Sleep state monitoring system based on multimodal signals

Abstract

This invention relates to the field of sleep state monitoring technology, specifically to a sleep state monitoring system based on multimodal signals. The system includes a data acquisition module, a wakefulness deviation index acquisition module, a wakefulness damping characteristic parameter acquisition module, and an evaluation index acquisition module. It acquires and aligns the subject's limb acceleration signals and frontal lobe EEG signals in real time; identifies limb stillness periods as evaluation windows, extracts corresponding stillness EEG sequences and segments them, calculates the wakefulness deviation index based on frequency domain morphology and temporal order, and uses nonlinear gating to remove artifact interference; fits the decay of the wakefulness deviation index to obtain the wakefulness damping characteristic parameter for a single sleep onset; finally, it analyzes the changing trend of this parameter across multiple sleep onsets to generate a long-term sleep improvement evaluation index. This invention achieves an objective and accurate assessment of the evolution of the subject's physiological state during sleep induction through time-frequency dual-domain feature fusion and process damping quantification.

Description

Technical Field

[0001] This invention relates to the field of sleep state monitoring technology, and more specifically to a sleep state monitoring system based on multimodal signals. Background Technology

[0002] Poor sleep quality is a common physiological phenomenon, manifested as difficulty falling asleep, sleep maintenance disorders, and early awakening, seriously affecting the quality of life and mental and physical health of subjects. Various sleep induction techniques, as auxiliary methods to improve sleep, are often used to regulate the physiological inhibitory state of subjects to improve sleep quality. In objective quantitative analysis of sleep state monitoring, accurately obtaining the evolution of subjects' sleep onset ability and the patterns of physiological state transitions is a core element.

[0003] Current objective assessment techniques primarily rely on polysomnography (PSG) or portable motion analyzers. However, in practice, motion analyzers only monitor limb movement, while some subjects remain physically still but their cerebral cortex remains in a high-frequency arousal state. Simple motion monitoring cannot identify this physiological characteristic of physical stillness but mental agitation, leading to discrepancies between monitoring results and the subject's actual experience, and failing to accurately reflect the influence of the sleep induction process on the physiological inhibitory state. Furthermore, while EEG-based analysis techniques can monitor physiological activity, existing methods often rely on single-frequency energy proportions. For subjects with difficulty falling asleep, their EEG signals during sleep onset are often accompanied by electromyographic artifacts or alpha wave intrusions. These interfering signals often manifest as increased low-frequency energy in the spectrum, easily confused with the slow waves of deep sleep. Therefore, current technologies lack effective means to distinguish between genuine deep sleep slow waves and chaotic low-frequency artifacts, leading to misjudgments of sleep depth during monitoring and an inability to accurately quantify the evolution of the subject's physiological characteristics under sleep induction intervention. Summary of the Invention

[0004] To address the technical problem of inaccurately distinguishing between genuine deep sleep slow waves and chaotic low-frequency artifacts, leading to inaccurate sleep depth assessment and consequently the inability to objectively quantify the temporal evolution of physiological states under the influence of sleep induction intervention, this invention aims to provide a sleep state monitoring system based on multimodal signals. The specific technical solution adopted is as follows: This invention provides a sleep state monitoring system based on multimodal signals, the system comprising: The data acquisition module is used to acquire the subject's limb acceleration signals and frontal lobe EEG signals in real time, and to perform time alignment on the limb acceleration signals and frontal lobe EEG signals; The arousal deviation index acquisition module is used to identify the period of continuous limb stillness based on limb acceleration signals as the assessment window, and to extract the resting period assessment sequence corresponding to the assessment window from the frontal lobe EEG signal; the resting period assessment sequence is divided into time-domain local sequences through a preset sliding window, and the arousal deviation index of each time-domain local sequence is obtained according to the frequency domain morphology and time-domain order of each time-domain local sequence. The arousal damping characteristic parameter acquisition module is used to acquire arousal damping characteristic parameters that characterize a single sleep-inducing process based on the decay of the arousal deviation index over time within the evaluation window. The assessment index acquisition module is used to obtain long-term sleep improvement assessment indicators based on the changing trends of arousal damping characteristic parameters during multiple sleep episodes.

[0005] Furthermore, the method for obtaining the evaluation window is as follows: Calculate the triaxial summation magnitude of the limb acceleration signal and perform low-pass filtering on the triaxial summation magnitude to obtain the limb activity intensity curve; The first time point in the limb activity intensity curve where the fluctuation of the detected value is less than the preset static judgment threshold and meets the preset duration is marked as the evaluation start time. Starting from the evaluation start time, a time period of preset length is extracted as the evaluation window.

[0006] Furthermore, the method for obtaining the arousal deviation index is as follows: For any local time-domain sequence, the difference between the maximum value and the minimum value in the local time-domain sequence is obtained and used as a reference analysis value. When the reference analysis value is greater than the preset eye-tracking threshold, the local time-domain sequence is removed. When the reference analysis value is less than or equal to the preset eye-tracking threshold, the degree of deviation of the frequency domain shape of the local time-domain sequence is obtained based on the frequency domain shape of the local time-domain sequence. Based on the temporal order of the local temporal sequence, the degree of non-uniformity of the connection of the local temporal sequence is obtained; The difference between the preset baseline non-uniformity of connection and the non-uniformity of connection is compared with a preset truncation threshold. The smaller value is then processed by a preset exponential function, and the result is used as the time-domain reference value of the local time-domain sequence. The result of multiplying the time-domain reference value by a preset reference weight and adding it to a first preset constant is taken as the effective reference value of the local time-domain sequence. The product of the frequency domain morphological deviation degree and the effective reference value is compared with a preset upper limit value, and the smaller value is taken as the awakening deviation index of the local time-domain sequence.

[0007] Furthermore, the method for obtaining the degree of frequency domain morphological deviation is as follows: The power spectral density is obtained by performing a fast Fourier transform on the local time-domain sequence. The power spectral density within the target frequency band is then extracted and normalized to obtain the cumulative spectral distribution of the local time-domain sequence. Based on the power law characteristics of EEG during deep sleep, a benchmark reference distribution of the ideal deep sleep power spectrum is constructed, and the probability density of the benchmark reference distribution is normalized to obtain the benchmark spectrum cumulative distribution. The arithmetic mean of the absolute values ​​of the differences between the cumulative spectral distribution and the reference cumulative spectral distribution at each frequency point is obtained using the one-dimensional Wasserstein distance algorithm, which is used as the degree of deviation of the frequency domain morphology of the local time-domain sequence.

