Sleep state evaluation system and method based on time series data analysis

Abstract

The application discloses a sleep state evaluation system and method based on time series data analysis, and relates to the technical field of intelligent sensors.The sleep state evaluation system and method based on time series data analysis comprise the following steps: S1, collecting, normalizing and time window reconstructing interlayer sensing data; S2, estimating a pillow core lag response, and inversing a pillow core lag residual error; S3, identifying a sleep behavior event, generating a candidate sleep disturbance chain, and screening an effective sleep disturbance chain; and S4, constructing a state feature set, generating a sleep state evaluation report, and binding a sensing evidence chain.The application effectively improves the explainability, time series consistency and state discrimination credibility of interlayer sensing data in intelligent pillow sleep state evaluation, and solves the problem of secondary deformation signal and human head and neck behavior signal aliasing caused by pillow core transmission.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, specifically to a sleep state assessment system and method based on time-series data analysis. Background Technology

[0002] With the increasing application of home sleep monitoring devices and smart sensors in pillow products, existing solutions often directly map interlayer pressure signals, micro-vibration signals, and sound signals to events such as turning over, breathing, snoring, pillow lifting, and awakening. In existing solutions, after the pillow core material experiences hysteresis recovery, neighborhood diffusion, and amplitude attenuation in response to human body load, the signals collected by the interlayer sensors are no longer a direct response to human head and neck behavior. If used directly for sleep state judgment, this can easily lead to misjudgment of respiratory cycle gaps, misidentification of snoring events, missegmentation of turning over recovery segments, and deviations in sleep state assessment results.

[0003] For example, the invention patent with announcement number CN113012719B discloses a method, system, electronic device, and storage medium for fatigue assessment of sleep sound signals. The method includes: acquiring audio signals during sleep; calculating the Mel-spectrum features of the audio signals; combining a Gaussian mixture model to separate the audio signals into snoring and non-snoring segments; calculating the proportion of snoring duration; inputting the Mel-spectrum features of the audio signals into an artificial intelligence speech model; the artificial intelligence speech model calculating and outputting a sound feature vector; obtaining the total sleep duration; and inputting the proportion of snoring duration, the sound feature vector, and the total sleep duration into a machine learning classifier to output a fatigue level. This method overcomes the limitations of traditional fatigue assessment methods and can objectively and accurately assess the body's daily fatigue state without relying on various medical devices. The machine learning algorithm fully extracts features from sleep sound signals, and the constructed assessment model has high accuracy.

[0004] For example, invention patent CN118675547B discloses a method and system for optimizing sleep quality based on a user's sleep pattern. The method includes: real-time monitoring and recording of the user's sounds during sleep using a sound acquisition device, followed by preprocessing; analyzing the acquired sound features; establishing a correspondence between sound features and sleep stages; evaluating the current sound features and mapping them to the correspondence to determine the user's current sleep quality state; controlling a sound output device to play sound stimuli based on the evaluated sleep quality state to optimize the user's sleep quality; continuously monitoring the user's sounds during sleep, analyzing changes in sleep quality, and adjusting the sound output in real time to form a closed-loop control. This invention adjusts the sound output based on the real-time monitored sleep state, providing personalized sound stimuli to the user, thereby optimizing sleep quality.

[0005] In existing technologies, pillow core materials can cause signal delays, attenuation, diffusion, and hysteresis recovery. The same human movement will produce different sensor responses under different pillow core thicknesses, different indentation depths, and different head and neck positions. If existing systems treat the interlayer signals as the original human signals, they will mistakenly write the pillow core material responses into sleep state judgments, making sleep staging and sleep quality scores lack a physically reliable basis.

[0006] Therefore, in order to address the above issues, there is an urgent need for a sleep state assessment system and method based on time-series data analysis. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a sleep state assessment system and method based on time-series data analysis, which solves the problem of the superposition of secondary deformation signals caused by pillow core transmission and human head and neck behavioral signals.

[0008] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a sleep state assessment method based on time-series data analysis, comprising: S1, collecting, normalizing, and reconstructing time windows of the interlayer sensor data during the user's pillow-lying process to generate a standard observation dataset of the pillow interlayer; S2, estimating the pillow core hysteresis response based on the standard observation dataset of the pillow interlayer, inverting the pillow core hysteresis residual to generate a head and neck behavior inversion dataset; S3, identifying sleep behavior events based on the head and neck behavior inversion dataset, generating candidate sleep disturbance chains, filtering valid sleep disturbance chains, and generating a sleep behavior event chain dataset; S4, constructing a set of state features based on the sleep behavior event chain dataset, generating a sleep state assessment report, and binding a sensor evidence chain.

[0009] Furthermore, the specific process of collecting, normalizing, and reconstructing the time window of the interlayer sensing data during the user's pillow-lying process to generate a standard observation dataset of the pillow interlayer is as follows: Pressure sensing units, vibration sensing units, and sound acquisition units are deployed within the smart pillow interlayer; data is collected during the user's pillow-lying process to form raw interlayer sensing data; the raw interlayer sensing data includes the pressure sensing area number, original pressure sampling value, sensor sampling time, pillow-end reception time, and sampling sequence number; idle data is collected when the user is not lying on the pillow, and a perturbation chain template set is obtained to form a pressure idle baseline; the raw data of each pressure sensing area is then processed... The pressure sample value is subtracted from the corresponding pressure idle baseline to obtain the net pressure value. When the absolute value of the difference between the net pressure value and the previous analysis time window is greater than the pressure jump threshold, the net pressure value is replaced with the median value of the net pressure values ​​of adjacent analysis time windows before and after the current analysis time window, and a jump correction flag is generated. When there is a gap in the sensor sampling sequence number, a missing sampling flag is generated in the corresponding analysis time window and the analysis time window position is retained. An analysis time window is established based on the sampling sequence number and the pillow end reception time. After the net pressure value is mapped to the analysis time window, the original interlayer sensing data is linearly normalized to generate a pillow interlayer standard observation dataset.

[0010] Furthermore, based on the standard observation dataset of the pillow interlayer, the specific process for estimating the pillow core hysteresis response is as follows: Based on the standard observation dataset of the pillow interlayer, the sensing areas whose net pressure value meets the contact judgment threshold are determined as effective head and neck bearing areas. The area numbers of all effective head and neck bearing areas are grouped into a set to obtain the set of sensing areas in the effective head and neck bearing state within the current time window. The uniquely bound sensing channel is directly indexed according to the sensing area number, and the sampled value of the channel within the time window is read as the interlayer sensing value. Based on the spatial coordinates of each sensing area within the pillow interlayer, the Euclidean distance between the center coordinates of the areas is calculated according to the rule that the Euclidean distance between the center coordinates of the areas is less than the adjacency distance threshold. Areas whose Euclidean distance meets the adjacency distance threshold are determined as nearby sensing areas. The region is divided into a set of neighboring sensing regions, composed of the region numbers of all adjacent sensing regions. Based on the propagation hysteresis window number of the sensing region, the mezzanine sensing values ​​of each neighboring sensing region within the set are read back within a time window. The median of the read-back mezzanine sensing values ​​is taken to obtain the neighboring hysteresis reference value of the sensing region. The original pressure sampling values ​​are processed using first-order difference, and the time window in which the absolute value of the first-order difference satisfies the pressure jump threshold is determined as the pressure jump point. Within the hysteresis search window after the pressure jump point, the mezzanine response sequence of the sensing region is read, and the time window in which the mezzanine response amplitude reaches its peak value is determined as the mezzanine response peak time window. The difference between the mezzanine response peak time window number and the pressure jump point time window number is determined as the propagation hysteresis window number of the sensing region.

[0011] Further, the specific process for inverting the pillow core hysteresis residual is as follows: The median of the interlayer sensing values ​​from the previous and two previous time windows, along with the adjacent hysteresis reference values, is taken within the same group to obtain the comprehensive hysteresis reference value. The time interval between adjacent time windows is divided by the sum of the pillow core rebound time constant and the anti-zero term to obtain the rebound time ratio. The rebound time ratio is negatively calculated and then exponentially calculated to obtain the pillow core rebound attenuation coefficient. The comprehensive hysteresis reference value is multiplied by the pillow core rebound attenuation coefficient to obtain the pillow core hysteresis estimate. The interlayer sensing value within the current time window is subtracted from the pillow core hysteresis estimate, and the absolute value of the subtraction is taken to obtain the absolute value of the hysteresis residual. The absolute value of the hysteresis residual is divided by the sum of the effective amplitude reference and the anti-zero term to obtain the scaled hysteresis residual. The scaled hysteresis residual is incremented by one and then the natural logarithm is taken to obtain the hysteresis residual compression term. The median of the hysteresis residual compression terms for all effective head and neck bearing sensing areas within the current time window is taken to obtain the pillow core hysteresis residual inversion value.

