -9 degrees. From 0 degrees to +90 degrees, the azimuth angle is defined as 0 degrees on the horizontal axis pointing directly forward from the origin, with clockwise as positive. The elevation angle is defined as 0 degrees on the vertical axis pointing directly upward from the origin, with downward as positive. Considering the actual use scenario in a bedroom, the bed area is located directly below and around the millimeter-wave radar inside the sleep light. After preliminary spatial calibration, the latitude and longitude range of the bed area in this spherical coordinate system is determined to be: azimuth angle from -60 degrees to +60 degrees, and elevation angle from 30 degrees to 60 degrees. This range can completely cover the area. A standard 1.8m x 2.0m bed area ensures no blind spots in sleep monitoring. The preset spatial resolution is 5 degrees. The latitude and longitude range of the bed area is divided into equal intervals according to the rule of dividing the azimuth and pitch angles into nodes every 5 degrees. The azimuth angle starts at -60 degrees and increases by 5 degrees each time until it reaches 60 degrees. Subtracting -60 degrees from 60 degrees gives a total span of 120 degrees. Dividing this by the 5-degree interval and adding 1 results in 25 azimuth angles. The pitch angle starts at 30 degrees and increases by 5 degrees each time until it ends at 60 degrees. Subtracting 30 degrees from 60 degrees gives a total span of 30 degrees. Dividing this by the 5-degree interval and adding 1 gives 7 pitch angle nodes. All azimuth and pitch angle nodes are matched one-to-one to generate 175 evenly distributed grid nodes in 25 rows and 7 columns. Each grid node corresponds to a unique actual spatial location within the bed area. The azimuth and pitch angle values ​​of each grid node are recorded.

[0087] The latitude and longitude coordinates of each grid node are converted into three-dimensional spatial coordinates to obtain the actual spatial position of the grid node in the bed area. Specifically, this step converts the azimuth and pitch angle values ​​of each grid node into the horizontal, vertical, and triangular axis values ​​of three-dimensional spatial rectangular coordinates. The conversion process is completed based on the conversion relationship between the spherical coordinate system and the rectangular coordinate system. Combined with the parameters of the virtual reference sphere, it is ensured that the converted coordinates can accurately correspond to the actual position of the bed. During the conversion, the azimuth and pitch angles of the grid node are first converted from angle values ​​to radians. The conversion method is to multiply the angle value by pi and then divide by 180. The three-dimensional horizontal axis coordinate is calculated by multiplying the radius of the virtual reference sphere by the cosine value corresponding to the pitch angle in radians, and then multiplying it by the sine value corresponding to the azimuth angle in radians. The three-dimensional vertical axis coordinate is calculated by multiplying the radius of the virtual reference sphere by the cosine value corresponding to the pitch angle in radians, and then multiplying it by the cosine value corresponding to the azimuth angle in radians. The three-dimensional vertical axis coordinate is calculated by multiplying the radius of the virtual reference sphere by the sine value corresponding to the pitch angle in radians, and then adding the vertical axis coordinate value of the center of the virtual reference sphere.

[0088] Taking a grid node with an azimuth angle of 0 degrees and an elevation angle of 45 degrees as an example, the angles are first converted to radians. 0 degrees is converted to 0 radians, and 45 degrees is converted to approximately 0.785 radians. Then, the coordinates of each axis are calculated sequentially. The horizontal axis coordinate is 0.12 meters multiplied by the cosine of 0.785 radians and then multiplied by the sine of 0 radians, resulting in 0 meters. The vertical axis coordinate is 0.12 meters multiplied by the cosine of 0.785 radians and then multiplied by the cosine of 0 radians, resulting in approximately 0.085 meters. The vertical axis coordinate is 0.12 meters multiplied by the sine of 0.785 radians and then multiplied by 0.05 meters, resulting in approximately 0.135 meters. Finally, the three-dimensional spatial coordinates of this grid node are 0 meters, 0.085 meters, and 0.135 meters.

[0089] Using the same calculation method, the latitude and longitude coordinates of all 175 grid nodes were converted to three-dimensional spatial coordinates in sequence. Each converted three-dimensional coordinate corresponds to an actual spatial location in the bed area, ensuring the accuracy of subsequent path construction.

[0090] Based on the three-dimensional spatial coordinates of the millimeter-wave radar and the three-dimensional spatial coordinates of the grid nodes, a spatial straight-line path from the radar to the grid nodes is constructed. Specifically, the phase center of the millimeter-wave radar antenna is fixed at the origin of the three-dimensional rectangular coordinate system, with coordinates of 0 meters, 0 meters, and 0 meters. Combining this with the three-dimensional spatial coordinates of each grid node obtained in the previous step, a spatial straight-line path from the millimeter-wave radar to the corresponding grid node is constructed. The spatial straight-line path starts at the phase center of the radar antenna and ends at the three-dimensional coordinates of the grid node. These two points determine a unique spatial straight line. The straight line is the path of the millimeter-wave signal propagating from the radar to the corresponding position on the bed, and it is also a key basis for determining whether the metal heat dissipation structure blocks the signal. To accurately describe the path, parameters are used to represent any position on the path. The parameters are set to... The value ranges from 0 to 1. The three-dimensional horizontal axis coordinate of any point on the path is equal to the parameter multiplied by the horizontal axis coordinate of the grid node, the vertical axis coordinate is equal to the parameter multiplied by the vertical axis coordinate of the grid node, and the vertical axis coordinate is equal to the parameter multiplied by the vertical axis coordinate of the grid node. When the parameter is 0, it corresponds to the radar starting point, and when the parameter is 1, it corresponds to the grid node ending point. This provides a calculation basis for subsequent judgment on whether the path intersects with the virtual reference sphere. Taking the grid nodes of 0 meters, 0.085 meters, and 0.135 meters in the previous step as an example, the horizontal axis coordinate of any point on its spatial straight path is always 0, the vertical axis coordinate is the parameter multiplied by 0.085 meters, and the vertical axis coordinate is the parameter multiplied by 0.135 meters. The parameter takes values ​​between 0 and 1, clearly showing the propagation trajectory of the millimeter wave signal from the radar to the position of the bed.

[0091] Determine whether a straight path in space intersects with a virtual reference sphere. If they intersect, calculate the path's crossing distance within the sphere and convert it into a theoretical attenuation value based on the attenuation characteristics of electromagnetic waves in metallic media. If they do not intersect, set the theoretical attenuation value to one. Use the calculated theoretical attenuation value as the attenuation compensation coefficient for the mesh nodes. Specifically, this involves: determining whether the path intersects with the virtual reference sphere; calculating the difference in three-dimensional coordinates between points on the path and the center of the virtual reference sphere; squaring and summing these differences; and comparing the result with the square of the sphere's radius to determine if they intersect. The specific calculation formula is as follows: ,in, For basic calculation values, , , These are the three-dimensional spatial coordinates of the grid nodes; ,in, These are intermediate calculation values. The vertical axis coordinate of the grid node; Fixed calculation values; in, The operation is a discriminant value; if If >0, the path intersects with the virtual reference sphere; if If the value is ≤0, the path is determined not to intersect with the virtual reference sphere. Taking grid nodes with coordinates of 0 meters, 0.085 meters, and 0.135 meters as examples, the result is obtained by substituting the values ​​into the above formula. If the value is greater than 0, the path is determined to intersect with the virtual reference sphere.

