Sleep staging method and system

Abstract

The invention discloses a sleep staging method and system. The method comprises the steps that physiological feature original signals of a target object are obtained; performing signal state detection on the physiological feature original signal; if the user leaves the bed or the signal is invalid, performing data cleaning on the sleep staging data of the day to obtain a final sleep staging result, and if the user is in the bed state and the signal is valid, performing preprocessing operation on the physiological feature original signal, and performing separation to obtain at least two physiological parameter signals related to the sleep state; performing multi-dimensional feature extraction on the physiological parameter signals to construct a target physiological feature array; based on the target physiological feature array, the target object information array and the sleep state information array, a sleep staging result corresponding to the current moment is obtained through a preset sleep staging model. According to the invention, non-inductive home monitoring can be realized by relying on non-intrusive equipment such as an intelligent mattress and an intelligent pillow in an intelligent home scene, and the signal anti-interference capability and the sleep staging accuracy are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart home, and in particular to a sleep staging method and system. BACKGROUND

[0002] With the rapid development of the smart home industry, sleep health management has become one of the core application scenarios of the smart home ecosystem. Smart mattress, smart pillow, wearable bracelet and other smart home devices are gradually penetrating the home sleep monitoring scenario. Users are increasingly demanding a non-invasive, convenient and accurate sleep staging and health assessment through smart home devices.

[0003] Current sleep staging technology has significant bottlenecks in signal acquisition and automatic staging methods. In the signal acquisition layer, the traditional gold standard polysomnography (PSG) needs to synchronously collect multiple signals such as electroencephalogram (EEG), electrooculogram (EOG) and electromyogram (EMG). Although the staging accuracy is high, it has problems such as complex operation, high equipment cost and poor wearing comfort, which makes it difficult to meet the daily home monitoring needs. While various simplified monitoring devices (such as wearable devices based on photoplethysmogram (PPG) and smart mattresses / pillows based on ballistocardiogram (BCG)) have the advantages of simple operation, low cost and strong comfort, but the signal acquisition dimension is single, which is easily affected by environmental interference and human body movement, resulting in insufficient sleep staging accuracy and limiting its application scenarios and user acceptance.

[0004] In the automatic staging method layer, traditional machine learning methods rely on expert hand-crafted time and frequency domain features, and then use SVM, XGBoost and other classifiers to realize staging. The model has low complexity, but the feature extraction is highly subjective and has weak generalization ability, and the staging accuracy is limited. Although deep learning methods can automatically learn signal features and reduce human dependence, they require a large amount of data and high scene coverage, and have large model parameter quantity and high complexity, which makes it difficult to adapt to the lightweight application needs in the home scenario. At the same time, existing technologies generally lack accurate judgment of signal effectiveness, do not fully consider the influence of body movement interference on physiological signals, and do not effectively fuse personalized information and sleep timing state of the target object, resulting in that the staging results are easily affected by noise, individual differences and other factors, and it is difficult to balance the convenience of monitoring and the accuracy of staging.

[0005] Therefore, there is an urgent need for a sleep staging technology that balances the convenience of operation, signal anti-interference ability and staging accuracy. SUMMARY

[0006] Therefore, it is necessary to provide a sleep staging method and system to solve at least one problem in the prior art.

[0007] In a first aspect, a sleep staging method is provided, comprising: acquire a physiological characteristic original signal of a target object; perform signal state detection on the physiological characteristic original signal, the signal state detection at least including bed exit state detection and signal validity detection; if the target object is in a bed exit state or the physiological characteristic original signal is invalid, and for a preset duration, acquire current-day sleep staging data of the target object, and perform data cleaning on the current-day sleep staging data to obtain a final sleep staging result; if the target object is in a bed exit state or the physiological characteristic original signal is invalid, and for a preset duration, acquire current-day sleep staging data of the target object, and perform data cleaning on the current-day sleep staging data to obtain a final sleep staging result; perform multi-dimensional feature extraction on the physiological parameter signals to construct a target physiological feature array; obtain a sleep staging result corresponding to a current time based on the target physiological feature array, a target object information array, and a sleep state information array through a preset sleep staging model.

[0008] In a possible implementation, the obtaining of the sleep staging result corresponding to the current time based on the target physiological feature array, the target object information array, and the sleep state information array through the preset sleep staging model comprises: input the target physiological feature array into a feature extraction model, perform feature extraction through the feature extraction model, and obtain a sleep depth feature array; splice the sleep depth feature array, the target object information array, and the sleep state information array to obtain a to-be-predicted feature array; input the to-be-predicted feature array into a classification prediction model, perform classification prediction through the classification prediction model, and obtain the sleep staging result corresponding to the current time.

[0009] In a possible implementation, the physiological characteristic original signal comprises a pressure signal and a piezoelectric signal, and the pre-processing of the physiological characteristic original signal to separate at least two physiological parameter signals related to a sleep state comprises: perform filter processing on the pressure signal based on a preset feature frequency range of the body motion signal to separate a body motion signal; perform feature analysis and quantization processing based on the body motion signal to obtain a body motion level reflecting the intensity of body motion interference; adaptively adjust filter parameters based on the body motion level to dynamically construct a respiratory signal filter set and / or a heart rate signal filter set; input the piezoelectric signal into the respiratory signal filter set and / or the heart rate signal filter set, and obtain a respiratory signal and / or a heart rate signal related to a sleep state after filter processing.

[0010] In a possible implementation, the adjusting the filter parameters adaptively based on the body movement level, dynamically constructing a breathing signal filter set and / or a heart rate signal filter set comprises: taking the body movement level as an adjustment parameter, calculating a breathing signal filter cutoff frequency and / or a heart rate signal filter cutoff frequency; based on a preset filter order and a signal sampling frequency, respectively calculating filter coefficients corresponding to the breathing signal filter cutoff frequency and / or filter coefficients corresponding to the heart rate signal filter cutoff frequency; based on the breathing signal filter cutoff frequency and the corresponding filter coefficients, constructing a breathing signal filter set; and / or based on the heart rate signal filter cutoff frequency and the corresponding filter coefficients, constructing the heart rate signal filter set.

[0011] In a possible implementation, the feature analysis and quantification processing based on the body movement signal to obtain a body movement level reflecting the intensity of body movement interference comprises: performing a preset window length sliding window absolute value summation processing on the body movement signal to obtain a body movement energy signal; determining a body movement energy maximum value of the body movement energy signal within the preset window length; counting an effective body movement number of the body movement energy maximum value greater than a preset body movement energy threshold value within the preset window length, and a time length ratio of a cumulative time length of the body movement energy signal greater than the preset body movement energy threshold value to the preset window length; based on the effective body movement number and the time length ratio, calculating the body movement level.

[0012] In a possible implementation, the physiological parameter signal comprises a breathing signal, a heart rate signal, and a body movement signal, and the multi-dimensional feature extraction on the physiological parameter signal to construct a target physiological feature array comprises: respectively determining feature peaks of the breathing signal and the heart rate signal, and respectively determining corresponding interval data and amplitude data based on the feature peaks; performing energy analysis on the body movement signal to obtain body movement energy data, and counting a body movement number within a preset time window; obtaining baseline physiological parameters in a pre-sleep resting state, the baseline physiological parameters comprising an average breathing rate, an average heart rate, and an average body movement energy value; based on the baseline physiological parameters, respectively performing normalization processing on the interval data and the body movement energy data to obtain normalized data; The variability features corresponding to the respiration signal and the heart rate signal are calculated respectively, and the variability features are normalized based on the reference physiological parameter to obtain normalized variability features; In a preset time range, the interval data, the amplitude data, the body movement energy data, the body movement times, the normalized data, the variability features, and the normalized variability features are integrated to construct the target physiological feature array.

[0013] In a possible implementation, the signal state detection on the physiological feature original signal includes: Based on the physiological feature original signal, the off-bed detection is performed on the target object; If the target object is in the in-bed state, it is determined whether the physiological feature original signal is valid; If the proportion of the length of time during which the physiological feature original signal is truncated or 0 in a preset window time is greater than a preset proportion threshold, it is determined that the physiological feature original signal is invalid.

[0014] In a possible implementation, the data cleaning on the sleep staging data of the day to obtain the final sleep staging result includes: The target sleep segment in the sleep staging data is determined, wherein the target sleep segment includes at least one of a short-time sleep segment, an island sleep segment, and a nap sleep segment; According to a preset data cleaning rule, the data cleaning is performed on different types of target sleep segments, and all sleep staging data are integrated to obtain the final sleep staging result.