[0008] Furthermore, the method for obtaining the degree of connection non-uniformity is as follows: The local time-domain sequence is downsampled to a preset low-frequency sampling rate to obtain a discrete time-point sequence; The discrete time point sequence is divided into equal-length sub-time series, and a complex network is constructed for each sub-time series using the horizontal visibility graph algorithm to obtain the connection degree of each node in the complex network. The mean of the Gini coefficients corresponding to the connectivity degree of each sub-segment of the time series is used as the degree of connectivity non-uniformity of the local time series.

[0009] Furthermore, the method for obtaining the arousal damping characteristic parameters is as follows: The arousal deviation indexes within the evaluation window are arranged according to the time order of the corresponding local time sequence to obtain the arousal deviation index sequence; Sliding variance detection is performed on the arousal deviation index sequence to obtain the target time when the arousal deviation index sequence enters a steady state; The time period from the start time to the target time is taken as the target time period. The arousal deviation index within the target time period is fitted into a straight line in chronological order. When the slope of the straight line is greater than or equal to 0, the arousal damping characteristic parameter of this sleep process is marked as an invalid null value. When the slope of the straight line is less than 0, the arousal deviation index within the target time period is processed by moving average filtering to obtain a macroscopic descending envelope sequence. Under the constraint that the decay time constant is greater than 0 and the baseline is within a preset range, the macroscopic descending envelope sequence is fitted nonlinearly using an exponential decay model. If the fitting converges and the fitting residual is less than a preset residual threshold, the decay time constant obtained from the fitting is obtained as the arousal damping characteristic parameter. If the fitting does not converge or the fitting residual is greater than or equal to the preset residual threshold, the arousal damping characteristic parameter of this sleep process is marked as an invalid null value.

[0010] Furthermore, the method for obtaining the target time is as follows: Obtain the sliding variance sequence of the arousal deviation index sequence, and detect the first time point in the sliding variance sequence where the value is lower than a preset steady-state variance threshold and the mean of the arousal deviation index up to the corresponding time point is less than a preset steady-state mean threshold, and take it as the target time point.

[0011] Furthermore, the method for obtaining the long-term sleep improvement assessment indicators is as follows: Arrange the arousal damping characteristic parameters according to the time sequence of the corresponding sleep-onset process, and remove the data marked as invalid null values ​​to obtain the arousal damping characteristic parameter sequence; The number of valid samples in the awakening damping feature parameter sequence is used as the cumulative count. When the cumulative count is equal to 1, 0 is set as the target slope. When the cumulative number of times is greater than 1 and less than the preset number of times threshold, the slope of the straight line fitted by the awakening damping characteristic parameter sequence is obtained by the least squares method and used as the target slope. When the cumulative number of times is greater than or equal to the preset number of times threshold, the slope of the straight line obtained by performing RANSAC robust regression analysis on the awakening damping characteristic parameter sequence will be used as the target slope. The negative of the target slope is used as an indicator for assessing long-term sleep improvement.

[0012] Furthermore, the method for obtaining the degree of non-uniformity of the preset benchmark connection is as follows: The average value of the connectivity unevenness of the subjects during their historical effective deep sleep periods was obtained, and the average value of the multiple connectivity unevenness was used as the preset baseline connectivity unevenness. If the cumulative duration of historical effective deep sleep periods is less than the preset sleep duration threshold, then the preset empirical constant will be used as the preset baseline for connection unevenness.

[0013] Furthermore, the exponential decay model is as follows: ;in, denoted as the arousal deviation index of the t-th local sequence in the target time period; t is the relative time corresponding to the t-th local sequence in the target time period; T is the decay time constant to be fitted; A is the initial amplitude to be fitted; C is the baseline to be fitted; and e is the natural constant.

[0014] The present invention has the following beneficial effects: This invention first identifies the period of sustained limb stillness based on limb acceleration signals as the assessment window, eliminating electromyographic interference caused by large movements such as turning over and limb shaking. Then, the assessment sequence corresponding to the resting period is extracted from the frontal lobe EEG signal, which is beneficial for obtaining electrophysiological data reflecting a pure physiological state and establishing a behavioral gating benchmark for sleep state monitoring. To capture the microscopic dynamic changes during sleep onset, the resting period assessment sequence is divided into temporal local sequences through a preset sliding window, achieving high-resolution temporal analysis of the continuous evolution of EEG signals. To prevent extremely low-frequency eye movement signals from deceiving subsequent algorithms and extreme noise from causing numerical overflow, eye movement artifacts are pre-emptively removed through amplitude peak-to-peak detection, and a numerical threshold truncation mechanism is introduced when calculating the temporal reference value and the arousal deviation index, effectively ensuring the numerical stability of monitoring indicators under complex physiological environments. To solve the technical problem that a single frequency domain feature cannot distinguish between slow waves of true deep sleep and low-frequency artifacts, the arousal deviation index of each temporal local sequence is obtained based on its frequency domain morphology and temporal order. A nonlinear gating model is constructed using temporal topological features to accurately reflect whether EEG signals simultaneously satisfy the energy distribution characteristics and structural order characteristics of deep sleep, effectively eliminating false sleep artifacts. To transform the physiological process of falling asleep into quantifiable engineering parameters, a macroscopic descending envelope is extracted based on the decay of the arousal deviation index over time within the evaluation window and fitted under parameter constraints to obtain arousal damping characteristic parameter characterizing a single sleep onset process. This accurately reflects the resistance to the transition from excitation to inhibition in the physiological state, objectively quantifying the evolutionary characteristics of a single sleep onset process, and avoiding the disruption of the overall statistical distribution by outlier data through a null value labeling and removal mechanism. To overcome the limitations of single measurements being susceptible to interference from random factors and to objectively evaluate long-term cumulative trends, long-term sleep improvement assessment indicators are obtained based on the changing trends of arousal damping characteristic parameters during multiple sleep onset processes, combined with sample size branching. This facilitates the automatic removal of outliers through robust regression, providing robust quantitative feedback on long-term sleep characteristics and effectively improving the objectivity and accuracy of monitoring the sleep state and physiological evolution patterns of subjects. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural block diagram of a sleep state monitoring system based on multimodal signals provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for obtaining an arousal deviation index according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a computer device provided according to an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the sleep state monitoring system based on multimodal signals proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The specific solution of the sleep state monitoring system based on multimodal signals provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Example 1: This invention proposes a sleep state monitoring system based on multimodal signals. Please refer to [link / reference]. Figure 1 The diagram shows a structural block diagram of a sleep state monitoring system based on multimodal signals provided in an embodiment of the present invention. The system includes: a data acquisition module 10, a wakefulness deviation index acquisition module 20, a wakefulness damping characteristic parameter acquisition module 30, and an evaluation index acquisition module 40.

