Sleep light awakening method based on user sleep curve

Abstract

The invention discloses a sleep mild wake-up method based on a user sleep curve, and belongs to the technical field of sleep monitoring, and the method comprises the steps: collecting a physiological signal of a user, generating a sleep stage curve by using a pre-trained sleep stage model, and carrying out the parallel analysis of heart rate variability and respiratory coordination to generate a mood index curve; overlapping and fusing the two curves to form a sleep-mood alignment feature set; in a preset wake-up time range, analyzing the sleep stage and psychological state of the user according to the feature set, and dynamically determining an optimal wake-up starting opportunity; when the clock arrives, an instruction is sent to the linkage alarm clock, and sound, light and touch multi-mode stimulation is triggered in sequence in a cooperative mode; in the wake-up process, the sleep depth of the user is continuously monitored, if it is detected that the depth recovery exceeds the critical threshold value, the stimulation intensity is adaptively adjusted until the depth falls back, it is ensured that the user naturally wakes up under the condition that the discomfort is the lowest, and intelligent mild wake-up is achieved.

Description

Technical Field

[0001] This invention belongs to the field of sleep monitoring technology, specifically a method for mild sleep arousal based on the user's sleep curve. Background Technology

[0002] With increasing attention to sleep quality, the wake-up process, as a crucial final step in the sleep cycle, directly impacts a user's physical and mental state upon waking and during the day. Current wake-up technologies often rely on fixed-time alarms, neglecting the user's actual sleep cycle and depth. For example, a sudden alarm interrupts deep sleep, leading to difficulty waking, dizziness, and lethargy – symptoms of "sleep inertia" that negatively affect daytime work and life efficiency. While some existing smart wake-up devices attempt to adjust wake-up timing by incorporating sleep monitoring, most rely solely on single body movement signals to determine sleep stage, resulting in low accuracy and difficulty in accurately capturing changes in sleep depth. Furthermore, the lack of real-time response to dynamic changes in sleep state during the wake-up process means that if the user's sleep depth rebounds during wake-up, the wake-up intensity cannot be adjusted promptly, leading to a poor wake-up experience. Therefore, there is an urgent need for a method for mild sleep arousal that can accurately monitor users' sleep status, generate high temporal resolution sleep curves and mood index curves, and dynamically adjust the timing and intensity of awakening based on the curves, in order to solve the problems of low awakening accuracy and poor user experience in existing technologies. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a mild sleep arousal method based on user sleep curves. It collects the user's physiological signals, generates sleep stage curves using a pre-trained sleep staging model, and analyzes heart rate variability and respiratory coordination in parallel to generate a mood index curve. The two curves are superimposed and fused to form a sleep-mood alignment feature set. Within a preset arousal time range, the user's sleep stage and psychological state are analyzed based on this feature set to dynamically determine the optimal arousal initiation time. When the time arrives, a command is sent to a linked alarm clock to sequentially trigger multimodal stimuli of sound, light, and touch. During the arousal process, the user's sleep depth is continuously monitored. If the detected depth rise exceeds a critical threshold, the stimulation intensity is adaptively adjusted until the depth falls back, ensuring the user wakes up naturally with minimal discomfort, thus achieving intelligent and gentle arousal.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] Methods for mild sleep arousal based on user sleep curves include:

[0006] Physiological signals are continuously collected by a sensor array in the sleeping mat and preprocessed to obtain a set of preprocessed physiological signals.

[0007] Based on the preprocessed set of physiological signals, a sleep stage curve is generated using a pre-trained sleep staging model. Simultaneously, heart rate variability and respiratory coordination are analyzed in parallel to generate a mood index curve. The sleep curve data and the mood index curve data are superimposed to form a sleep-mood alignment feature set.

[0008] Within the preset wake-up time range, the optimal wake-up initiation time is determined based on the sleep-mood alignment feature set. When the optimal wake-up initiation time is reached, a control command is sent to the linked alarm clock to trigger multimodal stimulation in a preset order.

[0009] During the wake-up process, the sleep pad continuously monitors the user's sleep status and obtains the real-time sleep depth level. If the user's sleep depth level is detected to rise above the critical threshold, the control terminal immediately adjusts the alarm clock intensity parameters until the sleep depth level is detected to fall back to the critical threshold again.

[0010] Specifically, the pre-trained sleep staging model is a hybrid model based on convolutional neural network-long short-term memory network; the training of the hybrid model is based on multi-channel sleep graph data labeled with N1, N2, N3, REM and wakefulness stages, and the training process uses cross-entropy loss function and Adam optimizer;

[0011] The hybrid model takes as input the preprocessed physiological signal set the body pressure signal, heart vibration signal and respiratory vibration signal, and outputs the sleep stage label at 30-second intervals. The sleep stage labels at consecutive time points are arranged in chronological order to generate a sleep stage curve. The sleep stages include wakefulness, REM sleep, light sleep and deep sleep.