[0092] The crossing distance is calculated when the path intersects the sphere. Coordinate calculations determine the two valid intersection points where the path crosses the sphere, and the straight-line distance between these two points is the crossing distance. The calculation formula is as follows: ,in, For the distance traveled, The difference in proportion between the two intersection points. , , These are the three-dimensional spatial coordinates of the grid nodes. Taking the above grid nodes as an example, the effective intersection point position ratio difference is 0.998. Substituting this into the formula, the result is 15.9 cm.

[0093] attenuation coefficient of millimeter waves in metal heat dissipation structures =0.8 dB / cm, the theoretical attenuation value is calculated using the following formula: in, This is the theoretical attenuation value. For the distance traveled, Let be the attenuation coefficient of millimeter waves in a metal heat dissipation structure. Substituting this into the numerical calculation, the result is 12.72 dB. The attenuation value in the decibel domain is converted into the attenuation compensation coefficient in the linear domain. The conversion formula is as follows: ,in, This is the attenuation compensation coefficient. The theoretical attenuation value is 4.33. If the path does not intersect with the virtual reference sphere, the attenuation compensation coefficient is directly set to 1.0. Following the above formula and calculation method, the intersection judgment, crossing distance calculation and attenuation compensation coefficient calculation of all 175 grid nodes are completed in sequence. Each grid node corresponds to a unique attenuation compensation coefficient.

[0094] In this embodiment of the invention, a spherical coordinate system is established with the center of a virtual reference sphere as the origin. The latitude and longitude range of the bed area is divided into grid nodes at equal intervals. The latitude and longitude coordinates of the nodes are converted into three-dimensional spatial coordinates, and a straight path for the radar to reach the corresponding spatial position is constructed. The crossing distance is calculated based on the intersection of the path with the virtual reference sphere. The corresponding theoretical attenuation value is obtained as the compensation coefficient by combining the attenuation characteristics of electromagnetic waves in the metal medium. Therefore, this invention overcomes the technical problems of uneven attenuation of millimeter-wave signals, low accuracy of compensation coefficient calculation, and inaccurate correspondence between spatial position and attenuation caused by the difference in shielding of metal heat dissipation structure at different spatial positions of the bed. This achieves precise quantification of the signal attenuation degree at each spatial position and realizes refined and accurate calculation of the attenuation compensation coefficient.

[0095] like Figure 2 As shown, in another preferred embodiment of the present invention, the corrected echo signal is spatially calibrated point-by-point using a two-dimensional spatial attenuation compensation matrix, and weak feature signals reflecting the human sleep state are extracted from the calibrated echo signal, including:

[0096] The corrected echo signal is subjected to range-dimensional Fast Fourier Transform (FFT) and angle-dimensional FFT to generate a range-angle spectrum. Each pixel in the range-angle spectrum corresponds to a spatial location in the bed region. Specifically, the process involves performing a two-dimensional Fast Fourier Transform (FFT) in both range and angle dimensions on the corrected echo signal to convert the time-domain echo signal into a frequency-domain range-angle spectrum, thereby achieving accurate matching between the spatial location of the bed region and the echo signal. The range-dimensional FFT points are set to 128 points, and the angle-dimensional FFT points are set to 64 points. The millimeter-wave radar signal sampling frequency is 2 GHz, the signal bandwidth is 1 GHz, the radar wavelength is 0.005 meters, and the corresponding radar operating frequency is 60 GHz.

[0097] When performing a range-dimensional Fast Fourier Transform (FFT), a 128-point range-dimensional FFT is performed on the corrected echo signal, and the range resolution is calculated using the following formula: In the formula The speed of light is 3 × 10⁻⁶. 8 meters per second Given the radar signal bandwidth as 1 GHz, the range resolution is 0.15 meters, meaning that each 0.15 meters in the range direction is divided into a range unit.

[0098] When performing an angle-dimensional Fast Fourier Transform, a 64-point angle-dimensional Fast Fourier Transform is performed on the signal after the distance-dimensional transform, and the angle resolution is calculated using the formula: In the formula Let 64 be the number of points for the Fast Fourier Transform in the angular dimension. Substituting the value, we can obtain an angular resolution of approximately 2.8125°, meaning that the angular direction is divided into one angular unit every 2.8125°.

[0099] After a two-dimensional Fourier transform, a distance-angle spectrum with a size of 128×64 is generated. Each pixel in the spectrum uniquely corresponds to a spatial location within the bed area, which is the intersection of the distance unit and the angle unit. The pixel amplitude represents the echo signal intensity at that spatial location.

[0100] The spatial position of each pixel in the distance-angle spectrum is mapped to the grid nodes of the two-dimensional spatial attenuation compensation matrix. The attenuation compensation coefficient corresponding to the pixel position is calculated by bilinear interpolation. Specifically, the number of pixels in the distance-angle spectrum is 128×64, which does not match the number of grid nodes of the two-dimensional spatial attenuation compensation matrix of 25×7. Therefore, they cannot be directly mapped one-to-one. Thus, the spatial position of each pixel in the distance-angle spectrum needs to be mapped to the grid nodes of the compensation matrix, and the accurate attenuation compensation coefficient corresponding to the pixel is calculated by bilinear interpolation.

[0101] When performing spatial location mapping, first determine the spatial azimuth and elevation angles corresponding to each pixel in the distance-angle spectrum, and match these angle parameters with the angle parameters of the compensation matrix grid nodes. The azimuth angle range of the grid nodes is -60° to 60° with an interval of 5°, and the elevation angle range is 30° to 60° with an interval of 5°. Find the four nearest grid nodes around the pixel and record the attenuation compensation coefficient and spatial coordinates of the nodes.

[0102] When performing bilinear interpolation, the attenuation compensation coefficient of the target pixel is obtained by weighting the compensation coefficients of four adjacent grid nodes. The calculation formula is as follows: ; In the formula The target pixel attenuation compensation coefficient. , , , These are the compensation coefficients for four adjacent nodes, where x and y are the relative distance ratios of the target pixels in the horizontal and vertical directions, respectively, and both range from 0 to 1.

[0103] Taking a pixel with an azimuth angle of 0° and an elevation angle of 45° as an example, the compensation coefficients of its four surrounding nodes are 4.33, 4.28, 4.25, and 4.20, respectively, with a horizontal relative ratio of 0.2 and a vertical relative ratio of 0.3. Substituting these values ​​into the formula, the attenuation compensation coefficient for this pixel is calculated to be 4.296. This method is used to complete the mapping and compensation coefficient calculation for all 8192 pixels within the distance-angle spectrum, ensuring that each pixel has a corresponding accurate compensation coefficient.

[0104] The amplitude of the corresponding pixel of the corrected echo signal is compensated point by point by the attenuation compensation coefficient to obtain the amplitude-calibrated echo signal; the phase unwrapping processing of the amplitude-calibrated echo signal is performed to extract the phase change information caused by the rise and fall of the human chest cavity. Specifically, this step first completes the amplitude calibration of the echo signal point by point by the attenuation compensation coefficient to eliminate the amplitude distortion caused by the metal heat dissipation structure, and then performs phase unwrapping on the calibration signal to extract the phase change information caused by the rise and fall of the human chest cavity.

