Sleep state recognition regulation and control method and system

Abstract

The invention is applied to the field of sleep management, and discloses a sleep state recognition regulation and control method and system.The method comprises the steps that piezoelectric signal data captured by piezoelectric sensing equipment and optical pulse signal data captured by wearable sensing equipment in the sleep process of a target user are obtained; body movement artifact removal processing is conducted on the piezoelectric signal data and the optical pulse signal data, and first piezoelectric signal data and first optical pulse signal data are obtained; performing fusion processing on the first piezoelectric signal data and the first optical pulse signal data to obtain a target fusion feature; determining a sleep type of the target user based on the target fusion feature, performing sleep state recognition on the target fusion feature based on the sleep type, and generating a sleep core index corresponding to a recognition result; and according to the sleep type and the sleep core index, generating and executing a sleep regulation strategy of the target user. According to the method, the sleep quality of the user is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sleep management, in particular to a sleep state recognition and regulation method and system. BACKGROUND

[0002] With the improvement of people's health management awareness, sleep quality monitoring and optimization has become a research hotspot in the health field. The accurate recognition of sleep state is the core prerequisite for scientific sleep management, which depends on the effective collection and analysis of physiological signals during sleep.

[0003] At present, sleep state recognition and regulation mainly rely on single sensing devices to obtain data, such as using electroencephalogram devices to collect brain wave signals for sleep state recognition and regulation. However, this method has obvious defects. Among them, the contact type device is easy to bind the user's sleep posture, interfere with the natural sleep state, and cause signal collection distortion; the non-contact type device is greatly affected by environmental light, distance and other factors, and has poor signal stability, and it is difficult to capture subtle physiological state changes. These defects make the data obtained by single sensing dimension unable to comprehensively and objectively reflect the physiological state difference in the sleep process, which further leads to insufficient sleep state recognition accuracy and low result reliability, and cannot provide accurate sleep regulation. SUMMARY

[0004] The present application provides a sleep state recognition and regulation method and system to solve the technical problem of how to improve the existing sleep state recognition and regulation method, and to achieve the effect of improving the sleep quality of users.

[0005] In order to solve the above technical problems, the present application provides a sleep state recognition and regulation method and system, which comprises: acquiring piezoelectric signal data captured by a piezoelectric sensing device and optical pulse signal data captured by a wearable sensing device during the sleep process of a target user; performing body motion artifact rejection processing on the piezoelectric signal data and the optical pulse signal data respectively to obtain first piezoelectric signal data and first optical pulse signal data; performing fusion processing on the first piezoelectric signal data and the first optical pulse signal data to obtain target fusion features; determining the sleep type of the target user based on the target fusion features, and performing sleep state recognition on the target fusion features based on the sleep type to generate sleep core indicators corresponding to the recognition results; generating and executing a sleep regulation strategy for the target user according to the sleep type and the sleep core indicators.

[0006] As one of the preferred solutions, the piezoelectric signal data and the optical pulse signal data are respectively subjected to body motion artifact removal processing to obtain first piezoelectric signal data and first optical pulse signal data, which includes: The piezoelectric signal data is sequentially subjected to baseline calibration and noise pre-filtering processing to obtain initial piezoelectric signal data. According to the frequency characteristic difference between the physiological signal and the body motion interference signal of the target user, the initial piezoelectric signal data is subjected to signal component separation to obtain a physiological signal component and a body motion interference reference signal; wherein the body motion interference reference signal is used to represent the interference intensity of body motion behavior on the piezoelectric signal during sleep; Based on the body motion interference reference signal, the physiological signal component is respectively subjected to Kalman filter iteration optimization to integrate and generate first piezoelectric signal data.

[0007] As one of the preferred solutions, the piezoelectric signal data and the optical pulse signal data are respectively subjected to body motion artifact removal processing to obtain first piezoelectric signal data and first optical pulse signal data, which includes: The optical pulse signal data is sequentially subjected to baseline drift suppression and noise pre-filtering processing to obtain initial optical pulse signal data. The initial optical pulse signal data is decomposed into a plurality of characteristic modal components based on a multi-scale decomposition algorithm. Each of the characteristic modal components is subjected to multi-dimensional characteristic analysis, and effective modal characteristic components are selected based on the analysis results; wherein the multi-dimensional characteristic analysis is designed to identify and remove components dominated by body motion artifacts in each of the characteristic modal components based on the prior frequency band characteristics of the physiological signal. The effective modal characteristic components are sequentially subjected to signal reconstruction and blind source separation to obtain first optical pulse signal data.

[0008] As one of the preferred solutions, the first piezoelectric signal data and the first optical pulse signal data are subjected to fusion processing to obtain target fusion features, which includes: The first piezoelectric signal data and the first optical pulse signal data are subjected to time alignment processing to obtain second piezoelectric signal data and second optical pulse signal data after time sequence synchronization; The signal-to-noise ratio of the first piezoelectric signal data and the first optical pulse signal data is determined, and the second piezoelectric signal data and the second optical pulse signal data are subjected to data layer fusion processing based on the signal-to-noise ratio to obtain first fusion data. Multi-dimensional physiological features are extracted from the first fusion data, and feature saliency analysis is performed on the multi-dimensional physiological features to obtain core features strongly related to sleep state. The core features in the first fused data are enhanced using an attention mechanism to obtain the target fused features.

[0009] As one preferred embodiment, the step of determining the sleep type of the target user based on the target fusion features, and performing sleep state identification on the target fusion features based on the sleep type to generate core sleep indicators corresponding to the identification results includes: The target fusion features are subjected to feature differentiation analysis to obtain the sleep type of the target user; wherein, the feature differentiation analysis determines the sleep type as difficult to fall asleep, difficult to maintain sleep, early awakening, or a combination of the sleep type by the differences in respiratory variability, heart rate variability, body movement amplitude, and blood oxygenation trend. A sleep state recognition model matching the sleep type is determined, and the target fusion features are input into the sleep state recognition model to perform sleep state recognition, thereby obtaining the temporal recognition results of the sleep period and the wakefulness period. Based on the time sequence recognition results, core sleep indicators are generated, including total sleep time, sleep efficiency, sleep latency, and wakefulness time after falling asleep.

[0010] Another aspect of the present invention provides a sleep state recognition and control system, comprising: The acquisition module is used to acquire piezoelectric signal data captured by the piezoelectric sensing device and optical pulse signal data captured by the wearable sensing device during the target user's sleep process. The preprocessing module is used to perform motion artifact removal processing on the piezoelectric signal data and the optical pulse signal data respectively to obtain the first piezoelectric signal data and the first optical pulse signal data; The fusion module is used to fuse the first piezoelectric signal data and the first optical pulse signal data to obtain the target fusion feature; The identification module is used to determine the sleep type of the target user based on the target fusion features, and to identify the sleep state of the target fusion features based on the sleep type, thereby generating core sleep indicators corresponding to the identification results. The control module is used to generate and execute the sleep control strategy for the target user based on the sleep type and the core sleep indicators.

