Quaternary sleep state monitoring method based on non-contact sensor

Abstract

The invention particularly relates to a quartered sleep state monitoring method and a quartered sleep state monitoring system based on a non-contact vibration sensor. The quartered sleep state monitoring method and the quartered sleep state monitoring system are particularly suitable for realizing non-sensitive sleep quality evaluation through an intelligent bed in a family environment. Comprising the following steps that 1, vibration signals of a human body in the sleep process are continuously collected through a non-contact vibration sensor; step 2, preprocessing the vibration signal, including denoising and segmentation processing, to obtain time domain vibration data; step 3, performing frequency domain transformation on the time domain vibration data, and extracting frequency domain features including respiratory rate, heart rate and heart variability parameters; 4, time alignment is carried out on the frequency domain features and sleep staging labels synchronously obtained through a polysomnography, and a training data set is constructed; an XGBoost classification model based on time-frequency domain feature fusion is provided, fundamental frequency and harmonic energy proportions of respiratory signals are extracted through short-time Fourier transform, and quartering classification is achieved in combination with the LF / HF ratio of heart rate variability.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent health monitoring, and particularly relates to a method and system for monitoring four-stage sleep states based on a non-contact vibration sensor, which is especially suitable for realizing non-invasive sleep quality assessment through an intelligent bed in a home environment. Background Art

[0002] Traditional polysomnography (PSG) needs to paste electrodes on the user's body surface to collect signals such as electroencephalogram and electrocardiogram, which has problems such as strong invasiveness and complex operation. Existing non-contact technologies such as patent CN109480787B use ultra-wideband radar to monitor breathing and body movement, but the radar equipment is costly (cost per unit > 2000 yuan), and the signal attenuation is serious when penetrating clothes (attenuation rate > 40%), resulting in a decline in monitoring accuracy. Patent CN111067503A analyzes heart rate variability (HRV) based on ECG signals, but requires contact electrodes and cannot avoid interfering with sleep.

[0003] In view of the above pain points, the present invention proposes a low-cost and high-precision solution: using a non-contact vibration sensor (such as a piezoelectric film, cost < 50 yuan per piece) embedded in an intelligent bed to collect human vibration signals, and combining an XGBoost model optimized by transfer learning to achieve four-stage sleep monitoring, with a classification accuracy rate of 92.3% (an increase of 12.5% compared to the radar solution), and no need for the user's active cooperation. Summary of the Invention

[0004] The present application provides a method for monitoring four-stage sleep states based on a non-contact sensor, including the following steps:

[0005] Step 1: Continuously collect human vibration signals during sleep through a non-contact vibration sensor;

[0006] Step 2: Preprocess the vibration signals, including denoising and segmentation, to obtain time-domain vibration data;

[0007] Step 3: Perform a frequency-domain transformation on the time-domain vibration data to extract frequency-domain features, including respiratory rate, heart rate, and cardiac variability parameters;

[0008] Step 4: Align the frequency-domain features with the sleep stage labels synchronously obtained through polysomnography in time to construct a training dataset;

[0009] Step 5: According to the label distribution of the collected dataset and the prior distribution of the true sleep stages, screen the dataset so that the number of each sleep stage state in the dataset conforms to the distribution of the true sleep states; use the XGBoost algorithm to train the training dataset to generate a four-stage sleep state classifier;

[0010] Step 6: In real-time monitoring, the newly collected vibration signals are processed through Step 2-3 to obtain sleep characteristic data. Time characteristics are added according to the sleep time and combined with the sleep characteristic data, and then input into the classifier to output the classification results of the wake, light sleep, deep sleep, and REM sleep stages;

[0011] Step 7: Post-process the classification results, including time series filtering and noise removal, and modify the classification results by combining the characteristic information such as getting out of bed given by the original sensor data to generate an all-night sleep staging status map.

[0012] Furthermore, the non-contact vibration sensor is a piezoelectric film sensor or an acceleration sensor, integrated in the mattress or bed frame of the smart bed, and is used to detect human body movements and chest vibration signals.

[0013] Furthermore, in Step 3, the frequency domain transformation is the short-time Fourier transform (STFT) or wavelet transform, and the frequency domain characteristics further include: 1) the fundamental frequency and harmonic energy distribution of the respiratory signal; 2) the power spectral density in the 0.5-4 Hz frequency band of the heart rate signal; 3) the time-frequency joint analysis results of cardiac variability parameters.