[0105] When performing amplitude point-by-point calibration, a pixel-by-pixel amplitude compensation method is adopted. The original amplitude of each pixel in the distance-angle spectrum is multiplied by the corresponding attenuation compensation coefficient to obtain the amplitude-calibrated echo signal. The calculation formula is as follows: In the formula The amplitude after pixel calibration. The original amplitude of the pixel. The pixel attenuation compensation coefficient is used as an example. Taking a pixel with a compensation coefficient of 4.296 and an original amplitude of 0.3 as an example, the amplitude after calibration is 1.2888 when substituted into the formula. The amplitude calibration of all pixels is completed in sequence to obtain an echo signal with uniform amplitude and no distortion.

[0106] When performing phase unwrapping and phase change information extraction, the echo signal after amplitude calibration suffers from phase wrapping, and the phase value is limited to − to The least squares phase unwrapping method needs to be used to handle this, and the unwrapping formula is as follows: In the formula The phase value after untangling. The original winding phase value, To ensure the continuity of the phase, integers, The value is 3.1416. After unwrapping, 100 frames of amplitude-calibrated echo signals are continuously acquired. The phase difference between each frame and the first frame is calculated to obtain the phase change sequence, which is the phase change information caused by the rise and fall of the human chest cavity.

[0107] Bandpass filtering is applied to the phase change information to separate the respiratory and heart rate components. Moving target detection is performed on multiple consecutive frames of range-angle spectra to extract micro-Doppler features corresponding to turning over and body movements. Specifically, this involves separating the respiratory and heart rate-related frequency components from the phase change information through bandpass filtering, and then performing moving target detection based on multiple consecutive frames of range-angle spectra to extract the micro-Doppler features corresponding to turning over and body movements. This achieves complete separation of various weak physiological features of the human body. In the process of separating the respiratory and heart rate components, combined with human physiological characteristics, the normal range of respiratory rate is 0.1 to 0.3 Hz, and the normal range of heart rate is 0.8 to 2.0 Hz. Based on this, two second-order infinite impulse response bandpass filters are specifically designed. The bandpass filter for respiratory rate has a passband range of 0.1 to 0.3 Hz. It utilizes the phase change information from the current moment and the previous two moments, combined with the filtering outputs from the previous two moments, to perform filtering operations and ultimately extract the pure respiratory rate component. The bandpass filter for heart rate has a passband range of 0.8 to 2.0 Hz, employing the same filtering logic to extract a stable heart rate component from the phase change information. After separation processing by the two sets of filters, the respiratory rate obtained in this embodiment is stabilized at 0.2 Hz, and the heart rate is stabilized at 1.2 Hz.

[0108] During the extraction of micro-Doppler features, thirty frames of distance-angle spectrum data were continuously acquired, with an interval of 0.1 seconds between frames and a total acquisition time of 3 seconds. For the same pixel location, the amplitude difference between the current frame and the previous frame after amplitude calibration was calculated. This difference was used to determine whether human movement existed. A threshold of 0.5 was set for the amplitude difference. When the calculated difference exceeded this threshold, it was determined that there was turning over or body movement in the bed area corresponding to that pixel. For the detected movement area, the corresponding micro-Doppler frequency was calculated by combining the actual human movement speed and the radar wavelength. This frequency was used as the micro-Doppler feature corresponding to the body movement. Taking a turning speed of 0.01 meters per second as an example, the calculated micro-Doppler frequency was 4 Hz. The micro-Doppler features corresponding to all body movements were extracted sequentially using this method, ultimately forming a complete micro-Doppler feature sequence.

[0109] The respiratory rate component, heart rate component, and micro-Doppler features corresponding to turning over and body movement are fused to form a weak feature signal reflecting the human sleep state. Specifically, the separated respiratory rate component, heart rate component, and extracted body movement micro-Doppler features are weighted and fused to construct a weak feature signal that can comprehensively reflect the human sleep state. The fusion weights are set to 0.3 for the respiratory rate component, 0.3 for the heart rate component, and 0.4 for the micro-Doppler features, with a total weight of 1.0.

[0110] The feature fusion calculation formula is as follows: In the formula, S(n) is a weak characteristic signal. For respiratory rate components, For heart rate components, The value is a micro-Doppler eigenvalue. When there is no body movement, the value is 0. Taking the state without body movement as an example, substituting the corresponding frequency value, the signal value is 0.42. When there is a turning movement, the signal value is 2.02. According to this fusion formula, the data of all continuously collected moments are processed to generate a continuous weak feature signal sequence. This sequence completely contains the sleep state features such as human breathing, heartbeat, and body movement. It is free from interference and distortion, and provides high-quality and high-reliability feature input for subsequent AI sleep state recognition.

[0111] In this embodiment of the invention, a distance-angle spectrum is generated by performing a fast Fourier transform on the corrected echo signal in both distance and angle dimensions. The pixel positions in the spectrum are mapped to a compensation matrix, and the corresponding attenuation compensation coefficient is calculated by bilinear interpolation to complete point-by-point compensation of the signal amplitude. Then, the phase change information is extracted by unwrapping the signal, and the respiratory and heart rate components are separated by bandpass filtering. The body motion-related micro-Doppler features are extracted by moving target detection and feature fusion is performed. Therefore, the technical problems of insufficient spatial calibration accuracy of echo signals, difficulty in phase information analysis, and difficulty in effectively separating and extracting weak physiological features of the human body are overcome. Thus, the technical problems of accurately completing signal spatial calibration, stably extracting weak physiological features related to human sleep, and constructing a complete and reliable sleep state feature signal are achieved.

[0112] In a preferred embodiment of the present invention, the extracted weak feature signals are input into a pre-trained AI sleep state recognition algorithm, which outputs the current sleep stage and indicators such as respiratory rate and number of times the user turns over, including:

[0113] The weak feature signal is segmented by a sliding window to obtain feature sequences corresponding to multiple time windows. Each feature sequence contains respiratory rate components, heart rate components, and micro-Doppler features. Specifically, the generated continuous weak feature signal is segmented by a sliding window to divide the continuous signal sequence into multiple independent time windows. Each time window corresponds to a fixed duration of sleep data, ensuring that the feature sequence in each window fully reflects the human sleep characteristics within that time period.

[0114] Core parameter settings: Based on the changing patterns of human sleep characteristics, features such as breathing and body movement are relatively stable in a short period of time. A window that is too long may miss details of body movement, while a window that is too short is susceptible to noise interference. The core parameters of the sliding window are set as follows: Window duration: 5 seconds, meaning each time window corresponds to 5 seconds of weak feature signals; Sliding step size: 2 seconds, meaning the time interval between two adjacent time windows is 2 seconds, ensuring continuous signal coverage and avoiding feature omissions; Weak feature signal sampling frequency: 10Hz, corresponding to a frame interval of 0.1 seconds, with 10 weak feature signal values ​​collected every second. Therefore, the number of weak feature signal samples contained in each time window is: 5 seconds × 10Hz = 50 samples.

[0115] Segmentation Process and Calculation: Sliding window segmentation uses a gradual sliding method with a fixed window duration. The formula for calculating the number of segments is as follows: ,in, The total number of time windows after segmentation. The total duration of the weak characteristic signal is set to 60 seconds in this embodiment, corresponding to a complete 1-minute sleep monitoring data. The window duration is 5 seconds. With a sliding step of 2 seconds, the result obtained from numerical calculation is that the 28 weak feature signals, which take 60 seconds, can be divided into 28 time windows, each containing a feature sequence of 50 samples.