[0021] The data acquisition module 10 is used to acquire the subject's limb acceleration signals and frontal lobe electroencephalogram (EEG) signals in real time.

[0022] Specifically, this embodiment uses a single subject as an example for analysis, and all subsequent references to this subject will refer to that subject. To obtain data on the subject's behavioral activities and electroencephalographic state during natural sleep, enabling subsequent assessments to have both behavioral and electrophysiological dimensions, this embodiment uses a three-axis accelerometer worn on the subject's non-dominant wrist to collect limb acceleration signals in real time, capturing subtle limb tremors and macroscopic turning movements; and a single-channel EEG sensor worn on the subject's forehead to collect frontal lobe EEG signals in real time, obtaining neuroelectrophysiological data reflecting the level of cortical arousal. This embodiment sets the sampling frequency of the limb acceleration signal to 50Hz and the frontal lobe EEG signal to 250Hz to ensure sufficient coverage of the effective bandwidth of the target signals (such as the high-frequency characteristics of slow EEG waves and EMG artifacts), while avoiding data redundancy. The implementer can set the sampling frequency of the limb acceleration signal and the frontal lobe EEG signal according to sensor bandwidth or storage limitations, which is not limited here.

[0023] Considering that limb acceleration signals and frontal lobe EEG signals are collected by different sensors, there may be transmission delays or clock deviations between them. Direct analysis may lead to a mismatch between behavioral and electrophysiological states (e.g., misjudging artifacts during limb movement as EEG waveforms during rest). To ensure accurate correspondence between the limb resting moment and the EEG capture moment in subsequent analysis, this embodiment uses a hardware tagging mechanism or a unified system clock to time-align the limb acceleration signals and frontal lobe EEG signals. Specifically, the system uses the same microsecond-level clock source to timestamp the two data streams and aligns the two signals based on the timestamps during the data buffering stage, controlling the maximum synchronization error within 100ms, thereby constructing a multimodal physiological data sequence with strict time synchronization. The hardware tagging mechanism and the unified system clock are well-known and will not be elaborated further.

[0024] The arousal deviation index acquisition module 20 is used to identify the continuous static period of the limb based on the limb acceleration signal as the assessment window, and to extract the static period assessment sequence corresponding to the assessment window from the frontal lobe EEG signal; the static period assessment sequence is divided into time-domain local sequences through a preset sliding window, and the arousal deviation index of each time-domain local sequence is obtained according to the frequency domain morphology and temporal order of each time-domain local sequence.

[0025] Specifically, to avoid interference from electromyographic artifacts (such as swallowing and turning over) caused by large limb movements on EEG signal analysis, this embodiment first identifies the period of sustained limb stillness based on limb acceleration signals. That is, by monitoring the fluctuation of limb activity intensity, the time period in which the subject is in a stable state is locked. In order to establish the time benchmark for subsequent EEG state analysis, the period of sustained limb stillness is used as the evaluation window. This evaluation window represents a pure analysis period that excludes interference from random movements. In order to obtain EEG data containing only spontaneous activity of the cerebral cortex, the evaluation sequence of the resting period corresponding to the evaluation window is extracted from the frontal lobe EEG signal, thus avoiding broadband noise introduced by limb movement from the source. To capture the microscopic dynamic changes during the sleep process, a preset sliding window is used to divide the resting phase assessment sequence into multiple temporal local sequences, ensuring that the continuous evolution of the EEG state can be analyzed with high temporal resolution. It should be noted that in this embodiment, the duration of the preset sliding window is set to 30 seconds and the sliding step size is 5 seconds, ensuring that the window area corresponding to each sliding contains at least three complete EEG slow wave cycles, while controlling the computational load while ensuring temporal resolution. Implementers can set the duration and sliding step size of the preset sliding window according to the accuracy requirements of signal analysis or processor performance, which is not limited here.

[0026] Considering that in practice, frontal lobe EEG signals are easily mixed in with eye movement signals, and these signals typically exhibit low-frequency and highly ordered waveforms, which can easily interfere with subsequent frequency domain and time domain analyses, leading to misjudgments of deep sleep, this embodiment performs pre-processing hard rejection before feature calculation. This involves calculating the difference between the maximum and minimum values ​​of each local temporal sequence as a reference analysis value, while setting a preset eye movement threshold of 150 μV. This value is based on the typical amplitude range of frontal lobe EMG and eye movements, and can be adjusted by the user according to sensor sensitivity. If the reference analysis value of any local temporal sequence is greater than the preset eye movement threshold, that local temporal sequence is marked as an eye movement artifact and rejected, and no further feature calculations are performed, ensuring the purity of the subsequent sleep curve fitting. If the reference analysis value of that local temporal sequence is less than or equal to the preset eye movement threshold, subsequent analysis is performed.

[0027] It is known that in insomnia assessment, both genuine deep sleep slow waves and certain artifact signals (such as eye movements and teeth grinding electromyography) exhibit low-frequency, high-energy characteristics in their frequency domain energy distribution. Therefore, it is difficult to distinguish between deep sleep slow waves and low-frequency artifacts (such as eye movements and teeth grinding electromyography) based solely on a single frequency domain feature. Furthermore, artifact signals often have a chaotic temporal structure. To achieve a high-confidence assessment of arousal level, this embodiment obtains an arousal deviation index for each local temporal sequence based on its frequency domain morphology and temporal order. By introducing temporal topological features as a nonlinear gating mechanism, it accurately reflects whether the EEG signal simultaneously satisfies the energy distribution characteristics and structural order characteristics of deep sleep. The larger the arousal deviation index, the more the EEG signal deviates from the ideal deep sleep state (it may be fast waves during wakefulness, or it may be artifacts with high low-frequency energy but chaotic temporal structure), indicating a higher arousal level or stronger interference. Conversely, the smaller the arousal deviation index, the more the cerebral cortex is in a state of true deep inhibition.

