Smart watch fragmentation sleep precision identification method based on multi-source sensor fusion

Abstract

The application discloses a kind of intelligent watch fragmentation sleep precision identification method based on multi-source sensor fusion.Through the collection heart rate variability, wrist temperature and micro-motion data, the physiological characteristic phase adaptive boundary search algorithm is used to construct the continuous multi-dimensional physiological characteristic sequence without fixed time window;Combined with sympathetic nervous active level matching and short-time mutation sensitive asymmetric hidden Markov decoding, high-precision physiological state transition path is generated;Then through micro-sleep cycle mapping and fifteen-minute duration screening, effective sleep segments are aggregated, and the fragmentation sleep identification result is output.The method breaks through the limitation of traditional fixed window, significantly improves the recognition accuracy and clinical reference value of complex and discontinuous sleep patterns in real-life scenarios.

Description

Technical Field

[0001] This invention relates to the technical field of sleep recognition, and in particular to a method for accurate recognition of fragmented sleep in smartwatches based on multi-source sensor fusion. Background Technology

[0002] In recent years, with the increasing awareness of health management and chronic disease prevention, sleep quality monitoring has become one of the core functions of wearable devices. Smartwatches, with their lightweight, unobtrusive wearing experience and all-day data collection capabilities, have become an ideal platform for home sleep monitoring. However, the sleep recognition technology currently mainstream in the smartwatch market still has significant limitations, making it difficult to meet the need for accurate assessment of fragmented and discontinuous sleep patterns.

[0003] Traditional sleep tracking methods primarily rely on accelerometers to capture user body movement information and combine this with simple threshold rules to determine sleep and wakefulness. These methods typically segment the raw signal using a fixed-length time window and then classify the state using pre-defined rules or shallow machine learning models. This single-modal analysis based on body movement ignores the complex physiological changes that occur during sleep, particularly failing to reflect the dynamic regulatory characteristics of the autonomic nervous system at different sleep stages. For example, during REM sleep, although the body is at rest, heart rate and respiration fluctuate irregularly; while in deep sleep, physiological indicators such as decreased body temperature and reduced heart rate variability are highly specific. Relying solely on body movement data easily leads to misclassifying such resting but non-sleep states as sleep, or mistaking brief awakenings for continuous sleep, resulting in distorted staging results.

[0004] While some high-end devices have begun to incorporate optical heart rate sensors to obtain heart rate variability data, existing fusion strategies remain crude. Most systems simply stitch together data from multiple sensors and input it into a fixed-structure classifier, failing to adequately consider the asynchronicity and dynamic coupling of different physiological signals over time. More importantly, these methods generally employ the fixed-cycle segmentation used in medical polysomnography (PSG), forcibly dividing continuous physiological processes into segments of equal length. This rigid windowing mechanism cannot adapt to individual differences and sudden changes in nocturnal physiological states. Especially when dealing with fragmented sleep scenarios such as afternoon naps, multiple nighttime awakenings, or short naps, these segments are often discarded entirely due to insufficient duration, severely underestimating the user's actual effective sleep volume.

[0005] Furthermore, existing algorithms often employ standard Hidden Markov Models (HMMs) when modeling sleep state transitions. These models typically assume a symmetric or stationary distribution for their state transition probability matrices, neglecting the asymmetry and abrupt sensitivity of sleep stage transitions. For example, the probability of suddenly transitioning from deep sleep to wakefulness is much higher than the reverse transition, and micro-arousals, which often last only a few seconds, have a significant impact on sleep continuity. Traditional models lack the ability to model such short-duration, high-frequency state jumps, resulting in smooth transitions in hidden state decoding and the loss of crucial physiological details.

[0006] A deeper problem lies in the fact that current technologies have failed to establish a refined mapping relationship between multi-source physiological signals and the level of autonomic nervous system activity. The dynamic balance between the sympathetic and parasympathetic nervous systems is the core mechanism regulating the sleep-wake cycle, and heart rate variability, wrist skin temperature, and micro-movement intensity are precisely the external manifestations of this balance. However, current methods mostly use these signals as independent features, failing to extract their co-evolutionary patterns from dynamic dimensions such as phase changes and transient slopes, and failing to adaptively adjust feature boundaries according to physiological states, thus limiting the model's generalization ability.

[0007] In summary, the industry urgently needs a novel sleep recognition method that can overcome the constraints of fixed time windows, deeply integrate multi-source dynamic physiological features, accurately model the process of sudden changes in sleep state, and effectively identify short-term effective sleep fragments. This invention addresses these pain points by proposing a fragmented sleep accuracy recognition system based on multi-source sensor fusion and adaptive physiological boundary search, aiming to significantly improve the accuracy and clinical reference value of sleep monitoring in real-life scenarios using smartwatches. Summary of the Invention

[0008] In view of this, the present invention provides a method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion. The purpose is to solve the problems of low identification accuracy, rigid time window, and insufficient state transition modeling in existing smartwatch sleep recognition technologies when dealing with fragmented sleep. By constructing a continuous multi-dimensional physiological feature sequence that is not limited by a fixed time window and introducing a state decoding mechanism sensitive to short-term mutations, the present invention aims to achieve refined perception and accurate staging of complex sleep patterns in real-life scenarios, thereby providing users with sleep quality assessment results that have more clinical reference value.

[0009] To achieve the above objectives, this invention provides a method for accurate identification of fragmented sleep patterns in smartwatches based on multi-source sensor fusion, comprising the following steps: S1: Based on heart rate variability data, wrist temperature data and micro-motion data collected by smartwatches, the transient change slope of each type of data over time is extracted, and the physiological feature phase adaptive boundary search algorithm is used to determine the physiological feature change boundary, forming a continuous multidimensional physiological feature sequence that is not constrained by a fixed time window. S2: Based on the continuous multidimensional physiological feature sequence without fixed time window restrictions, perform autonomic nerve activity level matching on the multidimensional physiological features at each sampling time, calculate the observation state emission probability corresponding to the respective autonomic nerve activity level at that sampling time, and generate an initial state probability distribution sequence. S3: Based on the initial state probability distribution sequence, an asymmetric hidden Markov decoding algorithm sensitive to short-term mutations is introduced to decode the state transition process between adjacent sampling times, determine the optimal hidden link under the condition of rapid switching between states, and obtain the physiological state transition path sequence. S4: Based on the physiological state transition path sequence, perform microsleep cycle mapping classification on the continuous hidden physiological states, and convert and reconstruct the hidden physiological states into wakefulness, light sleep, deep sleep and REM sleep stages to obtain a refined sleep stage segment sequence. S5: Screen the duration of the refined sleep stage fragment sequence, and aggregate the sleep stage fragments with a duration of 15 minutes or more according to the temporal continuity to obtain the fragmented sleep recognition result corresponding to the data collected by the smartwatch.