[0012] Furthermore, the specific process for generating the head and neck behavior inversion dataset is as follows: the inversion value of the pillow core hysteresis residual is compared with the behavior inversion threshold; when the inversion value of the pillow core hysteresis residual is greater than the behavior inversion threshold, the current time window is marked as an effective window for human head and neck behavior; when the inversion value of the pillow core hysteresis residual is less than or equal to the behavior inversion threshold, the current time window is marked as the dominant window for the pillow core material response, the inversion value of the pillow core hysteresis residual is retained, and the time window is not directly used as the basis for judging event types for which triggering features have not yet been defined; the head and neck behavior inversion dataset is generated by structurally binding according to a unified analysis time.

[0013] Furthermore, based on the head and neck behavior inversion dataset, the specific process for identifying sleep behavior events and generating candidate sleep disturbance chains is as follows: Based on the head and neck behavior inversion dataset, events whose absolute value of the first-order difference of net stress exceeds the behavior inversion threshold and whose time window number is greater than M are identified as sleep behavior events; sleep behavior events are sorted according to a unified analysis time, and the time interval, carrier zone number, and event type between adjacent sleep behavior events are read; when the difference in the number of time windows between sleep behavior events is less than or equal to the time interval window number threshold, and the carrier zone number is the same or the Euclidean distance between the center coordinates of two carrier zones is less than the adjacency distance threshold, the sleep behavior events are classified into the same candidate sleep disturbance chain; when adjacent sleep behavior events do not meet the above conditions... When the conditions are met, the current candidate sleep disturbance chain ends and the next candidate sleep disturbance chain begins; a reference time interval is generated based on the median of the time intervals of the corresponding events of all samples in the disturbance chain template set; the event type field of each sleep behavior event in the candidate sleep disturbance chain is read, the event type field is converted into a fixed event type code, and arranged according to the unified analysis time of the event start, to obtain the event type sequence of the candidate sleep disturbance chain; the template event type field in the disturbance chain template is read and arranged according to the template definition order, to obtain the event type sequence of the disturbance chain template; starting from the first event type code of the event type sequence, each event is counted until the last event type sequence, to obtain the number of events contained in the disturbance chain template.

[0014] Further, the specific process for screening effective sleep disturbance chains and generating a sleep behavior event chain dataset is as follows: Based on the inverted data fragments corresponding to each event in the candidate sleep disturbance chain, the contact state verification value of the events in the candidate sleep disturbance chain is obtained through the contact state evidence verification method; the longest common subsequence matching is performed between the candidate event type sequence and the template event type sequence, and the number of events with the same order in the two sequences is counted to obtain the length of the common subsequence; the length of the common subsequence is divided by the sum of the total number of events in the template event type sequence and the zero-prevention term to obtain the event order coverage term; the actual time interval of each group of adjacent events in the candidate sleep disturbance chain is subtracted from the reference time interval of the corresponding adjacent events in the disturbance chain template to obtain the difference between adjacent event intervals; the difference between adjacent event intervals is divided by the corresponding reference time interval and the length of the template event type sequence. The sum of zero-prevention terms is taken as the absolute value to obtain the time rhythm deviation value. The median of the time rhythm deviation values ​​of all adjacent events is taken, then the negative value is used for exponential operation to obtain the time rhythm adaptation term. The minimum value of all contact state verification values ​​is taken to obtain the contact state weakness term. The event sequence coverage term, time rhythm adaptation term, and contact state weakness term are multiplied sequentially to obtain the template coverage result. The maximum value of the template coverage result is taken to obtain the sleep disturbance chain sequence coverage value. The sleep disturbance chain sequence coverage value is compared with the chain formation threshold. When the sleep disturbance chain sequence coverage value is greater than the chain formation threshold, the current candidate sleep disturbance chain is retained as a valid sleep disturbance chain. When the sleep disturbance chain sequence coverage value is less than or equal to the chain formation threshold, the current candidate sleep disturbance chain is split into isolated event fragments to generate a sleep behavior event chain dataset.

[0015] Furthermore, the specific process of constructing a state feature set based on the sleep behavior event chain dataset is as follows: Based on the sleep behavior event chain dataset, a state feature set is constructed using the time window feature aggregation method. Based on the state feature set, the number of complete evidence items in the set is divided by the total number of required evidence items to obtain the sensor evidence completeness. The number of continuously satisfied windows is obtained using the continuous window backtracking method. The actual values ​​of key features corresponding to candidate sleep states are read window by window from the analysis time window backward and compared with the feature conditions corresponding to the state. The number of analysis time windows that continuously satisfy the feature conditions is accumulated. Backtracking stops when an analysis time window that does not satisfy the feature conditions is encountered, thus obtaining the number of satisfied windows. The number of analysis time windows of all candidate state segments identified in the current night under the same sleep state type is counted, and the median is taken as the duration scale. The feature numbers used to represent the state are used to form a key feature set. The data items that match the feature numbers within the analysis time window are read and the maximum value is calculated. Minimum normalization is used to obtain the actual values ​​of state features. A reference center value is obtained using in-state statistical methods. The absolute value of the difference between each actual value and the reference center value is calculated, and the median of all absolute differences is taken to obtain the scale value. The number of satisfying windows is divided by the sum of the persistence scale and the zero-prevention term to obtain the state persistence ratio. The state persistence ratio is negatively evaluated and then exponentially calculated. The result of the exponential calculation is subtracted from the actual value of the state feature to obtain the deviation distance. The deviation distance is divided by the sum of the scale value and the zero-prevention term to obtain the scaled deviation value. The scaled deviation value is exponentially calculated according to the deviation sensitivity coefficient, then one is added, and the reciprocal is taken to obtain the feature adaptation term. The minimum value of the feature adaptation terms for all state features within the key feature set is taken to obtain the core adaptation term. The completeness of sensor evidence, the fragment adaptation term, and the core adaptation term are multiplied sequentially to obtain the sleep state fragment evidence adaptation value.

[0016] Furthermore, the specific process of generating a sleep state assessment report and binding the sensor evidence chain is as follows: The sleep state segment evidence fit values ​​for each candidate sleep state within each analysis time window are read according to a unified analysis time; analysis time windows where the sleep state segment evidence fit values ​​for the same candidate sleep state meet the corresponding state threshold and continuously meet the corresponding number of confirmation windows are merged into the same state segment; the first analysis time window that meets the conditions in the state segment is taken as the segment start point, and the last analysis time window that continuously meets the conditions is taken as the segment end point, forming a sleep state segment; the starting unified analysis time of the first sleep segment is determined as the sleep onset start point; the start and end times, duration, and number of segments for each type of sleep state segment are statistically analyzed to generate sleep abnormality cause labels and a sleep state assessment report; the segment start point and segment end point of each sleep state segment are read and indexed in the head and neck behavior inversion dataset and the sleep behavior event chain dataset; the sensor evidence obtained from the index is structurally bound to the corresponding sleep state segment number to generate a sensor evidence chain.

[0017] Furthermore, a second aspect of the present invention provides a sleep state assessment system based on time-series data analysis, applied to a sleep state assessment method based on time-series data analysis, comprising: a pillow core time-series acquisition and preprocessing module, used to acquire, normalize, and reconstruct time windows of pillow core sensor data during the user's pillow-lying process, generating a pillow core standard observation dataset; a pillow core hysteresis inversion module, used to estimate the pillow core hysteresis response based on the pillow core standard observation dataset, invert the pillow core hysteresis residual, and generate a head and neck behavior inversion dataset; a sleep event chain construction module, used to identify sleep behavior events based on the head and neck behavior inversion dataset, generate candidate sleep disturbance chains, filter valid sleep disturbance chains, and generate a sleep behavior event chain dataset; and a state assessment and evidence binding module, used to construct a state feature set based on the sleep behavior event chain dataset, generate a sleep state assessment report, and bind sensor evidence chains.

[0018] Beneficial effects The present invention has the following beneficial effects: (1) In this invention, the main head bearing area, neck support area, left offset area, right offset area and occipital transition area are collected in separate areas, and the effective head and neck bearing area set is determined based on the normalized pressure value. This allows the calculation to be carried out only on the real bearing area, reducing the interference of short-term edge contact and non-bearing area signals on sleep state assessment.