[0028] Preferably, in one feasible embodiment, the evaluation window is obtained by: firstly calculating the triaxial resultant acceleration magnitude of the limb acceleration signal, that is, the triaxial component of each sampling point in the limb acceleration sequence. Euclidean norm (This is common knowledge and will not be repeated.) The triaxial combined acceleration magnitude can comprehensively reflect the overall motion intensity of the limbs in any direction. To eliminate high-frequency noise caused by sensor jitter or physiological muscle tremors, and to retain the low-frequency components reflecting macroscopic changes in limb posture, the triaxial combined acceleration magnitude is low-pass filtered (e.g., using a low-pass filter with a 1Hz cutoff frequency) to obtain a limb activity intensity curve, making the data curve smoother and reflecting the true trend of limb activity. Considering that some insomniacs frequently turn over or make minor adjustments to their posture during the early stages of sleep, these brief periods of stillness are insufficient as a stable EEG assessment benchmark. The first time point in the limb activity intensity curve where the fluctuation amplitude of the detected value is less than the preset static judgment threshold and meets the preset duration is marked as the assessment start time, indicating that the subject has entered a physiologically stable period suitable for monitoring EEG physiological characteristics. In this embodiment, the preset static judgment threshold is set to 0.01g (unit of gravitational acceleration) to ensure that only micro-movements such as breathing are captured, excluding obvious limb displacement. The preset duration is set to 5 minutes to ensure that the subject has completely transitioned from a state of awake agitation to a relatively peaceful resting state. The implementer can set the static judgment threshold and preset duration according to the sensor sensitivity or clinical scenario requirements, which are not limited here. To obtain sufficiently long and stable EEG data for analyzing the complete sleep onset process, this embodiment uses the assessment start time as the starting point and extracts a preset time period as the assessment window to provide a unified time reference for subsequent signal analysis. This embodiment sets the preset length to 40 minutes to ensure that it covers the entire process from rest to entering light sleep and even deep sleep for most insomniac subjects, avoiding the inability to capture the complete decline in arousal level due to an excessively short truncation. The implementer can set the preset length based on the expected average sleep latency; no limitation is imposed here.

[0029] Preferably, in one possible implementation of this embodiment, the method for obtaining the arousal deviation index is described in [reference needed]. Figure 2 The document presents a flowchart of a method for obtaining an arousal deviation index, as provided in this embodiment. The method includes the following steps: Step S201: For any local time-domain sequence, obtain the degree of deviation of the frequency domain shape of the local time-domain sequence based on the frequency domain shape of the local time-domain sequence.

[0030] To ensure clarity and reasonableness, this embodiment uses a local time-domain sequence as an example for analysis. Considering the significant differences in the frequency domain energy distribution of EEG signals under different physiological states (e.g., high high-frequency energy during wakefulness and high low-frequency energy during deep sleep), the frequency domain morphology of this local time-domain sequence is used to obtain the degree of deviation of the frequency domain morphology, accurately reflecting whether the distribution characteristics of the current EEG energy spectrum at different frequencies conform to the typical pattern of deep sleep. The greater the degree of deviation of the frequency domain morphology, the higher the proportion of high-frequency components in the current EEG signal, and the more it deviates from the slow-wave characteristics of deep sleep; conversely, the smaller the degree of deviation of the frequency domain morphology, the closer the frequency domain energy structure of the current EEG signal is to the ideal deep sleep.

[0031] In one possible implementation of this embodiment, the method for obtaining the frequency domain morphological deviation is as follows: To obtain the energy density distribution of the EEG signal at different frequencies, a Fast Fourier Transform is performed on the local time-domain sequence to obtain the power spectral density. The power spectral density within the target frequency band is extracted and its probability density is normalized to obtain the cumulative spectral distribution of the local time-domain sequence. This eliminates individual differences in the absolute signal intensity and retains only the relative proportion of frequency domain energy in different frequency bands. To establish a standard reference system for measuring the degree of deep sleep, the method is based on the power-law characteristics of deep sleep (which typically follow...). The decay law, in which, This paper constructs a benchmark reference distribution for the power spectrum of ideal deep sleep, and normalizes the probability density of the benchmark reference distribution to obtain the benchmark spectrum cumulative distribution, which represents the theoretical energy distribution in the purest and deepest slow-wave sleep state. It should be noted that this embodiment sets the target frequency band to 0.5Hz-30Hz, forcibly eliminating DC and extremely low-frequency components below 0.5Hz to prevent… The model's infinity near 0Hz causes the normalization denominator to overflow; To quantify the morphological distance between the current actual EEG spectrum and the ideal deep sleep spectrum, a one-dimensional Wasserstein distance algorithm is used to obtain the arithmetic mean of the absolute values ​​of the differences between the cumulative spectral distribution and the baseline cumulative spectral distribution at each frequency point. This arithmetic mean serves as the degree of frequency domain morphological deviation of the local time-domain sequence, effectively capturing the overall structural differences between probability distributions and exhibiting good sensitivity to minute spectral drifts. The Fast Fourier Transform, EEG power-law properties, and the one-dimensional Wasserstein distance algorithm are all well-known and will not be elaborated upon further.

[0032] Step S202: Based on the temporal order of the local temporal sequence, obtain the degree of non-uniformity of the connection of the local temporal sequence.

[0033] Considering the limitations of single-frequency domain feature analysis, such as the intrusion of alpha waves in a highly aroused state (ordered temporal morphology, small frequency domain morphological deviation) or electromyographic artifacts (chaotic temporal morphology, small frequency domain morphological deviation), frequency domain features are often similar, making it difficult to distinguish between deep sleep and highly disturbed states. In order to achieve accurate discrimination of the true state of EEG signals, this embodiment obtains the degree of connectivity inhomogeneity of the local temporal sequence based on its temporal order, accurately reflecting the structural orderliness of the EEG signal waveform in the temporal domain. The greater the degree of connectivity inhomogeneity, the more significant the slow wave component in the EEG signal (a typical feature of deep sleep, with ordered waveforms, few and prominent peaks, and extremely uneven connectivity distribution), indicating an ordered temporal structure. Conversely, the smaller the degree of connectivity inhomogeneity, the more random fluctuation artifacts or high-frequency, low-amplitude fast waves are mixed in the signal (a feature of the awake or light sleep stage, with relatively uniform node connectivity), indicating a chaotic temporal structure.