[0011] As one preferred embodiment, the preprocessing module is specifically used for: The piezoelectric signal data is subjected to baseline calibration and noise pre-filtering in sequence to obtain the initial piezoelectric signal data; Based on the frequency characteristics difference between the target user's physiological signals and body movement interference signals, the initial piezoelectric signal data is subjected to signal component separation to obtain physiological signal components and body movement interference reference signals; wherein the body movement interference reference signals are used to characterize the interference intensity of body movement behavior on piezoelectric signals during sleep. Based on the body motion interference reference signal, the physiological signal components are subjected to Kalman filtering iterative optimization and integrated to generate the first piezoelectric signal data.

[0012] As one preferred embodiment, the preprocessing module is specifically used for: The optical pulse signal data is sequentially subjected to baseline drift suppression and noise pre-filtering to obtain the initial optical pulse signal data; The initial optical pulse signal data is decomposed into multiple characteristic modal components based on a multi-scale decomposition algorithm; Multidimensional characteristic analysis is performed on each of the aforementioned characteristic modal components, and effective modal characteristic components are obtained by screening based on the analysis results; wherein, the multidimensional characteristic analysis is designed to identify and eliminate the components dominated by body motion artifacts in each of the aforementioned characteristic modal components based on the prior frequency band characteristics of physiological signals. The effective modal feature components are sequentially reconstructed and blind source separated to obtain the first optical pulse signal data.

[0013] As one preferred embodiment, the fusion module is specifically used for: The first piezoelectric signal data and the first optical pulse signal data are time-aligned to obtain the second piezoelectric signal data and the second optical pulse signal data after time synchronization. The signal-to-noise ratio of the first piezoelectric signal data and the first optical pulse signal data is determined, and the second piezoelectric signal data and the second optical pulse signal data are fused based on the signal-to-noise ratio to obtain the first fused data; Multidimensional physiological features are extracted from the first fused data, and feature significance analysis is performed on the multidimensional physiological features to obtain core features that are strongly correlated with sleep state. The core features in the first fused data are enhanced using an attention mechanism to obtain the target fused features.

[0014] As one preferred embodiment, the identification module is specifically used for: The target fusion features are subjected to feature differentiation analysis to obtain the sleep type of the target user; wherein, the feature differentiation analysis determines the sleep type as difficult to fall asleep, difficult to maintain sleep, early awakening, or a combination of the sleep type by the differences in respiratory variability, heart rate variability, body movement amplitude, and blood oxygenation trend. A sleep state recognition model matching the sleep type is determined, and the target fusion features are input into the sleep state recognition model to perform sleep state recognition, thereby obtaining the temporal recognition results of the sleep period and the wakefulness period. Based on the time sequence recognition results, core sleep indicators are generated, including total sleep time, sleep efficiency, sleep latency, and wakefulness time after falling asleep.

[0015] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: 1) The solution of this invention uses dual-modal data acquisition technology of piezoelectric sensing and wearable optical pulse sensing, combined with targeted body motion artifact removal processing, to completely solve the pain point of traditional single-modal monitoring being easily affected by turning over and device displacement, making physiological signal acquisition in home scenarios purer and more reliable.

[0016] 2) This invention combines multimodal data fusion algorithms with intelligent analysis models, which can not only accurately distinguish different sleep subtypes, but also generate quantitative core indicators such as sleep efficiency and latency, achieving an upgrade from "fuzzy perception" to "precise quantification," providing a scientific basis for subsequent regulation. At the same time, the non-intrusive monitoring design eliminates the need for users to wear complex devices, and coupled with low-power operation technology, significantly improves long-term user compliance and avoids the interference of traditional sleep monitoring on sleep. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a sleep state recognition and control method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the forced wake-up mode of the control strategy in one embodiment of the present invention; Figure 3 This is a schematic diagram of an adaptive intelligent control system in one embodiment of the present invention; Figure 4 This is a structural block diagram of a sleep state recognition and control system according to one embodiment of the present invention; Figure label: Among them, 11 is the acquisition module; 12 is the preprocessing module; 13 is the fusion module; 14 is the identification module; and 15 is the control module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] One embodiment of the present invention provides a method for sleep state recognition and regulation. For details, please refer to [link / reference]. Figure 1 , Figure 1 The diagram shown is a flowchart of a sleep state recognition and regulation method according to one embodiment of the present invention, which includes steps S1-S5: S1: Acquire piezoelectric signal data captured by the piezoelectric sensing device and optical pulse signal data captured by the wearable sensing device during the target user's sleep process.

[0022] In this embodiment, the piezoelectric sensing device uses a flexible piezoelectric monitoring pad, with its core sensing unit being a polyvinylidene fluoride (PVDF) piezoelectric film. This component possesses high sensitivity and good flexibility, enabling contactless monitoring and avoiding interference with the user's sleep. The wearable sensing device is a wristband that integrates a photoplethysmography (PPG) sensing module, along with a skin conductance sensor and a skin temperature sensor. PPG utilizes the characteristic that the absorption coefficient of hemoglobin in blood for specific wavelengths of light changes with the pulse to detect physiological parameters such as heart rate, respiration, and blood oxygenation. Both types of devices feature low-power operation, meeting the needs of continuous data acquisition throughout the night, and are compatible with Internet of Things (IoT) systems, supporting real-time data transmission.

[0023] During deployment, the piezoelectric monitoring pad is laid flat under the target user's mattress or on the bed surface, ensuring complete coverage of the user's main body contact areas during sleep to effectively capture heartbeat, respiration, and body movement information. The piezoelectric monitoring pad connects to the smart home gateway via a wired interface for stable data transmission. The wristband is worn on the inside of the user's non-dominant wrist, with a tightness that allows the sensor module to fit snugly against the skin without affecting blood circulation. It connects to the smart home gateway via Bluetooth wireless communication protocol to ensure real-time data transmission. After deployment, the system automatically initializes the devices, calibrates the sensor module's acquisition baseline, eliminates initial deviations, and ensures the accuracy of signal acquisition.

[0024] During the signal acquisition phase, the piezoelectric monitoring pad, based on the piezoelectric effect, captures heartbeat signals, respiratory signals, and body movement information in real time during the user's sleep. It can also deduce individual sleep posture from body movement information. When multiple people are sleeping, causing interference with the piezoelectric signal, a complementary interface with wearable devices is reserved to supplement missing signal information. The wristband's PPG sensor module simultaneously acquires optical pulse signals, combined with skin conductance and skin temperature sensors, to obtain multi-dimensional physiological information such as heart rate, respiration, blood oxygenation, skin conductance activity, and skin temperature. During acquisition, the system simultaneously records the sleep start time and wake-up control time, ensuring accurate correspondence between physiological signals and the timeline. All acquired data is temporarily stored in its raw format without additional processing, preserving the original signal characteristics and providing a complete data foundation for subsequent preprocessing steps such as body movement artifact removal.