[0019] (2) The present invention corrects the interlayer response of different sensing areas by transmitting the number of hysteresis windows, the pillow core rebound time constant and the effective amplitude reference, so that the same head and neck behavior can form a unified data caliber in different bearing areas, different indentation states and different rebound states, thereby improving the consistency of sleep state assessment results in continuous nighttime monitoring.

[0020] (3) In this invention, the event sequence, adjacent event time interval and contact state verification value of candidate sleep disturbance chains are jointly evaluated by the sleep disturbance chain sequence coverage value, isolated sound segments, isolated micro-motion segments and non-continuous bearing segments are screened out, thereby improving the recognition credibility of breathing disturbance chains, latent arousal chains and pillow-to-pillow chains.

[0021] (4) In this invention, by combining the completeness of sensor evidence, the number of consecutively satisfied windows and the key feature deficiency fitting items through the sleep state fragment evidence fitting value, the stable sleep state, breathing disturbance state, latent arousal state, posture adjustment state and pillow-off state all have corresponding data sources, avoiding the occasional fluctuations of a single analysis time window from directly changing the sleep state assessment results.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] Figure 1 This is a flowchart of the sleep state assessment method based on time-series data analysis of the present invention; Figure 2 This is a diagram illustrating the architecture of the sleep state assessment system based on time-series data analysis according to the present invention. Figure 3 This is a comparison diagram of the pillow core hysteresis compensation before and after the invention. Figure 4 This is the inversion heat map of the sensing region of the present invention; Figure 5 This is a curve showing the adaptation value of the multi-state evidence in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1-5This invention provides a technical solution: a sleep state assessment method based on time-series data analysis, comprising: S1, collecting, normalizing, and reconstructing time windows of the pillow interlayer sensor data during the user's pillow-lying process to generate a standard observation dataset of the pillow interlayer; S2, estimating the pillow core hysteresis response based on the standard observation dataset of the pillow interlayer, inverting the pillow core hysteresis residual to generate a head and neck behavior inversion dataset; S3, identifying sleep behavior events based on the head and neck behavior inversion dataset, generating candidate sleep disturbance chains, filtering valid sleep disturbance chains, and generating a sleep behavior event chain dataset; S4, constructing a set of state features based on the sleep behavior event chain dataset, generating a sleep state assessment report, and binding a sensor evidence chain.

[0026] Specifically, the process of collecting, normalizing, and reconstructing time windows of the interlayer sensing data during the user's pillow-lying process to generate a standard observation dataset of the pillow interlayer is as follows: Pressure sensing units, vibration sensing units, and sound acquisition units are deployed within the interlayer of the smart pillow. The pressure sensing units are arranged in zones according to the main head support area, neck support area, left offset area, right offset area, and pillow edge transition area. Vibration sensing units are located in the interlayer positions corresponding to the main head support area and neck support area. The sound acquisition unit is located on the pillow body near the side that contacts the head and neck. Data is collected during the user's lying down to form raw interlayer sensing data. The raw interlayer sensing data includes the pressure sensing area number, the original pressure sampling value, the sensor sampling time, the pillow end reception time, and the sampling sequence number. When the user is not lying down, idle data is collected and a set of disturbance chain templates is obtained to form a pressure idle baseline. The pressure idle baseline is the median value of the original pressure sampling value of the corresponding pressure sensing area within the idle sampling window.

[0027] The net pressure value is obtained by subtracting the corresponding pressure idle baseline from the original pressure sample value of each pressure sensing area. When the absolute value of the difference between the net pressure value and the previous analysis time window is greater than the pressure jump threshold, the net pressure value is replaced with the median value of the net pressure values ​​of adjacent analysis time windows before and after the current analysis time window, and a jump correction flag is generated. When there is a gap in the sensor sampling sequence number, a missing sampling flag is generated in the corresponding analysis time window and the analysis time window position is retained. An analysis time window is established based on the sampling sequence number and the pillow end receiving time. Since the sensor sampling time may be asynchronous in the distributed acquisition system, the pillow end receiving time is used in conjunction with the sampling sequence number for sequential calibration to eliminate the impact of transmission jitter on the alignment of the time window. After the net pressure value is mapped to the analysis time window, the median value of the net pressure value within the analysis time window is taken as the pressure value of the time window. The original data of the interlayer sensing is linearly normalized to generate a standard observation dataset of the pillow interlayer.

[0028] In this implementation plan, the interference of single-point abnormal sampling and missing sampling number intervals on the continuity of data in the analysis time window is reduced by pressure jump correction and missing sampling identification retention. The analysis time window is established by combining the sampling number and the pillow end reception time, so that the interlayer sensing data can be reconstructed in a unified time order, reducing the time misalignment caused by asynchronous distributed acquisition and transmission jitter. By aggregating and normalizing the net pressure value in the time window, a standard observation dataset of pillow interlayer with consistent data caliber, clear time order and complete anomaly identification is formed, providing a stable data foundation for pillow core hysteresis inversion, sleep behavior event identification and sleep state assessment.

[0029] Specifically, based on the standard observation dataset of pillow interlayer, the process of estimating the hysteresis response of the pillow core is as follows: Based on the standard observation dataset of the pillow interlayer, the sensing areas whose net pressure values ​​meet the contact determination threshold are identified as effective head and neck bearing areas. A set of region numbers for all effective head and neck bearing areas is created, resulting in the set of sensing areas in effective head and neck bearing state within the current time window. The contact determination threshold is set based on the fluctuation range of the unloaded baseline, taking twice the standard deviation of the unloaded net pressure value. The uniquely bound sensing channel is directly indexed by the sensing area number, and the sampled value of the channel within the time window is read as the interlayer sensing value. The raw values ​​of all sampling points of the channel within the current time window are not subject to additional filtering. Based on the spatial coordinates of each sensing area within the pillow interlayer, the Euclidean distance between the center coordinates of the areas is calculated according to the rule that the Euclidean distance between the center coordinates of the areas is less than the adjacency distance threshold. Areas whose Euclidean distance meets the adjacency distance threshold are identified as neighboring sensing areas of the sensing area. A set of neighboring sensing areas is created by grouping the region numbers of all neighboring sensing areas. The spatial coordinates are pre-calibrated and stored in the configuration file, and the adjacency distance threshold is 1.5 times the minimum distance between the centers of each area. Based on the number of propagation lag windows in the sensing region, the interlayer sensing values ​​of each neighboring sensing region within the set of neighboring sensing regions are read back within the time window. The median of the read-back interlayer sensing values ​​is taken to obtain the neighboring lag reference value of the sensing region. The interlayer sensing values ​​of several lag windows before the current time window are read back; if the number of lag windows is zero, the current time window is read back. The original pressure sampling values ​​are processed by first-order difference, and the time window in which the absolute value of the first-order difference satisfies the pressure jump threshold is determined as the pressure jump point. The first-order difference is the difference between the net pressure value of the current time window and the net pressure value of the previous time window, and the pressure jump threshold is 5% of the full scale of the net pressure value. Within the lag search window after the pressure jump point, the interlayer response sequence of the sensing region is read, and the time window in which the interlayer response amplitude reaches its peak value is determined as the interlayer response peak time window. The length of the lag search window is 5 to 20 time windows, and the window starts from the first time window after the pressure jump point. The difference between the interlayer response peak time window number and the pressure jump point time window number is determined as the number of propagation lag windows in the sensing region. The difference is the peak time window number minus the mutation point time window number, and the result is a non-negative integer. If no peak is found, the number of propagation lag windows is recorded as the default window number.

[0030] In this implementation scheme, the neighboring hysteresis reference value is obtained by hysteresis window readback and median fusion, which reduces the impact of abnormal sampling in a single neighboring area on the hysteresis response estimation; the number of transmission hysteresis windows is determined by the time window difference between the pressure mutation point and the interlayer response peak, so that the response delay of different sensing areas can be quantified and recorded; through the above processing, the correspondence between the effective bearing area, neighboring areas, neighboring hysteresis reference value and the number of transmission hysteresis windows can be formed before entering the pillow core hysteresis residual inversion, providing a stable data basis for stripping the pillow core hysteresis recovery, regional diffusion and sensing delay effects.