[0015] In a second aspect, a sleep staging system is provided, including: A physiological feature original signal acquisition unit is configured to acquire a physiological feature original signal of a target object; A signal state detection unit is configured to perform signal state detection on the physiological feature original signal, and the signal state detection at least includes off-bed state detection and signal validity detection; A final sleep staging result acquisition unit is configured to, if the target object is in the off-bed state or the physiological feature original signal is invalid and lasts for a preset length of time, acquire sleep staging data of the target object of the day, and perform data cleaning on the sleep staging data of the day to obtain a final sleep staging result; A physiological parameter signal acquisition unit is configured to, if the target object is in the in-bed state and the physiological feature original signal is valid, perform a preprocessing operation on the physiological feature original signal to separate at least two physiological parameter signals related to sleep states; A target physiological feature array construction unit is configured to perform multi-dimensional feature extraction on the physiological parameter signals to construct a target physiological feature array; a sleep staging result generation unit configured to obtain a sleep staging result corresponding to a current time instant based on the target physiological feature array, the target object information array, and the sleep state information array by using a preset sleep staging model.

[0016] The sleep staging method and system, the method implementation of which comprises: acquiring a physiological feature original signal of a target object; performing signal state detection on the physiological feature original signal, which at least includes off-bed state detection and signal validity detection; if the target object is in an off-bed state or the physiological feature original signal is invalid, and this state lasts for a preset time length, acquiring the target object's current day sleep staging data, and performing data cleaning on the current day sleep staging data to obtain a final sleep staging result; if the target object is in an on-bed state and the physiological feature original signal is valid, performing preprocessing operation on the physiological feature original signal to separate at least two physiological parameter signals related to sleep state; performing multi-dimensional feature extraction on the physiological parameter signals to construct a target physiological feature array; and obtaining a sleep staging result corresponding to a current time instant based on the target physiological feature array, a target object information array, and a sleep state information array by using a preset sleep staging model. In the embodiments of the present application, by collecting physiological feature original signals such as pressure signals and piezoelectric signals which are easy to obtain, high-quality data is screened through off-bed state and signal validity detection, and then key physiological parameter signals such as respiration, heart rate, and body movement are accurately separated through preprocessing operation, and then multi-dimensional feature extraction is performed on these signals to construct a target physiological feature array, while the target object information array and the sleep state information array are fused, and deep feature mining and accurate classification are realized by using a preset sleep staging model, which not only does not require complex wearing and massive training data, but also takes into account the convenience of home monitoring and the flexibility of engineering implementation in the smart home scenario, and effectively solves the problems of insufficient accuracy of traditional simplified monitoring equipment, physiological signals easily disturbed by body movement, and insufficient feature fusion, significantly improves the anti-interference ability and result accuracy of sleep staging, and meets the application requirements of sleep monitoring in multiple fields such as smart home health management and smart terminals. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0018] Figure 1 is an application environment schematic diagram of the sleep staging method in an embodiment of the present application; Figure 2is a flowchart of a sleep staging method in an embodiment of the present application; Figure 3 is a model structure diagram of a feature extraction model in an embodiment of the present application; Figure 4 is a model structure diagram of a discrimination prediction model in an embodiment of the present application; Figure 5 is an example diagram of heart rate feature point selection in an embodiment of the present application; Figure 6 is an example diagram of respiratory feature point selection in an embodiment of the present application; Figure 7 is a structure diagram of a sleep staging system in an embodiment of the present application; Figure 8 is a diagram of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0020] The sleep staging method provided in the embodiments of the present application can be applied in an application environment such as Figure 1 The application environment includes a device end S1, a client end S2 and a server end S3, which can be connected to each other in communication. The main carrier of the device end S1 can be a smart bed frame or a smart mattress and the like, which is mainly composed of a signal acquisition module, a signal state judgment module, a signal preprocessing module, a feature extraction module, a sleep staging module and a post-processing module. The client end S2 includes but is not limited to various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices, on which a physiological state detection related APP or a small program and the like is pre-installed. The server end S3 can be realized by an independent server or a server cluster composed of multiple servers, which can be a local server or a cloud server.

[0021] Specifically, the server can be used for user information management, such as user information storage and information delivery, and the user information includes but is not limited to personal information of the user and historical physiological parameter information. The historical physiological parameter information can include: resting heart rate and resting respiratory rate. The resting heart rate and the resting respiratory rate refer to the average heart rate value and the average respiratory rate in a resting state. The resting state refers to a time segment in which the user continuously has no body movement for more than a threshold TH1 (such as 5 minutes) and does not enter sleep. The resting heart rate and the resting respiratory rate can be automatically updated once a day at a fixed time point (such as after the user gets up) or updated in real time based on the latest resting state data of the user. The client is mainly used for result display of sleep staging, user personal information input, and related data uploading to the server. The device end is used for real-time calculation of sleep staging, and can dynamically update the sleep staging result stored in the server.

[0022] In an embodiment, as shown in Figure 2 , a sleep staging method is provided. The method is applied to the device end in Figure 1 , and includes the following steps: In step S110, a physiological characteristic original signal of a target object is acquired. Optionally, the physiological characteristic original signal can be collected by a data collection module. The data collection module includes at least one sensor for directly collecting original data reflecting the physiological function state or physiological activity characteristics of the target object (such as a user in a sleep state), such as: a pressure sensor and a piezoelectric sensor placed under the user's torso (such as directly below the shoulder, below the back, or below the hip) for collecting pressure signals and piezoelectric signals related to body movement, breathing, and heart rate; an electroencephalogram sensor configured on the forehead or other regions of the brain for collecting electroencephalogram signals reflecting brain activity; an optical volume pulse wave sensor worn on the finger, earlobe, or other parts for collecting optical volume pulse wave signals representing blood circulation state; in addition, it can also include acoustic sensors (for collecting acoustic signals such as breathing sound and snoring sound), temperature sensors (for collecting body surface or environmental temperature signals), acceleration sensors (for collecting limb movement acceleration signals) arranged in a sleep scene.

[0023] It should be noted that the data collection module can flexibly configure the type and deployment position of the sensor according to the actual application scene (such as home sleep monitoring and clinical sleep evaluation). The sampling frequency, signal resolution, and other parameters of the sensor can be adaptively adjusted based on the extraction requirements of the target physiological characteristics, to ensure that the collected original signals have completeness and effectiveness, and to provide high-quality data support for subsequent signal processing and sleep staging.

[0024] In step S120, signal state detection is performed on the physiological characteristic original signal, and the signal state detection at least includes bed leaving state detection and signal validity detection; Optionally, signal state detection is performed on the collected physiological characteristic original signal, and the detection process can include bed leaving state detection and signal validity detection. In the bed leaving state detection, the amplitude distribution range, energy fluctuation intensity and continuous stability of the physiological characteristic original signal such as the pressure signal and the piezoelectric signal are analyzed, and a preset bed staying determination threshold is combined (for example, the pressure signal amplitude is in the reasonable range of human body pressure, and the piezoelectric signal energy change conforms to the physiological characteristic of staying in bed), so as to accurately judge whether the target object is in the bed staying state, thereby excluding the invalid data interference of the non-sleep period collected when the target object leaves the bed. In the signal validity detection, the signal quality indicators are counted based on a preset time window (for example, 1 second to 5 seconds), specifically including the proportion of the length of time in which the signal is truncated or 0 in the window, the proportion of the signal amplitude exceeding the normal physiological range, the proportion of the signal noise energy, and the like. If any indicator exceeds the corresponding preset validity threshold, it is determined that the signal is invalid, otherwise it is valid. Through the detection, the problems of signal distortion, noise pollution or data loss caused by poor sensor contact, environmental interference, equipment failure and the like can be avoided, high-quality and processing-required effective data are selected for the subsequent preprocessing and feature extraction links, and the accuracy of the sleep staging result is ensured.