[0010] As a further improvement of the present invention: Optionally, step S1 further includes: A multi-source sensor data acquisition channel for a smartwatch is established, comprising a heart rate variability acquisition channel, a wrist temperature acquisition channel, and a micro-motion data acquisition channel; the heart rate variability acquisition channel outputs a heart rate variability sequence, the wrist temperature acquisition channel outputs a wrist temperature sequence, and the micro-motion data acquisition channel outputs a micro-motion data sequence. Define a unified time index set as follows: ; in, This represents the unified time index set corresponding to the data acquisition channels of the multi-source sensors in a smartwatch. This represents the last sampling time in the unified time index set. Represents any sampling time in the unified time index set. ; At any moment At this point, the sampled values ​​of the heart rate variability sequence are represented as The sampled values ​​of the wrist temperature sequence are represented as follows: The sampled values ​​of the micro-motion data sequence are represented as ; Time alignment processing is performed on the heart rate variability sequence, wrist temperature sequence, and micro-motion data sequence. The time alignment processing includes unifying the sampling time, removing duplicate sample values, and filling in missing sample values. The missing sample value filling is performed by linear interpolation of adjacent valid sample values. The completed heart rate variability sequence, wrist temperature sequence, and micro-motion data sequence all correspond to the unified time index set. Transient change slopes were calculated for the time-aligned heart rate variability sequence, wrist temperature sequence, and micromotion data sequence, respectively. These transient change slopes included the heart rate variability transient change slope, wrist temperature transient change slope, and micromotion transient change slope. The formula for calculating the transient change slope is as follows: ; in, Indicates the sensor data type identifier. , This represents a data type representing heart rate variability. This represents the data type of wrist temperature. This represents a micro-motion data type; Indicates time Corresponding to sensor data type identifier The sampled value; when hour, , Indicates time The slope of transient changes in heart rate variability; when hour, , Indicates time The slope of the transient change in wrist body temperature; when hour, , Indicates time The slope of the transient change of micro-motion; Indicates time The previous sampling time corresponds to the sensor data type identifier. The sampled values; for The sampling time is set. ; Build Time Transient slope vector: ; in, Indicates time The transient change slope vector, wherein the three components of the transient change slope vector are, in order, the heart rate variability transient change slope, the wrist body temperature transient change slope, and the micromotion transient change slope; The transient slope vector is searched for boundaries using a physiological feature phase adaptive boundary search algorithm. The physiological feature phase adaptive boundary search algorithm includes parallel stable phase boundary search paths, transitional phase boundary search paths, and abrupt phase boundary search paths. The stable phase boundary search path is used to handle sampling moments where the slopes of heart rate variability transient changes, wrist body temperature transient changes, and micro-motion transient changes do not exceed their respective stable slope thresholds; the transitional phase boundary search path is used to handle sampling moments where one or both of the slopes of heart rate variability transient changes, wrist body temperature transient changes, and micro-motion transient changes exceed their corresponding stable slope thresholds; the abrupt phase boundary search path is used to handle sampling moments where the slopes of heart rate variability transient changes, wrist body temperature transient changes, and micro-motion transient changes all exceed their corresponding stable slope thresholds. In the stable phase boundary search path, the current sampling time is incorporated into the previous continuous physiological feature segment; in the transitional phase boundary search path, the current sampling time is recorded as the candidate boundary time, and the consistency of the slope direction before and after the candidate boundary time is verified; in the abrupt phase boundary search path, the current sampling time is determined as the phase boundary time, and a new continuous physiological feature segment is established from the phase boundary time. The slope direction consistency verification includes: when the transient change slope sign corresponding to the same sensor data type identifier is consistent in the sampling time before the candidate boundary time, the candidate boundary time, and the sampling time after the candidate boundary time, the candidate boundary time is retained as the phase boundary time; when the transient change slope sign is not consistent, the candidate boundary time is merged into the previous continuous physiological feature segment. Based on the phase boundary time, the heart rate variability sampling values, wrist temperature sampling values, micromotion intensity values, heart rate variability transient change slope, wrist temperature transient change slope, and micromotion transient change slope are combined into a continuous multidimensional physiological feature sequence; time The continuous multidimensional physiological feature vector is represented as: ; in, Indicates time The continuous multidimensional physiological feature vectors are obtained; the continuous multidimensional physiological feature sequence without fixed time window restrictions is formed by the continuous multidimensional physiological feature vectors of all sampling times in the order of sampling times, and the continuous multidimensional physiological feature sequence without fixed time window restrictions is output.

[0011] Optionally, step S2 further includes: The set of autonomic nervous system activity levels is defined as follows: ; in, This represents a set of autonomic nervous system activity levels. Indicates a high level of sympathetic activity. Indicates the level of mixed sympathetic and parasympathetic activity. Indicates the level of higher parasympathetic activity. Indicates a level of rapid fluctuation in activity; For each moment Continuous multidimensional physiological feature vectors Calculate the observed state emission probability for each corresponding principal nerve activity level in the set of autonomic nerve activity levels; the formula for calculating the observed state emission probability is: ; in, Indicates time Continuous multidimensional physiological feature vectors Corresponding to the Individual autonomic nervous system activity levels The observed emission probability, ; Represents an exponential function with the natural constant as its base; Indicates time Continuous multidimensional physiological feature vectors With the Individual autonomic nervous system activity levels The standardized distance between the hierarchical feature centers; This represents the level index in the set of autonomic nervous system activity levels. ; Indicates time Continuous multidimensional physiological feature vectors With the Individual autonomic nervous system activity levels The standardized distance between the hierarchical feature centers; The standardized distance Calculate according to the following formula: ; in, The dimension index represents a continuous multidimensional physiological feature vector. ; Indicates time Continuous multidimensional physiological feature vectors The One component; Indicates the first Individual autonomic nervous system activity levels The hierarchical feature center in the first Central values ​​in each dimension; Indicates the first The distance weight coefficients for each dimension are equal to one. Time The emission probabilities of the four observed states are arranged in the order of their levels in the set of autonomic nervous system activity levels to form the initial state probability vector: ; in, Indicates time The initial state probability vector is obtained; the initial state probability distribution sequence is formed by the initial state probability vectors of all sampling times in the order of sampling times, and the initial state probability distribution sequence is output.

[0012] Optionally, step S3 further includes Establish a set of hidden physiological states: ; in, Represents the set of hidden physiological states. Indicates a state of being awake or in a candidate state. This indicates a candidate state of light sleep. Indicates a candidate state for deep sleep. This indicates a candidate state for rapid eye movement (REM). Indicates a short-term mutation candidate state; Construct the asymmetric state transition matrix: ; in, Represents the asymmetric state transition matrix; Indicates from the first A hidden physiological state Transfer to the A hidden physiological state The state transition probability, , ;when hour, and They have different values; the sum of the five state transition probabilities in the same row is one; The initial state probability vector is mapped to the observation probability of the hidden physiological state; for the candidate states of wakefulness, light sleep, deep sleep, and rapid eye movement, the observation state emission probabilities of high sympathetic activity level, mixed sympathetic and parasympathetic activity level, high parasympathetic activity level, and rapid fluctuation activity level are used as the observation probabilities, respectively; for the candidate states of short-term abrupt change, the observation probability is calculated by the change amplitude between the initial state probability vectors at adjacent sampling times. The formula for calculating the observation probability of the short-term mutation candidate state is as follows: ; in, Indicates time Corresponding to short-time mutation candidate state The probability of observation; Indicates the amplification factor for the sudden change; Indicates time Corresponding to the Individual autonomic nervous system activity levels The probability of emission under the observed state; Indicates time The previous sampling time corresponds to the first Individual autonomic nervous system activity levels The probability of emission under the observed state; This represents the absolute value operation; This means taking the smaller of the two input values; for The sampling time is set. ; For time The observation probabilities corresponding to the candidate states of wakefulness, light sleep, deep sleep, and rapid eye movement (REM) are respectively set as follows: , , , ,in, Indicates time Corresponding to the awake candidate state The probability of observation, Indicates time Corresponding to the candidate state of light sleep The probability of observation, Indicates time Corresponding to deep sleep candidate states The probability of observation, Indicates time Corresponding to rapid eye movement candidate states The probability of observation; A short-time mutation-sensitive asymmetric hidden Markov decoding algorithm is applied to calculate the optimal hidden link for rapid state switching. This algorithm uses a recursive scoring method to calculate the path score of each hidden physiological state at each time step and records the preceding hidden physiological state of each hidden physiological state at each time step. The recursive scoring formula is as follows: ; in, Indicates arrival time The A hidden physiological state Path score at time; Represent the natural logarithm function; Indicates time Corresponding to the A hidden physiological state The probability of observation; This represents a positive stable term that prevents the logarithmic function input from being zero. This represents the index of the hidden physiological state at the previous sampling time. ; Indicates arrival time The A hidden physiological state Path score at time; Indicates time From the A hidden physiological state Switch to the A hidden physiological state Short-term mutation-sensitive regulatory terms; The short-time mutation-sensitive regulation term is calculated according to the following formula: ; in, This represents the short-term mutation sensitivity regulation coefficient; This indicates that the hidden physiological state at the previous sampling time is different from the hidden physiological state at the current sampling time; This indicates that the hidden physiological state at the previous sampling time is the same as the hidden physiological state at the current sampling time. During the recursive scoring process, at each moment... and each current hidden physiological state Record the index of the preceding hidden physiological state that makes the maximum value in the recursive scoring formula true, and form a preceding index table; after completing the recursive scoring of all sampling times, start from the hidden physiological state with the largest path score in the last sampling time, and backtrack to the first sampling time according to the preceding index table to obtain the physiological state transition path sequence, and output the physiological state transition path sequence. During the execution of the short-term mutation-sensitive asymmetric hidden Markov decoding algorithm, a lower bound check for state transition probability and a lower bound check for observation probability are set. The lower bound check for state transition probability is used to correct state transition probabilities that are less than the lower bound for transition probability to the lower bound for transition probability, and to normalize the corrected asymmetric state transition matrix row by row. The lower bound check for observation probability is used to correct observation probabilities that are less than the lower bound for observation probability to the lower bound for observation probability, and to normalize the observation probabilities of the five hidden physiological states at the same time.