[0116] Feature sequence composition: The feature sequence of each time window fully contains three core features: respiratory rate component, heart rate component, and micro-Doppler feature corresponding to the 50 samples in that window. The three types of features correspond one-to-one and are matched synchronously to ensure that the feature data of each window is complete, comprehensive, and accurately reflects the human sleep state during that time period. For example: The first time window corresponds to the weak feature signal of 0-5 seconds, containing 50 samples, each of which corresponds to a set of respiratory rate component, heart rate component, and micro-Doppler feature data; the second time window corresponds to the weak feature signal of 2-7 seconds, also containing 50 samples, and so on, until all 28 windows are segmented, ensuring that the entire 60 seconds of sleep data is complete without omission or redundancy.

[0117] The feature sequences of each time window are input into a pre-trained AI sleep stage classifier and an AI breathing and turning regressor. These classifiers and regressors are trained based on machine learning algorithms. Specifically, the process involves inputting the feature sequences of each time window into the pre-trained AI sleep stage classifier and AI breathing and turning regressor. Both the classifier and regressor are AI models built based on machine learning algorithms, specifically adapted to the sleep monitoring scenario of this invention. They can accurately process weak feature signals and output corresponding recognition and regression results. Considering the requirements for sleep state recognition and monitoring indicator regression, a random forest algorithm is used to train the AI ​​sleep stage classifier and the AI ​​breathing and turning regressor models respectively. The specific training parameters are as follows:

[0118] AI Sleep Stage Classifier: 100 decision trees, maximum decision tree depth 10, minimum number of splits 5, minimum number of leaf nodes 2, training set accuracy 92%, validation set accuracy 90%. This AI model is specifically designed to identify human sleep stages, and the output results are divided into four stages: wakefulness, light sleep, deep sleep, and REM sleep.

[0119] AI Respiratory and Turning Regressor: 80 decision trees, maximum decision tree depth 8, minimum number of sample splits 4, minimum number of sample leaf nodes 1, training set regression error 0.02Hz respiratory rate, 0.1 turns per minute, validation set regression error 0.03Hz respiratory rate, 0.1 turns per minute. This AI model is specifically designed to regress and output respiratory rate values ​​and turns per minute detection values.

[0120] Feature sequence input process: The 50 sample feature sequences of each time window are simultaneously input into the sleep stage classifier and the breathing and turning regressor in the order of respiratory rate component, heart rate component, and micro-Doppler features. During input, the feature sequences are normalized preprocessed to ensure that the feature data range is consistent and to avoid the model output accuracy being affected by the difference in feature magnitude. For example, the 50 sample feature sequences of the first time window (0-5 seconds) are simultaneously input into the two AI models. The classifier begins to identify the sleep stage corresponding to the window, and the regressor begins to calculate the respiratory rate and number of turnings corresponding to the window. The feature input and model calculation of all 28 time windows are completed in sequence to ensure that each window has a corresponding preliminary output result.

[0121] The AI ​​sleep stage classifier outputs the probability distribution of the sleep stage to which the current time window belongs, and the AI ​​breathing and turning regressor outputs the respiratory rate value and turning number detection value corresponding to the current time window. Specifically, the AI ​​sleep stage classifier and the AI ​​breathing and turning regressor output the preliminary results for each time window. The classifier outputs the probability distribution of the sleep stage, and the regressor outputs the specific respiratory rate value and turning number detection value. All results correspond one-to-one with the time window.

[0122] The AI ​​sleep stage classifier outputs the following results: For each time window, the classifier outputs the probability distribution of the human body being in four stages: wakefulness, light sleep, deep sleep, and REM sleep. The sum of the probabilities of the four stages is 1.0. The higher the probability value, the higher the probability that the human body is in that sleep stage within that time window. For example: The first time window (0-5 seconds) outputs the following probability distribution: wakefulness 0.1, light sleep 0.7, deep sleep 0.1, REM sleep 0.1, indicating that the human body is most likely in the light sleep stage during this time period; The second time window (2-7 seconds) outputs the following probability distribution: wakefulness 0.05, light sleep 0.75, deep sleep 0.15, REM sleep 0.05, further confirming that the human body is in the light sleep stage during this time period; The eighth time window (14-19 seconds) outputs the following probability distribution: wakefulness 0.02, light sleep 0.18, deep sleep 0.7, REM sleep 0.1, indicating that the human body is most likely in the deep sleep stage during this time period.

[0123] The AI-powered respiration and turning regressor outputs the following results: For each time window, the regressor outputs preliminary detection values ​​for two specific monitoring indicators: respiratory rate (Hz) and the number of turns (times). The number of turns within each window is a non-negative integer or decimal, with decimals representing incomplete turning movements. For example: Time window 1 (0-5 seconds): respiratory rate 0.21Hz, 0 turns; Time window 2 (2-7 seconds): respiratory rate 0.19Hz, 0 turns; Time window 8 (14-19 seconds): respiratory rate 0.18Hz, 0 turns; Time window 15 (28-33 seconds): respiratory rate 0.22Hz, 1.0 turn, indicating that one complete turning movement was detected within this window.

[0124] The output results for each time window must meet the basic logic. The respiratory rate value must be within the normal human respiratory rate range of 0.1-0.3Hz, and the number of times the body turns over must be ≥0. If any abnormal results are found that are outside the range, they should be marked as invalid values.

[0125] The probability distribution of sleep stages across multiple consecutive time windows is weighted and smoothed, with the stage with the highest probability being taken as the final current sleep stage. Kalman filtering is applied to the respiratory rate and turning frequency values ​​across multiple consecutive time windows to remove random noise and obtain stable respiratory rate and turning frequency indices. Specifically, to address the issue of fluctuating sleep stage probabilities within a single time window, a weighted smoothing algorithm is used to process the probability distribution of sleep stages across multiple consecutive time windows, ensuring the stability of sleep stage determination. Three consecutive time windows are selected as the smoothing range: the current window to be processed is designated as window 0, the previous time window as window 1, and the two previous time windows as window 2. To reflect the dominance of the current sleep state, differentiated weights are set: window 0 has a weight of 0.6, window 1 has a weight of 0.3, and window... The weight of □2 is 0.1, and the total weight is 1.0. For the four sleep stages of wakefulness, light sleep, deep sleep, and REM sleep, the smoothed probability of each stage is calculated. Taking the current window as an example, its deep sleep probability is 0.7, the deep sleep probability of window 1 is 0.6, and the deep sleep probability of window 2 is 0.5. Substituting into the formula, the smoothed probability of deep sleep = 0.6 × 0.7 + 0.3 × 0.6 + 0.1 × 0.5 = 0.65. Simultaneously, the smoothed probabilities of wakefulness (0.025), light sleep (0.199), and REM sleep (0.126) are calculated. After completing the smoothing calculation for all sleep stages in a single window, the stage with the highest probability is selected as the initial sleep stage of that window. For example, the deep sleep probability of 0.65 in the 8th window is the highest, so the initial sleep stage of that window is determined to be deep sleep. This logic is followed to complete the weighted smoothing and initial sleep stage determination for all time windows.