[0025] In step S130, if the target object is in the bed leaving state or the physiological characteristic original signal is invalid and lasts for a preset time length, the sleep staging data of the target object on the same day is obtained, and the sleep staging data on the same day is cleaned to obtain the final sleep staging result. Optionally, after the signal state detection is completed, if the target object is in the bed leaving state or the physiological characteristic original signal is invalid and the state lasts for a preset time length (for example, 30 seconds), the sleep data ending processing procedure is triggered. First, the system obtains all the sleep staging data (including the stage staging results of the valid bed staying period) of the target object on the same day, and then performs cleaning operation on the data, for example, eliminating abnormal staging records in the signal invalid period, correcting redundant information in the data collection process, and supplementing the time sequence correlation data in the valid period, and finally outputting complete and accurate final sleep staging result, ensuring the integrity and reliability of the sleep data.

[0026] The sleep staging data on the same day refers to the sleep staging data in the time slice from the specified time point to the current time. The selection rule of the specified time point can be that if the current time point is after 18:00, the specified time point is 18:00 on the same day; if the current time point is not after 10:00 (for example, 2:00 in the morning or 8:00 in the morning), the specified time point is 18:00 of the previous day (that is, the starting reference point of the sleep cycle of the previous day).

[0027] In step S140, if the target object is in the in-bed state and the physiological characteristic original signal is valid, a pre-processing operation is performed on the physiological characteristic original signal to separate at least two physiological parameter signals related to the sleep state; Optionally, when it is confirmed that the target object is in the in-bed state and the collected physiological characteristic original signal (such as a pressure signal, a piezoelectric signal, etc.) meets the validity requirement, a pre-processing operation can be further performed on the valid physiological characteristic original signal. For example, the irrelevant influences such as environmental interference, equipment noise, and signal baseline drift can be removed through signal denoising (such as filter denoising, baseline correction, etc.) first, and then targeted processing can be performed based on the differences in the characteristic frequency intervals, amplitude characteristics, etc. of different physiological parameter signals by using a signal separation algorithm (such as adaptive filtering based on body movement level, frequency domain separation, etc.) to accurately separate at least two physiological parameter signals related to the sleep state from the original signal, including a body movement signal, a respiration signal, and a heart rate signal. The body movement signal reflects limb activity, the respiration signal reflects respiration state, and the heart rate signal represents heart rate change.

[0028] In step S150, multi-dimensional feature extraction is performed on the physiological parameter signals to construct a target physiological characteristic array. Optionally, the physiological parameter signals related to the sleep state, such as the heart rate signal, the respiration signal, and the body movement signal, are first screened for valid extreme points through peak point detection to construct a heart beat / respiration interval array and an amplitude array, and then the abnormal value is removed (the baseline deviation of the heart beat interval is more than 20%, and the baseline deviation of the respiration interval is more than 50%) and linear interpolation processing is performed to generate data with equal time intervals. Then, the average heart rate, the respiration rate, and the body movement energy value (determined according to the priority of the latest, the historical recent, and the first use preset value before falling asleep) of the user in the resting state are obtained, and the interval data, the body movement energy, and the heart rate / respiration variability features (standard deviation) within the preset window TH14 (such as 5 minutes) are normalized based on the above resting reference parameters. Finally, within the preset time window, the piezoelectric and pressure original signals, the amplitude arrays after processing, various normalized features, and the body movement level, the body movement frequency, etc. are integrated to construct a target physiological characteristic array with comprehensive dimensions and accurate representation.

[0029] It can be understood that the target physiological characteristic array can include but is not limited to the piezoelectric original signal, the pressure original signal, the normalized heart beat interval array, the normalized respiration interval array, the processed heart beat amplitude array, the processed respiration amplitude array, the normalized body movement energy array, the body movement level array, the respiration variability, the heart rate variability, the normalized respiration variability, the normalized heart rate variability, the body movement frequency, etc. within the window time TH15 (such as 30 seconds).

[0030] In step S160, based on the target physiological feature array, the target object information array, and the sleep state information array, a preset sleep staging model is used to obtain a sleep staging result corresponding to the current time.

[0031] It should be noted that the target object information array can include age, gender, body mass index (BMI), sleep type, and other information related to the user's personal information, wherein the age, gender, and BMI are set by the user on the APP side, and the sleep type acquisition method includes but is not limited to: user-initiated setting or obtaining through a specific scale on the APP, and the sleep type setting can be in the form of a grade (for example, 0-2, representing good sleep quality, general sleep quality, and poor sleep quality, respectively).

[0032] The sleep state information array can include the sleep staging result of the previous time, the sleep staging result duration of the previous time, the sleep time, the number of sleep cycles performed, the current sleep cycle duration, the time spent in bed, and other data related to the sleep timing.

[0033] Optionally, the preset sleep staging model can be a deep learning model, which can include a feature extraction model and a classification prediction model. After the target physiological feature array is deeply mined and dimensionally optimized by the feature extraction model, it is spliced with the target object information array and the sleep state information array, and then the spliced features are accurately classified by the classification prediction model, and finally the sleep staging result (such as wakefulness period, light sleep period, deep sleep period, and rapid eye movement period) corresponding to the current time is output, so as to provide data support for sleep quality evaluation and personalized intervention.

[0034] In addition, based on the sleep staging result corresponding to the current time, the sleep staging result of the previous time, the sleep staging result duration of the previous time, the sleep time, the number of completed sleep cycles, the current sleep cycle duration, and the related parameter information of the time spent in bed can be updated. For example, if the sleep staging result of the current time is consistent with the sleep staging result of the previous time, the sleep staging duration of the previous time is added by the current time step (e.g., 30 seconds, matching the feature extraction window time length), the accumulated sleep time is also added by the time step, and the current sleep cycle duration remains in the accumulation state. If the sleep staging result of the current time is inconsistent with the sleep staging result of the previous time, the sleep staging result of the previous time is updated to the current staging result, the sleep staging duration of the previous time is reset to the current time step, and the accumulated sleep time continues to be added by the time step. When the staging result switches from non-rapid eye movement period (light sleep / deep sleep) to rapid eye movement period, or from rapid eye movement period to light sleep period, it is determined that a complete sleep cycle is completed, the number of completed sleep cycles is increased by 1, the current sleep cycle duration is reset to the current time step and starts to be re-accumulated. The time spent in bed is continuously accumulated from the first time the target object is determined to be in bed, and is not affected by the change of the sleep staging result. Only when the off-bed state is detected and the pre-set time length is reached, the update is stopped. Through the real-time dynamic update of the above parameters, the sleep state information array always maintains the latest and accurate time sequence related data, so as to improve the continuity and accuracy of the staging result.

[0035] In the embodiments of the present application, a sleep staging method is provided, which comprises: acquiring a physiological characteristic original signal of a target object; performing signal state detection on the physiological characteristic original signal, the signal state detection at least comprising off-bed state detection and signal validity detection; if the target object is in an in-bed state and the physiological characteristic original signal is valid, performing a preprocessing operation on the physiological characteristic original signal to separate at least two physiological parameter signals related to sleep states; performing multi-dimensional feature extraction on the physiological parameter signals to construct a target physiological characteristic array; and based on the target physiological characteristic array, a target object information array and a sleep state information array, obtaining a sleep staging result corresponding to a current time through a preset sleep staging model. In the embodiments of the present application, by collecting physiological characteristic original signals such as pressure signals and piezoelectric signals which are easy to obtain, high-quality data is screened through off-bed state and signal validity detection, and then key physiological parameter signals such as respiration, heart rate and body movement are accurately separated through a preprocessing operation. Then, multi-dimensional feature extraction is performed on these signals to construct a target physiological characteristic array, while a target object information array and a sleep state information array are fused. Deep feature mining and accurate classification are achieved with the aid of a preset sleep staging model, which not only takes into account the convenience of home monitoring and the flexibility of engineering implementation in the smart home scenario without complex wearing and massive training data, but also effectively solves the problems of insufficient accuracy of traditional simplified monitoring equipment, physiological signal interference by body movement and insufficient feature fusion, significantly improves the anti-interference ability and result accuracy of sleep staging, and meets the sleep monitoring application requirements in multiple fields such as smart home health management and smart terminals.

[0036] In an embodiment of the present application, based on the target physiological characteristic array, the target object information array and the sleep state information array, the sleep staging result corresponding to the current time is obtained through the preset sleep staging model, which comprises: inputting the target physiological characteristic array into a feature extraction model to perform feature extraction through the feature extraction model to obtain a sleep depth feature array; splicing the sleep depth feature array, the target object information array and the sleep state information array to obtain a to-be-predicted feature array; inputting the to-be-predicted feature array into a classification prediction model to perform classification prediction through the classification prediction model to obtain the sleep staging result corresponding to the current time.