[0034] In one possible implementation of this embodiment, the method for obtaining the degree of connection non-uniformity is as follows: In order to reduce the computational complexity of subsequent complex network construction while preserving the main temporal contour of the EEG signal, this embodiment downsamples the local temporal sequence to a preset low-frequency sampling rate to obtain a discrete time point sequence, effectively reducing the amount of data and improving computational efficiency. In this embodiment, the preset low-frequency sampling rate is set to 50Hz to ensure that it is sufficient to capture the temporal characteristics of slow waves (0.5Hz-4Hz) and most sleep spindle waves (12Hz-14Hz). The implementer can set the preset low-frequency sampling rate according to the limitations of computing resources and the target frequency band, which is not limited here. To map a one-dimensional time series into a complex network that characterizes its topology, transforming the nonlinear dynamics of the time-domain signal into the topological connectivity of network nodes, and while significantly reducing the computational cost of single network construction while preserving local waveform structure features, thus avoiding the nonlinear growth of the number of edges and memory overflow risks caused by excessively long sequences, the discrete time-point sequence is divided into equal-length sub-sequences. Then, a horizontal visibility chart algorithm is used to construct a complex network for each sub-sequence, obtaining the connectivity degree of each node in the complex network. For example, for any two time points in a sub-sequence... and If any midpoint between the two satisfy Then, a connection edge is established between node i and node j to obtain the connection degree of each node in the complex network (i.e., the number of connection edges each node has); in this embodiment, the duration of each sub-segment time series is set to 5 seconds to ensure that each sub-segment sequence contains at least two complete EEG slow wave cycles (calculated at 0.5Hz, with a cycle of 2 seconds), so as to maintain the statistical stability of the topological connection relationship while reducing the computational complexity. The implementer can set the duration of the sub-segment time series according to the concurrent computing power of the hardware processor and the lower limit of the target analysis frequency band, which is not limited here; To quantify the non-uniformity of connectivity distribution, the mean Gini coefficient of the connectivity degree for all sub-segments of the time series is calculated as the degree of connectivity non-uniformity in the local time-domain sequence. This is because the Gini coefficient effectively reflects the clustering of connectivity degree distribution. During deep sleep, a few peak nodes have extremely high connectivity degrees (many points are visible), while most trough nodes have extremely low connectivity degrees, resulting in a high Gini coefficient. In awake or artifact states, waveform fluctuations are frequent and have similar amplitudes, leading to a relatively uniform distribution of node connectivity degrees and a lower Gini coefficient. The methods for obtaining discrete time-point sequences, the horizontal visibility chart algorithm, and the Gini coefficient acquisition methods are all well-known and will not be elaborated further.

[0035] Step S203: Based on the frequency domain morphological deviation degree and connectivity inhomogeneity, obtain the arousal deviation index of the local time-domain sequence.

[0036] It is known that deep sleep EEG signals have the dual characteristics of concentrated low-frequency energy (small frequency domain morphological deviation) and ordered waveform structure (high degree of connection non-uniformity). In order to improve the accuracy of recognizing the true deep sleep state and reduce the misjudgment rate of EMG artifacts and high arousal state, this embodiment obtains the arousal deviation index of the local time-domain sequence based on the degree of frequency domain morphological deviation and connection non-uniformity, so as to achieve effective recognition and elimination of artifact signals.

[0037] The specific method for obtaining the arousal deviation index is as follows: To establish a reference benchmark that reflects the individualized deep sleep characteristics of the subject, this embodiment first obtains multiple connectivity unevenness levels within the subject's historical effective deep sleep periods. The average of these multiple connectivity unevenness levels is used as the preset benchmark connectivity unevenness level, thereby achieving personalized adaptation of the assessment model. Considering that new subjects visiting for the first time may lack sufficient historical deep sleep data, if the cumulative duration of historical effective deep sleep periods is less than a preset sleep duration threshold, a preset empirical constant is used as the preset benchmark connectivity unevenness level, serving as a general cold-start parameter. This embodiment sets the preset sleep duration threshold to 60 minutes to ensure a sufficient sample size for calculating a stable average value; the preset empirical constant is set to 0.5, obtained based on the statistical average of deep sleep EEG data from healthy individuals, ensuring effective assessment even without historical data. Implementers can set the preset sleep duration threshold and preset empirical constant based on the statistical results of clinical databases; no limitations are imposed here. To construct a nonlinear gating mechanism that is extremely sensitive to temporal clutter, the difference between a preset baseline level of connection nonuniformity and the preset truncation threshold is compared with the smaller value, and a preset exponential function is applied. The processed result serves as the temporal reference value for this local temporal sequence. When the temporal order of the signal decreases (the degree of connection non-uniformity decreases), the difference between the preset baseline connection non-uniformity and the connection non-uniformity increases. After being amplified by the preset exponential function, the value increases sharply, achieving strong suppression of low-order signals. Considering that the preset exponential function has a very high nonlinear amplification effect on small fluctuations in input variables, if the EEG signal is subjected to extreme non-stationary noise (such as sudden electromagnetic pulse interference), the temporal structure will be completely deconstructed. To prevent unbounded exponential amplification from causing floating-point overflow and to ensure the computational continuity of the nonlinear gating model under severe interference, this embodiment compares the difference between the preset baseline connection non-uniformity and the connection non-uniformity with a preset truncation threshold. If the difference between the preset baseline connection non-uniformity and the connection non-uniformity is greater than the preset truncation threshold... This indicates that the temporal order of the current EEG signal has been severely distorted, and this deviation exceeds the reasonable dynamic range of physiological arousal level evolution. Directly incorporating it into the calculation will cause a numerical explosion, resulting in the arousal deviation index losing its discriminative power. In order to avoid computational collapse caused by extreme abnormal signals and to ensure that the fusion index can maintain numerical stability and logical monotonicity under strong interference, a preset truncation threshold is used in the subsequent preset exponential function processing. In this embodiment, the preset truncation threshold is set to 0.5. This value ensures that the exponential term can still be maintained within the resolvable numerical range under the maximum penalty. The implementer can set the preset truncation threshold according to the processor's floating-point operation precision and the expected anti-interference redundancy, which is not limited here. Wherein, w is the smaller value between the preset baseline connection non-uniformity and the connection non-uniformity and the preset truncation threshold; exp is an exponential function with the natural constant as the base. To adjust the weighting of time-domain features on the final evaluation result and avoid excessive oscillation of indicators due to minor fluctuations, the result of multiplying the time-domain reference value by a preset reference weight and adding it to a first preset constant is used as the effective reference value for the local time-domain sequence. In this embodiment, the preset reference weight is set to 1 as a balancing coefficient to ensure that the penalty for time-domain features is moderate. The first preset constant is set to 1 as a base multiplier to ensure that when the time-domain features are ideal (difference close to 0, exponent close to 1), the effective reference value is close to 2, resulting in a linear multiplicative effect on the frequency-domain features. Implementers can set the preset reference weight and the first preset constant according to their needs for sensitivity to artifacts, which is not limited here. To comprehensively reflect the degree of EEG signal deviation from deep sleep and to forcibly suppress artifact signals using time-domain features, the product of the frequency-domain morphological deviation degree and the effective reference value is compared with a preset upper limit, and the smaller value is taken as the arousal deviation index for that local time-domain sequence. This embodiment sets the preset upper limit to 1 because the exponential term in the nonlinear gating model may cause the calculation result to far exceed the reasonable logical range when dealing with extremely chaotic artifact signals, thus causing the subsequent steady-state variance threshold and steady-state mean threshold based on the normalized range to fail, resulting in a logical deadlock. To cut off the propagation chain of local numerical distortion to the subsequent time-series fitting stage, when the product of the frequency-domain morphological deviation degree and the effective reference value is greater than 1, the arousal deviation index is forcibly set to 1. Output saturation truncation ensures that the evaluation index strictly closes within the specified range. The range ensures the numerical stability of the threshold determination throughout the entire process. When encountering electromyographic artifacts, although the frequency domain morphological deviation may be small (high energy at low frequencies), the degree of connectivity inhomogeneity is extremely low, leading to an explosive increase in the effective reference value. This significantly raises the final arousal deviation index, correctly identifying it as a high arousal or high interference state.