[0031] Specifically, the process of inverting the hysteresis residual of the pillow core is as follows: The median of the interlayer sensing values ​​from the previous two time windows and the adjacent hysteresis reference values ​​in the same group is used to obtain the comprehensive hysteresis reference value. The time interval between adjacent time windows is divided by the sum of the pillow core rebound time constant and the anti-zero term to obtain the rebound time ratio. The negative value of the rebound time ratio is then used for exponential operation to obtain the pillow core rebound attenuation coefficient. The comprehensive hysteresis reference value is multiplied by the pillow core rebound attenuation coefficient to obtain the pillow core hysteresis estimate. The interlayer sensing value in the current time window is subtracted from the pillow core hysteresis estimate, and the absolute value of the subtraction result is taken to obtain the absolute value of the hysteresis residual. The absolute value of the hysteresis residual is divided by the sum of the effective amplitude reference and the anti-zero term to obtain the scaled hysteresis residual. The scaled hysteresis residual is incremented by one and then the natural logarithm is taken to obtain the hysteresis residual compression term. The median of the hysteresis residual compression terms of all effective head and neck bearing sensing areas in the current time window is taken to obtain the pillow core hysteresis residual inversion value.

[0032] The specific formula for calculating the inversion value of the pillow core hysteresis residual is as follows: ; In the formula, The value represents the pillow core hysteresis residual inversion value for the t-th time window, used to characterize the intensity of changes in human head and neck behavior remaining after removing the effects of pillow core hysteresis recovery and neighborhood diffusion within the current time window; t represents the analysis time window number, which is obtained by slicing the pressure sequence, micro-vibration sequence and sound sequence according to a fixed time window length, and the value range is positive integer, used to identify the time position corresponding to the current calculation. This indicates the number of the r-th sensing area within the pillow layer, which is determined by the partitioning arrangement of the pressure sensing units within the pillow layer. This indicates the total number of sensing areas within the pillow's interlayer, used to distinguish different sensing areas within the main head support area, neck support area, left offset area, right offset area, and pillow edge transition area; This represents the set of sensing regions in the effective head and neck bearing state within the t-th time window, used to define the effective bearing region participating in the calculation of the pillow core hysteresis residual inversion value; This means that the median of the calculation results for each sensing region within the effective head and neck bearing area set is obtained by sorting the hysteresis residual compression terms of each sensing region according to their numerical values ​​and taking the median value. This is used to reduce the impact of sudden jumps, local depressions and short-term missing sampling in a single region on the overall inversion results. This represents the mezzanine sensing value of the r-th sensing region within the t-th time window, used to represent the actual mezzanine response intensity of the current sensing region within the current time window; This represents the set of neighboring sensing regions of the r-th sensing region, used to provide a reference for the diffusion response of the region surrounding the r-th sensing region; This represents the qth neighboring sensing region number in the set of neighboring sensing regions. It is obtained by traversing the set of neighboring sensing regions and is used to read the interlayer sensing value of the neighboring region after hysteresis. This indicates that the q-th neighbor sensing region is in the q-th... The interlayer sensing value within a time window is used to represent the hysteresis diffusion reference generated by the neighboring area on the current area after being transmitted through the pillow core; This represents the neighboring hysteresis reference value of the r-th sensing region, used to estimate the diffusion response of the pillow core material in the spatial neighborhood; The number of transmission lag windows in the r-th sensing area within the t-th time window is determined by the time window difference between the pressure mutation point and the peak value of the interlayer response in the r-th sensing area. It is used to describe the lag time of human body load being transmitted from the pillow core to the interlayer sensor. Let represent the pillow core rebound time constant of the r-th sensing region within the t-th time window. This is obtained by measuring the time required for the pressure release segment to decrease from 80% to 20% of the effective bearing pressure reference. Its value range is [...]. , [This is used to characterize the time scale of a pillow core material recovering from a compressed state to a low-load state;] This represents the time interval between adjacent time windows, used to substitute the pillow core rebound time constant into the exponential decay term; The effective amplitude benchmark of the r-th sensing area in the t-th time window is determined by the 90th percentile value of the interlayer sensing value of the r-th sensing area within ten consecutive time windows after the user enters the pillow monitoring state. The value range is (0,1], which is used to scale the hysteresis residual to avoid the influence of amplitude differences in different sensing areas on the inversion value. This term represents the zero-prevention term. It is a decimal value greater than zero with a range of (0, 0.01) and is used to prevent the denominator from being zero and to ensure the stability of the calculation of the exponential, proportional, and logarithmic terms.

[0033] Table 1 shows the inversion data of the pillow core hysteresis characteristic parameters of the sensing areas. Multi-dimensional characteristic parameters were quantified and characterized for each of the seven sensing areas: Sensing area R1 has a transmission hysteresis window number of 1, a pillow core rebound time constant of 3.2, an effective amplitude baseline of 0.86, a neighboring hysteresis reference value of 0.73, a pillow core hysteresis estimate of 0.61, and a pillow core hysteresis residual inversion value of 0.118; Sensing area R2 has a transmission hysteresis window number of 1 and a pillow core rebound time constant of 3. 5. The effective amplitude baseline is 0.84, the nearest hysteresis reference value is 0.69, the pillow core hysteresis estimate is 0.58, and the pillow core hysteresis residual inversion value is 0.126; the number of transmission hysteresis windows in the sensing region R3 is 2, the pillow core rebound time constant is 4.1, the effective amplitude baseline is 0.82, the nearest hysteresis reference value is 0.65, the pillow core hysteresis estimate is 0.54, and the pillow core hysteresis residual inversion value is 0.154; the number of transmission hysteresis windows in the sensing region R4 is 2, and the pillow core... The rebound time constant is 4.4, the effective amplitude reference is 0.79, the nearest hysteresis reference value is 0.62, the pillow core hysteresis estimate is 0.51, and the pillow core hysteresis residual inversion value is 0.169; the sensing region R5 has a transmission hysteresis window number of 1, a pillow core rebound time constant of 3.8, an effective amplitude reference of 0.81, a nearest hysteresis reference value of 0.67, a pillow core hysteresis estimate of 0.56, and a pillow core hysteresis residual inversion value of 0.141; the sensing region R6 transmission... The data shows that in the sensing region R7, with a hysteresis window of 1, the pillow core rebound time constant is 3.1, the effective amplitude baseline is 0.85, the nearest hysteresis reference value is 0.72, the pillow core hysteresis estimate is 0.60, and the pillow core hysteresis residual inversion value is 0.121. In the sensing region R7, with a hysteresis window of 2, the data shows a larger pillow core rebound time constant, lower effective amplitude baseline and nearest hysteresis reference values, smaller pillow core hysteresis estimate, and higher hysteresis residual inversion value. In the region with a hysteresis window of 1, the overall rebound response is faster, the amplitude baseline is higher, and the residual inversion error is smaller. This data can provide a quantitative basis for modeling and error correction of pillow core hysteresis characteristics in different sensing regions.

[0034] Table 1. Inversion data of pillow core hysteresis characteristic parameters in the sensing area.

[0035] like Figure 3The comparison diagram before and after pillow core hysteresis compensation shows that the original interlayer sensing value, pillow core hysteresis estimate, and pillow core hysteresis residual inversion value are compared within the same analysis time window. The original interlayer sensing value represents the secondary deformation signal directly acquired by the interlayer sensor. The pillow core hysteresis estimate represents the material hysteresis response estimated by the historical time window, the adjacent sensing area, and the pillow core rebound time constant. The pillow core hysteresis residual inversion value represents the behavioral residual signal formed by subtracting the pillow core hysteresis estimate from the original interlayer sensing value. The filled area in the figure corresponds to the difference range between the original interlayer sensing value and the pillow core hysteresis estimate, used to illustrate the influence of pillow core material hysteresis recovery and neighborhood diffusion on the interlayer signal. As can be seen from this figure, this invention does not directly use the original interlayer sensing value to determine sleep state, but first strips away the pillow core hysteresis response and then uses the residual signal as head and neck behavior inversion data.

[0036] In this implementation plan, by comparing the results before and after compensation, the differences in transmission lag, rebound, and residual inversion of each sensing area can be intuitively reflected, providing a data foundation that is consistent in time, uniform in scale, and clear in source for the subsequent construction of head and neck behavior inversion dataset, identification of sleep behavior events, and generation of sleep state assessment results.

[0037] Specifically, the process of generating the head and neck behavior inversion dataset is as follows: The pillow core hysteresis residual inversion value is compared with the behavior inversion threshold. The behavior inversion threshold is derived from the upper quantile of the pillow core hysteresis residual inversion value within the stable bearing segment after the start of the current monitoring. If the number of stable bearing segments in the current monitoring is insufficient, the initial calibration threshold of the device is used. When the pillow core hysteresis residual inversion value is greater than the behavior inversion threshold, the current time window is marked as a valid window for human head and neck behavior. When the pillow core hysteresis residual inversion value is less than or equal to the behavior inversion threshold, the current time window is marked as the pillow core material response dominant window. The pillow core hysteresis residual inversion value is retained, but the time window is not directly used as the basis for judging event types for which trigger features have not yet been defined. The time window marked as the pillow core material response dominant window does not output turning over, abnormal breathing, or awakening events separately, but the pillow core hysteresis residual inversion value within the time window can be used as an input feature for these rules in the calculation, rather than directly determining the occurrence of the event. The data is structured and bound according to a unified analysis time to generate a head and neck behavior inversion dataset.