[0014] Furthermore, in Step 4, the time alignment is specifically as follows:

[0015] Through timestamp matching, the sleep staging labels marked by the polysomnograph are synchronized with the feature extraction window of the vibration signal at the millisecond level to ensure the temporal consistency of the features and labels in the dataset.

[0016] Furthermore, in Step 5, the loss function of the XGBoost algorithm is defined as:

[0017] L = ∑i = 1nI(f(xi) ≠ yi) + λΩ(f)

[0018] where I is the indicator function, Ω(f) is the model complexity regularization term, and λ is the hyperparameter.

[0019] Furthermore, in Step 7, the post-processing includes:

[0020] Adopt moving window mean filtering with a window length of 5 minutes;

[0021] Based on the state transition probability matrix (the probability of wake → light sleep is 0.6, the probability of light sleep → deep sleep is 0.3, and the probability of deep sleep → REM is 0.4), correct the abnormal classification results.

[0022] Furthermore, it includes:

[0023] A non-contact vibration sensor array is arranged in the support structure of the smart bed;

[0024] A signal processing module for performing the preprocessing and feature extraction of the vibration signals.

[0025] A machine learning module embedded with an XGBoost classifier for real-time sleep stage prediction.

[0026] A user interface for displaying the sleep stage status graph and health advice through a mobile terminal APP.

[0027] Furthermore, the system further includes an environmental sensor for collecting environmental temperature, humidity and noise data, and fusing the vibration signal features and inputting them into the classifier to improve the recognition accuracy of the REM stage.

[0028] Furthermore, the method adopts a transfer learning strategy in the training stage: pre-training stage: training a general classifier based on a public sleep dataset (such as Sleep-EDF); fine-tuning stage: incrementally training the classifier using the user's personalized data and fine-tuning the hyperparameter λ and the tree depth.

[0029] Furthermore, the method also includes abnormal sleep event detection: triggering apnea or arrhythmia alarms according to a sudden drop in the breathing rate (drop amplitude > 30%) or a sudden change in the heart rate (standard deviation > 15%); the alarm signal is transmitted to the user terminal and the emergency contact device via Bluetooth.

[0030] The beneficial effects of this solution are:

[0031] 1) Hardware layer: Design a low-power piezoelectric film sensor array, which is embedded in the upper polyurethane foam of the smart bed mattress in a 5×5 grid form (spacing 15 cm), and synchronously extracts the breathing rate (error ±0.5 times / minute), heart rate (error ±2 bpm) and body movement signals by detecting the chest micro-vibration (frequency 0.1 - 20 Hz), with a sensitivity of 0.01 m / s 2 .

[0032] 2) Algorithm layer: Propose an XGBoost classification model based on time-frequency domain feature fusion, extract the fundamental frequency (0.2 - 0.33 Hz) of the breathing signal and the harmonic energy ratio (> 60% is a deep sleep feature) through short-time Fourier transform (STFT, window length 2 seconds, overlap rate 50%), and combine the LF / HF ratio of heart rate variability (HRV) (> 1.5 is a REM marker) to achieve four-stage classification.

[0033] 3) System layer: Integrate an environmental temperature and humidity sensor (SHT35, accuracy ±0.2 °C) and a noise detection module (MAX9814, frequency response 20 - 20 kHz), and improve the REM stage recognition rate from 78% to 89% through multi-modal data fusion (vibration + environment). Description of the Drawings

[0034] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0035] Figure 1 Flowchart of the four-stage sleep staging monitoring method based on a non-contact sensor according to the present invention

[0036] Figure 2 Flowchart of the four-stage sleep staging monitoring method in the embodiments of the present invention Detailed implementation manners

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application. In addition, it should be understood that the specific implementation manners described herein are only used to explain and illustrate the present application, and are not used to limit the present application. In the present application, unless otherwise stated, the orientation terms such as "upper", "lower", "left", and "right" generally refer to the upper, lower, left, and right in the actual use or working state of the device, specifically the drawing direction in the drawings.

[0038] As Figure 1 、 2 shown, a four-stage sleep state monitoring method based on a non-contact sensor includes the following steps:

[0039] Step 1: Continuously collect vibration signals of the human body during sleep through a non-contact vibration sensor; in this embodiment, 16 PVDF piezoelectric film sensors (size 5 cm × 5 cm, thickness 0.2 mm) are embedded in the mattress in a 4×4 array, with a sampling frequency of 100 Hz and a dynamic range of ±2 g. The sensor output is preprocessed by an AD623 amplifier (gain 100 times) and a fourth-order Butterworth low-pass filter (cutoff frequency 25 Hz) to eliminate high-frequency noise.