[0025] During the data collection process, the sampling frequency and data format of both types of devices remain consistent. Timestamp synchronization is achieved through a smart home gateway, ensuring consistency between piezoelectric signal data and optical pulse signal data in the time dimension. All collected data is temporarily stored encrypted on a local gateway or in the cloud to protect the privacy and security of user physiological data. This embodiment's signal acquisition scheme is entirely based on a non-intrusive monitoring design, requiring no complex user operations, effectively improving long-term user compliance. Furthermore, the collected multi-dimensional physiological data reduces the impact of environmental interference in home settings, providing high-quality data input for accurate sleep state assessment.

[0026] S2: Perform motion artifact removal processing on the piezoelectric signal data and optical pulse signal data respectively to obtain the first piezoelectric signal data and the first optical pulse signal data.

[0027] It should be noted that in home sleep monitoring scenarios, the target user's body movements during sleep, such as turning over, leg movements, and device relocation, can strongly interfere with piezoelectric signal data and optical pulse (PPG) signal data, forming body movement artifacts and causing distortion of the original physiological signals. To address this technical challenge, this embodiment provides a path-specific body movement artifact removal scheme based on the differences in characteristics between the two types of signals. Through standardized processing procedures and precise algorithm configuration, it effectively separates physiological signals from body movement interference, providing high-quality data support for sleep state assessment.

[0028] Preferably, in one embodiment of the present invention, the piezoelectric signal data is subjected to volumetric artifact removal processing to obtain first piezoelectric signal data, including: Baseline calibration and noise pre-filtering are performed on the raw piezoelectric signal data to eliminate baseline drift and high-frequency noise. Specifically, for each raw piezoelectric signal... The mean is subtracted, and then the amplitude is normalized using the z-score normalization method. The z-score normalization method refers to converting the original signal into a normalized signal with a mean of 0 and a standard deviation of 1 by calculating the mean and standard deviation of the signal, thereby eliminating the difference in signal amplitude at different acquisition time periods. Subsequently, a wide passband filter of 0.1Hz-20Hz is applied to filter out high-frequency electromagnetic interference and extremely low-frequency environmental drift to obtain the initial piezoelectric signal data.

[0029] Based on the difference in frequency characteristics between the target user's physiological signals and body movement interference signals, the initial piezoelectric signal data is separated into signal components to obtain physiological signal components and body movement interference reference signals; among which, the body movement interference reference signals are used to characterize the interference intensity of body movement behavior on piezoelectric signals during sleep.

[0030] Based on the difference in frequency characteristics between physiological signals and body movement interference, this embodiment employs a dual-filter separation method. Specifically, a linear-phase FIR (finite-length impulse response) low-pass filter with a cutoff frequency of 0.8Hz is used to separate the respiratory signal components in the frequency range of 0.1Hz-0.8Hz. Then, the heartbeat signal component with a frequency range of 0.8Hz-3.5Hz is separated by a bandpass filter. This allows for the initial extraction of physiological signal components. Simultaneously, the original piezoelectric signal... A high-pass filter with a cutoff frequency of 5Hz is applied, and the average absolute amplitude within each 0.5-second window is calculated to generate a motion interference reference signal. This signal is used to quantitatively characterize the intensity of interference of body movement behavior on the piezoelectric signal during sleep.

[0031] Furthermore, based on the body motion interference reference signal, the physiological signal components are subjected to Kalman filtering iterative optimization and integrated to generate the first piezoelectric signal data.

[0032] Specifically, an autoregressive model is used to describe the dynamic changes of physiological signals and body motion artifacts, with a state vector. The physiological component is the respiratory or heartbeat signal to be extracted. The bandpass-filtered physiological signal component is combined with the body motion reference signal to generate the observation vector. The observation equation is defined as ,in, Matrix design is The first row indicates that the observed signal is a superposition of the physiological component and the body motion artifact component; the second row indicates the body motion reference signal. It is proportional to the volumetric artifact component in the state vector ( (This is a proportionality coefficient, determined through equipment factory calibration). To observe noise.

[0033] Adopting information-based The covariance matching method is used to estimate the observation noise covariance matrix online. Introducing a body motion reference signal As a regulating factor: Calculation Mean over a 30-second sliding window ,when At this point, it is determined to be a large-scale physical movement event. At this time, the observed noise covariance matrix will be... Middle and physiological observation channels The corresponding noise variance term multiplied by a factor =10, artificially reducing the weight of distorted observations, forcing the filter to rely more on model predictions to update state estimates, thus avoiding interference from motion artifacts in the extraction of physiological signals.

[0034] The separated respiratory signal components and heartbeat signal components Run the adaptive Kalman filter process configured above, respectively. Based on the state transition model and the adaptively updated noise covariance matrix... The signal at each time step is iteratively filtered, and the filtered state is estimated. Extract the first state variable (i.e., the pure physiological component) and obtain the following: (Pure heartbeat signal) and (Pure respiratory signal). The two pure physiological signals are integrated to generate the first piezoelectric signal data after removing motion artifacts.

[0035] Preferably, in one embodiment of the present invention, the optical pulse signal data is subjected to body motion artifact removal processing to obtain first optical pulse signal data, including: The optical pulse signal data was sequentially processed with baseline drift suppression and noise pre-filtering to obtain the initial optical pulse signal data. Specifically, the original optical pulse (PPG) was subjected to polynomial detrending processing, fitting a polynomial function to subtract the slow baseline drift component; then, a fourth-order Butterworth zero-phase low-pass filter with a frequency range of 0.05Hz-10Hz was applied to filter out high-frequency noise and extremely low-frequency drift (the Butterworth filter has a flat passband amplitude-frequency response, and the zero-phase design avoids signal phase distortion); finally, the amplitude was calibrated using the z-score normalization method to obtain the initial optical pulse signal data.

[0036] The initial optical pulse signal data is decomposed into multiple characteristic mode components based on a multi-scale decomposition algorithm. Specifically, the CEEMDAN algorithm is used to decompose the initial optical pulse signal data, decomposing the single-channel PPG signal into multiple intrinsic mode functions (IMFs). Each IMF represents the oscillation component of the signal at different time scales. This decomposition constructs a pseudo-multi-channel representation that satisfies the basic assumption of blind source separation.

[0037] Multidimensional characteristic analysis was performed on each feature modal component, and effective modal feature components were selected based on the analysis results. Among them, the multidimensional characteristic analysis was designed to identify and remove the components dominated by body motion artifacts in each feature modal component based on the prior frequency band features of physiological signals.

[0038] Specifically, a multidimensional characteristic analysis was performed on each IMF component, including power spectral characteristics (dominant frequency position), statistical characteristics (spectral entropy), and periodic stability. Based on the prior frequency band characteristics of physiological signals (heart rate-related components are concentrated in 0.8Hz-3.0Hz, respiratory modulation components are distributed in 0.1Hz-0.5Hz, and body motion artifact components exhibit wide spectral distribution, high spectral entropy, and poor periodicity), IMF components dominated by body motion artifacts were identified and eliminated, and effective modal characteristic components were obtained. The effective modal characteristic components were then reconstructed to obtain the preliminarily denoised PPG signal.