[0037] Optionally, after the target physiological characteristic array is input into the feature extraction model to perform feature extraction, a sleep depth feature array of a current window time TH15 can be obtained. The feature extraction model can be as follows: Figure 3As shown, it can include a first input layer 1, a second input layer 2 and a third input layer 3, wherein the first input layer 1 is composed of piezoelectric raw signals, pressure raw signals within the window time TH15, if TH15 takes 30 seconds, the sampling rate of the piezoelectric raw signals and the pressure raw signals is 200Hz, then the size corresponding to the first input layer 1 is 6000*2*1; the second input layer 2 is composed of physiological signal feature arrays within the window time TH15, including: normalized interbeat interval array, normalized interbreath interval array, processed heart beat amplitude array, processed breath amplitude array, normalized body movement energy array, body movement level array, if TH15 takes 30 seconds, then the size corresponding to the second input layer 2 is 30*6*1; the third input layer 3 is composed of single-value physiological features within the window time TH15 (a total of 5 features), including: breath variability, heart rate variability, normalized breath variability, normalized heart rate variability, body movement frequency. As shown in Figure 3 As shown, the first convolutional layer 1 to the eleventh convolutional layer 11 are all convolutional layers, the specific operation includes three steps of Conv1d (convolution processing), BN (batch normalization processing) and ReLU (activation processing), wherein K7, K8, K9, K50 and K400 represent the convolution kernel size of 7*1, 8*1, 9*1, 50*1 and 400*1 respectively; wherein s1, s2, s3, s4, s8 and s50 represent the step length of 1, 2, 3, 4, 8 and 50 respectively; wherein p0, p1 and p2 represent the number of left and right padding 0 of the padding type maxpooling; c8, c16 and c32 represent the number of convolution kernels of 8, 16 and 32 respectively; taking the first convolutional layer 1 as an example, the output parameter quantity of this layer is 745*2*16; the convolutional layer 1*1 is a 1*1 size convolutional layer, which is used for input size shape adjustment and nonlinearity increase (which also includes ReLU operation); is an array splicing operation; the first fully connected layer 1 and the second fully connected layer 2 are both fully connected layers, and the activation function thereof is ReLU, and the third fully connected layer 3 is an output layer, and the activation function thereof is softmax, so as to obtain the probability of each sleep stage for the data segment; the acquisition process of the feature extraction model is to train the model structure based on Figure 3 As shown, the model structure is trained, and after the training is completed, the third fully connected layer 3 is removed; the feature array 2 is the output corresponding to the second fully connected layer 2. After the training is completed, the output layer (such as the third fully connected layer 3, the softmax activation) is removed, and the structure of the second fully connected layer 2 is retained.

[0038] Specifically, the first input layer 1, the second input layer 2 and the third input layer 3 can each include multiple groups of convolutional layers composed of Conv1d (convolution processing), BN (batch normalization processing) and ReLU (activation processing) to cover features of different time scales. The first input layer 1 corresponds to convolutional layers 1-4, can use a large convolution kernel (such as K400x1, step s50) to capture long-term physiological signal trends, and use a small convolution kernel (such as K7x1, step s1) to extract short-term details, and finally adjust the dimension to 36x64 through convolutional layer 1-1 (1x1 convolution). The second input layer 2 corresponds to convolutional layers 5-8: uses a medium convolution kernel (such as K50x1, step s8) to extract the time sequence correlation of the feature array, and finally adjusts the dimension to 5x76 through convolutional layer 1-1. The third input layer 3 corresponds to convolutional layers 9-11: uses a small convolution kernel (such as K5x1, step s1) to extract the nonlinear correlation of single-value features, and finally adjusts the dimension to 4x80 through convolutional layer 1-1. After the feature arrays processed by the first input layer 1, the second input layer 2 and the third input layer 3 are spliced and combined, they are sequentially input into the first fully connected layer 1 (output dimension is 128) and the second fully connected layer 2 (output dimension is 32 dimensions), and the output of the second fully connected layer 2 is a sleep depth feature array with a length of 32 (i.e., the number of elements of the sleep depth feature array is 32).

[0039] Then, the sleep depth feature array extracted by the feature extraction model can be spliced with the target object information array and the sleep state information array to obtain a to-be-predicted feature array, and the to-be-predicted feature array is input into the classification prediction model. After classification by the classification prediction model, the sleep staging result corresponding to the current time can be output.

[0040] The structure of the classification prediction model can be as shown in Figure 4 The sleep depth feature array has a length of 32; the target object information array can include age, gender, body mass index (BMI), sleep type, and has a length of 4; the sleep state information array includes the previous sleep staging result, the duration of the previous sleep staging result, the sleep time, the number of sleep cycles performed, the current sleep cycle duration, and the duration of being in bed, and has a length of 6. The fourth full connection layer 4 and the fifth full connection layer 5 are full connection layers, and the activation functions thereof are ReLU. The sixth full connection layer 6 is an output layer, and the activation function thereof is softmax. Specifically, the sleep depth feature array with a length of 32, the target object information array with a length of 4 (containing age, etc.), and the sleep state information array with a length of 6 (containing the sleep stage result of the previous moment, etc.) are spliced into a to-be-predicted feature array with a dimension of 42, and then are sequentially input into the fourth full connection layer 4 (ReLU activation, outputting 64 dimensions), the fifth full connection layer 5 (ReLU activation, outputting 32 dimensions) for feature transformation and compression, and finally the sixth full connection layer 6 (softmax activation) outputs the probabilities of each sleep stage (such as the wake stage, the light sleep stage, the deep sleep stage, and the rapid eye movement stage), and the category with the maximum probability is taken as the sleep stage result of the current moment.

[0041] It should be noted that the acquisition process of the classification prediction model is based on the to-be-predicted sample array with sleep stage annotation using the model structure shown in the figure for training. In the training process, the cross-entropy loss function can be used to optimize the full connection layer parameters, and after the training is completed, the classification prediction of the input features can be realized. Figure 4

[0042] In an embodiment of the present application, the physiological characteristic original signal includes a pressure signal and a piezoelectric signal, and the pre-processing operation on the physiological characteristic original signal separates at least two physiological parameter signals related to the sleep state, including: The pressure signal is filtered based on a preset characteristic frequency range of the body movement signal to separate a body movement signal; The body movement signal is analyzed and quantified to obtain a body movement level reflecting the body movement interference intensity; The filter parameters are adaptively adjusted based on the body movement level to dynamically construct a breathing signal filter set and / or a heart rate signal filter set; The piezoelectric signal is input into the breathing signal filter set and / or the heart rate signal filter set, and after the filtering processing, the breathing signal and / or the heart rate signal related to the sleep state are obtained.

[0043] ​Optionally, the physiological characteristic raw signal can include a pressure signal and a piezoelectric signal, and the physiological parameter signal to be extracted can include a body movement signal, a respiration signal and / or a heart rate signal, and the specific preprocessing operation process can be: first, the pressure signal can be band-pass filtered according to the frequency interval characteristic of the body movement signal (such as the low-to-medium frequency characteristic interval corresponding to human limb activity), so as to separate the [4Hz, 40Hz] body movement signal reflecting the limb activity state from the mixed raw signal; then, time domain feature analysis (such as calculating the signal energy and amplitude fluctuation range in a preset window) is performed on the body movement signal, and the body movement signal is quantized into a body movement level reflecting the body movement interference intensity (the higher the level, the more significant the interference) through a pre-designed calculation rule; subsequently, the filter parameters are adaptively adjusted based on the obtained body movement level, for example, when the body movement level is high, the passband width of the filter is reduced to strengthen the suppression of high-frequency interference, and when the body movement level is low, the passband width is expanded to retain more physiological signal details, so as to dynamically construct a respiration signal filter set and a heart rate signal filter set adapted to the current interference state; finally, the piezoelectric signal is input into the two specifically constructed filter sets, and after filtering, the respiration signal (reflecting the respiration rhythm) and the heart rate signal (representing the heart rate change) directly related to the sleep state can be accurately extracted from the piezoelectric signal.