[0038] like Figure 4The heatmap shown illustrates the inversion of the sensing regions. The pillow core hysteresis residual inversion values ​​for different sensing regions within different analysis time windows are displayed in matrix form. The horizontal axis represents the analysis time window segment number, and the vertical axis represents the sensing region number within the pillow layer. Color levels correspond to the magnitude of the inversion value for each region within its respective time segment. This figure illustrates that the pillow core hysteresis inversion results are not a single time series, but rather possess both temporal and regional distributions. This figure can identify the differences in inversion response between the main head-bearing area, neck support area, and lateral offset area within different time segments, and is used to support the calculation of the effective head and neck support area set, the adjacent sensing region set, and the median of the regional residuals. This figure demonstrates that the pillow layer sensing data, after inversion, forms spatially regionalized sleep observation results.

[0039] In this implementation scheme, by retaining the pillow core material response hysteresis residual inversion value within the dominant window as a rule input feature, event recognition can utilize the change information within the time window while avoiding event output triggered directly by a single time window. By generating a head and neck behavior inversion dataset through structured binding according to a unified analysis time, the inversion values, window labels, and event judgment inputs corresponding to each time window can maintain a one-to-one correspondence, providing a data foundation with clear sources, continuous time sequence, and well-defined judgment boundaries for sleep behavior event recognition, sleep disturbance chain construction, and sleep state assessment.

[0040] Specifically, the process of identifying sleep behavior events and generating candidate sleep perturbation chains based on the head and neck behavior inversion dataset is as follows: Based on the head and neck behavior inversion dataset, events whose absolute value of the first-order difference of net stress exceeds the behavior inversion threshold and whose time window number is greater than M are identified as sleep behavior events; M is the duration window number threshold, with a value of 2; sleep behavior events include neck support descent events, respiratory cycle gap events, effective snoring events, head position fine-tuning events, turning over and recovery events, latent arousal events, pillow departure events, and pillow return events; sleep behavior events are sorted according to a unified analysis time, and the time interval, bearing area number, and event type between adjacent sleep behavior events are read; when the difference in the time window number between sleep behavior events is less than or equal to the interval window number threshold, and the bearing area numbers are the same or the Euclidean distance between the center coordinates of two bearing areas is less than the adjacency distance threshold, the sleep behavior events are classified into the same candidate sleep disturbance chain; when adjacent sleep behavior events do not meet the above conditions, the current candidate sleep disturbance chain is terminated and the next candidate sleep disturbance chain is started.

[0041] A reference time interval is generated based on the median of the time intervals between corresponding events for all samples in the perturbation chain template set. The perturbation chain template set is generated from public datasets or batch user sample statistics. The template field structure includes: event type code, event sequence position, reference time interval between adjacent events, and allowed time window interval range. The event type field of each sleep behavior event in the candidate sleep perturbation chain is read, the event type field is converted into a fixed event type code, and arranged according to the unified analysis time of the event start, to obtain the event type sequence of the candidate sleep perturbation chain. The template event type field in the perturbation chain template is read and arranged according to the template definition order, to obtain the event type sequence of the perturbation chain template. Starting from the first event type code in the event type sequence, each event is counted until the last event type sequence, to obtain the number of events contained in the perturbation chain template.

[0042] In this implementation scheme, candidate sleep disturbance chains are standardized by using event type codes, event sequence positions, reference time intervals between adjacent events, and allowed time window intervals in the disturbance chain template set. This provides a reference for event order and temporal rhythm for calculating the sleep disturbance chain sequence coverage value. By converting the event type field in the candidate sleep disturbance chain into a fixed event type code and arranging them according to the unified analysis time of the event start, a comparable sequence structure between the candidate chain and the template chain can be achieved. By statistically analyzing the number of events contained in the disturbance chain template, a clear denominator can be provided for the normalization calculation of the event sequence coverage item, thereby improving the stability and verifiability of the sleep disturbance chain screening results.

[0043] Specifically, the process of filtering effective sleep disturbance chains and generating a sleep behavior event chain dataset is as follows: Based on the inverted data fragments corresponding to each event in the candidate sleep perturbation chain, the contact state verification value of the events in the candidate sleep perturbation chain is obtained through the contact state evidence verification method. To address the link matching instability caused by disordered event order or missed detections, the longest common subsequence method is adopted. The candidate event type sequence is matched with the template event type sequence, and the longest common subsequence matching is performed between the candidate event type sequence and the template event type sequence. The number of events with consistent order in the two sequences is counted to obtain the length of the common subsequence. The length of the common subsequence is divided by the sum of the total number of events in the template event type sequence and the zero-prevention term to obtain the event order coverage term. The event order coverage term is then calculated for each pair of adjacent events in the candidate sleep perturbation chain. The actual time interval is subtracted from the reference time interval of the corresponding adjacent event in the perturbation chain template to obtain the interval difference between adjacent events. The interval difference between adjacent events is divided by the sum of the corresponding reference time interval and the zero-prevention term, and the absolute value is taken to obtain the time rhythm deviation value. The median of the time rhythm deviation values ​​of all adjacent events is taken, and then the negative value is used for exponential operation to obtain the time rhythm adaptation term. The minimum value of all contact state verification values ​​is taken to obtain the contact state bottleneck term. The event sequence coverage term, the time rhythm adaptation term, and the contact state bottleneck term are multiplied in sequence to obtain the template coverage result. The result can suppress the invalid chain misjudgment caused by the disorder of event sequence or missed detection. The maximum value of the template coverage result is taken to obtain the sleep perturbation chain sequence coverage value.

[0044] The specific formula for calculating the sleep disturbance chain sequence coverage value is as follows: ; In the formula, The sleep disturbance chain sequence coverage value represents the c-th candidate sleep disturbance chain, which is used to characterize the degree of matching between the candidate sleep disturbance chain and the preset sleep disturbance evolution link; c represents the candidate sleep disturbance chain number, which is obtained by sequentially numbering the candidate sleep disturbance chains according to a unified analysis time, and is used to distinguish different candidate sleep disturbance chains formed during this monitoring process. This represents the set of perturbation chain templates, used to provide a sequence matching benchmark for candidate sleep perturbation chains; This represents the template number in the perturbation chain template set, obtained by traversing the perturbation chain template set, and is used to determine the template object currently participating in the matching. This represents the event type sequence of the c-th candidate sleep disturbance chain, used to record the order in which events occur within the candidate sleep disturbance chain; This represents the event type sequence of the p-th perturbation chain template, used as a reference for the order matching of candidate event type sequences; This represents the length of the longest common subsequence between the candidate event type sequence and the template event type sequence, used to measure the degree to which the candidate sleep perturbation chain covers the template event sequence; This represents the number of events contained in the p-th perturbation chain template, obtained by counting the number of events in the template event type sequence, and is used to normalize the length of the longest common subsequence. The time interval between adjacent events in the i-th group in the c-th candidate sleep disturbance chain is obtained by subtracting the end time of the previous event from the start time of the unified analysis of the subsequent event. The value range is a time length greater than or equal to zero, which is used to characterize the actual occurrence rhythm of adjacent events in the candidate sleep disturbance chain. This represents the reference time interval between adjacent events in the i-th group within the p-th perturbation chain template. It is obtained by reading the median value of the time interval between adjacent events of the same type in the effective perturbation chain. When there are insufficient effective perturbation chains, it is obtained by reading the median value of the time interval between adjacent events of the same type in this monitoring. The value range is the time length greater than zero, which is used to provide a reference benchmark for the time rhythm of adjacent events. This indicates the sequence number of adjacent event groups in the event chain. It is obtained by numbering adjacent event pairs in the candidate sleep disturbance chain in chronological order and is used to locate adjacent event groups or event positions that participate in time interval comparison and contact state verification. This represents the contact state verification value of the i-th event in the c-th candidate sleep disturbance chain, used to characterize whether the event has sensor evidence consistent with the user's head and neck contact state; This indicates a zero-prevention term. By setting it to a decimal greater than zero, with a value range of (0, 0.01), it is used to prevent the denominator from being zero in the calculation of the number of template events, the reference time interval, and the proportion, and to improve the stability of the formula calculation.