[0040] Step 2: Preprocess the vibration signals, including denoising and segmentation processing, to obtain time-domain vibration data; in this embodiment, the original signal is subjected to moving average filtering (window length 1 second) to remove instantaneous interference. Through three-layer decomposition and reconstruction based on the DB4 wavelet, the respiratory signal (0.1 - 0.5 Hz) and the heartbeat signal (0.5 - 4 Hz) are separated.

[0041] Step 3: Perform frequency-domain transformation on the time-domain vibration data to extract frequency-domain features, including respiratory rate, heart rate, and heart rate variability parameters. In this embodiment, for respiratory features, calculate the number of breaths per minute (BR) and its standard deviation (SDBR). During deep sleep, SDBR < 1.2 breaths per minute. For heart rate variability, extract SDNN (normal value > 50 ms) and RMSSD (normal value > 30 ms) of the RR interval. During REM sleep, SDNN decreases by 15%-20%. For body movement index, calculate the root mean square (RMS) of the acceleration signal, and the threshold > 0.1 m / s 2 It is determined as a body movement event.

[0042] Step 4: Align the frequency-domain features with the sleep stage labels synchronously obtained by a polysomnograph in time to construct a training dataset. In this embodiment, the polysomnograph (Compumedics Grael) and the vibration sensor data achieve time synchronization through the NTP protocol, with an error < 10 ms. The labeled data is segmented by 30-second windows and matched frame by frame with the feature matrix.

[0043] Step 5: According to the label distribution of the collected dataset and the prior distribution of the true sleep stages, perform screening on the dataset so that the number of each sleep stage state in the dataset conforms to the distribution of the true sleep states. Use the XGBoost algorithm to train the training dataset to generate a four-stage sleep state classifier. XGBoost parameters: tree depth = 6, learning rate = 0.1, λ = 1.0 (L2 regularization), and 200 training iterations. Use 5-fold cross-validation, with an average accuracy of 92.3% (85.7% for the control group SVM).

[0044] Step 6: In real-time monitoring, process the newly collected vibration signal through Steps 2-3 to obtain sleep feature data, add time features according to the sleep time, combine them with the sleep feature data, input them into the classifier, and output the classification results of the awake, light sleep, deep sleep, and REM sleep stages. Update the classification results every 30 seconds, with a delay < 0.5 seconds. Output probability vector: P = [awake, light sleep, deep sleep, REM], and a threshold > 0.7 is determined as a valid state.

[0045] Step 7: Post-process the classification results, including time series filtering and noise removal, and modify the classification results by combining feature information such as getting out of bed given by the original sensor data to generate an all-night sleep stage state map. Apply the Viterbi algorithm to correct abnormal jumps based on the state transition probability matrix (awake → light sleep = 0.6, light sleep → deep sleep = 0.3, deep sleep → REM = 0.4). Generate a sleep structure diagram, and mark the duration and transition times of each stage.

[0046] The above steps use a non-contact vibration sensor to collect human vibration signals, avoiding the discomfort of traditional contact sensors. By performing frequency-domain transformation, respiratory rate, heart rate, and heart rate variability parameters are extracted, and sleep staging is achieved by combining machine learning algorithms. The XGBoost algorithm is used for classification, which has high classification accuracy and real-time performance. Traditional sleep monitoring methods require users to wear multiple contact sensors, with complex operations and affecting the natural sleep state. Existing non-contact technologies (such as radar or optoelectronic sensors) are costly and have complex signal processing, making it difficult to popularize and apply. This application provides a non-invasive and low-cost sleep monitoring method suitable for daily home use. High-precision four-stage sleep state monitoring is achieved through machine learning algorithms, improving the user experience.

[0047] Further, the non-contact vibration sensor is a piezoelectric film sensor or an acceleration sensor, integrated into the mattress or bed frame of the intelligent bed, and is used to detect human body movements and thoracic vibration signals. In this embodiment, by using a piezoelectric film sensor or an acceleration sensor, human vibration signals can be detected with high sensitivity. The sensor is integrated into the intelligent bed, eliminating the need for users to wear additional devices, which is convenient to use. The sensor array is integrated using a flexible FPC circuit board, allowing the mattress to have a bending radius > 50 mm without affecting the signal quality, and is suitable for adjustable bed types. Traditional sensors need to be worn by users, affecting sleep comfort. Existing non-contact sensors (such as radar) are costly and difficult to popularize. This application provides a low-cost and high-sensitivity non-contact vibration sensor solution. The sensor is integrated into the intelligent bed and is suitable for daily home use.