[0044] It should be noted that the physiological parameter signal can include at least two of the respiration signal, the heart rate signal and the body movement signal, wherein the respiration signal and the heart rate signal can also be extracted without relying on the body movement level: when there is no need to adapt to body movement interference (such as a laboratory static monitoring environment) or the sensor deployment scene has no obvious body movement interference, a respiration signal filter set and a heart rate signal filter set can be constructed using preset fixed filter parameters (for example, the fixed band-pass filtering range of the respiration signal is 0.15Hz-4Hz, and the fixed band-pass filtering range of the heart rate signal is 0.5Hz-3Hz), and the piezoelectric signal is directly filtered to obtain the target physiological parameter signal. This design makes the scheme support both dynamic anti-interference scenes and static non-interference scenes, and improves the scene adaptation flexibility of the technical scheme.

[0045] In addition, the preset characteristic frequency interval [4Hz, 40Hz] of the body movement signal can be flexibly adjusted according to the actual application scene, for example, the upper and lower limits of the frequency interval can be fine-tuned according to the limb activity characteristics of different groups of people such as children and the elderly, or the signal propagation characteristics of different sensor deployment positions such as mattresses and pillows, to ensure the accuracy of body movement signal separation; the band-pass filtering process can use conventional digital filtering algorithms such as Butterworth filtering and Chebyshev filtering, and the specific algorithm selection can be determined according to the signal processing accuracy requirement and the calculation resource adaptability, without departing from the technical concept of the present application.

[0046] In an embodiment of the present application, the feature analysis and quantization processing based on the body movement signal obtains a body movement level reflecting the body movement interference intensity, including: performing preset window length sliding window absolute value summation processing on the body movement signal to obtain a body movement energy signal; determining a body movement energy maximum value of the body movement energy signal within the preset window length; counting the number of effective body movements in which the body movement energy maximum value is greater than a preset body movement energy threshold value within the preset window length, and a time length ratio of a cumulative time length in which the body movement energy signal is greater than the preset body movement energy threshold value to the preset window length; calculating the body movement level based on the number of effective body movements and the time length ratio.

[0047] Optionally, the separated body movement signal is subjected to preset window length TH8 sliding window absolute value summation processing to obtain a body movement energy signal. The value of TH8 can be 1 second, and the sliding window step can be 1 sampling point, so as to realize continuous and high-resolution characterization of the body movement signal energy. Then, the body movement energy maximum value within TH8 is searched. If the body movement energy maximum value is greater than a preset body movement energy threshold value TH9 (determined based on the body movement energy statistical value in a quiet bed scene), it is considered that the maximum value is caused by body movement. At this time, the number of body movements is incremented by 1. After the body movement energy maximum value is traversed, the number of effective body movements s and the time length ratio of the time length in which the body movement energy is greater than the preset threshold value TH9 within the window time TH1 are counted. The body movement level of the current window time TH1 is calculated based on the number of effective body movements s and the time length ratio , and the specific calculation method is as follows: ; wherein, is a smaller value operation of the two, to ensure that the body movement level does not exceed 100, so that the result is within a unified quantization interval, facilitating subsequent determination of the body movement interference degree, e is a natural constant, is a preset value, which can be 2. For example, when the number of effective body movements is large and the time length ratio is high, the body movement level is high, representing that the current body movement interference has a more significant impact on the physiological signal.

[0048] In an embodiment of the present application, the body movement level is used to adaptively adjust the filter parameters, dynamically construct a respiratory signal filter set and / or a heart rate signal filter set, including: using the body movement level as an adjustment parameter to calculate the respiratory signal filter cutoff frequency and / or the heart rate signal filter cutoff frequency; ​Based on the preset filter order and the signal sampling frequency, the filter coefficients corresponding to the respiratory signal filter cutoff frequency and / or the filter coefficients corresponding to the heart rate signal filter cutoff frequency are calculated respectively; Based on the respiratory signal filter cutoff frequency and the corresponding filter coefficients, a respiratory signal filter set is constructed; and / or Based on the heart rate signal filter cutoff frequency and the corresponding filter coefficients, the heart rate signal filter set is constructed.

[0049] Optionally, based on the filter key parameters, the filter key parameters can include the cutoff frequency of the respiratory signal filter and / or the cutoff frequency of the heart rate signal filter, and specifically, the body movement level reflecting the body movement interference intensity can be taken as the core adjustment parameter, and the respiratory signal filter cutoff frequency and / or the heart rate signal filter cutoff frequency adapted to the current interference state are calculated in combination with the physiological characteristic frequency range of the respiratory signal and the heart rate signal, wherein the lower limit value of the respiratory signal band-pass filter cutoff frequency is and the upper limit value of the cutoff frequency is , which can be calculated by the following formula: ; ; The lower limit value of the heart rate signal band-pass filter cutoff frequency is and the upper limit value of the cutoff frequency is , which can be calculated by the following formula: ; ; Subsequently, based on the preset filter order (such as 4th order, which is adapted to the physiological signal processing accuracy requirement and takes into account the filtering effect and the calculation complexity) and the signal sampling frequency (such as 200Hz, which is consistent with the sensor sampling rate in the foregoing), the respiratory signal filter coefficients and the heart rate signal filter coefficients corresponding to the above dynamic cutoff frequencies are solved by a digital filtering algorithm (such as Butterworth filtering algorithm), and the calculation process of the related filter coefficients can refer to the implementation logic of the butter function in MATLAB (including the steps of normalized cutoff frequency calculation, transfer function solving, and coefficient discretization).

[0050] Finally, taking the calculated respiratory signal filter dynamic cutoff frequency and its corresponding filter coefficient as the core parameter, a respiratory signal filter set (including a band-pass filter unit, a signal amplitude calibration unit) is constructed; and / or the heart rate signal filter dynamic cutoff frequency and its corresponding filter coefficient are taken as the core parameters to construct a heart rate signal filter set. Through this dynamic construction method, the frequency response characteristics of the filter set are adjusted in real time with the body motion level, realizing dynamic adaptation to body motion interference, ensuring effective noise suppression in strong interference scenarios and preserving physiological signal details in weak interference scenarios.

[0051] In an embodiment of the present application, the physiological parameter signal includes a respiratory signal, a heart rate signal, and a body motion signal, and the multi-dimensional feature extraction of the physiological parameter signal to construct a target physiological feature array includes: The feature peak values of the respiratory signal and the heart rate signal are determined respectively, and based on the feature peak values, the corresponding interval data and amplitude data are determined respectively; The body motion signal is subjected to energy analysis to obtain body motion energy data, and the number of body motions in the preset time window is counted; The baseline physiological parameters in the pre-sleep resting state are obtained, including the average respiratory rate, the average heart rate, and the average body motion energy value; Based on the baseline physiological parameters, the interval data and the body motion energy data are normalized to obtain normalized data; The variability features corresponding to the respiratory signal and the heart rate signal are calculated respectively, and the variability features are normalized based on the baseline physiological parameters to obtain normalized variability features; Within a preset time range, the interval data, the amplitude data, the body motion energy data, the number of body motions, the normalized data, the variability features, and the normalized variability features are integrated to construct the target physiological feature array.