[0126] To address the issues of minor fluctuations in respiratory rate and false detections of turning over, a Kalman filter algorithm is used to denoise the respiratory rate and turning over detection values ​​across multiple consecutive time windows. The core parameters are set as follows: process noise variance Q = 0.001, observation noise variance R = 0.01, initial respiratory rate x(0) = 0.2Hz, and initial covariance P(0) = 0.01. Based on these parameters, the initial respiratory rate detection value for each time window is Kalman filtered to obtain the filtered respiratory rate value for each window. For example, the initial detection value for the first window is 0.21Hz, and after filtering, it is 0.205Hz; the second window... The initial detection value was 0.19Hz, and after filtering it was 0.199375Hz; the initial detection value in the third window was 0.20Hz, and after filtering it was 0.199574Hz; the core parameters were set as process noise variance Q=0.005, observation noise variance R=0.02, initial number of rollovers x(0)=0 times, and initial covariance P(0)=0.02; based on the above parameters, the initial detection value of the number of rollovers in each time window was subjected to Kalman filtering to obtain the number of rollovers after filtering in each window. For example, the initial detection value in the 15th window was 1.0 times, and after filtering it was 0.286 times; subsequent windows continued to be filtered, and the final value stabilized at around 1.0 times.

[0127] After smoothing and denoising all time windows, the data is integrated to determine the final core monitoring indicators. The sleep stage with the highest probability after weighted smoothing of the last time window is taken as the final current sleep stage. For example, in the 28th smoothed window of the last 60-second monitoring data, the probabilities are: wakefulness 0.02, light sleep 0.83, deep sleep 0.12, and REM sleep 0.03. Light sleep has the highest probability, so the current sleep stage is determined to be light sleep. The respiratory rate values ​​after Kalman filtering of all time windows are summed and then divided by the total number of windows to obtain the average value, which is taken as the final respiratory rate indicator. For example, the total respiratory rate after filtering 28 windows is approximately 5.6Hz, and the average value = 5.6 ÷ 28 = 0.2Hz, or 12 times per minute, which is the final respiratory rate indicator. The number of times of turning over after Kalman filtering of all time windows is summed and the sum is rounded to the nearest integer, which is taken as the final number of times of turning over indicator. For example, the total number of times of turning over after filtering 28 windows is 1.2 times, and after rounding, the final number of times of turning over indicator is 1 time.

[0128] In this embodiment of the invention, by performing sliding window segmentation on weak feature signals to obtain corresponding feature sequences, and inputting the feature sequences into a pre-trained AI sleep stage classifier and AI breathing and turning regressor for recognition and calculation, and by performing weighted smoothing on the probability distribution of sleep stages in multiple consecutive time windows and Kalman filtering on the respiratory rate and turning frequency detection values, the technical problems of sleep state recognition results being easily affected by single window data and jumping, random noise interference in physiological indicator output, and insufficient recognition accuracy and data stability are overcome. Thus, the current sleep stage is accurately determined, and stable and reliable respiratory rate and turning frequency indicators are output.

[0129] In a preferred embodiment of the present invention, control commands are generated based on the output sleep stage to trigger the corresponding sleep linkage strategy of the smart bedroom device, including:

[0130] The system acquires the current sleep stage, respiratory rate, and number of turns. Based on the current sleep stage, it matches the corresponding device control parameters from a preset linkage strategy library. This library stores mapping relationships between different sleep stages and light brightness, color temperature, sleep-aid sound type and volume, and smart curtain opening / closing degree. Specifically, it acquires previously output core sleep monitoring data, matches the device control parameters corresponding to the current sleep stage based on the preset linkage strategy library, and accurately acquires the three core monitoring data points. The specific values ​​are as follows: the current sleep stage is light sleep, and the respiratory rate is... The suction frequency index is 0.2 Hz, or 12 times per minute, and the turning frequency index is 1 time per minute, which translates to 6 times per hour. All indicators have undergone preliminary smoothing and noise reduction processing, ensuring accurate and stable data that can be directly used as the core basis for matching linkage strategies. In the linkage strategy library construction phase, a pre-set linkage strategy library is used, storing a one-to-one correspondence between different sleep stages and the control parameters of smart bedroom devices. This covers three core device categories: the brightness and color temperature of sleep lights, the sound type and volume of sleep aid speakers, and the opening and closing degree of smart curtains, clearly defining the specific control parameters corresponding to each sleep stage. The parameter settings are based on human sleep physiology, ensuring adaptation to the needs of the human body in different sleep states. The specific correspondences are as follows, all of which are directly callable values:

[0131] The parameters corresponding to the sleep guidance period are: the brightness of the sleep lamp decreases linearly from 80 lux to 20 lux, the color temperature is 2700 Kelvin warm yellow, the sleep aid sound type is white noise, the volume is 30 decibels, and the opening degree of the smart curtains is 50%.

[0132] The parameters corresponding to the deep sleep maintenance period are: the sleep light brightness is 0 lux completely off or 5 lux minimum brightness, the color temperature is kept at 2700 Kelvin, the sleep aid sound type remains unchanged, the volume is reduced to 15 decibels background volume, and the smart curtain opening degree is 0% completely closed.

[0133] The parameters corresponding to the light sleep stage are: a sleep lamp brightness of 20 lux in the normal light sleep mode, a color temperature of 2700 Kelvin, a sleep aid sound type of white noise, a volume of 25 decibels, and a smart curtain opening degree of 0%.

[0134] The parameters corresponding to the waking period are: the sleep light brightness is 100 lux in daytime mode, the color temperature is 5000 Kelvin natural light color, the sleep aid sound stops playing, and the smart curtains are 100% fully open.

[0135] In the linkage parameter matching stage, the current sleep stage is accurately matched with the sleep stages in the linkage strategy library, and all device control parameters corresponding to the light sleep stage are extracted, namely, the sleep light brightness is 20 lux and the color temperature is 2700 Kelvin, the sleep aid speaker plays white noise and the volume is 25 decibels, and the smart curtain opening degree is 0%. After the parameter matching is completed, the subsequent control command generation step is entered.

[0136] If the current sleep stage is the sleep induction period, a first control command is generated to gradually dim the brightness of the sleep light according to a linear decrease curve and switch the color temperature to a warm yellow temperature. Simultaneously, the sleep aid speaker is triggered to play preset white noise. Specifically, if the current sleep stage is the sleep induction period, meaning the body is preparing to fall asleep but has not yet entered sleep, the core goal at this time is to guide the body to fall asleep quickly. Therefore, a first control command is generated, linking the sleep light and sleep aid speaker to perform corresponding adjustment operations. Regarding sleep light adjustment, a linear decrease method is used to adjust the brightness of the sleep light to avoid sudden brightness changes interfering with the body's sleep. The initial brightness of the sleep light is set to 80 lux, which is suitable for pre-sleep activities; the target brightness is 20 lux, which is a soft brightness suitable for guiding the body to fall asleep; the total duration of brightness decrease is 5 minutes, or 300 seconds. The formula for calculating the brightness using the linear decrease curve is: ,in, The brightness of the sleep light at second t is... The initial brightness is 80 lux. The target brightness is 20 lux. To reduce the total duration by 300 seconds, The value range is 0 to 300 seconds.

[0137] Sleep lamp color temperature control: The sleep lamp color temperature is switched to 2700K warm yellow temperature. This color temperature light is soft and has no glaring blue light. It can inhibit the reduction of melatonin secretion in the human body and help guide sleep. The color temperature switching adopts an instant switching mode with a switching time of ≤1 second to avoid interference caused by gradual color temperature change.

[0138] Sleep Aid Sound Control: Trigger the sleep aid sound to play preset white noise in the frequency range of 20~20000Hz. The sound is uniform and can mask environmental noise. Set the playback volume to 30dB to meet the human body's auditory comfort before sleep and do not affect falling asleep. The playback time is 30 minutes. If the human body has not entered the light sleep stage after 30 minutes, the playback time will be automatically extended by 15 minutes until the sleep stage is detected to have switched to the light sleep stage.