[0013] Optionally, step S4 further includes: Establish a set of sleep stages: ; in, This represents a set of sleep stages. Indicating a state of wakefulness, This indicates the light sleep stage. Indicates the deep sleep stage. Indicates the rapid eye movement (REM) phase; Perform microsleep cycle mapping classification to classify wakefulness candidate states. Mapped to the waking stage Candidate state of light sleep Mapped to light sleep stage Candidate state for deep sleep Mapped to deep sleep stage Rapid eye movement candidate states Mapped to REM (Rapid Eye Movement) stage ; For short-time mutation candidate states The system performs a joint mapping of preceding and following neighborhoods. This joint mapping includes: mapping the short-term mutation candidate state to the same sleep stage when the hidden physiological state corresponding to the previous sampling time and the hidden physiological state corresponding to the next sampling time are mapped to the same sleep stage; mapping the short-term mutation candidate state to the awake stage when the hidden physiological state corresponding to the previous sampling time and the hidden physiological state corresponding to the next sampling time are mapped to different sleep stages; mapping the short-term mutation candidate state to the awake stage when it is located at the first sampling time; mapping the short-term mutation candidate state to the sleep stage corresponding to the hidden physiological state at the next sampling time; and mapping the short-term mutation candidate state to the sleep stage corresponding to the hidden physiological state at the previous sampling time when it is located at the last sampling time. Sampling times that are consecutive in sampling time sequence and have the same sleep stage are merged into sleep stage segments; each sleep stage segment includes segment start time, segment end time, segment duration, and segment sleep stage; the segment duration is obtained by multiplying the number of sampling intervals between the segment end time and the segment start time by the sampling time interval; After sleep stage segments are formed, segment boundary consistency checks are performed. These checks include: merging adjacent sleep stage segments into one segment when the sleep stages of two adjacent segments are the same; deleting a sleep stage segment when its duration is less than a sampling time interval and incorporating the sampling time corresponding to the deleted segment into the preceding segment; and incorporating the sampling time corresponding to the deleted segment into the next sleep stage segment when the deleted segment does not have a preceding segment. After the microsleep cycle mapping classification and segment boundary consistency verification, a refined sleep stage segment sequence is obtained and output.

[0014] Optionally, step S5 further includes: Each sleep stage segment in the refined sleep stage segment sequence is screened for duration. Segments lasting 15 minutes or more and belonging to the light sleep, deep sleep, or REM sleep stages are marked as valid sleep stage segments; segments lasting less than 15 minutes or belonging to the awake stage are marked as invalid sleep stage segments. Continuous aggregation is performed on valid sleep stage segments. The continuous aggregation includes a first aggregation path and a second aggregation path. The first aggregation path is used to handle the case where there is no waking phase between adjacent valid sleep stage segments; in the first aggregation path, valid sleep stage segments that are adjacent in time and have no waking phase in between are aggregated into the same fragmented sleep recognition segment; The second aggregation path handles the case where there is a waking phase between adjacent valid sleep stages. In the second aggregation path, the duration of the waking phase sleep stage between two adjacent valid sleep stages is calculated. When the duration of the waking phase sleep stage is less than five minutes, the two adjacent valid sleep stages before and after the waking phase sleep stage, along with the waking phase sleep stage itself, are aggregated into the same fragmented sleep identification segment. When the duration of the waking phase sleep stage is five minutes or more, the two adjacent valid sleep stages before and after the waking phase sleep stage are divided into two different fragmented sleep identification segments. Each fragmented sleep segment includes the start time, end time, total duration, cumulative wakefulness time, cumulative light sleep time, cumulative deep sleep time, and cumulative REM sleep time. The total duration is determined by the time interval between the end and start times of the sleep segment. The cumulative wakefulness time is the sum of the durations of the wakefulness sleep stages within the fragmented sleep segment. The cumulative light sleep time is the sum of the durations of the light sleep stages within the fragmented sleep segment. The cumulative deep sleep time is the sum of the durations of the deep sleep stages within the fragmented sleep segment. The cumulative REM sleep time is the sum of the durations of the REM sleep stages within the fragmented sleep segment. After continuous aggregation is completed, a validity check is performed on each fragmented sleep recognition segment. The validity check includes: if the sum of the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM sleep stage in the fragmented sleep recognition segment reaches 15 minutes or more, the fragmented sleep recognition segment is retained; if the sum of the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM sleep stage in the fragmented sleep recognition segment does not reach 15 minutes, the fragmented sleep recognition segment is deleted. After duration screening, continuous aggregation, and validity verification, fragmented sleep recognition results are obtained. The fragmented sleep recognition results include the number of retained fragmented sleep segments, the start time of each fragmented sleep segment, the end time of each fragmented sleep segment, the total duration of each fragmented sleep segment, the cumulative time of the awake stage, the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM sleep stage, thereby achieving accurate fragmented sleep recognition for smartwatches based on multi-source sensor fusion.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: This invention significantly improves the ability and coverage of fragmented sleep identification. Traditional methods typically exclude or classify sleep segments lasting less than 30 minutes as wakefulness, leading to a severe underestimation of the user's actual effective sleep. This invention, by establishing a 15-minute effective sleep threshold and designing an aggregation strategy that includes short wakefulness intervals, can effectively identify and integrate fragmented sleep events composed of multiple short sleep segments.

[0016] This invention significantly enhances the physiological rationality and temporal sensitivity of sleep state discrimination through dynamic fusion and adaptive boundary division of multi-source physiological signals. Unlike the traditional fixed window segmentation method, this invention employs an adaptive boundary search algorithm based on transient change slope and physiological feature phase. It dynamically determines the physiological state boundary based on the coordinated mutation characteristics of three types of signals: heart rate variability, body temperature, and micro-motion, ensuring that the feature sequence remains highly synchronized with the actual physiological rhythm.