[0039] The effective modal feature components are sequentially reconstructed and blind source separated to obtain the first optical pulse signal data. Specifically, a blind source separation algorithm is performed on the reconstructed PPG signal to further separate residual motion artifacts and physiological signal components, ultimately generating the first optical pulse signal data.

[0040] The first piezoelectric signal data and the first optical pulse signal data processed by this embodiment can effectively eliminate the interference of body movement artifacts such as turning over, leg movements, and equipment displacement. The purity and reliability of physiological signals are significantly improved, fully meeting the accuracy requirements of subsequent sleep feature extraction, sleep subtype differentiation, and sleep regulation.

[0041] S3: The first piezoelectric signal data and the first optical pulse signal data are fused to obtain the target fusion feature.

[0042] In home-based, non-intrusive sleep monitoring and CBT-I adaptive regulation, the first piezoelectric signal data (including pure heartbeat and respiratory signals) and the first optical pulse (PPG) signal data (including precise heart rate, respiratory rate, and blood oxygenation signals) carry different dimensions of physiological information. When used independently, they suffer from limited information content and interference resistance, making it difficult to comprehensively characterize the user's sleep state. Existing signal fusion methods often employ simple splicing, failing to consider data temporal consistency, signal quality differences, and feature correlations. This results in high feature redundancy and a lack of emphasis on core information, affecting the accuracy of subsequent sleep state identification and insomnia subtype differentiation. To address this technical challenge, this embodiment provides a standardized dual-modal signal fusion scheme that fully leverages the complementary value of the two types of data to generate highly recognizable target fusion features, providing core support for accurate sleep assessment and sleep regulation.

[0043] The core technologies of this embodiment include time alignment, signal-to-noise ratio (SNR) based data layer fusion, multi-dimensional physiological feature extraction, Fisher score (F-score) feature saliency analysis, and attention mechanism enhancement. Specifically, time alignment eliminates temporal deviations during dual-modal signal acquisition by unifying timestamp benchmarks, ensuring data consistency across time. SNR-based data layer fusion assigns fusion weights based on the quality differences between the two signal types, prioritizing the retention of physiological information from high-quality signals and improving the reliability of the fused data. Multi-dimensional physiological feature extraction mines temporal, frequency, and morphological features from the fused data, comprehensively covering physiological representations related to sleep states. Fisher score feature saliency analysis quantifies the distinguishing ability of features across different sleep states, selecting core features strongly correlated with sleep states and eliminating redundant information. Attention mechanism enhancement assigns dynamic weights based on the contribution of features to sleep state recognition, highlighting the dominant role of core features and improving the accuracy of subsequent model recognition.

[0044] Preferably, in one embodiment of the present invention, the first piezoelectric signal data and the first optical pulse signal data are fused to obtain target fusion features, including: The first piezoelectric signal data and the first optical pulse signal data are time-aligned to obtain the second piezoelectric signal data and the second optical pulse signal data after time synchronization. Specifically, due to slight differences in the acquisition triggering mechanism and transmission delay between the piezoelectric monitoring pad and the wristband, the first piezoelectric signal data and the first optical pulse signal data need to be time-calibrated. Using the system time of the smart home gateway as a reference, the timestamp information of the two types of signal data is extracted (the acquisition accuracy is 10ms for both). Linear interpolation is used to complete and calibrate signal segments with a time deviation exceeding 5ms, ensuring that at the same time node, the timestamp error between the second piezoelectric signal data (time-synchronized piezoelectric signal) and the second optical pulse signal data (time-synchronized PPG signal) is ≤3ms. The core function of this step is to eliminate the physiological information mismatch caused by time misalignment, laying a foundation for time consistency for subsequent fusion processing.

[0045] The signal-to-noise ratio of the first piezoelectric signal data and the first optical pulse signal data is determined. Based on the signal-to-noise ratio, the second piezoelectric signal data and the second optical pulse signal data are fused at the data layer to obtain the first fused data.

[0046] Specifically, the signal-to-noise ratio (SNR) of the second piezoelectric signal data and the second optical pulse signal data is calculated. The SNR is calculated as the ratio of the effective signal amplitude to the noise amplitude. The effective amplitude is determined by extracting the average amplitude of the main frequency band of the signal (0.1Hz-3.5Hz for the piezoelectric signal and 0.1Hz-3.0Hz for the PPG signal), and the noise amplitude is determined by calculating the average amplitude of the high-frequency band of the signal (≥20Hz). A SNR threshold of 15dB is set. When the SNR of the second piezoelectric signal data is ≥15dB and higher than that of the second optical pulse signal data, a fusion weight of 0.6-0.8 is assigned to the piezoelectric signal and a weight of 0.2-0.4 to the PPG signal. When the SNR of the second optical pulse signal data is ≥15dB and higher, a fusion weight of 0.6-0.8 is assigned to the PPG signal and a weight of 0.2-0.4 to the piezoelectric signal. When the SNR of both types of signals is below 15dB, equal weighting (0.5 for each) is used for fusion. The synchronized signals are fused at the data layer using a weighted summation method to generate the first fused data.

[0047] Multidimensional physiological features were extracted from the first fused data, and feature significance analysis was performed on the multidimensional physiological features to obtain core features that are strongly correlated with sleep state.

[0048] Specifically, multidimensional physiological features covering the time domain, frequency domain, and morphology are extracted from the first fused data: For heartbeat and heart rate signals, time domain features (mean, standard deviation, peak interval) and frequency domain features (heart rate variability, including low-frequency components, high-frequency components, and their ratio) are extracted; for respiratory signals, time domain features (respiratory frequency, respiratory depth amplitude) and frequency domain features (respiratory variability) are extracted; for blood oxygenation signals, waveform morphology features (peak width, rising slope), blood oxygenation variability, and trend information (overnight blood oxygenation mean, minimum value, and fluctuation amplitude) are extracted; simultaneously, the amplitude and rate of change of skin conductance activity and the amplitude change information of skin temperature signal are integrated to construct an initial multidimensional physiological feature set containing 20-30 dimensions. The technical effect of this step is to comprehensively capture physiological representations related to sleep state, providing a rich data foundation for subsequent feature selection.

[0049] Furthermore, the Fisher score (F-score) feature evaluation method was used to perform significance analysis on the initial multidimensional physiological feature set to screen core features strongly correlated with sleep state. The Fisher score is calculated as follows: in, This represents the score of a certain feature i in the classification; and Representing the first The feature in the total sample set and the first The mean of the sample set; It represents the first The feature in the first The first in the class dataset A number; This represents the total number of categories; It represents the first The number of samples in the class dataset.

[0050] In this embodiment, the preferred F-score threshold is 1.8. Features with scores higher than the threshold are selected as core features, including respiratory rate variability, low-frequency / high-frequency ratio of heart rate variability, skin temperature change rate, blood oxygen fluctuation amplitude, and body movement amplitude-related features. Redundant features with scores lower than the threshold (such as some signal mean features) are removed.