[0038] The arousal damping characteristic parameter acquisition module 30 is used to acquire arousal damping characteristic parameter characterizing a single sleep-inducing process based on the decay of the arousal deviation index over time within the evaluation window.

[0039] Specifically, the process of falling asleep is physiologically characterized by a dynamic evolution of arousal levels transitioning from a high to a low state. This evolution can be analogized to a damped energy dissipation dynamic system. To objectively quantify the regulatory evolution of the subject's physiological state during sleep induction and avoid the loss of time-series information due to relying solely on the sleep latency as a single indicator, a wakefulness damping characteristic parameter is obtained based on the decay of the wakefulness deviation index over time within the assessment window. This parameter accurately reflects the resistance characteristics during the transition from wakefulness to inhibition by fitting the rate of decrease in arousal level curve. A larger wakefulness damping characteristic parameter indicates a slower decrease in arousal level and a more sluggish transition to inhibition, indicating greater resistance to falling asleep. Conversely, a smaller wakefulness damping characteristic parameter indicates a rapid decay in arousal level, a more significant improvement effect of the sleep induction process on the physiological state, and less resistance to falling asleep.

[0040] Preferably, in one feasible embodiment, the method for obtaining the arousal damping characteristic parameter is as follows: First, the arousal deviation index within the evaluation window is arranged according to the time sequence of the corresponding local time domain sequence to obtain the arousal deviation index sequence, so as to construct a dynamic curve that fully reflects the sleep-onset process; in order to avoid the overall fitting rate of decline being lowered by the stable data (tailing) of the arousal level at a low level after falling asleep, the arousal deviation index sequence is then subjected to sliding variance detection, that is, the variance change of the arousal deviation index sequence within the specified sliding window is calculated, thereby obtaining the target time when the arousal deviation index sequence enters the steady state, which is beneficial to accurately capture the dynamic process of a significant decrease in arousal level, that is, to evaluate the time period corresponding to the start time to the target time, thereby improving the accuracy and representativeness of subsequent nonlinear fitting. The method for obtaining the target time is as follows: Obtain the sliding variance sequence of the arousal deviation index sequence. This involves setting a specified sliding window and sliding it across the arousal deviation index sequence, obtaining the sequence composed of the variances of the arousal deviation index within each sliding window. In this embodiment, the duration of the specified sliding window is set to 5 minutes, and the sliding step size is 30 seconds to ensure that the macroscopic trend of arousal level fluctuations is captured, while avoiding statistical noise caused by an excessively short window. The implementer can set the duration and sliding step size of the specified sliding window according to the average timescale of the sleep onset process; no limitation is imposed here. To automatically identify unstable states of the nervous system using statistical features... The transition point from a high-excitation state to a stable low-inhibition state is identified. The first time point in the sliding variance sequence where the value is below a preset steady-state variance threshold and the mean of the arousal deviation index up to the corresponding time point is less than a preset steady-state mean threshold is then used as the target time. In this embodiment, the preset steady-state variance threshold is set to 0.005 to ensure that periods still in the arousal fluctuation phase are excluded. The preset steady-state mean threshold is set to 0.3 (under normalized range) to ensure that the captured steady state is indeed a deep sleep state rather than a continuous awake and still state. The implementer can set the preset steady-state variance threshold and the preset steady-state mean threshold according to the distribution characteristics of the normalized data; no limitation is imposed here. For ease of description, this embodiment uses the time period from the start time to the target time as the target time period. In order to quickly screen for abnormal or ineffective adjustment processes, the arousal deviation index within the target time period is fitted into a straight line in chronological order. When the slope of the straight line is greater than or equal to 0, it indicates that the arousal level has not decreased or has even increased during the assessment period, that is, the subject has failed to fall asleep or the physiological high arousal state has intensified. At this time, the arousal damping characteristic parameter of this sleep process is marked as an invalid null value to prevent abnormal fixed extreme values ​​from interfering with the long-term evolution analysis of the subsequent time series. When the slope of the straight line is less than 0, it indicates that the arousal level is declining and the sleep-inducing process is effective. Considering that real EEG signals are affected by physiological rhythm regulation, their temporal characteristics have local high-frequency random oscillations. If the original sequence carrying oscillation noise is directly input into the nonlinear optimizer, the least squares algorithm is very likely to get stuck in a local minimum and fail to converge. Therefore, the arousal deviation index within the target time period is subjected to moving average filtering to obtain a macroscopic descending envelope sequence that removes high-frequency random fluctuations. In order to avoid the algorithm generating negative damping or floating baselines that violate physical laws due to abnormal data jumps during boundless optimization, under the constraint that the decay time constant is greater than 0 and the baseline is within a preset range, the macroscopic descending envelope sequence is fitted nonlinearly using an exponential decay model. In this embodiment, the preset range of the baseline is set as follows: No specific restrictions are imposed here; If the fit converges and the fitting residual is less than the preset residual threshold, it indicates that the extracted envelope curve highly conforms to the exponential decay dynamics, and the confidence level of this evaluation data is high. The decay time constant obtained from the fit is then used as a wakefulness damping characteristic parameter to accurately characterize the magnitude of physiological inhibitory resistance during the sleep onset process. If the fit does not converge or the fitting residual is greater than or equal to the preset residual threshold, it indicates that the subject's physiological state fluctuates drastically during the current monitoring period and does not possess typical sleep onset decay characteristics (e.g., frequent micro-awakening). In this case, the wakefulness damping characteristic parameter for this sleep onset process is marked as an invalid null value to ensure the rigor of the output parameters. The exponential decay model is as follows: ;in, Let be the arousal deviation index of the t-th local sequence in the target time period; t be the relative time corresponding to the t-th local sequence in the target time period; T be the decay time constant to be fitted; A be the initial amplitude to be fitted; C be the baseline to be fitted; and e be the natural constant. Linear fitting and nonlinear least squares fitting are well-known and will not be elaborated further.

[0041] The assessment index acquisition module 40 is used to obtain long-term sleep improvement assessment indicators based on the changing trend of the arousal damping characteristic parameters during multiple sleep episodes.