[0139] Control command execution: The parameters of brightness reduction, color temperature switching, and white noise playback are integrated into the first control command and sent to the sleep lamp and sleep aid speaker. The device executes the command synchronously after receiving it, without the need for manual intervention, thus realizing intelligent control of sleep guidance.

[0140] If the current sleep stage is deep sleep maintenance, a second control command is generated to turn off the sleep light or adjust it to the lowest brightness, reduce the volume of the sleep aid sound to the preset background volume, and detect the status of the smart curtains. If the curtains are not closed, a closing command is generated, specifically including: adjusting the sleep light brightness to the lowest level or turning it off directly. The minimum brightness is set to 5 lux, retaining only a faint light to avoid the discomfort of complete darkness without affecting deep sleep. If the user prefers a completely dark environment, the sleep light brightness can be directly turned off to 0 lux. The two modes can be selected according to the user's preset. In this embodiment, the sleep light brightness of 0 lux is selected to be turned off. lux; Reduce the volume of the sleep-aid sound to a preset background volume of 15dB, a negligible volume that maintains the auxiliary effect of white noise without interfering with deep sleep; Maintain the sleep-aid sound type as white noise and continue playing until the sleep stage switches to light sleep or wakefulness; Detect the current opening degree of the smart curtains, with a detection frequency of 1 time / 10 seconds. If the curtains are detected as not fully closed (opening degree > 0%), immediately generate a curtain closing command to control the curtains to close completely at a uniform speed, set to 1% / second. If the current curtain opening degree is 50%, the closing time is: ,in, The time required for closure, The current opening degree is 50%. The target opening degree is 0%. Assuming a closing speed of 1% / second, the calculation yields 50 seconds. After 50 seconds, the curtains are fully closed, creating a dark environment conducive to deep sleep. The parameters for turning off the sleep light, adjusting the sleep aid volume, and closing the curtains are integrated into a second control command, which is sent to three types of smart devices. The devices execute synchronously, and the device status is monitored in real time throughout the process. If abnormal curtain closure or volume deviation is detected, the control command is immediately resent.

[0141] If the current sleep stage is light sleep and the number of turning over exceeds a preset threshold, a third control command is generated to fine-tune the sleep light to a low-light mode and adjust the frequency of the sleep aid sound to assist in the sleep transition. Specifically, if the current sleep stage is light sleep, meaning the body is in a light sleep state and is easily disturbed by external factors, an increase in the number of turning over indicates unstable sleep, requiring assistance to transition. The core objective is to help maintain a light sleep state and reduce sleep fluctuations. Therefore, when the number of turning over exceeds the preset threshold, a third control command is generated, linking the sleep light and sleep aid sound to perform corresponding fine-tuning. Combining the physiological characteristics of the human body during light sleep, the preset threshold for the number of turning over is set to 2 times / hour, i.e., no more than once every 30 minutes. The calculation method for the number of turning over is: converting the output number of turning over 1 time / 60 seconds into the number of turning over per hour. The calculation formula is: ,in, The number of times per hour. If a sleeper turns over once in 60 seconds, the calculated number is 6 times per hour. Since 6 times per hour is greater than 2 times per hour, exceeding the preset threshold, a third control command is triggered. The sleep light is then fine-tuned from the normal light sleep mode (20 lux) to a low-light mode. The low-light mode brightness is set to 10 lux, using a linear fine-tuning method with a 2-second adjustment time to avoid sudden brightness changes that could disrupt sleep. The fine-tuning calculation formula is as follows: ,in, The normal light level for light sleep is 20 lux. The brightness is 10 lux in low-light mode. To adjust the duration to 2 seconds, t is set to a value between 0 and 2 seconds. Substituting the values ​​into the calculation, the brightness steadily drops to 10 lux after 2 seconds, providing a dim light to aid sleep without disturbing light sleep.

[0142] Fine-tuning the sleep-aid sound: Adjusting the frequency of the sleep-aid sound to aid sleep transition. The original white noise frequency of 400Hz is adjusted to 300Hz, making it lower and softer. The frequency adjustment uses a linear gradient method, with an adjustment time of 10 seconds. The gradient calculation formula is as follows: ,in, The original frequency is 400Hz. The target frequency is 300Hz. To adjust the duration to 10 seconds, Set the value for 0-10 seconds, and after 10 seconds the frequency will steadily drop to 300Hz; at the same time, keep the sleep aid volume constant at 25dB to avoid interference caused by volume changes.

[0143] Control command execution: The parameters for fine-tuning the brightness of the sleep lamp and the frequency of the sleep aid sound are integrated into a third control command and sent to the sleep lamp and the sleep aid sound. The devices execute synchronously. After execution, the number of times the user turns over is continuously monitored. If the number of times the user turns over drops below the threshold (≤2 times / hour), a callback command is generated to restore the device parameters to the normal state of light sleep.

[0144] If the current sleep stage is a waking period and the respiratory rate index exceeds the wake-up threshold, a fourth control command is generated to gradually increase the brightness of the sleep light to daytime mode according to the preset wake-up time, while stopping the sleep aid sound. Specifically, regarding the setting of the respiratory rate wake-up threshold, based on the physiological characteristics of the human body during the waking period, the respiratory rate wake-up threshold is set to 0.3 Hz, which is 18 times per minute. When the respiratory rate index is detected to exceed this threshold, it indicates that the human body has fully woken up, and the wake-up linkage strategy is immediately triggered. If the human respiratory rate index is detected to rise from 0.2 Hz to 0.32 Hz, exceeding the 0.3 Hz wake-up threshold, the fourth control command is immediately triggered.

[0145] Regarding the sleep light wake-up control, based on the preset 15-minute wake-up time, the sleep light brightness is gradually increased from the lowest brightness of 5 lux to 100 lux in daytime mode, using a linear increment curve to avoid sudden brightness changes that could irritate the eyes and ensure a gentle wake-up process. The increment calculation formula is as follows: ,in, For the first The brightness of the sleep light in seconds, The initial brightness is 5 lux. The target brightness is 100 lux. The total wake-up time is 900 seconds, and t takes values ​​from 0 to 900 seconds.

[0146] Sleep lamp color temperature control: The sleep lamp color temperature is gradually switched from a warm yellow temperature of 2700K to a natural light color of 5000K. The switching time is the same as the wake-up time, 900 seconds, with a linear gradient to avoid sudden color temperature changes that may irritate the eyes. The formula for calculating the color temperature gradient is: ,in, Let be the color temperature at second t. The initial color temperature is 2700K. The target color temperature is 5000K. The total wake-up time is 900 seconds. The color temperature ranges from 0 to 900 seconds, and after 900 seconds, it smoothly switches to 5000K.

[0147] Sleep Aid Sound Control: Immediately stop playing sleep-aid white noise. The stopping process will be gradual, lasting 10 seconds, with the volume linearly decreasing from the current 15dB or 25dB to 0dB to avoid interference from a sudden stop. The volume reduction calculation formula is as follows: ,in, For the first Volume in seconds, Assuming the current volume is 25dB, The target volume is 0dB. The fade-out duration is 10 seconds. The duration is 0 to 10 seconds, after which playback will stop completely.