[0051] Attention mechanisms are used to enhance the core features in the first fused data to obtain the target fused features.

[0052] Specifically, the core features from the first fused data are input into an attention mechanism network to construct a feature weight allocation model. The attention mechanism network dynamically allocates weights by calculating the correlation score between each core feature and the sleep state recognition task: high weights (0.15-0.25) are assigned to features strongly correlated with sleep stages, such as respiratory rate variability and heart rate variability; medium weights (0.05-0.15) are assigned to auxiliary features, such as skin temperature change rate and blood oxygen fluctuation amplitude; and weights related to body movement amplitude are dynamically adjusted based on the real-time intensity of body movement interference (0.1-0.15 for small body movement interference, and reduced to 0.03-0.08 for large body movement interference). The core features are then integrated into a one-dimensional target fusion feature vector through weighted summation, completing the attention mechanism enhancement process.

[0053] S4: Determine the sleep type of the target user based on the target fusion features, and identify the sleep state based on the sleep type of the target fusion features to generate the core sleep indicators corresponding to the identification results.

[0054] In home-based, non-intrusive sleep monitoring and regulation systems, the accuracy of sleep state identification directly determines the targeting and effectiveness of regulation strategies. Existing sleep state identification technologies often use a uniform model to perform indiscriminate analysis on all users, failing to consider the characteristic differences of insomnia subtypes. This results in low accuracy in matching the identification results with the user's actual sleep status, making it difficult to support personalized sleep regulation. Furthermore, traditional identification methods only output qualitative sleep state labels, lacking standardized quantitative indicators, which cannot comprehensively reflect the user's sleep quality and affect the evaluation and optimization of regulation effects.

[0055] This embodiment uses differential analysis of target fusion features to achieve accurate determination of insomnia subtypes and fine identification of sleep states, generating standardized core sleep indicators.

[0056] Preferably, in one embodiment of the present invention, the sleep type of the target user is determined based on the target fusion features, and sleep state identification is performed on the target fusion features based on the sleep type to generate core sleep indicators corresponding to the identification results, including: Feature differentiation analysis is performed on the target fusion features to obtain the sleep type of the target user. Among them, feature differentiation analysis determines the sleep type as difficulty falling asleep, difficulty maintaining sleep, early awakening, or a combination of these by observing the differences in respiratory variability, heart rate variability, body movement amplitude, and blood oxygenation trend.

[0057] Specifically, based on the respiratory variability, heart rate variability, body movement amplitude, and blood oxygenation trend information extracted from the target fusion features, a multi-dimensional feature performance difference determination is made: If the sleep latency (preliminary estimate) in the target fusion features exceeds 30 minutes, and the proportion of high-frequency components of heart rate variability is less than 0.3 and the fluctuation amplitude of respiratory variability is less than 5%, it is determined to be the difficulty falling asleep type. The core feature of this type is poor stability of physiological signals during the sleep onset stage, making it difficult to quickly enter a sleep state; If the feature shows that the frequency of events with body movement amplitude ≥0.5V after falling asleep exceeds 5 times / night, and respiratory variability and heart rate variability fluctuate frequently during the middle of sleep (1-3 hours after falling asleep), it is determined to be the difficulty maintaining sleep type. The core feature is that sleep is easily disturbed and awakened during sleep; If the feature shows that the preliminary estimate of total sleep time is less than 6 hours, and the blood oxygenation trend in the later stage of sleep (more than 4 hours after falling asleep) shows a continuous increase or increased fluctuation, it is determined to be the early awakening type. The core feature is that the sleep cycle terminates prematurely and cannot maintain a complete sleep duration; If the feature determination conditions of two or more subtypes are met at the same time, it is determined to be the mixed type.

[0058] A sleep state recognition model matching the sleep type is determined, and the target fusion features are input into the sleep state recognition model to perform sleep state recognition, thereby obtaining the temporal recognition results of the sleep period and the wakefulness period.

[0059] Specifically, the corresponding recognition model is matched according to the feature complexity of sleep type: for sleep difficulty and early awakening, which have single feature dimensions, the K-Nearest Neighbor (KNN) model is matched, where the number of neighbors K is set to the optimal value in the range of 3-15 through training set validation (K=5 is preferred in this embodiment), and the distance metric is the standardized Mahalanobis distance to ensure sensitive recognition of single feature differences; for sleep maintenance difficulty, which has frequent feature fluctuations, the Support Vector Machine (SVM) model is matched, and the Radial Basis Function (RBF) is selected as the kernel function. The penalty parameter C and the kernel function parameter γ are optimized by grid search combined with cross-validation. In this embodiment, after optimization, C=10 and γ=0.1 to improve the classification accuracy of fluctuating features; for mixed sleep, which has complex feature combinations, the bagging tree model is matched, with the number of base learners set to 50-200 (100 is preferred in this embodiment). The maximum tree depth is limited to within 15 according to the size of the training set to avoid overfitting. The recognition stability of complex features is improved by integrating multiple learners.

[0060] The model training employs a one-night cross-validation strategy. Specifically, the target fusion features for multiple nights (≥7 nights) of the target user are split along the time dimension. One night's data is selected as the test set each time, and the remaining data serves as the training set. Each 30-second time slice in the training set corresponds to a set of target fusion features and manually labeled sleep / wake phase tags. Training and testing are performed iteratively until all nightly data has been tested, ensuring the model adapts to the changes in sleep characteristics across different nights. After training, the target fusion features for the current user are split into 30-second time slices, input into the matching recognition model, and the output is a "sleep phase" or "wake phase" tag for each time slice, forming the time-series recognition result.

[0061] Based on the time sequence recognition results, core sleep indicators are generated, including total sleep time, sleep efficiency, sleep latency, and wakefulness time after falling asleep.

[0062] Specifically, based on the time sequence identification results and the control time (sleep start time, wake-up time) recorded by the system, core indicators are calculated according to the definition of macro sleep structure indicators.

[0063] 1) Total Sleep Time (TST): The cumulative duration of all time slices marked as "sleep period" in the statistical time sequence identification results, that is, the sum of the time of each sleep period (N1, N2, N3, R). 2) Sleep efficiency (SE): Calculated using the formula "SE = (TST / total recording time) × 100%", where the total recording time is the total duration from the time the lights were turned off to the time the lights were turned on, as recorded by the system. 3) Sleep latency (SL): The time interval from the moment the lights are turned off to the start time of the first frame in the time sequence recognition result marked as "sleep period"; 4) Waking Time After Falling Asleep (WASO): The total duration of all time slices marked as "awake" between the start of sleep (the beginning of the first sleep frame) and the time the lights are turned on.