[0048] Further, the frequency-domain transformation in step 3 is the short-time Fourier transform (STFT) or wavelet transform, and the frequency-domain features further include: 1) the fundamental frequency and harmonic energy distribution of the respiratory signal; 2) the power spectral density of the heart rate signal in the 0.5 - 4 Hz frequency band; 3) the time-frequency joint analysis result of the heart rate variability parameters. In this embodiment, by using the short-time Fourier transform (STFT) or wavelet transform for frequency-domain analysis, richer features can be extracted. The frequency-domain features include the fundamental frequency of the respiratory signal, the power spectral density of the heart rate signal, and the heart rate variability parameters. The STFT parameters are optimized to a Hanning window (window length 2 seconds), with a frequency-domain resolution of 0.5 Hz, and the proportion of respiratory harmonic energy (1 - 3 harmonics) is extracted as a deep sleep stage feature. Traditional time-domain analysis methods are difficult to capture the subtle changes in the sleep state. Existing frequency-domain analysis methods have insufficient feature extraction, affecting the classification accuracy. This application extracts richer features through frequency-domain transformation, improving the classification accuracy. Combining time-frequency joint analysis can more accurately reflect the changes in the sleep state.

[0049] Further, the time alignment in step 4 is specifically:

[0050] Through timestamp matching, the sleep stage labels marked by the polysomnograph are synchronized with the feature extraction window of the vibration signal at the millisecond level, ensuring the temporal consistency of features and labels in the dataset. In this embodiment, a timestamp matching method is adopted to achieve millisecond-level synchronization of multi-source data. The time alignment uses a hardware trigger signal to establish a GPIO synchronization interface between the PSG and the vibration sensor to ensure frame-level alignment of data (error < 1 ms). Ensure the temporal consistency of features and labels, and improve the model training effect. It can solve the problem in the prior art that the time of multi-source data (such as vibration signals and polysomnograph data) is out of sync, affecting the model training effect. Through millisecond-level time alignment, ensure the temporal consistency of the dataset. Improve the model training effect and improve the classification accuracy.

[0051] Further, the loss function of the XGBoost algorithm described in step 5 is defined as:

[0052] L = ∑i = 1nI(f(xi) ≠ yi) + λΩ(f)

[0053] where II is the indicator function, Ω(f)Ω(f) is the model complexity regularization term, and λλ is the hyperparameter. In this embodiment, the loss function introduces class weights, weights the REM samples 1.5 times, solves the problem of class imbalance, and the recall rate is increased by 18%. Using the XGBoost algorithm for classification has high classification accuracy and real-time performance. The loss function includes an indicator function and a model complexity regularization term to prevent overfitting. Traditional classification algorithms (such as SVM) have high computational complexity when dealing with high-dimensional features and are difficult to apply in real time. Existing ensemble learning algorithms (such as random forests) are prone to overfitting, affecting the classification effect. Achieve high-precision and low-latency sleep staging through the XGBoost algorithm. The loss function design effectively prevents overfitting and improves the generalization ability of the model.

[0054] Further, the post-processing described in step 7 includes:

[0055] Adopt sliding window mean filtering with a window length of 5 minutes;

[0056] Based on the state transition probability matrix (the probability of waking → light sleep is 0.6, the probability of light sleep → deep sleep is 0.3, and the probability of deep sleep → REM is 0.4), correct the abnormal classification results. In this embodiment, sliding window mean filtering is adopted to smooth the classification results.

[0057] Based on the state transition probability matrix, correct the abnormal classification results and improve the classification accuracy. In the previous technology, there are noise and outliers in the classification results, affecting the user experience. Existing post-processing methods are difficult to effectively correct abnormal classification results. This application improves the smoothness and accuracy of the classification results through sliding window mean filtering and state transition probability matrix correction. Improve the user experience and provide more reliable sleep staging results.