[0045] The sleep disturbance chain sequential coverage value is compared with the chain formation threshold. The chain formation threshold is derived from the upper quantile of the sequential coverage values ​​of all candidate sleep disturbance chains in this monitoring. If the number of candidate sleep disturbance chains is insufficient, the device's initial calibration threshold is used. When the sleep disturbance chain sequential coverage value is greater than the chain formation threshold, the current candidate sleep disturbance chain is retained as a valid sleep disturbance chain. When the sleep disturbance chain sequential coverage value is less than or equal to the chain formation threshold, the current candidate sleep disturbance chain is split into isolated event fragments to generate a sleep behavior event chain dataset.

[0046] In this implementation scheme, candidate sleep disturbance chains that meet the link coverage conditions are retained as valid sleep disturbance chains, while candidate chains that do not meet the conditions are split into isolated event fragments. This can improve the continuity, credibility, and verifiability of the sleep behavior event chain dataset. As a result, it provides an event chain data foundation with clear order, stable temporal relationship, and clear evidence source for sleep state segmentation, sleep abnormality cause label generation, and sensor evidence chain binding.

[0047] Specifically, the process of constructing a set of state features based on the sleep behavior event chain dataset is as follows: Based on the sleep behavior event chain dataset, a state feature set is constructed using the time window feature aggregation method. Based on the state feature set, the completeness of the sensor evidence is obtained by dividing the number of complete evidence items in the set by the total number of required evidence items. The evidence item set is predefined as: whether the net pressure value meets the contact judgment threshold, whether the vibration frequency band energy exceeds the noise floor, and whether the contact state verification value is a valid value, for a total of three items. The number of complete evidence items is the number of items that actually meet the above conditions, and the total number of required evidence items is 3.

[0048] The number of consecutively satisfied windows is obtained through a continuous window backtracking method. Starting from the analysis time window, the actual values ​​of key features corresponding to candidate sleep states are read window by window and compared with the feature conditions corresponding to the state. The number of analysis time windows that continuously satisfy the feature conditions is accumulated, and backtracking stops when an analysis time window that does not satisfy the feature conditions is encountered, thus obtaining the number of satisfied windows. The number of analysis time windows for all candidate state segments identified in the current night under the same sleep state type is counted, and the median is taken as the persistence scale. The statistical scope includes all state segments identified and marked as this sleep state in this monitoring, excluding monitoring data. The feature numbers used to characterize the state are used to form a key feature set. Data items matching the feature numbers within the analysis time window are read, and maximum-minimum normalization is performed to obtain the actual values ​​of the state features. The maximum and minimum values ​​used for maximum-minimum normalization are taken as the theoretical value boundaries of the features. The theoretical maximum value of the net pressure is the full-scale pressure value, and the minimum value is 0. A reference center value is obtained through state-based statistical methods. The absolute value of the difference between each actual value and the reference center value is calculated, and the median of all absolute differences is taken to obtain the scale value.

[0049] Divide the number of satisfying windows by the sum of the duration scale and the zero-prevention term to obtain the state duration ratio; take the negative value of the state duration ratio and perform an exponential operation, then subtract the result of the exponential operation from the value to obtain the fragment fitting term; subtract the reference center value from the actual value of the state feature and take the absolute value of the subtraction result to obtain the deviation distance; divide the deviation distance by the sum of the scale value and the zero-prevention term to obtain the scaled deviation value; exponentiate the scaled deviation value according to the deviation sensitivity coefficient, add one and take the reciprocal to obtain the feature fitting term; take the minimum value of the feature fitting terms of all state features in the key feature set to obtain the core fitting term; multiply the sensor evidence completeness, fragment fitting term and core fitting term in sequence to obtain the sleep state fragment evidence fitting value.

[0050] The specific formula for calculating the fit value of sleep state fragment evidence is as follows: ; In the formula, The evidence fit value of the sleep state segment corresponding to the candidate sleep state s in the t-th analysis time window is used to characterize the degree of evidence matching between the current analysis time window and the candidate sleep state s. This represents any sleep state in the candidate sleep state set, used to determine the state category corresponding to the current formula calculation; This indicates the current analysis time window number, which is obtained by slicing the pressure sequence, microvibration sequence, and sound sequence according to the analysis time window. The value range is positive integer, and it is used to identify the time position corresponding to the current sleep state segment evidence fit value. It represents the completeness of sensor evidence for state s in the t-th analysis time window, and is used to characterize the completeness of the event chain evidence in the current analysis time window; This represents the number of windows that meet the criteria, used to characterize the temporal continuity of candidate sleep states; The duration of state s is represented and used as a time scale reference for the segment duration fit. This represents the set of key features corresponding to state s, used to limit the state features participating in the adaptation calculation of the current candidate sleep state; This represents the j-th state feature number in the key feature set, which is obtained by traversing the key feature set corresponding to state s, and is used to locate the specific state feature involved in the deviation calculation. This represents the actual value of the j-th state feature within the t-th analysis time window. It is used to represent the actual state performance corresponding to the continuous value of the respiratory cycle, the continuous value of neck support, the stable value of the center of pressure, the effective snoring density, the number of low-amplitude continuous disturbances, the time to turn over and recover, or the sequence coverage value of the sleep disturbance chain in the current analysis time window. This represents the reference center value of the j-th state feature under state s, which is used as a benchmark for comparing the degree of deviation of the current state feature. This represents the scale value of the j-th state feature under state s, used to scale the difference between the current state feature and the reference center value. The deviation sensitivity coefficient of the j-th state feature under state s is obtained through the state discrimination mapping method. The value range is a real number greater than zero. It is used to control the weakening of the fit value of the sleep state fragment evidence after the j-th state feature deviates from the reference center value. This represents the zero-prevention term, which is pre-defined and has a value range of (0, 0.01). It is used to prevent the denominator from being zero and to ensure the stability of the calculation of the proportional and exponential terms in the formula.

[0051] like Figure 5The multi-state evidence fit curves shown illustrate that stable sleep, respiratory disturbance, latent arousal, postural adjustment, and head-off states each have corresponding sleep state fragment evidence fit values ​​within their respective analysis time windows. This curve illustrates that the state assessment and evidence binding module calculates fit results for each candidate sleep state separately, rather than directly outputting sleep conclusions based on a single sensor channel. The fit values ​​for stable sleep states are jointly constrained by continuous respiratory cycle values, stable pressure center values, continuous neck support values, and effective snoring density; the fit values ​​for respiratory disturbance states are jointly constrained by respiratory cycle gaps, effective snoring density, sound contact verification values, and sleep disturbance chain sequence coverage values; the fit values ​​for latent arousal states are jointly constrained by the number of low-amplitude continuous disturbances, the number of pressure center round trips, the number of neck support interruptions, and the respiratory cycle reconstruction results; the fit values ​​for postural adjustment states are jointly constrained by the pressure center migration distance, turning recovery time, new bearing area stability value, and respiratory cycle recovery results; and the fit values ​​for head-off states are jointly constrained by the duration of head and neck bearing loss and the head-off pressure establishment time.

[0052] This implementation scheme can quantify the degree of matching between the current time window and stable sleep state, respiratory disturbance state, latent arousal state, posture adjustment state, and pillow-off state; avoid abnormally increasing the fit value of candidate states by a single feature; intuitively distinguish the fit changes of different sleep states within each analysis time window, and provide stable and traceable data basis for forming sleep state segments, generating sleep state assessment reports, and binding sensor evidence chains.

[0053] Specifically, the process of generating a sleep state assessment report and linking it to the sensor evidence chain is as follows: The evidence fit values ​​of sleep state segments within each analysis time window for each candidate sleep state are read according to a unified analysis time. Analysis time windows in which the evidence fit values ​​of sleep state segments for the same candidate sleep state meet the corresponding state threshold and are consecutively satisfied for the corresponding number of confirmation windows are merged into the same state segment. The first analysis time window in the state segment that meets the conditions is taken as the segment start point, and the last analysis time window that consecutively meets the conditions is taken as the segment end point, forming a sleep state segment. The unified analysis time of the first sleep segment is determined as the sleep onset start point. The start and end times, duration, and number of segments of each type of sleep state segment are counted respectively. The evidence fit values ​​of sleep state segments, continuous values ​​of respiratory cycles, stable values ​​of pressure centers, continuous values ​​of neck support, and effective snoring density of stable sleep states are read. When the above features meet the feature conditions corresponding to stable sleep states and are consecutively satisfied within the stable sleep confirmation window, the corresponding analysis time windows are merged into stable sleep segments.