[0064] After calculation, the core indicators are validated. If the sleep duration (SL) exceeds 120 minutes or the sleep energy expenditure (SE) is below 40%, adjustments are made based on body movement data from the piezoelectric monitoring pad and blood oxygen saturation data from the pulse oximeter (PPG) to ensure the accuracy of the indicators. This step generates standardized and quantifiable core sleep indicators that comprehensively reflect the user's sleep quality and provide clear quantitative basis for sleep regulation.

[0065] S5: Generate and execute sleep regulation strategies for target users based on sleep type and core sleep indicators.

[0066] In home-based, unobtrusive sleep monitoring systems, the core objective of sleep regulation is to achieve precise intervention based on the user's personalized sleep characteristics, thereby improving insomnia symptoms and establishing a regular biological clock. Existing sleep regulation technologies mostly employ universal solutions with fixed parameters, failing to consider the differences in insomnia subtypes (difficulty falling asleep, difficulty maintaining sleep, etc.) and core sleep indicators (sleep efficiency SE, sleep latency SL, etc.), resulting in insufficient targeting and limited effectiveness. Furthermore, traditional regulation lacks intelligent optimization mechanisms, cannot dynamically adjust strategies based on the user's sleep improvement progress, and relies on manual user cooperation, leading to poor adherence.

[0067] This embodiment, based on dual inputs of sleep type and core indicators, combines Internet of Things (IoT) technology and reinforcement learning algorithms to generate personalized and adaptive sleep regulation strategies and execute them automatically, achieving closed-loop regulation of type adaptation, indicator-driven, and intelligent optimization.

[0068] In this embodiment, a multimodal state vector is first constructed. Specifically, the state vector is constructed by integrating four types of core data. To provide input for reinforcement learning agents: 1) Sleep type characteristics (labels such as difficulty falling asleep and corresponding physiological characteristic weights); 2) Statistical characteristics of core sleep indicators (mean SE value of the past 7 nights, maximum SL value, cumulative WASO value, etc.); 3) Real-time environmental parameters (light intensity, color temperature, sound volume, room temperature, etc.); 4) Historical control parameters (threshold of bedtime wakefulness in the past 30 days, lighting control parameters, wake-up music type, etc.).

[0069] The aforementioned data is input into an attention mechanism network, where dynamic weights are assigned based on their contribution to sleep regulation. For example, when sleep efficiency (SE) is below 80%, its weight is increased to 0.3, while the weight of color temperature in the environmental parameters is reduced to 0.1, highlighting the dominant role of core indicators. This ultimately forms a weighted multimodal state vector. The technical effect of this step is to achieve effective fusion of multi-source data, providing comprehensive and focused state input for policy generation.

[0070] Furthermore, personalized control strategies are generated based on reinforcement learning. The reinforcement learning agent adopts the Actor-Critic framework and receives state vectors. Then, discrete or continuous control actions are output through the policy network (Actor). This includes adjusting the World Wake-Up Time (WTOB) threshold, configuring sleep microenvironment parameters (light, sound), and setting forced wake-up trigger conditions. The Critic network simultaneously evaluates the long-term rewards of the current state-action pair to guide strategy optimization.

[0071] Specifically, the reward function R is calculated according to the formula " = Sleep index improvement × 0.6 + User wake-up satisfaction × 0.3 - Energy consumption cost × 0.1”. The improvement of sleep index is quantified by the increase in SE and the decrease in SL (e.g., if SE increases from 75% to 85%, it will get 10 points). User wake-up satisfaction is collected through the feedback interface (1-5 points). Energy consumption cost is calculated based on the device's operating power and duration (e.g., if the light runs at full power for 10 minutes, deduct 1 point).

[0072] When generating a regulation strategy, it is necessary to consider the specific performance of sleep type and core indicators. For details, please refer to [link / reference needed]. Figure 2 ,like Figure 2 The diagram shown is a schematic diagram of a forced wake-up mode of a control strategy provided in an embodiment of the present invention.

[0073] Specifically, for those with difficulty falling asleep and SL>40 minutes and SE<70%, the core strategy is to "shorten the WTOB threshold + optimize the sleep microenvironment." The initial WTOB threshold is set to 15 minutes (lower than the default 20 minutes), while configuring low color temperature (2700K), low illuminance (<50 lux), long wavelength (>550nm) light and white noise (loudness 30-40dB). For those with difficulty maintaining sleep and WASO>60 minutes, with frequent excessive body movement amplitude, the core strategy is to "stabilize the sleep environment + reduce nighttime disturbances." Lighting parameters are kept constant, the sound environment is switched to continuous pink noise, and the bed is kept at a fixed tilt angle (0°) to avoid accidental vibration at night. For those with early awakening and TST<6 hours and SE<75%, the core strategy is to "fix the wake-up time + extend the effective sleep cycle." The same wake-up time is set every day (error ≤15 minutes), and nighttime lighting and sound parameters are kept stable to avoid premature awakening signals. For mixed types, the core regulatory actions corresponding to the subtype are integrated, and the regulatory intensity is allocated according to the characteristic weight of each subtype.

[0074] Furthermore, the regulatory actions on the output of the policy network. Safety restraints are implemented as follows: the rate of change in light brightness is ≤10 lux / minute to avoid strong light stimulation; the maximum threshold for sound loudness is ≤60dB to prevent hearing damage; the rate of change in bed tilt angle is ≤10° / second, the maximum tilt angle is 90°, and the vibration intensity is limited to the range that the human body can tolerate (0.3g-0.5g); the threshold for awake time in bed is adjusted within a range of 10-30 minutes, with each adjustment increment of ±3 minutes, to avoid excessive restriction or leniency.

[0075] The controlled actions are stored according to environmental parameters and control parameters. The environmental parameters include light color temperature, illuminance, sound type and loudness, while the control parameters include WTOB threshold, wake-up time, bed adjustment parameters, etc. All of them establish communication protocols (Bluetooth 5.0 or wired connection) with IoT execution devices (smart lights, speakers, adjustable bed, timers) to ensure real-time transmission of instructions.

[0076] In this embodiment, the sleep regulation process is divided into three stages: sleep preparation, intervention during sleep, and forced awakening. 1) Sleep preparation stage (30 minutes before turning off the lights), the system activates microenvironment control, configures the lighting (2700K, <50lux) and sound (white noise or mindfulness meditation audio) according to the strategy, and adjusts the bed to a horizontal position (0° tilt) to guide the user to relax; 2) During the sleep intervention phase, the piezoelectric monitoring pad identifies the user's bed state and wake-sleep state in real time. If the WTOB exceeds the strategy's set threshold, a mild intervention is immediately triggered (the light brightness is slightly adjusted to 5-10 lux, and the sound is switched to a gentle guiding tone). If the user still has not fallen asleep after 5 minutes, a stronger intervention is initiated (the bed vibrates slightly at a frequency of 5 Hz). For users who have difficulty maintaining sleep, the body movement amplitude and blood oxygen trend are monitored every hour at night. If the body movement amplitude is ≥0.5V and lasts for 10 seconds, the ambient noise level is automatically reduced by 5 dB to prevent awakening. 3) Forced wake-up phase: Upon reaching the wake-up time set by the strategy (which remains unchanged), multimodal collaborative wake-up is initiated: the lights gradually brighten at a rate of 10 lux / minute until the maximum illuminance is reached; upbeat music is played, with the loudness gradually increasing from 30dB to 50dB; the bed tilts to 90° at a rate of 10° / second and starts the vibration mode until the piezoelectric monitoring pad detects that the user has left the bed (the signal amplitude drops below the threshold), at which point all wake-up actions stop, and the lights and bed are restored to their initial state.