[0058] Further, it includes:

[0059] A non-contact vibration sensor array arranged in the support structure of the smart bed;

[0060] A signal processing module for performing the preprocessing and feature extraction of the vibration signals;

[0061] A machine learning module embedded with an XGBoost classifier for real-time sleep staging prediction;

[0062] A user interaction interface for displaying the sleep staging status graph and health suggestions through a mobile terminal APP. In this embodiment, the system integrates a non-contact vibration sensor, a signal processing module, a machine learning module, and a user interaction interface to achieve full-process sleep monitoring. The user interaction interface displays the sleep staging status graph and health suggestions through the mobile terminal APP, enhancing the user experience. Existing sleep monitoring systems have a single function and are difficult to meet the diverse needs of users. The user interaction interface is not user-friendly, affecting the usage experience. This application provides a full-process and multi-functional sleep monitoring system to meet the diverse needs of users. Through a friendly user interaction interface, the user experience is enhanced. In this embodiment, the system integrates a non-contact vibration sensor, a signal processing module, a machine learning module, and a user interaction interface to achieve full-process sleep monitoring. The user interaction interface displays the sleep staging status graph and health suggestions through the mobile terminal APP, enhancing the user experience. Existing sleep monitoring systems have a single function and are difficult to meet the diverse needs of users. The user interaction interface is not user-friendly, affecting the usage experience. This application provides a full-process and multi-functional sleep monitoring system to meet the diverse needs of users. Through a friendly user interaction interface, the user experience is enhanced.

[0063] Further, the system further includes an environmental sensor for collecting environmental temperature, humidity, and noise data, and fusing it with the vibration signal features and inputting them into the classifier to improve the recognition accuracy of the REM stage. In this embodiment, the environmental sensor data is fused with the vibration features through Kalman filtering, and when the change in temperature and humidity > 10%, the model parameter adaptive adjustment is triggered, and the REM recognition rate is increased by 11%. The system integrates an environmental sensor to collect environmental temperature, humidity, and noise data. The environmental data is fused with the vibration signal features and input into the classifier to improve the recognition accuracy of the REM stage. Existing sleep monitoring systems do not consider the influence of environmental factors on sleep states. The recognition accuracy of the REM stage is relatively low, affecting the overall classification effect. This application improves the recognition accuracy of the REM stage by fusing environmental data with vibration signal features. It provides a more comprehensive sleep monitoring service and improves the user experience.

[0064] Furthermore, the method adopts a transfer learning strategy in the training phase: Pre-training phase: Train a general classifier based on a publicly available sleep dataset (such as Sleep-EDF); Fine-tuning phase: Use the user's personalized data to incrementally train the classifier, and fine-tune the hyperparameter λ and the tree depth. In this embodiment, a transfer learning strategy is adopted to train the model by combining the publicly available dataset and the user's personalized data. In the fine-tuning phase, the hyperparameter λ and the tree depth are adjusted to enhance the personalization ability of the model. Existing model training methods rely on a large amount of labeled data and are difficult to quickly adapt to personalized needs. The generalization ability of the model is insufficient, which affects the classification effect. This application can quickly adapt to personalized needs and enhance the generalization ability of the model through the transfer learning strategy, providing a more accurate sleep staging service and improving the user experience.

[0065] Furthermore, the method further includes the detection of abnormal sleep events: Trigger an apnea or arrhythmia alarm according to a sudden drop in the respiratory rate (the drop amplitude > 30%) or a sudden change in the heart rate (the standard deviation > 15%); The alarm signal is transmitted to the user terminal and the emergency contact device via Bluetooth. In this embodiment, the sliding window CUSUM algorithm is used for apnea detection, and the alarm is triggered when the respiratory amplitude drops > 30% continuously for 5 times, with a false alarm rate < 2 times / night. The system monitors the respiratory rate and heart rate in real time to detect abnormal sleep events. The alarm signal is transmitted to the user terminal and the emergency contact device via Bluetooth to provide timely warning. Existing sleep monitoring systems do not provide the function of detecting abnormal sleep events. Abnormal sleep events (such as apnea and arrhythmia) are difficult to detect in time, posing a health risk. This application can detect abnormal sleep events in time and reduce the health risk through real-time monitoring and alarm functions, providing a more comprehensive sleep health management service and improving the user experience.

[0066] It can be understood that the meanings of the nouns in each embodiment of this application are the same. For the content not detailed in a certain embodiment, the specific implementation details can refer to the descriptions in other embodiments. The example descriptions and technical effects shown in the foregoing embodiments can be correspondingly realized. For the repeated parts, this embodiment will not be elaborated further.