[0052] Optionally, when the physiological parameter signal includes a respiratory signal, a heart rate signal, and a body motion signal, the multi-dimensional feature extraction process to construct a target physiological feature array is as follows: First, feature peak extraction is performed on the respiratory signal and the heart rate signal: taking the heart rate signal peak point (J point) as an example, first, a signal minimum point (which needs to satisfy being less than a preset threshold TH10 and having a time interval with a previous valid minimum value within a preset range) is detected, a valid minimum value point is marked (the minimum value flag is set to 1) and a cache and TH10 are updated; then, a maximum value point (which needs to satisfy being greater than a preset threshold TH11 and there being a matching valid minimum value point in sequence) is detected, a J point is determined and a cache, TH11 and the minimum value flag (such as setting the valid minimum value point flag to 0) are updated. The respiratory signal peak point extraction process is consistent with that of the heart rate signal, and initial TH10 (trough determination threshold) and TH11 (peak determination threshold) can be determined based on pre-acquired data statistics of different groups of people in a quiet bed scene (for example, TH11 takes the lower quartile of the pre-acquired data peak value x 1.2, and TH10 takes the upper quartile of the pre-acquired data trough value x 0.8, and the coefficients can be fine-tuned according to the actual scene). Based on the extracted peak points, a heart beat interval array, a respiratory interval array, a heart beat amplitude array and a respiratory amplitude array are respectively constructed; linear interpolation and other methods are used on the amplitude array to generate an equal time interval array (such as 1 point per 1 second), and the interval array is first low-pass filtered to obtain a baseline array, and abnormal values deviating by more than 20% (heart beat interval) or 50% (respiratory interval) are removed, and then linear interpolation is used to generate an equal time interval array. Among them, the J point is selected as shown in FIG. 8, and the respiratory signal peak point is selected as shown in FIG. 9. Figure 5 Figure 6 ​​Secondly, the body motion signal is subjected to energy analysis to obtain body motion energy data, and the number of body motions in a preset time window is counted; at the same time, reference physiological parameters (including average respiratory rate, average heart rate and average body motion energy value) in a pre-sleep resting state are obtained, the resting state determination criterion is that the body motion level is less than 1 and the duration exceeds a preset threshold TH12 (such as 9 minutes), and the reference parameter priority is: the latest resting state before falling asleep > the latest resting state in a preset time period TH13 (such as 7 days) > the parameter in the TH12 time period after falling asleep in the previous effective sleep > the population statistical value or fixed preset value (such as average heart rate 60 times / minute, average respiratory rate 12 times / minute) when used for the first time. Specifically, if no resting state meeting the requirements is found before falling asleep, there is a resting state in the preset time period TH13, TH13 is taken as 7 days for example, the average heart rate, average respiration and average body motion energy value in the latest resting state are selected as the reference physiological parameters in the resting state; if no resting state is detected in the preset time period TH13, the average heart rate, average respiratory rate and average body motion energy value in the TH12 time period after falling asleep for the first time in the previous effective sleep process are used as the reference physiological parameters in the resting state; if the user uses it for the first time, the reference physiological parameters in the resting state can use the population statistical value or the fixed preset value, for example, the average heart rate is taken as 60 times / minute, the average respiratory rate is taken as 12 times / minute, and the average body motion energy / minute, it should be noted that the default threshold of the average body motion energy is related to the type of the sensor actually used, the circuit amplification coefficient, the window length and other factors.

[0053] Then, the processed inter-beat intervals, the processed inter-breath intervals and the body motion energy can be normalized based on the obtained reference physiological parameters, the normalization process can be that the processed inter-beat intervals * average heart rate in the resting state / 60000 (unit: millisecond), the processed inter-breath intervals * average respiratory rate in the resting state / 60000 (unit: millisecond), and the body motion energy / average body motion energy in the resting state, so as to obtain the normalized inter-beat interval array, the normalized inter-breath interval array and the normalized body motion energy array.

[0054] At the same time, the normalized respiratory variability feature and the heart rate variability feature can also be calculated, the respiratory variability feature refers to the standard deviation of the inter-breath interval array in a preset window time TH14, TH14 can be taken as 5 minutes, the heart rate variability feature refers to the standard deviation of the inter-beat interval array in the preset window time TH14, and the normalization process can be realized by the following way: respiratory variability feature * average respiratory rate in the resting state / 60000 (unit: millisecond), heart rate variability feature * average heart rate in the resting state / 60000 (unit: millisecond).

[0055] Finally, within a window time TH15 (e.g., 30 seconds), the piezoelectric raw signal, the pressure raw signal, the normalized inter-beat interval array, the normalized inter-breath interval array, the processed beat amplitude array, the processed breath amplitude array, the normalized body movement energy array, the body movement level array, the breath variability, the heart rate variability, the normalized breath variability, the normalized heart rate variability, and the body movement count are integrated to construct a target physiological feature array.

[0056] In an embodiment of the present application, the signal state detection on the physiological feature raw signal comprises: Based on the physiological feature raw signal, the off-bed detection is performed on the target object. If the target object is in the in-bed state, it is determined whether the physiological feature raw signal is valid. If the proportion of the length of time during which the physiological feature raw signal is truncated or 0 within a preset window time is greater than a preset proportion threshold, it is determined that the physiological feature raw signal is invalid.

[0057] Optionally, the piezoelectric signal and the pressure signal are taken as examples to specifically illustrate the physiological feature raw signal. First, the piezoelectric signal is subjected to band-pass filtering processing in the frequency band of [0.15 Hz, 4 Hz] to extract the physiological signal in the frequency band. Then, the sum of the absolute values of the physiological signal within a window time TH1 (e.g., 8 seconds) is calculated to obtain the physiological signal energy intensity. It is determined whether the physiological signal energy intensity is greater than a preset threshold TH2 and whether the pressure value is greater than a preset threshold TH3. If both conditions are met, it is determined that the target object is in the in-bed state, otherwise, it is determined that the target object is in the off-bed state.

[0058] It should be noted that if the physiological signal energy intensity is less than a preset threshold TH4 and the duration exceeds TH5, and the difference between the maximum and minimum values of the pressure value within the window time TH1 is less than a preset threshold TH6, the pressure threshold TH3 is updated to the average value of the pressure value within the current window time TH1. That is, when it is detected that the user has no obvious physiological activity (the filtered physiological signal energy is weak and stable for a long time) and the pressure signal is in a stable state, the core pressure determination standard TH3 for off-bed detection is dynamically adjusted to adapt to the actual pressure level under the current environment, thereby avoiding the decrease in adaptability of the original fixed threshold due to factors such as changes in environmental temperature and humidity, slight shifts in the installation position of the sensor, and sensor drift caused by long-term use, and thus ensuring the accuracy of the in-bed state determination (avoiding false positives caused by environmental changes and sensor drift).

[0059] When it is detected that the current target object is in the bed state, the piezoelectric signal can be subjected to abnormal state detection, and the abnormal state determination rule is that the length ratio r of the piezoelectric signal being truncated or being 0 in the window time TH1 is calculated. If the length ratio exceeds a preset threshold TH7 (such as 50%), it is considered that the piezoelectric signal state is abnormal. If the signal state is abnormal, the signal state is adjusted to the piezoelectric signal state abnormal. If the piezoelectric signal state is normal, the pressure signal can be subjected to abnormal state detection, and the abnormal state determination rule is that the length ratio of the pressure signal being truncated or being 0 in the window time TH1 is calculated. If the length ratio exceeds the preset threshold TH7, it is considered that the pressure signal state is abnormal. If the pressure signal state is normal, the signal state is set to normal, otherwise the signal state is adjusted to the pressure signal state abnormal. That is, if both the piezoelectric signal and the pressure signal are normal, the overall state of the physiological characteristic original signal is set to normal. If any signal is abnormal, it is marked as the piezoelectric signal state abnormal or the pressure signal state abnormal.

[0060] In an embodiment of the present application, the data cleaning on the sleep staging data of the day is performed to obtain a final sleep staging result, including: determining a target sleep segment in the sleep staging data, wherein the target sleep segment includes at least one of a short sleep segment, an island sleep segment, and a nap sleep segment; According to a preset data cleaning rule, the different types of target sleep segments are subjected to data cleaning, and all sleep staging data are integrated to obtain a final sleep staging result.

[0061] It should be noted that the short sleep segment refers to a sleep segment with a too short duration, such as a sleep segment with a non-bed sleep stage duration less than 2 minutes. The nap sleep segment refers to a sleep segment with a time interval from a main sleep segment exceeding 3 hours, and the sleep segment appears at a time not in a designated time segment. The island sleep segment refers to a sleep segment before and after which are both non-sleep states, and the sleep segment has a duration less than 10 minutes.

[0062] Optionally, from the entire sleep staging data recorded on the same day, target sleep segments that need to be processed are screened out, and the target sleep segments include at least one of a short sleep segment, an island sleep segment, and a nap sleep segment. Then, different types of target sleep segments are processed according to preset data cleaning rules. For example, for a short sleep segment, a forward merging rule is used to correct the short sleep segment to the previous valid sleep stage (for example, if the previous stage is light sleep, the short sleep segment is merged into the light sleep stage); for an island sleep segment (unrelated to the main sleep cycle), the cleaning method is to directly mark the segment as a wake-up state, and to remove the isolated short period record without actual sleep significance (to avoid false sleep staging caused by signal noise); for a nap sleep segment, the segment is separately marked as a nap (rather than included in the sleep staging data of the main sleep cycle), to ensure the purity of the main sleep cycle data, while retaining the nap record for separate analysis. Finally, after the targeted cleaning of various types of target sleep segments is completed, the remaining valid sleep staging data (including the main sleep cycle data and the target sleep segment data retained / corrected) are re-integrated in time sequence order, to complete the data time sequence association, remove redundancy, abnormality and invalid information, and finally form a complete, accurate and actual sleep state final sleep staging result (including the wake-up period, light sleep period, deep sleep period, rapid eye movement period distribution of the main sleep cycle, and the separately marked nap record), to ensure the scientificity and usability of the sleep data.