[0077] Furthermore, at the end of each sleep cycle, this embodiment collects three types of data for strategy updates: 1) Objective sleep indicators (new SE, SL, WASO, etc.) to assess the effectiveness of regulation; 2) User subjective feedback (collecting satisfaction ratings of 1-5 points and post-wake-up discomfort such as dizziness and fatigue through the feedback interface); 3) Equipment energy consumption data (operating power and duration of lights, bed, and audio equipment). Input this data into the reward function for calculation. Generate experience tuples And store it in the experience replay buffer. Periodically sample data from the buffer, optimize the parameters of the policy network and value network through gradient descent, and introduce a target network to enhance training stability. For new users, based on the base model pre-trained on the clinical insomnia dataset, fine-tune it by transferring and learning the sleep data of the user's first three nights to quickly adapt to personalized features; in a multi-user scenario, each user's local device independently trains the model and only uploads the parameters to the federal server for aggregation to ensure data privacy.

[0078] When the continuous monitoring time exceeds 6 days, the system adjusts the WTOB threshold according to the sleep efficiency SE: if SE > 90%, it means that the current WTOB threshold is too loose, increase the threshold by 3 minutes, and gradually extend the time in bed; if 80% < SE < 90%, it means that the threshold adaptability is good and remains unchanged; if SE < 80%, it means that the current WTOB threshold is too long, resulting in low sleep efficiency, reduce the threshold by 3 minutes, and improve sleep concentration by restricting the ineffective time in bed. The adjusted threshold is input into the system corresponding to the control scheme as a new control parameter and participates in the generation of the next round of strategies, forming a cycle of "index feedback, parameter adjustment, and strategy optimization". See specifically Figure 3 , Figure 3 which is a schematic diagram of an adaptive intelligent control system provided by an embodiment of the present invention.

[0079] Another embodiment of the present invention provides a sleep state recognition and control system. Specifically, please refer to Figure 4 , Figure 4 which shows a structural block diagram of a sleep state recognition and control system in one embodiment of the present invention, and it includes: An acquisition module 11, configured to acquire piezoelectric signal data captured by a piezoelectric sensing device and optical pulse signal data captured by a wearable sensing device during the sleep process of a target user; A preprocessing module 12, configured to perform body movement artifact removal processing on the piezoelectric signal data and the optical pulse signal data respectively to obtain first piezoelectric signal data and first optical pulse signal data; A fusion module 13, configured to perform fusion processing on the first piezoelectric signal data and the first optical pulse signal data to obtain target fusion features; An identification module 14, configured to determine the sleep type of the target user based on the target fusion features, and perform sleep state recognition on the target fusion features based on the sleep type to generate sleep core indicators corresponding to the recognition results; A control module 15, configured to generate and execute a sleep control strategy for the target user according to the sleep type and the sleep core indicators.

[0080] Preferably, in an embodiment of the present invention, the preprocessing module is specifically configured to: The piezoelectric signal data is subjected to baseline calibration and noise pre-filtering in sequence to obtain the initial piezoelectric signal data; Based on the frequency characteristics difference between the target user's physiological signals and body movement interference signals, the initial piezoelectric signal data is separated into signal components to obtain the physiological signal components and the body movement interference reference signal; the body movement interference reference signal is used to characterize the interference intensity of body movement behavior on the piezoelectric signal during sleep. Based on the body motion interference reference signal, the physiological signal components are iteratively optimized by Kalman filtering and integrated to generate the first piezoelectric signal data.

[0081] Preferably, in one embodiment of the present invention, the preprocessing module is specifically used for: The optical pulse signal data is sequentially subjected to baseline drift suppression and noise pre-filtering to obtain the initial optical pulse signal data; The initial optical pulse signal data is decomposed into multiple characteristic modal components based on a multi-scale decomposition algorithm; Multidimensional characteristic analysis was performed on each feature modal component, and effective modal feature components were selected based on the analysis results. Among them, the multidimensional characteristic analysis was designed to identify and remove the components dominated by body motion artifacts in each feature modal component based on the prior frequency band features of physiological signals. The effective modal feature components are sequentially reconstructed and blind source separated to obtain the first optical pulse signal data.

[0082] Preferably, in one embodiment of the present invention, the fusion module is specifically used for: The first piezoelectric signal data and the first optical pulse signal data are time-aligned to obtain the second piezoelectric signal data and the second optical pulse signal data after time synchronization. The signal-to-noise ratio of the first piezoelectric signal data and the first optical pulse signal data is determined, and the second piezoelectric signal data and the second optical pulse signal data are fused based on the signal-to-noise ratio to obtain the first fused data; Multidimensional physiological features were extracted from the first fused data, and feature significance analysis was performed on the multidimensional physiological features to obtain core features that are strongly correlated with sleep state. Attention mechanisms are used to enhance the core features in the first fused data to obtain the target fused features.

[0083] Preferably, in one embodiment of the present invention, the identification module is specifically used for: Feature differentiation analysis is performed on the target fusion features to obtain the sleep type of the target user; among them, feature differentiation analysis determines the sleep type as difficulty falling asleep, difficulty maintaining sleep, early awakening, or a combination of the sleep type by the differences in respiratory variability, heart rate variability, body movement amplitude, and blood oxygenation trend. A sleep state recognition model matching the sleep type is determined, and the target fusion features are input into the sleep state recognition model to identify the sleep state, thereby obtaining the temporal recognition results of the sleep period and the wakefulness period. Based on the time sequence recognition results, core sleep indicators are generated, including total sleep time, sleep efficiency, sleep latency, and wakefulness time after falling asleep.

[0084] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for sleep state recognition and regulation, characterized in that, include: Acquire piezoelectric signal data captured by piezoelectric sensing devices and optical pulse signal data captured by wearable sensing devices during the target user's sleep process; The piezoelectric signal data and the optical pulse signal data are subjected to body motion artifact removal processing to obtain the first piezoelectric signal data and the first optical pulse signal data; The first piezoelectric signal data and the first optical pulse signal data are fused together to obtain the target fusion feature; Based on the target fusion features, the sleep type of the target user is determined, and based on the sleep type, the sleep state is identified by the target fusion features, generating core sleep indicators corresponding to the identification results. Based on the sleep type and the core sleep indicators, a sleep regulation strategy for the target user is generated and executed.