[0148] Control command execution: The parameters of increasing sleep light brightness, gradual color temperature change, and gradual fading and stopping of sleep aid sound are integrated into the fourth control command and sent to the sleep light and sleep aid sound. The devices execute synchronously. After waking up, the device status is switched to daytime mode to provide users with a comfortable wake-up environment.

[0149] In this embodiment of the invention, based on the current sleep stage, respiratory rate, and number of times the user turns over, corresponding device control parameters are matched in a preset linkage strategy library. Control commands are generated for different sleep stages and physiological indicator thresholds to adjust bedroom smart devices such as sleep lights, sleep aid speakers, and smart curtains to perform differentiated linkage operations. This overcomes the technical problems of existing smart sleep lights having a single linkage strategy, being unable to adaptively adjust based on real-time sleep status, and being difficult to adapt to the sleep assistance needs of the entire sleep cycle. As a result, it achieves the technical effect of realizing refined and adaptive smart device linkage control throughout the entire sleep cycle, improving sleep assistance effects and user sleep experience, and enhancing the synergy and intelligence of the smart home system.

[0150] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0151] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0152] Experimental example:

[0153] I. Experimental Background

[0154] This experiment validates a sleep state perception system for a sleep lamp that integrates millimeter-wave radar and AI. The experimental scenario is a standard bedroom environment (20 square meters) in a smart home experience center. The experiment aims to verify the accuracy of the system's sleep state perception via millimeter-wave radar, as well as the effectiveness of AI-based intelligent device linkage control.

[0155] II. Experimental Equipment and Parameters

[0156] The experimental equipment included: a sleep lamp integrating millimeter-wave radar (radar operating frequency of 60 GHz, bandwidth of 4 GHz), an AI edge computing unit (computing power of 4 trillion operations per second), a sleep aid speaker, and a set of smart curtains. The subjects were one hundred healthy adults (aged 18 to 55), and the experiment lasted for seven consecutive nights per participant.

[0157] Experimental parameter settings: millimeter-wave radar transmit power of 10 milliwatts, detection distance of 0.3 to 3 meters, angular resolution of 5 degrees; the AI ​​sleep state recognition algorithm adopts a deep neural network model, and the input features include respiratory rate, body movement amplitude, heart rate variability, etc.; the sleep stages are divided into four stages: wakefulness, light sleep, deep sleep, and REM sleep.

[0158] Figure 3 This demonstrates multi-stage filtering of the original echo signal using electromagnetic interference (EMI) filtering and multipath reflection suppression. EMI filtering involves acquiring the operating current waveform of the LED driver circuit as a reference interference signal. An adaptive filter is used to cancel out components related to the reference interference from the original echo signal, increasing the signal-to-noise ratio (SNR) from 12.5 dB to 18.3 dB. Multipath reflection suppression is then performed using range-dimensional fast Fourier transform and spatial adaptive beamforming to eliminate multipath reflection clutter caused by the lampshade material, further improving the SNR to 22.7 dB.

[0159] Figure 4 The accuracy of sleep stage recognition was demonstrated, and the trained AI model was tested and validated, with the recognition accuracy for each sleep stage being statistically analyzed. Experimental results showed that the accuracy rate was 96.2% during wakefulness, 94.8% during light sleep, 93.5% during deep sleep, and 95.1% during REM sleep. The overall average recognition accuracy was 94.9%, which is superior to traditional accelerometer-based sleep monitoring methods (average accuracy of 85.3%).

[0160] Figure 5 This study demonstrates a comparison of sleep quality before and after system implementation, analyzing sleep quality indicators in both cases. Before implementation, the subjects had an average sleep latency of 28 minutes, a deep sleep duration of 65 minutes, a sleep efficiency of 78%, an average of 4.5 awakenings per night, and a subjective sleep score of 72. After implementing the system, the sleep latency decreased to 15 minutes, the deep sleep duration increased to 95 minutes, the sleep efficiency improved to 91%, the number of awakenings per night decreased to 1.8, and the sleep score improved to 88.

[0161] 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 sleep state sensing system for a sleep lamp that integrates millimeter wave and AI, characterized in that: include: The acquisition module is used to transmit millimeter-wave signals to the bed area via millimeter-wave radar and receive the reflected echoes to obtain the original echo signals. Electromagnetic interference filtering is performed on the original echo signal to eliminate electromagnetic interference generated by the LED bead driving circuit, resulting in the first filtered signal. Multipath reflection suppression is applied to the first filtered signal to eliminate multipath reflection clutter caused by the lampshade material, thus obtaining the second filtered signal; The calculation module is used to perform metal shielding compensation on the second filtered signal to correct the signal attenuation caused by the metal heat dissipation structure of the lamp body and obtain the corrected echo signal. The metal shielding compensation includes collecting the spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertex of the inner wall of the lamp cover, and multiple auxiliary points on the surface of the lamp body. By minimizing the sum of squared distances from each point to the fitted sphere, a virtual reference sphere is fitted. The virtual reference sphere is then meshed to generate a two-dimensional spatial attenuation compensation matrix covering the bed area. The calibration module is used to perform spatial point-by-point calibration on the calibrated echo signal using a two-dimensional spatial attenuation compensation matrix, and to extract weak feature signals reflecting the human sleep state from the calibrated echo signal. The processing module is used to input the extracted weak feature signals into the pre-trained AI sleep state recognition algorithm, and output the current sleep stage, breathing rate, and number of times the user turns over. Based on the output sleep stage, it generates control commands to link the smart bedroom devices to execute corresponding sleep linkage strategies.

2. The millimeter-wave and AI-coordinated sleep lamp sleep state perception system according to claim 1, characterized in that, The original echo signal is obtained by transmitting millimeter-wave signals into the bed area using millimeter-wave radar and receiving the reflected echoes. Electromagnetic interference filtering is performed on the original echo signal to eliminate electromagnetic interference generated by the LED bead driving circuit, resulting in a first filtered signal, including: Collect the operating current waveform of the LED bead driver circuit and use the operating current waveform as a reference interference signal; The original echo signal and the reference interference signal are input into an adaptive filter. The adaptive filter dynamically updates the filter coefficients based on the least squares criterion to cancel out the interference components related to the reference interference signal from the original echo signal, thus obtaining the preliminary filtered signal. The initial filtered signal is subjected to band-stop filtering to remove fixed-band interference caused by the switching frequency of the driving circuit and its harmonics, thus obtaining the first filtered signal.

3. The sleep state perception system for a sleep lamp that integrates millimeter wave and AI according to claim 2, characterized in that, The first filtered signal is subjected to multipath reflection suppression to eliminate multipath reflection clutter caused by the lampshade material, resulting in the second filtered signal, which includes: A range-dimensional fast Fourier transform is performed on the first filtered signal to obtain a range image sequence. Based on the geometric positional relationship between the millimeter-wave radar and the inner wall of the lampshade, the range cell corresponding to each reflection point on the inner wall of the lampshade in the range image is calculated, and the time-domain echo component in the range cell is extracted as the original estimate of the multipath reflection of the lampshade. The original estimate of the multipath reflection of the lampshade is convolved with the pre-measured reflectance coefficient of the lampshade material to reconstruct the time-domain waveform of the multipath interference; the time-domain waveform of the multipath interference is subtracted from the first filtered signal to obtain the preliminary multipath suppression signal. The initial multipath suppression signal is processed by spatial adaptive beamforming. By pointing the main lobe of the beam to the bed area and forming a null in the angular direction of the lamp cover, residual lamp cover directional clutter is filtered out to obtain the second filtered signal.