[0017] The short-term mutation-sensitive asymmetric hidden Markov decoding mechanism proposed in this invention effectively overcomes the shortcomings of traditional state transition models in characterizing the dynamic process of sleep. By constructing an asymmetric state transition matrix and introducing a mutation regulation term based on the probability change of adjacent time points, this mechanism can sensitively respond to rapid switching between sleep stages, especially showing stronger robustness when dealing with non-stationary processes such as sudden transition from deep sleep to wakefulness or the insertion of micro-awakening. Simultaneously, the system performs joint mapping of the preceding and following neighborhoods and boundary consistency verification on short-term mutation candidate states, further optimizing the continuity and physiological logic of the hidden state sequence. This series of technological innovations ensures that the final sleep staging results not only conform to the basic rules of medical sleep staging but also accurately reflect the complex sleep behavior patterns of individuals in real life. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the probability of autonomic nervous system activity levels. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0020] Example 1: A method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion, such as... Figure 1 As shown, it includes the following steps: S1: Based on heart rate variability data, wrist temperature data, and micro-motion data collected by a smartwatch, the transient change slope of each data type over time is extracted. A physiological feature phase adaptive boundary search algorithm is then used to determine the boundaries of physiological feature changes, forming a continuous multidimensional physiological feature sequence unconstrained by a fixed time window, including: A multi-source sensor data acquisition channel for a smartwatch is established, comprising a heart rate variability acquisition channel, a wrist temperature acquisition channel, and a micro-motion data acquisition channel. The heart rate variability acquisition channel outputs a heart rate variability sequence, the wrist temperature acquisition channel outputs a wrist temperature sequence, and the micro-motion data acquisition channel outputs a micro-motion data sequence. In this embodiment, the original sampling frequency of the heart rate variability acquisition channel is set to 1Hz, the original sampling frequency of the wrist temperature acquisition channel is set to 0.2Hz, and the original sampling frequency of the micro-motion data acquisition channel is set to 25Hz. The data from each channel are time-aligned and then unified to a sampling frequency of 1Hz. Define a unified time index set as follows: ; in, This represents the unified time index set corresponding to the data acquisition channels of the multi-source sensors in a smartwatch. This represents the last sampling time in the unified time index set. Represents any sampling time in the unified time index set. ; At any moment At this point, the sampled values ​​of the heart rate variability sequence are represented as The sampled values ​​of the wrist temperature sequence are represented as follows: The sampled values ​​of the micro-motion data sequence are represented as ; Time alignment processing is performed on the heart rate variability sequence, wrist temperature sequence, and micro-motion data sequence. This time alignment processing includes unifying sampling times, removing duplicate samples, and filling in missing samples. Missing samples are filled using linear interpolation of adjacent valid samples. The filled heart rate variability sequence, wrist temperature sequence, and micro-motion data sequence all correspond to the unified time index set. In this embodiment, the original samples from the micro-motion data acquisition channel are aggregated using root mean square (RMS) aggregation in 1-second increments, merging 25 original samples per second into a single representative value to obtain a 1Hz micro-motion data sequence. The original samples from the wrist temperature acquisition channel are expanded from 0.2Hz to 1Hz using linear interpolation; that is, for every 5 consecutive unified times, 3 missing times are filled using linear interpolation of two adjacent valid temperature samples. Transient change slopes were calculated for the time-aligned heart rate variability sequence, wrist temperature sequence, and micromotion data sequence, respectively. These transient change slopes included the heart rate variability transient change slope, wrist temperature transient change slope, and micromotion transient change slope. The formula for calculating the transient change slope is as follows: ; in, Indicates the sensor data type identifier. , This represents a data type representing heart rate variability. This represents the data type of wrist temperature. This represents a micro-motion data type; Indicates time Corresponding to sensor data type identifier The sampled value; when hour, , Indicates time The slope of transient changes in heart rate variability; when hour, , Indicates time The slope of the transient change in wrist body temperature; when hour, , Indicates time The slope of the transient change of micro-motion; Indicates time The previous sampling time corresponds to the sensor data type identifier. The sampled values; for The sampling time is set. ; Build Time Transient slope vector: ; in, Indicates time The transient change slope vector, wherein the three components of the transient change slope vector are, in order, the heart rate variability transient change slope, the wrist body temperature transient change slope, and the micromotion transient change slope; The transient slope vector is searched for boundaries using a physiological feature phase adaptive boundary search algorithm. The physiological feature phase adaptive boundary search algorithm includes parallel stable phase boundary search paths, transitional phase boundary search paths, and abrupt phase boundary search paths. The stable phase boundary search path is used to handle sampling moments where the slopes of heart rate variability transient changes, wrist body temperature transient changes, and micro-motion transient changes do not exceed their respective stable slope thresholds; the transitional phase boundary search path is used to handle sampling moments where one or both slopes of the heart rate variability transient change slope, wrist body temperature transient change slope, and micro-motion transient change slope exceed their corresponding stable slope thresholds; the abrupt phase boundary search path is used to handle sampling moments where the slopes of the heart rate variability transient change slope, wrist body temperature transient change slope, and micro-motion transient change slope all exceed their corresponding stable slope thresholds; in this embodiment, the stable slope threshold for heart rate variability is set to 8, the stable slope threshold for wrist body temperature is set to 0.05, and the stable slope threshold for micro-motion is set to 0.03; In the stable phase boundary search path, the current sampling time is incorporated into the previous continuous physiological feature segment; in the transitional phase boundary search path, the current sampling time is recorded as the candidate boundary time, and the consistency of the slope direction before and after the candidate boundary time is verified; in the abrupt phase boundary search path, the current sampling time is determined as the phase boundary time, and a new continuous physiological feature segment is established from the phase boundary time. The slope direction consistency verification includes: when the transient change slope sign corresponding to the same sensor data type identifier is consistent in the sampling time before the candidate boundary time, the candidate boundary time, and the sampling time after the candidate boundary time, the candidate boundary time is retained as the phase boundary time; when the transient change slope sign is not consistent, the candidate boundary time is merged into the previous continuous physiological feature segment. Based on the phase boundary time, the heart rate variability sampling values, wrist temperature sampling values, micromotion intensity values, heart rate variability transient change slope, wrist temperature transient change slope, and micromotion transient change slope are combined into a continuous multidimensional physiological feature sequence; time The continuous multidimensional physiological feature vector is represented as: ; in, Indicates time A continuous multidimensional physiological feature vector; a continuous multidimensional physiological feature sequence without fixed time window restriction is formed by the continuous multidimensional physiological feature vectors of all sampling times in the order of sampling times; the continuous multidimensional physiological feature sequence without fixed time window restriction is output. Optionally, to improve the adaptability of the physiological feature phase adaptive boundary search algorithm to individual physiological differences, an individualized stable slope threshold adaptive calibration step is performed before the formal identification process begins: data is collected from the subject in a known awake and quiet state. Heart rate variability sequences, wrist temperature sequences, and micro-motion data sequences at each sampling time point were used as calibration segments; data types for each sensor were labeled. Calculate the standard deviation of the slope of all transient changes within the calibration segment. That is, for each of the calibration segments The population standard deviation is calculated; the individualized stable slope threshold is calculated using the following formula: ; in, Indicates sensor data type identifier The corresponding individualized stable slope threshold; Indicates sensor data type identifier The corresponding calibration scaling factor; Indicates the sensor data type identifier within the calibration segment. The overall standard deviation of the corresponding transient change slope; when the calibration segment exists and its length is not less than At that time, the fixed stable slope threshold in the physiological feature phase adaptive boundary search algorithm is used. Replace with the individualized stable slope threshold When the calibration segment is missing or insufficient in length At that time, retain a fixed stable slope threshold. , and In this embodiment, Set to 600, which corresponds to 600 seconds of calibration data at a uniform sampling frequency of 1Hz; calibration scaling factor. Set to 2.0. Set to 2.5. Set to 2.0.