[0063] In the embodiments of the present application, the filter parameters (such as the cutoff frequency and the filter coefficient) are dynamically adjusted based on the body movement level to construct an adaptive filter bank to extract the respiratory and heart rate signals, and different scale information mining strategies are combined to resist body movement interference while completely retaining physiological signal details. In combination with the normalization processing of the heart rate, respiratory rate and body movement energy, the individual difference influence is effectively reduced, so that the scheme has good adaptability to different people and different time periods of the same person. Meanwhile, by introducing a user information array (compatible with personal attributes) and a sleep state information array (retaining time sequence association), and using a deep learning model for feature extraction and classification, the sensor real-time parameters and time dimension information are fully fused under the premise of controlling the model parameter size, so that the feature richness and time sequence association of each sleep stage are significantly improved. Meanwhile, based on the easily obtained original signals such as pressure signals and piezoelectric signals, the scheme is subjected to processes such as off-bed detection, signal effectiveness screening, accurate preprocessing and multi-source array fusion, without the need for complex wearing and massive training data, and the convenience of home monitoring in the smart home scenario and the flexibility of engineering implementation are taken into account. The scheme effectively solves the problems of insufficient accuracy of traditional simplified monitoring equipment, physiological signals easily disturbed by body movement, and insufficient feature fusion, significantly improves the anti-interference ability and result accuracy of sleep staging, and adapts to the sleep monitoring application requirements in multiple fields such as smart home health management and smart terminals.

[0064] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0065] In one embodiment, a sleep staging system is provided, which corresponds one-to-one with the sleep staging methods described in the above embodiments. For example... Figure 7 As shown, the sleep staging system includes a physiological feature raw signal acquisition unit 10, a signal state detection unit 20, a final sleep staging result acquisition unit 30, a physiological parameter signal acquisition unit 40, a target physiological feature array construction unit 50, and a sleep staging result generation unit 60. Detailed descriptions of each functional module are as follows: The physiological characteristic raw signal acquisition unit 10 is used to acquire the physiological characteristic raw signals of the target object; Signal state detection unit 20 is used to perform signal state detection on the raw physiological characteristic signals, and the signal state detection includes at least bed exit state detection and signal validity detection; The final sleep staging result acquisition unit 30 is used to acquire the sleep staging data of the target object for the day if the target object is out of bed or the original physiological characteristic signal is invalid, and for a preset duration, and to perform data cleaning on the sleep staging data of the day to obtain the final sleep staging result. The physiological parameter signal acquisition unit 40 is used to perform preprocessing operations on the original physiological characteristic signal to separate at least two physiological parameter signals related to sleep state if the target object is in a bed state and the original physiological characteristic signal is valid. The target physiological feature array construction unit 50 is used to extract multi-dimensional features from the physiological parameter signals to construct a target physiological feature array. The sleep staging result generation unit 60 is used to obtain the sleep staging result corresponding to the current time based on the target physiological feature array, the target object information array, and the sleep state information array, through a preset sleep staging model.

[0066] In one embodiment of this application, the sleep staging result generation unit 60 is further configured to: The target physiological feature array is input into the feature extraction model, and the feature extraction model is used to extract features to obtain the sleep depth feature array. The sleep depth feature array is concatenated with the target object information array and the sleep state information array to obtain the feature array to be predicted; The feature array to be predicted is input into the classification prediction model, and the classification prediction model is used to perform classification prediction to obtain the sleep stage result corresponding to the current time.

[0067] In an embodiment of the present application, the physiological characteristic original signal includes a pressure signal and a piezoelectric signal, and the physiological parameter signal acquisition unit 40 is further configured to: filter the pressure signal based on a preset characteristic frequency interval of the body motion signal to separate a body motion signal; perform feature analysis and quantization processing based on the body motion signal to obtain a body motion level reflecting the intensity of body motion interference; adaptively adjust filter parameters based on the body motion level to dynamically construct a breathing signal filter set and / or a heart rate signal filter set; input the piezoelectric signal into the breathing signal filter set and / or the heart rate signal filter set to obtain a breathing signal and / or a heart rate signal related to the sleep state after filter processing.

[0068] In an embodiment of the present application, the physiological parameter signal acquisition unit 40 is further configured to: use the body motion level as an adjustment parameter to calculate a breathing signal filter cutoff frequency and / or a heart rate signal filter cutoff frequency; based on a preset filter order and a signal sampling frequency, calculate filter coefficients corresponding to the breathing signal filter cutoff frequency and / or filter coefficients corresponding to the heart rate signal filter cutoff frequency, respectively; based on the breathing signal filter cutoff frequency and the corresponding filter coefficients, construct a breathing signal filter set; and / or based on the heart rate signal filter cutoff frequency and the corresponding filter coefficients, construct the heart rate signal filter set.

[0069] In an embodiment of the present application, the physiological parameter signal acquisition unit 40 is further configured to: perform a preset window length sliding window absolute value summation processing on the body motion signal to obtain a body motion energy signal; determine a body motion energy maximum value of the body motion energy signal within the preset window length; statistically determine an effective body motion number of the body motion energy maximum value greater than a preset body motion energy threshold value within the preset window length, and a time length ratio of a cumulative time length of the body motion energy signal greater than the preset body motion energy threshold value to the preset window length; based on the effective body motion number and the time length ratio, calculate the body motion level.

[0070] In an embodiment of the present application, the physiological parameter signal includes a breathing signal, a heart rate signal, and a body motion signal, and the target physiological characteristic array construction unit 50 is further configured to: determine feature peaks of the breathing signal and the heart rate signal, respectively, and based on the feature peaks, determine corresponding interval data and amplitude data, respectively. performing energy analysis on the body motion signal to obtain body motion energy data, and counting the number of body motions in the preset time window; obtaining a reference physiological parameter in a pre-sleep resting state, the reference physiological parameter including an average respiratory rate, an average heart rate, and an average body motion energy value; based on the reference physiological parameter, respectively normalizing the interval data and the body motion energy data to obtain normalized data; respectively calculating variability features corresponding to the respiratory signal and the heart rate signal, and normalizing the variability features based on the reference physiological parameter to obtain normalized variability features; In a preset time range, the interval data, the amplitude data, the body motion energy data, the body motion frequency, the normalized data, the variability features and the normalized variability features are integrated to construct the target physiological feature array.

[0071] In an embodiment of the present application, the signal state detection unit 20 is also used for: based on the physiological feature raw signal, detecting whether the target object is out of bed; if the target object is in bed, determining whether the physiological feature raw signal is valid; wherein, if the length of time of the physiological feature raw signal being truncated or 0 in the preset window time accounts for more than a preset proportion threshold, the physiological feature raw signal is invalid.

[0072] In an embodiment of the present application, the final sleep staging result acquisition unit 30 is also used for: determining a target sleep segment in the sleep staging data, wherein the target sleep segment includes at least one of a short sleep segment, an island sleep segment, and a nap sleep segment; According to a preset data cleaning rule, data cleaning is performed on different types of target sleep segments, and all sleep staging data is integrated to obtain a final sleep staging result.

[0073] In the embodiments of the present application, the filtering parameters (such as the cutoff frequency and the filtering coefficient) are dynamically adjusted based on the body motion level to construct an adaptive filter bank to extract the respiratory and heart rate signals, and the different scale information mining strategies are combined to resist the body motion interference while completely retaining the physiological signal details, and the normalization processing of the heart rate, respiratory rate and body motion energy is performed to effectively reduce the individual difference influence, so that the scheme has good adaptability in different populations and different time periods of the same person; meanwhile, the user information array (compatible with the personal attributes) and the sleep state information array (retaining the time sequence correlation) are introduced, and the deep learning model is used for feature extraction and classification, so that the real-time sensor parameters and the time dimension information are fully fused under the premise of controlling the model parameter size, and the feature richness and the time sequence correlation of each sleep stage are significantly improved; meanwhile, the scheme is based on the easily obtained original signals such as the pressure signal and the piezoelectric signal, and the processes such as off-bed detection, signal effectiveness screening, accurate preprocessing and multi-source array fusion are performed, so that the scheme does not need complex wearing and massive training data, and the convenience of home monitoring in the smart home scene and the flexibility of engineering implementation are considered, and the problems of insufficient accuracy of the traditional simplified monitoring equipment, physiological signals easily disturbed by body motion, and insufficient feature fusion are effectively solved, and the anti-interference ability and the result accuracy of the sleep staging are significantly improved, and the scheme adapts to the sleep monitoring application requirements in multiple fields such as smart home health management and smart terminal.