[0054] Read the sleep state fragment evidence fit value, respiratory cycle continuity value, effective snoring density, sound contact state verification value, and sleep disturbance chain sequence coverage value for the respiratory disturbance state. When the above features meet the feature conditions corresponding to the respiratory disturbance state and are continuously established within the respiratory disturbance confirmation window, the corresponding analysis time window is merged into a respiratory disturbance fragment.

[0055] Read the evidence fit values ​​of sleep state segments in the latent wakefulness state, the number of low-amplitude continuous perturbations, the number of pressure center round trips, the number of neck support interruptions, and the respiratory cycle reconstruction results; when the above features meet the feature conditions corresponding to the latent wakefulness state and are continuously established within the latent wakefulness confirmation window, the corresponding analysis time windows are merged into latent wakefulness segments.

[0056] Read the sleep state fragment evidence fit value, pressure center migration distance, turning recovery time, new bearing area stability value and respiratory cycle recovery result of the posture adjustment state; when the above features meet the feature conditions corresponding to the posture adjustment state and are continuously true within the posture adjustment confirmation window, the corresponding analysis time window is merged into the posture adjustment fragment.

[0057] Read the evidence fit values ​​of sleep state segments in the occipital separation state, the pressure values ​​of each head and neck bearing area, the duration of head and neck bearing disappearance, and the establishment time of occipital pressure. When the above features meet the feature conditions corresponding to the occipital separation state and are continuously established within the occipital separation confirmation window, the corresponding analysis time windows are merged into occipital separation segments.

[0058] The starting time of the first sleep segment was determined as the sleep onset point. The durations of stable sleep segments, respiratory disturbance segments, and posture adjustment segments were summed to obtain the effective sleep duration. The start and end times, durations, and number of segments for each type of sleep state were counted. Based on the ranking of the number of respiratory disturbance segments, the number of latent arousal segments, the number of neck support descent events, the number of effective snoring events, and the number of sleep disturbance segments, a predefined set of sleep abnormality cause labels was used. This set of sleep abnormality cause labels must include at least the following: snoring-related, sleep disturbance-related, frequent turning over-the-body-related, respiratory abnormality-related, and arousal-related. Based on the ranking of the number of respiratory disturbance segments, the number of latent arousal segments, the number of neck support descent events, the number of effective snoring events, and the number of sleep disturbance segments, the top three event types were selected to generate sleep abnormality cause labels and a sleep state assessment report.

[0059] Read the start and end points of each sleep state segment, index them in the head and neck behavior inversion dataset and the sleep behavior event chain dataset, and structurally bind the indexed sensor evidence with the corresponding sleep state segment number. The bound fields include at least: segment number, start and end unified analysis time, dataset index path of the sensor evidence, associated event type number and state feature number, and generate a sensor evidence chain.

[0060] In this implementation plan, each sleep state segment can be structurally bound to the corresponding inverted pressure segment, inverted vibration segment, sound contact state verification result, and disturbance chain evidence, so that the sleep onset point, effective sleep duration, state segment, and abnormal cause label in the sleep state assessment report all have traceable data sources.

[0061] Specifically, the second aspect of this invention provides a sleep state assessment system based on time-series data analysis, applied to a sleep state assessment method based on time-series data analysis, comprising: a pillow interlayer time-series acquisition and preprocessing module, used to acquire, linearly normalize, and reconstruct time windows for interlayer sensing data during the user's pillow-lying process, aligning the sampling sequence number and reception time with a fixed window length to generate a pillow interlayer standard observation dataset; and a pillow core hysteresis inversion module, used to estimate the pillow core hysteresis response based on the pillow interlayer standard observation dataset, and compensate for the original residual using a nearby hysteresis reference value, thereby improving the pillow core hysteresis response. The residuals are inverted to generate a head and neck behavior inversion dataset; the sleep event chain construction module is used to identify sleep behavior events based on the head and neck behavior inversion dataset, generate candidate sleep disturbance chains according to time intervals and spatial adjacency conditions of the carrying area, filter valid sleep disturbance chains, and generate a sleep behavior event chain dataset; the state assessment and evidence binding module is used to construct a set of state features based on the sleep behavior event chain dataset, generate a sleep state assessment report by comparing the fragment evidence fit value with the state threshold, and structurally bind sensor evidence with state fragment numbers and time sequence information, and bind the sensor evidence chain.

[0062] In this implementation plan, through the coordinated setup of the interlayer time-series acquisition and preprocessing module, the pillow core hysteresis inversion module, the sleep event chain construction module, and the state assessment and evidence binding module, the pressure signals, vibration signals, and sound signals during the user's pillow-laying process can be standardized into data input within a unified analysis time window.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0064] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A sleep state assessment method based on time-series data analysis, characterized in that, Includes the following steps: S1 collects, normalizes, and reconstructs time windows for the interlayer sensing data during the user's pillow-lying process, generating a standard observation dataset of pillow interlayer. S2, based on the standard observation dataset of pillow interlayer, estimates the pillow core hysteresis response, inverts the pillow core hysteresis residual, and generates a head and neck behavior inversion dataset; S3, based on the head and neck behavior inversion dataset, identifies sleep behavior events, generates candidate sleep disturbance chains, filters valid sleep disturbance chains, and generates a sleep behavior event chain dataset; S4, based on the sleep behavior event chain dataset, constructs a set of state features, generates a sleep state assessment report, and binds a sensor evidence chain.

2. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process of collecting, normalizing, and reconstructing the time window of the interlayer sensing data during the user's pillow-lying process to generate a standard observation dataset of the pillow interlayer is as follows: Pressure sensing units, vibration sensing units, and sound acquisition units are deployed within the interlayer of the smart pillow; data is collected during the user's lying down process to form raw interlayer sensing data; the raw interlayer sensing data includes pressure sensing area number, raw pressure sampling value, sensor sampling time, pillow end receiving time, and sampling sequence number; idle data is collected when the user is not lying down and a set of disturbance chain templates is obtained to form a pressure idle baseline; The net pressure value is obtained by subtracting the corresponding pressure idle baseline from the original pressure sample value of each pressure sensing area. When the absolute value of the difference between the net pressure value and the previous analysis time window is greater than the pressure jump threshold, the net pressure value is replaced with the median value of the net pressure values ​​of adjacent analysis time windows before and after the current analysis time window, and a jump correction flag is generated. When there is a gap in the sensor sampling sequence number, a missing sampling flag is generated in the corresponding analysis time window and the analysis time window position is retained. An analysis time window is established based on the sampling sequence number and the pillow end reception time. After the net pressure value is mapped to the analysis time window, the original data of the interlayer sensing is linearly normalized to generate a standard observation dataset of the pillow interlayer.

3. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process for estimating the pillow core hysteresis response based on the standard observation dataset of pillow interlayer is as follows: Based on the standard observation dataset of pillow interlayer, the sensing areas whose net pressure value meets the contact judgment threshold are identified as effective head and neck bearing areas. The area numbers of all effective head and neck bearing areas are compiled into a set, resulting in the set of sensing areas in the effective head and neck bearing state within the current time window. The uniquely bound sensing channel is directly indexed by the sensing area number, and the sampled value of the channel within the time window is read as the interlayer sensing value. Based on the spatial coordinates of each sensing area within the pillow interlayer, the Euclidean distance between the center coordinates of the areas is calculated according to the rule that the Euclidean distance between the center coordinates of the areas is less than the adjacency distance threshold. Areas whose Euclidean distance meets the adjacency distance threshold are identified as neighboring sensing areas. The area numbers of all neighboring sensing areas are... A set of neighboring sensing regions is formed to constitute the sensing region. Based on the number of propagation hysteresis windows in the sensing region, the interlayer sensing values ​​of each neighboring sensing region within the set are read back within a time window. The median of the read-back interlayer sensing values ​​is taken to obtain the neighboring hysteresis reference value of the sensing region. The original pressure sampling values ​​are processed using first-order difference. The time window in which the absolute value of the first-order difference satisfies the pressure jump threshold is determined as the pressure jump point. Within the hysteresis search window after the pressure jump point, the interlayer response sequence of the sensing region is read. The time window in which the interlayer response amplitude reaches its peak value is determined as the interlayer response peak time window. The difference between the interlayer response peak time window number and the pressure jump point time window number is determined as the number of propagation hysteresis windows in the sensing region.

4. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process for inverting the hysteresis residual of the pillow core is as follows: The median of the interlayer sensing values ​​from the previous and two previous time windows, along with the nearest hysteresis reference values, is taken within the same group to obtain the comprehensive hysteresis reference value. The time interval between adjacent time windows is divided by the sum of the pillow core rebound time constant and the anti-zero term to obtain the rebound time ratio. The negative value of the rebound time ratio is then used for exponential operation to obtain the pillow core rebound attenuation coefficient. The comprehensive hysteresis reference value is multiplied by the pillow core rebound attenuation coefficient to obtain the pillow core hysteresis estimate. The interlayer sensing value within the current time window is subtracted from the pillow core hysteresis estimate, and the absolute value of the subtraction is taken to obtain the absolute value of the hysteresis residual. The absolute value of the hysteresis residual is divided by the sum of the effective amplitude reference and the anti-zero term to obtain the scaled hysteresis residual. The lag residual is scaled up by one and then the natural logarithm is taken to obtain the lag residual compression term; The median of the hysteresis residual compression term for all effective head and neck bearing sensing regions within the current time window is taken to obtain the pillow core hysteresis residual inversion value.

5. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process for generating the head and neck behavior inversion dataset is as follows: The pillow core hysteresis residual inversion value is compared with the behavior inversion threshold. When the pillow core hysteresis residual inversion value is greater than the behavior inversion threshold, the current time window is marked as an effective window for human head and neck behavior. When the pillow core hysteresis residual inversion value is less than or equal to the behavior inversion threshold, the current time window is marked as the dominant window for pillow core material response. The pillow core hysteresis residual inversion value is retained, and the time window is not directly used as the basis for judging event types that have not yet defined trigger features. The head and neck behavior inversion dataset is generated by structurally binding according to a unified analysis time.

6. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process of identifying sleep behavior events and generating candidate sleep perturbation chains based on the head and neck behavior inversion dataset is as follows: Based on the head and neck behavior inversion dataset, events whose absolute value of the first-order difference of net stress exceeds the behavior inversion threshold and whose time window number is greater than M are identified as sleep behavior events. Sleep behavior events are sorted according to a unified analysis time, and the time interval, carrier zone number, and event type between adjacent sleep behavior events are read. When the difference in the number of time windows between sleep behavior events is less than or equal to the interval window number threshold, and the carrier zone numbers are the same or the Euclidean distance between the center coordinates of two carrier zones is less than the adjacency distance threshold, the sleep behavior events are grouped into the same candidate sleep disturbance chain. When adjacent sleep behavior events do not meet the above conditions, the current candidate sleep disturbance chain is terminated and the next candidate sleep disturbance chain is started. A reference time interval is generated based on the median of the time intervals between corresponding events for all samples in the perturbation chain template set. Read the event type field of each sleep behavior event in the candidate sleep disturbance chain, convert the event type field into a fixed event type code, and arrange them according to the unified analysis time of the event start to obtain the event type sequence of the candidate sleep disturbance chain; read the template event type field in the disturbance chain template and arrange it according to the template definition order to obtain the event type sequence of the disturbance chain template; count each event starting from the first event type code in the event type sequence until the last event type sequence to obtain the number of events contained in the disturbance chain template.

7. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process for filtering valid sleep perturbation chains and generating a sleep behavior event chain dataset is as follows: Based on the inverted data fragments corresponding to each event in the candidate sleep perturbation chain, the contact state verification value of the events in the candidate sleep perturbation chain is obtained through the contact state evidence verification method; the longest common subsequence matching is performed between the candidate event type sequence and the template event type sequence, and the number of events with the same order in the two sequences is counted to obtain the length of the common subsequence; the length of the common subsequence is divided by the sum of the total number of events in the template event type sequence and the zero-prevention term to obtain the event order coverage term; The actual time interval of each group of adjacent events in the candidate sleep perturbation chain is subtracted from the reference time interval of the corresponding adjacent events in the perturbation chain template to obtain the adjacent event interval difference. The adjacent event interval difference is divided by the sum of the corresponding reference time interval and the zero-prevention term, and the absolute value is taken to obtain the time rhythm deviation value. The median of the time rhythm deviation values ​​of all adjacent events is taken, and then the negative value is used for exponential operation to obtain the time rhythm adaptation term. The minimum value of all contact state verification values ​​is taken to obtain the contact state weakness term. The event sequence coverage term, the time rhythm adaptation term, and the contact state weakness term are multiplied in turn to obtain the template coverage result. The maximum value of the template coverage result is taken to obtain the sleep perturbation chain sequence coverage value. The sleep disturbance chain sequence coverage value is compared with the chain formation threshold. When the sleep disturbance chain sequence coverage value is greater than the chain formation threshold, the current candidate sleep disturbance chain is retained as a valid sleep disturbance chain. When the sleep disturbance chain sequence coverage value is less than or equal to the chain formation threshold, the current candidate sleep disturbance chain is split into isolated event fragments to generate a sleep behavior event chain dataset.

8. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process of constructing the state feature set based on the sleep behavior event chain dataset is as follows: Based on the sleep behavior event chain dataset, a set of state features is constructed using the time window feature aggregation method. Based on the set of state features, the number of complete evidence items in the set is divided by the total number of required evidence items to obtain the completeness of the sensor evidence. The number of consecutively satisfied windows is obtained by using the continuous window backtracking method. The actual values ​​of key features corresponding to candidate sleep states are read window by window from the analysis time window backward and compared with the feature conditions corresponding to the state. The number of analysis time windows that continuously meet the feature conditions is accumulated. Backtracking stops when an analysis time window that does not meet the feature conditions is encountered, and the number of satisfying windows is obtained. The number of analysis time windows of all candidate state segments identified in the current night under the same sleep state type is counted, and the median is taken as the persistence scale. The feature numbers used to represent the state are used to form a key feature set. The data items that match the feature numbers in the analysis time window are read, and the maximum and minimum normalization processes are performed to obtain the actual values ​​of the state features. The reference center value is obtained through the statistical method within the state. The absolute value of the difference between each actual value and the reference center value is calculated, and the median of all the absolute values ​​of the difference is taken to obtain the scale value. Divide the number of satisfying windows by the sum of the duration scale and the zero-prevention term to obtain the state duration ratio; take the negative value of the state duration ratio and perform an exponential operation, then subtract the result of the exponential operation from the value to obtain the segment fitting term; subtract the reference center value from the actual value of the state feature, and take the absolute value of the subtraction result to obtain the deviation distance; divide the deviation distance by the sum of the scale value and the zero-prevention term to obtain the scaled deviation value. The scaled deviation value is exponentially calculated according to the deviation sensitivity coefficient, then one is added and the reciprocal is taken to obtain the feature fitting term. The minimum value of the feature fit terms for all state features within the key feature set is taken to obtain the core fit term; the completeness of sensor evidence, the fragment fit term, and the core fit term are multiplied in sequence to obtain the fit value of sleep state fragment evidence.

9. The sleep state assessment method based on time-series data analysis according to claim 1, characterized in that: The specific process of generating a sleep state assessment report and binding it with a sensor evidence chain is as follows: The evidence fit values ​​of sleep state segments within each analysis time window for each candidate sleep state are read according to a unified analysis time. Analysis time windows in which the evidence fit values ​​of sleep state segments of the same candidate sleep state meet the corresponding state threshold and are consecutively satisfied with the corresponding number of confirmation windows are merged into the same state segment. The first analysis time window in the state segment that meets the conditions is taken as the segment start point, and the last analysis time window that consecutively meets the conditions is taken as the segment end point, thus forming a sleep state segment. The unified analysis time of the first sleep segment is determined as the sleep onset start point. The start and end times, duration, and number of segments of each type of sleep state segment are counted respectively, and sleep abnormality cause labels and sleep state assessment reports are generated. The start and end points of each sleep state segment are read and indexed in the head and neck behavior inversion dataset and the sleep behavior event chain dataset. The sensor evidence obtained from the index is structurally bound to the corresponding sleep state segment number to generate a sensor evidence chain.

10. A sleep state assessment system based on time-series data analysis, employing the sleep state assessment method based on time-series data analysis as described in any one of claims 1-9, characterized in that, include: The interlayer time series acquisition and preprocessing module is used to acquire, normalize, and reconstruct time windows of interlayer sensing data during the user's pillow-lying process, and generate a standard observation dataset of pillow interlayer. The pillow core hysteresis inversion module is used to estimate the pillow core hysteresis response based on the standard observation dataset of pillow interlayer, invert the pillow core hysteresis residual, and generate a head and neck behavior inversion dataset. The sleep event chain construction module is used to identify sleep behavior events, generate candidate sleep disturbance chains, filter valid sleep disturbance chains, and generate a sleep behavior event chain dataset based on the head and neck behavior inversion dataset. The state assessment and evidence binding module is used to construct a set of state features based on the sleep behavior event chain dataset, generate a sleep state assessment report, and bind the sensor evidence chain.

Images