[0021] S2: Based on the continuous multidimensional physiological feature sequence without fixed time window restrictions, perform autonomic nervous system activity level matching on the multidimensional physiological features at each sampling time, calculate the observed state emission probability corresponding to the respective autonomic nervous system activity level at that sampling time, and generate an initial state probability distribution sequence, including: The set of autonomic nervous system activity levels is defined as follows: ; in, This represents a set of autonomic nervous system activity levels. Indicates a high level of sympathetic activity. Indicates the level of mixed sympathetic and parasympathetic activity. Indicates the level of higher parasympathetic activity. Indicates a level of rapid fluctuation in activity; For each moment Continuous multidimensional physiological feature vectors Calculate the observed state emission probability for each corresponding principal nerve activity level in the set of autonomic nerve activity levels; the formula for calculating the observed state emission probability is: ; in, Indicates time Continuous multidimensional physiological feature vectors Corresponding to the Individual autonomic nervous system activity levels The observed emission probability, ; Represents an exponential function with the natural constant as its base; Indicates time Continuous multidimensional physiological feature vectors With the Individual autonomic nervous system activity levels The standardized distance between the hierarchical feature centers; This represents the level index in the set of autonomic nervous system activity levels. ; Indicates time Continuous multidimensional physiological feature vectors With the Individual autonomic nervous system activity levels The standardized distance between the hierarchical feature centers; The standardized distance Calculate according to the following formula: ; in, The dimension index represents a continuous multidimensional physiological feature vector. ; Indicates time Continuous multidimensional physiological feature vectors The One component; Indicates the first Individual autonomic nervous system activity levels The hierarchical feature center in the first Central values ​​in each dimension; Indicates the first The distance weight coefficients for each of the six dimensions are set to a sum of one; in this embodiment, the distance weight coefficients for the six dimensions are set sequentially as follows: , , , , , The first to sixth dimensions correspond to the heart rate variability sampling value, wrist temperature sampling value, micromotion intensity value, heart rate variability transient change slope, wrist temperature transient change slope, and micromotion transient change slope, respectively; the four autonomic nervous activity level grading feature centers. It is obtained by clustering and statistical averaging historical physiological data labeled with sleep stages, and can be updated online based on users' historical data during deployment; Time The emission probabilities of the four observed states are arranged in the order of their levels in the set of autonomic nervous system activity levels to form the initial state probability vector: ; in, Indicates time The initial state probability vector; the initial state probability distribution sequence is composed of the initial state probability vectors at all sampling times in the order of sampling times, and the initial state probability distribution sequence is output, such as... Figure 2 As shown.

[0022] S3: Based on the initial state probability distribution sequence, an asymmetric hidden Markov decoding algorithm sensitive to short-term mutations is introduced to decode the state transition process between adjacent sampling times, determine the optimal hidden link under the condition of rapid state switching, and obtain the physiological state transition path sequence, including: Establish a set of hidden physiological states: ; in, Represents the set of hidden physiological states. Indicates a state of being awake or in a candidate state. This indicates a candidate state of light sleep. Indicates a candidate state for deep sleep. This indicates a candidate state for rapid eye movement (REM). Indicates a short-term mutation candidate state; Construct the asymmetric state transition matrix: ; in, Represents the asymmetric state transition matrix; Indicates from the first A hidden physiological state Transfer to the A hidden physiological state The state transition probability, , ;when hour, and They have different values; the sum of the five state transition probabilities in the same row is one; in this embodiment, the asymmetric state transition matrix... The values ​​of each element are set as follows: Awake Candidate State The five state transition probabilities of the current row are as follows: , , , , Light sleep candidate state The five state transition probabilities of the current row are as follows: , , , , Deep sleep candidate state The five state transition probabilities of the current row are as follows: , , , , Rapid eye movement candidate states The five state transition probabilities of the current row are as follows: , , , , Short-term mutation candidate state The five state transition probabilities of the current row are as follows: , , , , ; The initial state probability vector is mapped to the observation probability of the hidden physiological state; for the candidate states of wakefulness, light sleep, deep sleep, and rapid eye movement, the observation state emission probabilities of high sympathetic activity level, mixed sympathetic and parasympathetic activity level, high parasympathetic activity level, and rapid fluctuation activity level are used as the observation probabilities, respectively; for the candidate states of short-term abrupt change, the observation probability is calculated by the change amplitude between the initial state probability vectors at adjacent sampling times. The formula for calculating the observation probability of the short-term mutation candidate state is as follows: ; in, Indicates time Corresponding to short-time mutation candidate state The probability of observation; Indicates the amplification factor for the sudden change; Indicates time Corresponding to the Individual autonomic nervous system activity levels The probability of emission under the observed state; Indicates time The previous sampling time corresponds to the first Individual autonomic nervous system activity levels The probability of emission under the observed state; This represents the absolute value operation; This means taking the smaller of the two input values; for The sampling time is set. In this embodiment, the amplification factor of the sudden change is... Set to 4.0; For time The observation probabilities corresponding to the candidate states of wakefulness, light sleep, deep sleep, and rapid eye movement (REM) are respectively set as follows: , , , ,in, Indicates time Corresponding to the awake candidate state The probability of observation, Indicates time Corresponding to the candidate state of light sleep The probability of observation, Indicates time Corresponding to deep sleep candidate states The probability of observation, Indicates time Corresponding to rapid eye movement candidate states The probability of observation; A short-time mutation-sensitive asymmetric hidden Markov decoding algorithm is applied to calculate the optimal hidden link for rapid state switching. This algorithm uses a recursive scoring method to calculate the path score of each hidden physiological state at each time step and records the preceding hidden physiological state of each hidden physiological state at each time step. The recursive scoring formula is as follows: ; in, Indicates arrival time The A hidden physiological state Path score at time; Represent the natural logarithm function; Indicates time Corresponding to the A hidden physiological state The probability of observation; This represents a positive stable term that prevents the logarithmic function input from being zero. This represents the index of the hidden physiological state at the previous sampling time. ; Indicates arrival time The A hidden physiological state Path score at time; Indicates time From the A hidden physiological state Switch to the A hidden physiological state The short-term mutation-sensitive regulation term; in this embodiment, the positive stable term. Set as ; For the first sampling time The path scores of the five hidden physiological states are initialized using the following formula: ; in, Indicates the first The initial prior probability of a hidden physiological state. In this embodiment, the initial prior probability adopts an equal probability distribution, that is... ; The short-time mutation-sensitive regulation term is calculated according to the following formula: ; in, This represents the short-term mutation sensitivity regulation coefficient; This indicates that the hidden physiological state at the previous sampling time is different from the hidden physiological state at the current sampling time; This indicates that the hidden physiological state at the previous sampling time is the same as the hidden physiological state at the current sampling time; in this embodiment, the short-term mutation sensitivity adjustment coefficient Set to 2.5; During the recursive scoring process, at each moment... and each current hidden physiological state Record the index of the preceding hidden physiological state that makes the maximum value in the recursive scoring formula true, and form a preceding index table; after completing the recursive scoring of all sampling times, start from the hidden physiological state with the largest path score in the last sampling time, and backtrack to the first sampling time according to the preceding index table to obtain the physiological state transition path sequence, and output the physiological state transition path sequence. During the execution of the short-term mutation-sensitive asymmetric hidden Markov decoding algorithm, a lower bound check for state transition probability and a lower bound check for observation probability are set. The lower bound check for state transition probability is used to correct state transition probabilities that are less than the lower bound to the lower bound, and to normalize the corrected asymmetric state transition matrix row by row. The lower bound check for observation probability is used to correct observation probabilities that are less than the lower bound to the lower bound, and to normalize the observation probabilities of the five hidden physiological states at the same time. In this embodiment, the lower bound for transition probability is set to 0.005, and the lower bound for observation probability is set to 0.005.