[0074] The specific limitations of the sleep staging system can be referred to the limitations of the sleep staging method in the above, which will not be repeated here. Each module in the above sleep staging system can be realized by software, hardware and their combination in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above modules.

[0075] In one embodiment, a computer device is provided, which can be a terminal device, and an internal structure diagram thereof can be as shown in Figure 8 The computer device includes a processor, a memory and a network interface connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a readable storage medium. The readable storage medium stores computer readable instructions. The network interface of the computer device is used to communicate with external terminals through network connection. The computer readable instructions are executed by the processor to implement a sleep staging method. The readable storage medium provided in the embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.

[0076] In the embodiments of the present application, a computer device is provided, which comprises a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, and the processor implements the steps of the sleep staging method as described above when executing the computer readable instructions.

[0077] In the embodiments of the present application, a readable storage medium is provided, which stores computer readable instructions, and the computer readable instructions implement the steps of the sleep staging method as described above when executed by a processor.

[0078] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by computer readable instructions instructing related hardware, and the computer readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0079] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0080] The above examples are only used to illustrate the technical solutions of the present application, but not limit the same; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent ones; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for sleep staging, characterized in that, The method includes: Obtain raw physiological characteristic signals of the target object; Perform signal state detection on the raw physiological characteristic signals, the signal state detection including at least out-of-bed state detection and signal validity detection; If the target object is out of bed or the original physiological characteristic signal is invalid, and continues for a preset duration, the sleep stage data of the target object for the day is obtained, and the sleep stage data for the day is cleaned to obtain the final sleep stage result. If the target object is in bed and the original physiological characteristic signal is valid, the original physiological characteristic signal is preprocessed to separate at least two physiological parameter signals related to sleep state. Multi-dimensional feature extraction is performed on the physiological parameter signals to construct a target physiological feature array; Based on the target physiological feature array, target object information array, and sleep state information array, the sleep stage result corresponding to the current moment is obtained through a preset sleep stage model.

2. The sleep staging method as described in claim 1, characterized in that, The process of obtaining the sleep staging result corresponding to the current moment based on the target physiological feature array, the target object information array, and the sleep state information array, through a preset sleep staging model, includes: The target physiological feature array is input into the feature extraction model, and the feature extraction model is used to extract features to obtain the sleep depth feature array. The sleep depth feature array is concatenated with the target object information array and the sleep state information array to obtain the feature array to be predicted; The feature array to be predicted is input into the classification prediction model, and the classification prediction model is used to perform classification prediction to obtain the sleep stage result corresponding to the current time.

3. The sleep staging method as described in claim 1, characterized in that, The raw physiological characteristic signals include pressure signals and piezoelectric signals. The preprocessing operation on the raw physiological characteristic signals separates at least two physiological parameter signals related to sleep state, including: Based on a preset characteristic frequency range of the body motion signal, the pressure signal is filtered to separate the body motion signal. Based on the body motion signal, feature analysis and quantification are performed to obtain the body motion level reflecting the intensity of body motion interference; Based on the body movement level, the filtering parameters are adaptively adjusted to dynamically construct a respiratory signal filter group and / or a heart rate signal filter group. The piezoelectric signal is input to the respiratory signal filter group and / or the heart rate signal filter group, and after filtering, the respiratory signal and / or heart rate signal related to the sleep state are obtained respectively.

4. The sleep staging method as described in claim 3, characterized in that, The step of adaptively adjusting filtering parameters based on the body movement level to dynamically construct a respiratory signal filter bank and / or a heart rate signal filter bank includes: Using the body movement level as an adjustment parameter, calculate the cutoff frequency of the respiratory signal filter and / or the cutoff frequency of the heart rate signal filter. Based on the preset filter order and signal sampling frequency, calculate the filter coefficients corresponding to the cutoff frequency of the respiratory signal filter and / or the filter coefficients corresponding to the cutoff frequency of the heart rate signal filter. Based on the cutoff frequency of the respiratory signal filter and the corresponding filter coefficients, a respiratory signal filter bank is constructed; and / or The heart rate signal filter bank is constructed based on the cutoff frequency of the heart rate signal filter and the corresponding filter coefficients.

5. The sleep staging method as described in claim 3, characterized in that, The step of performing feature analysis and quantification based on the body motion signal to obtain the body motion level reflecting the intensity of body motion interference includes: The body motion signal is obtained by summing the absolute values ​​of the sliding window for a preset window duration. Determine the maximum value of the body kinetic energy signal within the preset window duration; The number of effective body movements in which the maximum value of the body energy is greater than the preset body energy threshold within the preset window duration is counted, as well as the percentage of the total duration in which the body energy signal is greater than the preset body energy threshold within the preset window duration. The physical activity level is calculated based on the ratio of the effective number of physical movements to the duration.

6. The sleep staging method as described in claim 1, characterized in that, The physiological parameter signals include respiratory signals, heart rate signals, and body movement signals. The multi-dimensional feature extraction of the physiological parameter signals to construct a target physiological feature array includes: The characteristic peak values ​​of the respiratory signal and heart rate signal are determined respectively, and the corresponding interval data and amplitude data are determined based on the characteristic peak values ​​respectively; Energy analysis is performed on the body movement signal to obtain body movement energy data, and the number of body movements within a preset time window is counted. Obtain baseline physiological parameters at rest before sleep, including average respiratory rate, average heart rate, and average kinetic energy. Based on the aforementioned baseline physiological parameters, the interphase data and body energy data were normalized to obtain normalized data. The variability features corresponding to respiratory signals and heart rate signals are calculated separately, and the variability features are normalized based on the baseline physiological parameters to obtain normalized variability features. Within a preset time range, the interphase data, amplitude data, body energy data, number of body movements, normalized data, variability features, and normalized variability features are integrated to construct the target physiological feature array.

7. The sleep staging method as described in claim 1, characterized in that, The step of performing signal state detection on the original physiological characteristic signal includes: Based on the original physiological characteristic signals, the target object is detected after leaving the bed; If the target object is in a bed state, determine whether the original physiological characteristic signal is valid; If the percentage of time during which the original physiological feature signal is truncated or zero within a preset window period is greater than a preset percentage threshold, then the original physiological feature signal is invalid.

8. The sleep staging method as described in claim 1, characterized in that, The process of cleaning the sleep stage data for the day to obtain the final sleep stage result includes: Identify the target sleep segment in the sleep stage data, wherein the target sleep segment includes at least one of short sleep segment, isolated sleep segment, and nap sleep segment; According to the preset data cleaning rules, the data of different types of target sleep segments are cleaned, and then all sleep stage data are integrated to obtain the final sleep stage results.

9. A sleep staging system, characterized in that, The system includes: The physiological characteristic raw signal acquisition unit is used to acquire the physiological characteristic raw signals of the target object; A signal state detection unit is used to perform signal state detection on the raw physiological characteristic signals, wherein the signal state detection includes at least out-of-bed state detection and signal validity detection. The final sleep staging result acquisition unit is used to acquire the sleep staging data of the target object for the day if the target object is out of bed or the original physiological characteristic signal is invalid, and for a preset duration, and to perform data cleaning on the sleep staging data of the day to obtain the final sleep staging result. A physiological parameter signal acquisition unit is used to preprocess the original physiological characteristic signal to separate at least two physiological parameter signals related to sleep state if the target object is in a bed state and the original physiological characteristic signal is valid. The target physiological feature array construction unit is used to extract multi-dimensional features from the physiological parameter signals to construct the target physiological feature array. The sleep staging result generation unit is used to obtain the sleep staging result corresponding to the current moment based on the target physiological feature array, the target object information array, and the sleep state information array, through a preset sleep staging model.

10. A readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by a processor, they implement the steps of the sleep staging method as described in any one of claims 1 to 8.

Images