[0042] Specifically, to overcome the limitations of single physiological state measurements being susceptible to interference from occasional factors such as the daily environment (e.g., noise and temperature) or the subject's emotions (e.g., anxiety), this study aims to objectively evaluate the sleep improvement status of subjects over a long-term monitoring period. Furthermore, based on the changing trends of arousal damping characteristic parameters during multiple sleep episodes, long-term sleep improvement assessment indicators can be obtained. This integrates the microscopic single-episode feature evolution over time into macroscopic long-term change patterns, which is beneficial for providing quantitative and robust physiological state assessment feedback. It can also assist in evaluating the effectiveness of long-term sleep guidance methods and whether dynamic adjustments to the monitoring subjects' work and rest environment or guidance methods are necessary.

[0043] Preferably, in one possible implementation of this embodiment, the method for obtaining the long-term sleep improvement assessment index is as follows: First, the arousal damping characteristic parameters are arranged according to the time sequence of the corresponding sleep onset process. After removing data marked as invalid null values, a sequence of arousal damping characteristic parameters is obtained, constructing a time series reflecting the change of sleep onset resistance with the monitoring period, which is beneficial for extracting long-term evolution trends from the time series data. Considering that the accumulated sample size is small in the early stage of monitoring, in order to provide preliminary feedback on the evolution trend under limited data conditions, the effective sample size of the arousal damping characteristic parameter sequence is further obtained as the accumulation count. When the accumulation count is... When the number of occurrences equals 1 (i.e., the first effective monitoring), it indicates that there is only one element in the sequence, and the slope cannot be calculated. Therefore, the target slope is directly set to 0. When the cumulative number of occurrences is greater than 1 but less than the preset number threshold, it indicates that the sample size is too small to support a complex robust regression algorithm. Therefore, the slope of the straight line fitted by the wakefulness damping feature parameter sequence is obtained by the least squares method as the target slope to provide preliminary trend feedback. In this embodiment, the preset number threshold is set to 3 to ensure that there are at least three data points to form a stable basis for linear relationship estimation. Implementers can set the preset number threshold according to the minimum sample requirement of the regression algorithm, which is not limited here. When the cumulative number of measurements exceeds or equals a preset threshold, to eliminate the potential bias of outliers (such as a subject experiencing extreme difficulty falling asleep due to sudden external noise, forming an outlier) on the overall characteristic evolution assessment, the slope of the straight line obtained from RANSAC (Random Sample Consensus) robust regression analysis of the wakefulness damping characteristic parameter sequence is used as the target slope. This automatically identifies and removes outliers that do not conform to the overall trend, ensuring the robustness of the assessment results. To visually demonstrate sleep improvement to the subjects (it is generally expected that sleep resistance will decrease as monitoring progresses), the inverse of the target slope is used as a long-term sleep improvement assessment indicator. A positive value indicates a decreasing trend in sleep resistance and positive improvement in sleep status; the larger the value, the faster the improvement rate. RANSAC robust regression analysis is well-known and will not be elaborated further.

[0044] It is known that long-term sleep improvement assessment indicators are positively correlated with the degree of improvement in sleep characteristics. If the long-term sleep improvement assessment indicator is significantly positive, it indicates that as the number of monitoring sessions increases, the arousal damping characteristic parameter shows a stable downward trend (slope is negative), that is, the subject's resistance to falling asleep is effectively reduced, and this trend is not affected by abnormal data on a single day, indicating that the subject's ability to regulate sleep onset gradually returns to normal over a long period of time. If the long-term sleep improvement assessment indicator is close to zero or negative, it indicates that the evolution of sleep characteristics has not improved significantly or has relapsed, suggesting that the current sleep guidance method or sleep environment has not effectively alleviated the cortical hyperarousal state, and can be used as an objective reference for adjusting sleep intervention strategies.

[0045] In summary, this embodiment includes a data acquisition module, a wakefulness deviation index acquisition module, a wakefulness damping characteristic parameter acquisition module, and an evaluation index acquisition module. It collects and aligns the subject's limb acceleration signals and frontal lobe EEG signals in real time; identifies limb stillness periods as evaluation windows, extracts corresponding stillness EEG sequences and segments them, calculates the wakefulness deviation index based on frequency domain morphology and temporal order, and uses nonlinear gating to remove artifact interference; obtains the wakefulness damping characteristic parameter for a single sleep onset based on the decay of the wakefulness deviation index; and finally analyzes the changing trend of this parameter across multiple sleep onsets to generate a long-term sleep improvement evaluation index. This invention achieves an objective and accurate assessment of the evolution of the subject's physiological state during sleep induction through time-frequency dual-domain feature fusion and process damping quantification.

[0046] Example 2: This invention also proposes a sleep state monitoring device based on multimodal signals. The device includes a memory and a processor. The memory stores executable program code, and the processor calls and executes the executable program code to perform the sleep state monitoring system based on multimodal signals provided in the embodiments of this application. Specifically, the device may be a chip, component, or module. The chip may include a connected processor and memory; the memory stores instructions, and when the processor calls and executes the instructions, the chip can perform the sleep state monitoring system based on multimodal signals provided in the above embodiments.

[0047] Furthermore, this application also protects a computer device; please refer to [link to relevant documentation]. Figure 3 The computer device includes a memory 401, a processor 402, and a computer program 403 stored in the memory 401 and running on the processor 402. When the processor 402 executes the computer program 403, the computer device can execute any of the aforementioned sleep state monitoring systems based on multimodal signals.

[0048] Example 3: The present invention also provides a computer-readable storage medium storing computer program code, which, when executed on a computer, causes the computer to perform the aforementioned method steps to implement the sleep state monitoring system based on multimodal signals provided in the above embodiments.

[0049] Example 4: The present invention also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to realize the sleep state monitoring system based on multimodal signals provided in the above embodiments.

[0050] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0051] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0052] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A sleep state monitoring system based on multimodal signals, characterized in that, The system includes: The data acquisition module is used to acquire the subject's limb acceleration signals and frontal lobe EEG signals in real time, and to perform time alignment on the limb acceleration signals and frontal lobe EEG signals; The arousal deviation index acquisition module is used to identify the period of continuous limb stillness based on limb acceleration signals as the assessment window, and to extract the resting period assessment sequence corresponding to the assessment window from the frontal lobe EEG signal; the resting period assessment sequence is divided into time-domain local sequences through a preset sliding window, and the arousal deviation index of each time-domain local sequence is obtained according to the frequency domain morphology and time-domain order of each time-domain local sequence. The arousal damping characteristic parameter acquisition module is used to acquire arousal damping characteristic parameters that characterize a single sleep-inducing process based on the decay of the arousal deviation index over time within the evaluation window. The assessment index acquisition module is used to obtain long-term sleep improvement assessment indicators based on the changing trends of arousal damping characteristic parameters during multiple sleep episodes.