[0023] S4: Based on the physiological state transition path sequence, perform microsleep cycle mapping classification on the continuous hidden physiological states, transform and reconstruct the hidden physiological states into wakefulness, light sleep, deep sleep, and REM sleep stages, to obtain a refined sleep stage segment sequence, including: Establish a set of sleep stages: ; in, This represents a set of sleep stages. Indicating a state of wakefulness, This indicates the light sleep stage. Indicates the deep sleep stage. Indicates the rapid eye movement (REM) phase; Perform microsleep cycle mapping classification to classify wakefulness candidate states. Mapped to the waking stage Candidate state of light sleep Mapped to light sleep stage Candidate state for deep sleep Mapped to deep sleep stage Rapid eye movement candidate states Mapped to REM (Rapid Eye Movement) stage ; For short-time mutation candidate states The system performs a joint mapping of preceding and following neighborhoods. This joint mapping includes: mapping the short-term mutation candidate state to the same sleep stage when the hidden physiological state corresponding to the previous sampling time and the hidden physiological state corresponding to the next sampling time are mapped to the same sleep stage; mapping the short-term mutation candidate state to the awake stage when the hidden physiological state corresponding to the previous sampling time and the hidden physiological state corresponding to the next sampling time are mapped to different sleep stages; mapping the short-term mutation candidate state to the awake stage when it is located at the first sampling time; mapping the short-term mutation candidate state to the sleep stage corresponding to the hidden physiological state at the next sampling time; and mapping the short-term mutation candidate state to the sleep stage corresponding to the hidden physiological state at the previous sampling time when it is located at the last sampling time. Sampling times that are consecutive in sampling time sequence and belong to the same sleep stage are merged into sleep stage segments; each sleep stage segment includes a segment start time, a segment end time, a segment duration, and a segment sleep stage; the segment duration is obtained by multiplying the number of sampling intervals between the segment end time and the segment start time by the sampling time interval; in this embodiment, the sampling time interval is 1 second, and the unit of segment duration is seconds; let the time index corresponding to the segment start time of the sleep stage segment be . The time index corresponding to the end of the segment is The duration of the fragment The calculation formula is: ; in, The duration of a sleep stage segment is indicated in seconds. This indicates the time index corresponding to the end of the segment; This indicates the time index corresponding to the start time of the segment; In this embodiment, the sampling time interval is indicated. Second; After sleep stage segments are formed, segment boundary consistency checks are performed. These checks include: merging adjacent sleep stage segments into one segment when the sleep stages of two adjacent segments are the same; deleting a sleep stage segment when its duration is less than a sampling time interval and incorporating the sampling time corresponding to the deleted segment into the preceding segment; and incorporating the sampling time corresponding to the deleted segment into the next sleep stage segment when the deleted segment does not have a preceding segment. After the microsleep cycle mapping classification and segment boundary consistency verification, a refined sleep stage segment sequence is obtained and output.

[0024] S5: Screen the duration of the refined sleep stage fragment sequence, and aggregate sleep stage fragments with a duration of 15 minutes or more according to their temporal continuity to obtain the fragmented sleep recognition results corresponding to the data collected by the smartwatch, including: Each sleep stage segment in the refined sleep stage segment sequence is screened for duration. Segments lasting 15 minutes or more and belonging to the light sleep, deep sleep, or REM sleep stages are marked as valid sleep stage segments; segments lasting less than 15 minutes or belonging to the awake stage are marked as invalid sleep stage segments. Continuous aggregation is performed on valid sleep stage segments. The continuous aggregation includes a first aggregation path and a second aggregation path. The first aggregation path is used to handle the case where there is no waking phase between adjacent valid sleep stage segments; in the first aggregation path, valid sleep stage segments that are adjacent in time and have no waking phase in between are aggregated into the same fragmented sleep recognition segment; The second aggregation path handles the case where there is a waking phase between adjacent valid sleep stages. In the second aggregation path, the duration of the waking phase sleep stage between two adjacent valid sleep stages is calculated. When the duration of the waking phase sleep stage is less than five minutes, the two adjacent valid sleep stages before and after the waking phase sleep stage, along with the waking phase sleep stage itself, are aggregated into the same fragmented sleep identification segment. When the duration of the waking phase sleep stage is five minutes or more, the two adjacent valid sleep stages before and after the waking phase sleep stage are divided into two different fragmented sleep identification segments. Each fragmented sleep segment includes the start time, end time, total duration, cumulative wakefulness time, cumulative light sleep time, cumulative deep sleep time, and cumulative REM sleep time. The total duration is determined by the time interval between the end and start times of the sleep segment. The cumulative wakefulness time is the sum of the durations of the wakefulness sleep stages within the fragmented sleep segment. The cumulative light sleep time is the sum of the durations of the light sleep stages within the fragmented sleep segment. The cumulative deep sleep time is the sum of the durations of the deep sleep stages within the fragmented sleep segment. The cumulative REM sleep time is the sum of the durations of the REM sleep stages within the fragmented sleep segment. After continuous aggregation is completed, a validity check is performed on each fragmented sleep recognition segment. The validity check includes: if the sum of the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM sleep stage in the fragmented sleep recognition segment reaches 15 minutes or more, the fragmented sleep recognition segment is retained; if the sum of the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM sleep stage in the fragmented sleep recognition segment does not reach 15 minutes, the fragmented sleep recognition segment is deleted. After duration screening, continuous aggregation, and validity verification, fragmented sleep recognition results are obtained. These results include the number of retained fragmented sleep segments, the start time of each segment, the end time of each segment, the total duration of each segment, the cumulative time of the waking phase, the cumulative time of the light sleep phase, the cumulative time of the deep sleep phase, and the cumulative time of the REM (Rapid Eye Movement) phase. This enables accurate fragmented sleep recognition on a smartwatch based on multi-source sensor fusion. Optionally, to improve the quantifiability of fragmented sleep recognition results, after completing the validity review, a path decoding confidence scoring step is performed on each retained fragmented sleep recognition segment: for each sampling time within the fragmented sleep recognition segment... Extract the decoding state obtained at that moment during the execution of the asymmetric hidden Markov decoding algorithm, which is sensitive to short-term mutations. In the hidden physiological state set The corresponding state index And read the observation probability corresponding to the state index. ,in Indicates time The state index of the decoding state; let the start time index of the fragmented sleep recognition segment be... The end time index is Then the path decoding confidence score of the fragmented sleep recognition segment. Calculate according to the following formula: ; in, The confidence score for path decoding of fragmented sleep recognition segments; Indicates the start time index of the fragmented sleep recognition segment; Indicates the end time index of the fragmented sleep recognition segment; Indicates time The observation probability corresponding to the decoding state; the confidence score of the path decoding. The range of values ​​is A higher value indicates a higher degree of agreement between the multidimensional physiological features at each sampling time within the segment and the decoded hidden physiological state, resulting in a more reliable identification result. In this embodiment, the path decoding confidence score is used to... As an additional output field for fragmented sleep recognition results; when When the fragmented sleep recognition segment is marked as a high-confidence segment; when When this occurs, the fragmented sleep segment is marked as a low-confidence segment and noted in the recognition result to prompt the user to pay attention; the path decoding confidence score, together with the total duration of the sleep segment and the cumulative time of each stage, constitutes the complete descriptive information of the fragmented sleep segment.

[0025] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0026] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0027] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion, characterized in that, Includes the following steps: S1: Based on heart rate variability data, wrist temperature data and micro-motion data collected by smartwatches, the transient change slope of each type of data over time is extracted, and the physiological feature phase adaptive boundary search algorithm is used to determine the physiological feature change boundary, forming a continuous multidimensional physiological feature sequence that is not constrained by a fixed time window. S2: Based on the continuous multidimensional physiological feature sequence that is not constrained by a fixed time window, the multidimensional physiological features at each sampling time are matched with the autonomic nerve activity level, the observation state emission probability corresponding to the respective autonomic nerve activity level at that sampling time is calculated, and an initial state probability distribution sequence is generated. S3: Based on the initial state probability distribution sequence, an asymmetric hidden Markov decoding algorithm sensitive to short-term mutations is introduced to decode the state transition process between adjacent sampling times, determine the optimal hidden link under the condition of rapid switching between states, and obtain the physiological state transition path sequence. S4: Based on the physiological state transition path sequence, perform microsleep cycle mapping classification on the continuous hidden physiological states, and convert and reconstruct the hidden physiological states into wakefulness, light sleep, deep sleep and REM sleep stages to obtain a refined sleep stage segment sequence. S5: Screen the duration of the refined sleep stage fragment sequence, and aggregate the sleep stage fragments with a duration of 15 minutes or more according to the temporal continuity to obtain the fragmented sleep recognition result corresponding to the data collected by the smartwatch.