4. The millimeter-wave and AI-coordinated sleep lamp sleep state sensing system according to claim 3, characterized in that, Metal shielding compensation is applied to the second filtered signal to correct the signal attenuation caused by the metal heat dissipation structure of the lamp body, and the corrected echo signal is obtained. Metal shielding compensation includes collecting the spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertex of the inner wall of the lamp cover, and multiple auxiliary points on the surface of the lamp body. By minimizing the sum of squared distances from each point to the fitted sphere, a virtual reference sphere is fitted. The virtual reference sphere is meshed to generate a two-dimensional spatial attenuation compensation matrix covering the bed region, including: Metal shielding compensation is applied to the second filtered signal. The spatial coordinates of the center point of the LED light-emitting surface, the geometric centroid of the metal heat sink, the vertex of the inner wall of the lamp cover, and multiple auxiliary points on the surface of the lamp body are collected by a three-dimensional scanning device to form an initial point set. By setting the objective function of spherical fitting to minimize the sum of squared distances from each point to the fitted sphere, the initial point set is iteratively optimized to determine the coordinates of the center of the virtual reference sphere and the radius of the sphere, thus obtaining the virtual reference sphere. The virtual reference sphere is divided into equally spaced latitude and longitude grids in the bed area. Each grid node corresponds to a spatial position in the bed area. Based on the relative geometric relationship between the spatial position and the millimeter-wave radar, the theoretical attenuation value caused by the metal heat dissipation structure is calculated. The theoretical attenuation value is used as the attenuation compensation coefficient of the grid node. The attenuation compensation coefficients of each node are arranged in the row and column order of the grid to construct a two-dimensional spatial attenuation compensation matrix covering the bed area.

5. The millimeter-wave and AI-coordinated sleep lamp sleep state perception system according to claim 4, characterized in that, The virtual reference sphere is divided into equally spaced latitude and longitude grids across the bed area. Each grid node corresponds to a spatial location within the bed area. Based on the relative geometric relationship between the spatial location and the millimeter-wave radar, the theoretical attenuation value caused by the metal heat dissipation structure's obstruction is calculated. This theoretical attenuation value is used as the attenuation compensation coefficient for the grid node, including: A spherical coordinate system is established with the center of the virtual reference sphere as the origin. The latitude and longitude range of the bed area under the spherical coordinate system is determined. The latitude and longitude range is divided into equal intervals according to the preset spatial resolution to generate the latitude and longitude coordinates of each grid node. The latitude and longitude coordinates of each grid node are converted into three-dimensional spatial coordinates to obtain the actual spatial position of the grid node in the bed area; Based on the three-dimensional spatial coordinates of the millimeter-wave radar and the three-dimensional spatial coordinates of the grid nodes, a straight spatial path from the radar to the grid nodes is constructed. Determine whether the spatial straight path intersects with the virtual reference sphere. If they intersect, calculate the path's crossing distance within the sphere and convert the crossing distance into a theoretical attenuation value based on the attenuation characteristics of electromagnetic waves in a metallic medium. If they do not intersect, set the theoretical attenuation value to one. Use the calculated theoretical attenuation value as the attenuation compensation coefficient for the mesh nodes.

6. The millimeter-wave and AI-coordinated sleep lamp sleep state sensing system according to claim 5, characterized in that, The corrected echo signal is spatially calibrated point-by-point using a two-dimensional spatial attenuation compensation matrix. Weak feature signals reflecting human sleep states are then extracted from the calibrated echo signal, including: The corrected echo signal is subjected to distance-dimensional fast Fourier transform and angle-dimensional fast Fourier transform to generate a distance-angle spectrum. Each pixel in the distance-angle spectrum corresponds to a spatial location in the bed area. The spatial position of each pixel in the distance-angle spectrum is mapped to the grid nodes of the two-dimensional spatial attenuation compensation matrix, and the attenuation compensation coefficient corresponding to the pixel position is calculated by bilinear interpolation. The amplitude of the corresponding pixel of the corrected echo signal is compensated point by point by the attenuation compensation coefficient to obtain the amplitude-calibrated echo signal; the amplitude-calibrated echo signal is subjected to phase unwrapping processing to extract the phase change information caused by the rise and fall of the human chest cavity; Bandpass filtering was applied to the phase change information to separate the respiratory frequency component and the heart rate component. Moving target detection was performed on multiple consecutive frames of distance-angle spectrum to extract the micro-Doppler features corresponding to turning over and body movement. By fusing respiratory rate components, heart rate components, and micro-Doppler features corresponding to turning over and body movement, a weak feature signal reflecting the human sleep state is constructed.

7. The sleep state sensing system for a sleep lamp that integrates millimeter wave and AI according to claim 6, characterized in that, The extracted weak feature signals are input into a pre-trained AI sleep state recognition algorithm, which outputs the current sleep stage and indicators such as breathing rate and number of times the user turns over, including: Weak feature signals are segmented by a sliding window to obtain feature sequences corresponding to multiple time windows. Each feature sequence contains respiratory rate components, heart rate components, and micro-Doppler features. The feature sequence of each time window is input into a pre-trained AI sleep stage classifier and an AI breathing and turning regressor, which are trained based on machine learning algorithms; The AI ​​sleep stage classifier outputs the probability distribution of the sleep stage to which the current time window belongs, and the AI ​​breathing and turning regressor outputs the respiratory rate value and the number of turnings detected for the current time window. The probability distribution of sleep stages in multiple consecutive time windows is weighted and smoothed, and the stage with the highest probability is taken as the final current sleep stage. Kalman filtering is applied to the respiratory rate and turning frequency values ​​in multiple consecutive time windows to remove random noise and obtain stable respiratory rate and turning frequency indices.

8. The sleep state sensing system for a sleep lamp that integrates millimeter wave and AI according to claim 7, characterized in that, Based on the output sleep stage, control commands are generated to trigger corresponding sleep-related strategies to be executed by smart bedroom devices, including: The system acquires the current sleep stage, as well as the respiratory rate and number of times the user turns over. Based on the current sleep stage, it matches the corresponding device control parameters in a preset linkage strategy library. The linkage strategy library stores the mapping relationship between different sleep stages and light brightness, color temperature, sleep aid sound type and volume, and smart curtain opening degree. If the current sleep stage is the sleep induction period, the first control command is generated to gradually dim the brightness of the sleep light according to a linear decreasing curve, switch the color temperature to a warm yellow temperature, and trigger the sleep aid speaker to play preset white noise. If the current sleep stage is deep sleep maintenance, a second control command is generated to turn off the brightness of the sleep light or adjust it to the lowest brightness, reduce the volume of the sleep aid sound to the preset background volume, and detect the status of the smart curtains. If the curtains are not closed, a closing command is generated. If the current sleep stage is light sleep and the number of times the sleeper turns over exceeds a preset threshold, a third control command is generated to fine-tune the sleep light to a low-light mode and adjust the frequency of the sleep aid sound to assist in the transition to sleep. If the current sleep stage is a waking period and the respiratory rate index exceeds the wake-up threshold, a fourth control command is generated to gradually increase the brightness of the sleep lamp to daytime mode according to the preset wake-up time, while stopping the sleep aid sound.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to perform the system as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, performs the system as described in any one of claims 1 to 8.

Images