[0067] The above has introduced in detail the sonic massage device, electric bed, double combination bed, sonic massage method and storage medium provided by this application. Specific examples are used in this article to elaborate on the principle and implementation manner of this application. The descriptions of the above embodiments are only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A four-stage sleep state monitoring method based on a non-contact sensor, characterized in that: The following steps are involved: Step 1: Continuously collect vibration signals of the human body during sleep through a non-contact vibration sensor; Step 2: preprocessing the vibration signal, including denoising and segmentation processing, to obtain time domain vibration data; Step 3: performing frequency domain transformation on the time domain vibration data to extract frequency domain features, including respiratory rate, heart rate and heart variability parameters; Step 4: Time-aligning the frequency domain features with the sleep stage labels synchronously acquired through the polysomnography to construct a training data set; Step 5: According to the collected data set label distribution and the prior distribution of the real sleep stage, the data set is screened so that the number of each sleep stage state in the data set conforms to the distribution of the real sleep state; the training data set is trained using the XGBoost algorithm to generate a four-stage sleep state classifier; Step 6: In real-time monitoring, the newly collected vibration signal is processed through steps 2-3 to obtain sleep characteristic data, and a time feature is added according to the sleep time, combined with the sleep characteristic data, input into the classifier, and outputs the classification results of awake, light sleep, deep sleep and REM sleep stages; Step 7: Post-process the classification results, including time series filtering and noise removal, and modify the classification results based on the characteristic information such as leaving the bed given by the original sensor data to generate a whole night sleep stage state diagram.

2. The method according to claim 1, characterized in that: The non-contact vibration sensor is a piezoelectric film sensor or an acceleration sensor, which is integrated into the mattress or bed frame of the smart bed and is used to detect human body movement and chest vibration signals.

3. The method according to claim 1, characterized in that The frequency domain transformation in step 3 is short-time Fourier transform (STFT) or wavelet transform, and the frequency domain features further include: 1) fundamental frequency and harmonic energy distribution of respiratory signal; 2) power spectrum density of 0.5-4Hz frequency band of heart rate signal; 3) time-frequency joint analysis results of cardiac variability parameters.

4. The method according to claim 1, characterized in that The time alignment in step 4 is specifically as follows: Through timestamp matching, the sleep stage labels marked by the polysomnography and the feature extraction window of the vibration signal are synchronized at the millisecond level to ensure the temporal consistency of the features and labels in the dataset.

5. The method according to claim 1, characterized in that: The loss function of the XGBoost algorithm described in step 5 is defined as: L=∑i=1nI(f(xi)≠yi)+λΩ(f) Among them, II is the indicator function, Ω(f)Ω(f) is the model complexity regularization term, and λλ is a hyperparameter.

6. The method according to claim 1, characterized in that The post-processing described in step 7 includes: Sliding window mean filtering is used, and the window length is 5 minutes; The abnormal classification results were corrected based on the state transition probability matrix (the probability of awake→light sleep is 0.6, the probability of light sleep→deep sleep is 0.3, and the probability of deep sleep→REM is 0.4).

7. A four-stage sleep state monitoring system based on a non-contact sensor, characterized in that: include: A non-contact vibration sensor array is arranged in the support structure of the smart bed; A signal processing module, used to perform vibration signal preprocessing and feature extraction as described in claims 1-3; A machine learning module, embedded with XGBoost classifier, for real-time sleep stage prediction; The user interaction interface displays sleep stage status diagrams and health suggestions through a mobile terminal APP.

8. The system according to claim 7, characterized in that The system further includes an environmental sensor for collecting environmental temperature, humidity and noise data, and fusing the data with vibration signal features into a classifier to improve the recognition accuracy of the REM stage.

9. The method according to claim 1, characterized in that: The method adopts a transfer learning strategy in the training phase: Pre-training phase: a general classifier is trained based on a public sleep dataset (such as Sleep-EDF); Fine-tuning stage: Use user personalized data to incrementally train the classifier and fine-tune the hyperparameter λλ and tree depth.

10. The method according to claim 1, characterized in that The method also includes abnormal sleep event detection: triggering a apnea or arrhythmia alarm based on a sudden drop in respiratory rate (drop > 30%) or a sudden change in heart rate (standard deviation > 15%); the alarm signal is transmitted to the user terminal and emergency contact device via Bluetooth.

Images