2. The method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion according to claim 1, characterized in that, Step S1 includes: A multi-source sensor data acquisition channel for a smartwatch is established, comprising a heart rate variability acquisition channel, a wrist temperature acquisition channel, and a micro-motion data acquisition channel; the heart rate variability acquisition channel outputs a heart rate variability sequence, the wrist temperature acquisition channel outputs a wrist temperature sequence, and the micro-motion data acquisition channel outputs a micro-motion data sequence. Define a unified time index set as follows: ； in, This represents the unified time index set corresponding to the data acquisition channels of the multi-source sensors in a smartwatch. This represents the last sampling time in the unified time index set. Represents any sampling time in the unified time index set. ; At any moment At this point, the sampled values ​​of the heart rate variability sequence are represented as The sampled values ​​of the wrist temperature sequence are represented as follows: The sampled values ​​of the micro-motion data sequence are represented as ; Time alignment processing is performed on the heart rate variability sequence, wrist temperature sequence, and micro-motion data sequence. The time alignment processing includes unifying the sampling time, removing duplicate sample values, and filling in missing sample values. The missing sample value filling is performed by linear interpolation of adjacent valid sample values. The completed heart rate variability sequence, wrist temperature sequence, and micro-motion data sequence all correspond to the unified time index set. Transient change slopes were calculated for the time-aligned heart rate variability sequence, wrist temperature sequence, and micromotion data sequence, respectively. These transient change slopes included the heart rate variability transient change slope, wrist temperature transient change slope, and micromotion transient change slope. The formula for calculating the transient change slope is as follows: ； in, Indicates the sensor data type identifier. , This represents a data type representing heart rate variability. This represents the data type of wrist temperature. This represents a micro-motion data type; Indicates time Corresponding to sensor data type identifier The sampled value; when hour, , Indicates time The slope of transient changes in heart rate variability; when hour, , Indicates time The slope of the transient change in wrist body temperature; when hour, , Indicates time The slope of the transient change of micro-motion; Indicates time The previous sampling time corresponds to the sensor data type identifier. The sampled values; for The sampling time is set. ; Build Time Transient slope vector: ； in, Indicates time The transient change slope vector, wherein the three components of the transient change slope vector are, in order, the heart rate variability transient change slope, the wrist body temperature transient change slope, and the micromotion transient change slope; The physiological characteristic phase adaptive boundary search algorithm is used to search for the boundary of the transient slope vector; the physiological characteristic phase adaptive boundary search algorithm includes parallel stable phase boundary search paths, transitional phase boundary search paths and abrupt phase boundary search paths.

3. The method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion according to claim 2, characterized in that, Step S1 further includes: The stable phase boundary search path is used to handle sampling moments where the slopes of heart rate variability transient changes, wrist body temperature transient changes, and micro-motion transient changes do not exceed their respective stable slope thresholds; the transitional phase boundary search path is used to handle sampling moments where one or both of the slopes of heart rate variability transient changes, wrist body temperature transient changes, and micro-motion transient changes exceed their corresponding stable slope thresholds; the abrupt phase boundary search path is used to handle sampling moments where the slopes of heart rate variability transient changes, wrist body temperature transient changes, and micro-motion transient changes all exceed their corresponding stable slope thresholds. In the stable phase boundary search path, the current sampling time is incorporated into the previous continuous physiological feature segment; in the transitional phase boundary search path, the current sampling time is recorded as the candidate boundary time, and the consistency of the slope direction before and after the candidate boundary time is verified; in the abrupt phase boundary search path, the current sampling time is determined as the phase boundary time, and a new continuous physiological feature segment is established from the phase boundary time. The slope direction consistency verification includes: when the transient change slope sign corresponding to the same sensor data type identifier is consistent in the sampling time before the candidate boundary time, the candidate boundary time, and the sampling time after the candidate boundary time, the candidate boundary time is retained as the phase boundary time; when the transient change slope sign is not consistent, the candidate boundary time is merged into the previous continuous physiological feature segment. Based on the phase boundary time, the heart rate variability sampling values, wrist temperature sampling values, micromotion intensity values, heart rate variability transient change slope, wrist temperature transient change slope, and micromotion transient change slope are combined into a continuous multidimensional physiological feature sequence; time The continuous multidimensional physiological feature vector is represented as: ； in, Indicates time The continuous multidimensional physiological feature vectors are obtained; the continuous multidimensional physiological feature sequence without fixed time window restrictions is formed by the continuous multidimensional physiological feature vectors of all sampling times in the order of sampling times, and the continuous multidimensional physiological feature sequence without fixed time window restrictions is output.

4. The method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion according to claim 3, characterized in that, Step S2 includes: The set of autonomic nervous system activity levels is defined as follows: ； in, This represents a set of autonomic nervous system activity levels. Indicates a high level of sympathetic activity. Indicates the level of mixed sympathetic and parasympathetic activity. Indicates the level of higher parasympathetic activity. Indicates a level of rapid fluctuation in activity; For each moment Continuous multidimensional physiological feature vectors Calculate the observed state emission probability for each corresponding principal nerve activity level in the set of autonomic nerve activity levels; the formula for calculating the observed state emission probability is: ； in, Indicates time Continuous multidimensional physiological feature vectors Corresponding to the Individual autonomic nervous system activity levels The probability of emission under the observed state. ; Represents an exponential function with the natural constant as its base; Indicates time Continuous multidimensional physiological feature vectors With the Individual autonomic nervous system activity levels The standardized distance between the hierarchical feature centers; This represents the level index in the set of autonomic nervous system activity levels. ; Indicates time Continuous multidimensional physiological feature vectors With the Individual autonomic nervous system activity levels The standardized distance between the hierarchical feature centers; The standardized distance Calculate according to the following formula: ； in, The dimension index represents a continuous multidimensional physiological feature vector. ; Indicates time Continuous multidimensional physiological feature vectors The One component; Indicates the first Individual autonomic nervous system activity levels The hierarchical feature center in the first Central values ​​in each dimension; Indicates the first The distance weight coefficients for each dimension are equal to one. Time The emission probabilities of the four observed states are arranged in the order of their levels in the set of autonomic nervous system activity levels to form the initial state probability vector: ； in, Indicates time The initial state probability vector is obtained; the initial state probability distribution sequence is formed by the initial state probability vectors of all sampling times in the order of sampling times, and the initial state probability distribution sequence is output.

5. The method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion according to claim 4, characterized in that, Step S3 includes: Establish a set of hidden physiological states: ； in, Represents the set of hidden physiological states. Indicates a conscious candidate state. Indicates a candidate state of light sleep. Indicates a candidate state for deep sleep. This indicates a candidate state for rapid eye movement (REM). Indicates a short-term mutation candidate state; Construct the asymmetric state transition matrix: ； in, Represents the asymmetric state transition matrix; Indicates from the first A hidden physiological state Transfer to the A hidden physiological state The state transition probability, , ;when hour, and They have different values; the sum of the five state transition probabilities in the same row is one; The initial state probability vector is mapped to the observation probability of the hidden physiological state; for the candidate states of wakefulness, light sleep, deep sleep, and rapid eye movement, the observation state emission probabilities of high sympathetic activity level, mixed sympathetic and parasympathetic activity level, high parasympathetic activity level, and rapid fluctuation activity level are used as the observation probabilities, respectively; for the candidate states of short-term abrupt change, the observation probability is calculated by the change amplitude between the initial state probability vectors at adjacent sampling times. The formula for calculating the observation probability of the short-term mutation candidate state is as follows: ； in, Indicates time Corresponding to short-term mutation candidate state The probability of observation; Indicates the amplification factor of the sudden change; Indicates time Corresponding to the Individual autonomic nervous system activity levels The probability of emission under the observed state; Indicates time The previous sampling time corresponds to the first Individual autonomic nervous system activity levels The probability of emission under the observed state; This represents the absolute value operation; This means taking the smaller of the two input values; for The sampling time is set. ; For time The observation probabilities corresponding to the candidate states of wakefulness, light sleep, deep sleep, and rapid eye movement (REM) are respectively set as follows: , , , ,in, Indicates time Corresponding to the awake candidate state The probability of observation, Indicates time Corresponding to the candidate state of light sleep The probability of observation, Indicates time Corresponding to deep sleep candidate states The probability of observation, Indicates time Corresponding to rapid eye movement candidate states The probability of observation; The short-time mutation-sensitive asymmetric hidden Markov decoding algorithm is used to calculate the optimal hidden link for fast switching between states. The short-time mutation-sensitive asymmetric hidden Markov decoding algorithm uses a recursive scoring method to calculate the path score of each hidden physiological state at each time step, and records the predecessor hidden physiological state of each hidden physiological state at each time step.