2. The sleep state monitoring system based on multimodal signals as described in claim 1, characterized in that, The method for obtaining the evaluation window is as follows: Calculate the triaxial summation magnitude of the limb acceleration signal and perform low-pass filtering on the triaxial summation magnitude to obtain the limb activity intensity curve; The first time point in the limb activity intensity curve where the fluctuation of the detected value is less than the preset static judgment threshold and meets the preset duration is marked as the evaluation start time. Starting from the evaluation start time, a time period of preset length is extracted as the evaluation window.

3. The sleep state monitoring system based on multimodal signals as described in claim 1, characterized in that, The method for obtaining the arousal deviation index is as follows: For any local time-domain sequence, the difference between the maximum value and the minimum value in the local time-domain sequence is obtained and used as a reference analysis value. When the reference analysis value is greater than the preset eye-tracking threshold, the local time-domain sequence is removed. When the reference analysis value is less than or equal to the preset eye-tracking threshold, the degree of deviation of the frequency domain shape of the local time-domain sequence is obtained based on the frequency domain shape of the local time-domain sequence. Based on the temporal order of the local temporal sequence, the degree of non-uniformity of the connection of the local temporal sequence is obtained; The difference between the preset baseline non-uniformity of connection and the non-uniformity of connection is compared with a preset truncation threshold. The smaller value is then processed by a preset exponential function, and the result is used as the time-domain reference value of the local time-domain sequence. The result of multiplying the time-domain reference value by a preset reference weight and adding it to a first preset constant is taken as the effective reference value of the local time-domain sequence. The product of the frequency domain morphological deviation degree and the effective reference value is compared with a preset upper limit value, and the smaller value is taken as the awakening deviation index of the local time-domain sequence.

4. The sleep state monitoring system based on multimodal signals as described in claim 3, characterized in that, The method for obtaining the degree of deviation in the frequency domain shape is as follows: The power spectral density is obtained by performing a fast Fourier transform on the local time-domain sequence. The power spectral density within the target frequency band is then extracted and normalized to obtain the cumulative spectral distribution of the local time-domain sequence. Based on the power law characteristics of EEG during deep sleep, a benchmark reference distribution of the ideal deep sleep power spectrum is constructed, and the probability density of the benchmark reference distribution is normalized to obtain the benchmark spectrum cumulative distribution. The arithmetic mean of the absolute values ​​of the differences between the cumulative spectral distribution and the reference cumulative spectral distribution at each frequency point is obtained using the one-dimensional Wasserstein distance algorithm, which is used as the degree of deviation of the frequency domain morphology of the local time-domain sequence.

5. The sleep state monitoring system based on multimodal signals as described in claim 3, characterized in that, The method for obtaining the degree of unevenness in the connection is as follows: The local time-domain sequence is downsampled to a preset low-frequency sampling rate to obtain a discrete time-point sequence; The discrete time point sequence is divided into equal-length sub-time series, and a complex network is constructed for each sub-time series using the horizontal visibility graph algorithm to obtain the connection degree of each node in the complex network. The mean of the Gini coefficients corresponding to the connectivity degree of each sub-segment of the time series is used as the degree of connectivity non-uniformity of the local time series.

6. The sleep state monitoring system based on multimodal signals as described in claim 2, characterized in that, The method for obtaining the awakening damping characteristic parameters is as follows: The arousal deviation indexes within the evaluation window are arranged according to the time order of the corresponding local time sequence to obtain the arousal deviation index sequence; Sliding variance detection is performed on the arousal deviation index sequence to obtain the target time when the arousal deviation index sequence enters a steady state; The time period from the start time to the target time is taken as the target time period. The arousal deviation index within the target time period is fitted into a straight line in chronological order. When the slope of the straight line is greater than or equal to 0, the arousal damping characteristic parameter of this sleep process is marked as an invalid null value. When the slope of the straight line is less than 0, the arousal deviation index within the target time period is subjected to moving average filtering to obtain a macroscopic descending envelope sequence. Under the constraint that the decay time constant is greater than 0 and the baseline is within the preset range, the macroscopic descending envelope sequence is fitted with nonlinear least squares using an exponential decay model. If the fitting converges and the fitting residual is less than the preset residual threshold, the decay time constant obtained from the fitting is obtained as an arousal damping characteristic parameter. If the fit does not converge or the fitting residual is greater than or equal to the preset residual threshold, the arousal damping characteristic parameter of this sleep process is marked as an invalid null value.

7. The sleep state monitoring system based on multimodal signals as described in claim 6, characterized in that, The method for obtaining the target time is as follows: Obtain the sliding variance sequence of the arousal deviation index sequence, and detect the first time point in the sliding variance sequence where the value is lower than a preset steady-state variance threshold and the mean of the arousal deviation index up to the corresponding time point is less than a preset steady-state mean threshold, and take it as the target time point.

8. The sleep state monitoring system based on multimodal signals as described in claim 1, characterized in that, The method for obtaining the long-term sleep improvement assessment indicators is as follows: Arrange the arousal damping characteristic parameters according to the time sequence of the corresponding sleep-onset process, and remove the data marked as invalid null values ​​to obtain the arousal damping characteristic parameter sequence; The number of valid samples in the awakening damping feature parameter sequence is used as the cumulative count. When the cumulative count is equal to 1, 0 is set as the target slope. When the cumulative number of times is greater than 1 and less than the preset number of times threshold, the slope of the straight line fitted by the awakening damping characteristic parameter sequence is obtained by the least squares method and used as the target slope. When the cumulative number of times is greater than or equal to the preset number of times threshold, the slope of the straight line obtained by performing RANSAC robust regression analysis on the awakening damping characteristic parameter sequence will be used as the target slope. The negative of the target slope is used as an indicator for assessing long-term sleep improvement.

9. The sleep state monitoring system based on multimodal signals as described in claim 3, characterized in that, The method for obtaining the degree of non-uniformity of the preset benchmark connection is as follows: The average value of the connectivity unevenness of the subjects during their historical effective deep sleep periods was obtained, and the average value of the multiple connectivity unevenness was used as the preset baseline connectivity unevenness. If the cumulative duration of historical effective deep sleep periods is less than the preset sleep duration threshold, then the preset empirical constant will be used as the preset baseline for connection unevenness.

10. The sleep state monitoring system based on multimodal signals as described in claim 6, characterized in that, The exponential decay model is as follows: ;in, denoted as the arousal deviation index of the t-th local sequence in the target time period; t is the relative time corresponding to the t-th local sequence in the target time period; T is the decay time constant to be fitted; A is the initial amplitude to be fitted; C is the baseline to be fitted; and e is the natural constant.

Images