2. The sleep state recognition and regulation method as described in claim 1, characterized in that, The step of performing motion artifact removal processing on the piezoelectric signal data and the optical pulse signal data respectively to obtain first piezoelectric signal data and first optical pulse signal data includes: The piezoelectric signal data is subjected to baseline calibration and noise pre-filtering in sequence to obtain the initial piezoelectric signal data; Based on the frequency characteristics difference between the target user's physiological signals and body movement interference signals, the initial piezoelectric signal data is subjected to signal component separation to obtain physiological signal components and body movement interference reference signals; wherein the body movement interference reference signals are used to characterize the interference intensity of body movement behavior on piezoelectric signals during sleep. Based on the body motion interference reference signal, the physiological signal components are subjected to Kalman filtering iterative optimization and integrated to generate the first piezoelectric signal data.

3. The sleep state recognition and regulation method as described in claim 1, characterized in that, The step of performing motion artifact removal processing on the piezoelectric signal data and the optical pulse signal data respectively to obtain first piezoelectric signal data and first optical pulse signal data includes: The optical pulse signal data is sequentially subjected to baseline drift suppression and noise pre-filtering to obtain the initial optical pulse signal data; The initial optical pulse signal data is decomposed into multiple characteristic modal components based on a multi-scale decomposition algorithm; Multidimensional characteristic analysis is performed on each of the aforementioned characteristic modal components, and effective modal characteristic components are obtained by screening based on the analysis results; wherein, the multidimensional characteristic analysis is designed to identify and eliminate the components dominated by body motion artifacts in each of the aforementioned characteristic modal components based on the prior frequency band characteristics of physiological signals. The effective modal feature components are sequentially reconstructed and blind source separated to obtain the first optical pulse signal data.

4. The sleep state recognition and regulation method as described in claim 1, characterized in that, The step of fusing the first piezoelectric signal data and the first optical pulse signal data to obtain target fusion features includes: The first piezoelectric signal data and the first optical pulse signal data are time-aligned to obtain the second piezoelectric signal data and the second optical pulse signal data after time synchronization. The signal-to-noise ratio of the first piezoelectric signal data and the first optical pulse signal data is determined, and the second piezoelectric signal data and the second optical pulse signal data are fused based on the signal-to-noise ratio to obtain the first fused data; Multidimensional physiological features are extracted from the first fused data, and feature significance analysis is performed on the multidimensional physiological features to obtain core features that are strongly correlated with sleep state. The core features in the first fused data are enhanced using an attention mechanism to obtain the target fused features.

5. The sleep state recognition and regulation method as described in claim 1, characterized in that, The process of determining the sleep type of the target user based on the target fusion features, and performing sleep state identification on the target fusion features based on the sleep type to generate core sleep indicators corresponding to the identification results includes: The target fusion features are subjected to feature differentiation analysis to obtain the sleep type of the target user; wherein, the feature differentiation analysis determines the sleep type as difficult to fall asleep, difficult to maintain sleep, early awakening, or a combination of the sleep type by the differences in respiratory variability, heart rate variability, body movement amplitude, and blood oxygenation trend. A sleep state recognition model matching the sleep type is determined, and the target fusion features are input into the sleep state recognition model to perform sleep state recognition, thereby obtaining the temporal recognition results of the sleep period and the wakefulness period. Based on the time sequence recognition results, core sleep indicators are generated, including total sleep time, sleep efficiency, sleep latency, and wakefulness time after falling asleep.

6. A sleep state recognition and control system, characterized in that, include: The acquisition module is used to acquire piezoelectric signal data captured by the piezoelectric sensing device and optical pulse signal data captured by the wearable sensing device during the target user's sleep process. The preprocessing module is used to perform motion artifact removal processing on the piezoelectric signal data and the optical pulse signal data respectively to obtain the first piezoelectric signal data and the first optical pulse signal data; The fusion module is used to fuse the first piezoelectric signal data and the first optical pulse signal data to obtain the target fusion feature; The identification module is used to determine the sleep type of the target user based on the target fusion features, and to identify the sleep state of the target fusion features based on the sleep type, thereby generating core sleep indicators corresponding to the identification results. The control module is used to generate and execute the sleep control strategy for the target user based on the sleep type and the core sleep indicators.

7. The sleep state recognition and control system as described in claim 6, characterized in that, The preprocessing module is specifically used for: The piezoelectric signal data is subjected to baseline calibration and noise pre-filtering in sequence to obtain the initial piezoelectric signal data; Based on the frequency characteristics difference between the target user's physiological signals and body movement interference signals, the initial piezoelectric signal data is subjected to signal component separation to obtain physiological signal components and body movement interference reference signals; wherein the body movement interference reference signals are used to characterize the interference intensity of body movement behavior on piezoelectric signals during sleep. Based on the body motion interference reference signal, the physiological signal components are subjected to Kalman filtering iterative optimization and integrated to generate the first piezoelectric signal data.

8. The sleep state recognition and control system as described in claim 6, characterized in that, The preprocessing module is specifically used for: The optical pulse signal data is sequentially subjected to baseline drift suppression and noise pre-filtering to obtain the initial optical pulse signal data; The initial optical pulse signal data is decomposed into multiple characteristic modal components based on a multi-scale decomposition algorithm; Multidimensional characteristic analysis is performed on each of the aforementioned characteristic modal components, and effective modal characteristic components are obtained by screening based on the analysis results; wherein, the multidimensional characteristic analysis is designed to identify and eliminate the components dominated by body motion artifacts in each of the aforementioned characteristic modal components based on the prior frequency band characteristics of physiological signals. The effective modal feature components are sequentially reconstructed and blind source separated to obtain the first optical pulse signal data.

9. The sleep state recognition and control system as described in claim 6, characterized in that, The fusion module is specifically used for: The first piezoelectric signal data and the first optical pulse signal data are time-aligned to obtain the second piezoelectric signal data and the second optical pulse signal data after time synchronization. The signal-to-noise ratio of the first piezoelectric signal data and the first optical pulse signal data is determined, and the second piezoelectric signal data and the second optical pulse signal data are fused based on the signal-to-noise ratio to obtain the first fused data; Multidimensional physiological features are extracted from the first fused data, and feature significance analysis is performed on the multidimensional physiological features to obtain core features that are strongly correlated with sleep state. The core features in the first fused data are enhanced using an attention mechanism to obtain the target fused features.

10. The sleep state recognition and control system as described in claim 6, characterized in that, The identification module is specifically used for: The target fusion features are subjected to feature differentiation analysis to obtain the sleep type of the target user; wherein, the feature differentiation analysis determines the sleep type as difficult to fall asleep, difficult to maintain sleep, early awakening, or a combination of the sleep type by the differences in respiratory variability, heart rate variability, body movement amplitude, and blood oxygenation trend. A sleep state recognition model matching the sleep type is determined, and the target fusion features are input into the sleep state recognition model to perform sleep state recognition, thereby obtaining the temporal recognition results of the sleep period and the wakefulness period. Based on the time sequence recognition results, core sleep indicators are generated, including total sleep time, sleep efficiency, sleep latency, and wakefulness time after falling asleep.

Images