6. The method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion according to claim 5, characterized in that, The recursive scoring in step S3 includes: The recursive scoring formula is: ； in, Indicates arrival time The A hidden physiological state Path score at time; Represents the natural logarithm function; Indicates time Corresponding to the A hidden physiological state The probability of observation; This represents a positive stable term that prevents the logarithmic function input from being zero. This represents the index of the hidden physiological state at the previous sampling time. ; Indicates arrival time The A hidden physiological state Path score at time; Indicates time From the A hidden physiological state Switch to the A hidden physiological state Short-term mutation-sensitive regulatory terms; The short-time mutation-sensitive regulation term is calculated according to the following formula: ； in, This represents the short-term mutation sensitivity regulation coefficient; This indicates that the hidden physiological state at the previous sampling time is different from the hidden physiological state at the current sampling time; This indicates that the hidden physiological state at the previous sampling time is the same as the hidden physiological state at the current sampling time. During the recursive scoring process, at each moment... and each current hidden physiological state Record the index of the preceding hidden physiological state that makes the maximum value in the recursive scoring formula true, and form a preceding index table; after completing the recursive scoring of all sampling times, start from the hidden physiological state with the largest path score in the last sampling time, and backtrack to the first sampling time according to the preceding index table to obtain the physiological state transition path sequence, and output the physiological state transition path sequence. During the execution of the short-term mutation-sensitive asymmetric hidden Markov decoding algorithm, a lower bound check for state transition probability and a lower bound check for observation probability are set. The lower bound check for state transition probability is used to correct state transition probabilities that are less than the lower bound for transition probability to the lower bound for transition probability, and to normalize the corrected asymmetric state transition matrix row by row. The lower bound check for observation probability is used to correct observation probabilities that are less than the lower bound for observation probability to the lower bound for observation probability, and to normalize the observation probabilities of the five hidden physiological states at the same time.

7. The method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion according to claim 5, characterized in that, The recursive scoring in step S4 includes: Establish a set of sleep stages: ； in, This represents a set of sleep stages. Indicating a state of wakefulness, This indicates the light sleep stage. Indicates the deep sleep stage. Indicates the rapid eye movement (REM) phase; Perform microsleep cycle mapping classification to classify wakefulness candidate states. Mapped to the waking stage Candidate state of light sleep Mapped to light sleep stage Candidate state for deep sleep Mapped to deep sleep stage Rapid eye movement candidate states Mapped to REM (Rapid Eye Movement) stage ; For short-time mutation candidate states The system performs a joint mapping of preceding and following neighborhoods, which includes: when the hidden physiological state corresponding to the previous sampling time and the hidden physiological state corresponding to the next sampling time of a short-term mutation candidate state are mapped to the same sleep stage, the short-term mutation candidate state is mapped to the same sleep stage; when the hidden physiological state corresponding to the previous sampling time and the hidden physiological state corresponding to the next sampling time are mapped to different sleep stages, the short-term mutation candidate state is mapped to the wake stage; when the short-term mutation candidate state is located at the first sampling time, the short-term mutation candidate state is mapped to the sleep stage of the hidden physiological state corresponding to the next sampling time; when the short-term mutation candidate state is located at the last sampling time, the short-term mutation candidate state is mapped to the sleep stage of the hidden physiological state corresponding to the previous sampling time. Sampling times that are consecutive in sampling time sequence and have the same sleep stage are merged into sleep stage segments; each sleep stage segment includes segment start time, segment end time, segment duration, and segment sleep stage; the segment duration is obtained by multiplying the number of sampling intervals between the segment end time and the segment start time by the sampling time interval; After sleep stage segments are formed, segment boundary consistency checks are performed. These checks include: merging adjacent sleep stage segments into one segment when the sleep stages of two adjacent segments are the same; deleting a sleep stage segment when its duration is less than a sampling time interval and incorporating the sampling time corresponding to the deleted segment into the previous segment; and incorporating the sampling time corresponding to the deleted segment into the next segment when the deleted segment does not have a preceding segment. After the microsleep cycle mapping classification and segment boundary consistency verification, a refined sleep stage segment sequence is obtained and output.

8. The method for accurate identification of fragmented sleep in smartwatches based on multi-source sensor fusion according to claim 7, characterized in that, The recursive scoring in step S5 includes: Each sleep stage segment in the refined sleep stage segment sequence is screened for duration. Segments with a duration of 15 minutes or more and belonging to the light sleep stage, deep sleep stage, or REM sleep stage are marked as valid sleep stage segments; segments with a duration of less than 15 minutes or belonging to the awake stage are marked as invalid sleep stage segments. Perform sequential aggregation on valid sleep stage segments; the sequential aggregation includes a first aggregation path and a second aggregation path; The first aggregation path is used to handle the case where there is no waking phase between adjacent valid sleep stage segments; in the first aggregation path, valid sleep stage segments that are adjacent in time and have no waking phase in between are aggregated into the same fragmented sleep recognition segment; The second aggregation path is used to handle the situation where there is a waking phase between adjacent valid sleep stages. In the second aggregation path, the duration of the waking phase sleep stage between two adjacent valid sleep stages is calculated. When the duration of the waking phase sleep stage is less than five minutes, the two adjacent valid sleep stages before and after the waking phase sleep stage, as well as the waking phase sleep stage, are aggregated into the same fragmented sleep recognition segment. When the duration of the waking phase sleep stage is five minutes or more, the two adjacent valid sleep stages before and after the waking phase sleep stage are divided into two different fragmented sleep recognition segments. Each fragmented sleep segment includes the start time of the sleep segment, the end time of the sleep segment, the total duration of the sleep segment, the cumulative time of the awake stage, the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM (Rapid Eye Movement) stage. The total duration of the sleep segment is determined by the time interval between the end time and the start time of the sleep segment. The cumulative time of the awake stage is the sum of the durations of the awake sleep stages within the fragmented sleep segment. The cumulative time of the light sleep stage is the sum of the durations of the light sleep stages within the fragmented sleep segment. The cumulative time of the deep sleep stage is the sum of the durations of the deep sleep stages within the fragmented sleep segment. The cumulative time of the REM stage is the sum of the durations of the REM (Rapid Eye Movement) sleep stages within the fragmented sleep segment. After continuous aggregation is completed, a validity check is performed on each fragmented sleep recognition segment. The validity check includes: if the sum of the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM stage in the fragmented sleep recognition segment reaches 15 minutes or more, the fragmented sleep recognition segment is retained; if the sum of the cumulative time of the light sleep stage, the cumulative time of the deep sleep stage, and the cumulative time of the REM stage in the fragmented sleep recognition segment does not reach 15 minutes, the fragmented sleep recognition segment is deleted. After duration screening, continuous aggregation, and validity verification, fragmented sleep recognition results are obtained. These results include the number of retained fragmented sleep segments, the start time of each segment, the end time of each segment, the total duration of each segment, the cumulative time of the awake phase, the cumulative time of the light sleep phase, the cumulative time of the deep sleep phase, and the cumulative time of the REM sleep phase. This enables accurate fragmented sleep recognition by smartwatches based on multi-source sensor fusion.

Images