[0012] Specifically, the parallel analysis of heart rate variability and respiratory coordination to generate a mood index curve includes:

[0013] Peak detection is performed on the heartbeat vibration signals in the preprocessed physiological signal set, the time difference between two adjacent heartbeat intervals is calculated to obtain the heartbeat interval sequence, and the heart rate variability index is calculated based on the heartbeat interval sequence; the heart rate variability index includes time domain index and frequency domain index; the frequency domain index includes low-frequency component, high-frequency component, and the ratio of low-frequency component to high-frequency component;

[0014] The time-domain and frequency-domain indicators are input into the pre-trained first mood regression model, and the first mood index sub-score is output.

[0015] Envelope extraction and peak detection are performed on the respiratory vibration signals in the preprocessed physiological signal set to generate a continuous respiratory cycle sequence, and the variance and disorder index of the respiratory cycle sequence are calculated.

[0016] Analyze the phase synchronicity between respiration and heartbeat to generate a coordination index for respiratory sinus arrhythmia;

[0017] The variance, disorder index, and respiratory sinus arrhythmia coordination index of the respiratory cycle sequence are input into the pre-trained second mood regression model, and the second mood index sub-score is output.

[0018] The first and second mood index sub-scores are weighted and merged according to preset weights to generate a mood index value at a time node of 30 seconds. The mood index values ​​at consecutive time nodes are arranged in chronological order to generate a mood index curve.

[0019] Specifically, the step of overlaying sleep curve data and mood index curve data to form a sleep-mood alignment feature set includes:

[0020] Align the sleep stage curve and the mood index curve on the time axis;

[0021] For each identical timestamp, a feature vector is constructed; the feature vector includes at least the sleep stage code, mood index value, and the sleep stage change trend and mood index change trend extracted from the previous preset time window for that timestamp, and the interval of the timestamp is consistent with the time node interval of the sleep stage curve and the mood index curve.

[0022] The feature vectors of all timestamps are arranged in chronological order to form the sleep-mood alignment feature set.

[0023] Specifically, determining the optimal wake-up initiation time based on the sleep-mood alignment feature set includes setting wake-up initiation criteria: the sleep stage label value in the feature pair is greater than or equal to 3, and the mood index value is... satisfy ;

[0024] Within the preset wake-up time range, the feature pairs in the sleep-mood alignment feature set are traversed. When the feature pairs of three consecutive time nodes meet the wake-up start judgment conditions for the first time, the first time node is determined as the best wake-up start time.

[0025] Specifically, in the process of sending control commands to the linked alarm clock, the control unit of the sleep mat establishes a communication connection with the control module of the linked alarm clock via Bluetooth protocol; when the optimal wake-up start time is reached, the control unit generates control commands; the control commands are sent to the control module of the linked alarm clock via Bluetooth communication; after receiving the control commands, the control module of the linked alarm clock parses the parameter information in the commands and starts multimodal stimulation in a preset order.

[0026] Specifically, multimodal stimuli are triggered in a pre-defined sequence, including:

[0027] The control instructions include at least the stimulus type, intensity parameter, and triggering sequence;

[0028] The linked alarm clock first triggers the first modal stimulus according to the control command; the first modal stimulus is a gradually increasing white noise auditory stimulus;

[0029] After the white noise auditory stimulus has lasted for a first preset duration, a second modal stimulus is triggered by superimposing a control command; the second modal stimulus is a gradually increasing warm-colored light visual stimulus.

[0030] If the user is not detected leaving the bed after the warm light visual stimulation has lasted for a second preset time, the third modal stimulation is triggered by superimposing the control command; the third modal stimulation is the tactile stimulation of the micro-vibration unit built into the sleep pad.

[0031] Specifically, the control terminal immediately adjusts the alarm clock intensity parameters, including:

[0032] After the control unit of the sleep mat detects that the sleep depth level has risen above the critical threshold, it sends an adjustment request signal to the control terminal through the communication protocol. The adjustment request signal includes the current sleep depth level, the type of multimodal stimulation that has been triggered, and the current intensity parameters. The intensity parameters include the intensity of light, sound, and vibration. The critical threshold is a preset sleep depth level value, which corresponds to the critical state in which the user is easily awakened and difficult to fall asleep again.

[0033] After receiving the adjustment request signal, the control terminal calculates the new alarm clock intensity parameters according to the preset intensity adjustment rules.

[0034] The control terminal encapsulates the new intensity parameters into adjustment instructions and sends them to the control module of the linked alarm clock;

[0035] After receiving the adjustment command, the control module of the alarm clock immediately updates the intensity parameters and executes the adjusted stimulus output.

[0036] Specifically, during the wake-up process, the sleep pad continuously monitors the user's sleep state and obtains real-time sleep depth levels, including:

[0037] After the arousal stimulus is triggered, the sensor array continues to collect real-time physiological signals;

[0038] The real-time physiological signals are preprocessed in real time and input into a preset lightweight real-time sleep depth assessment model.

[0039] The real-time sleep depth assessment model outputs a quantified real-time sleep depth level; the real-time sleep depth level is a continuous value or discrete score from deep sleep to wakefulness.

[0040] Specifically, the sleep-based mild wake-up method based on the user's sleep curve also includes: recording data of the wake-up process; the control unit of the sleep mat records the wake-up initiation time, the real-time sleep depth level at each time point, the alarm intensity adjustment time, the intensity parameters before and after adjustment, and the changes in sleep depth level in real time, forming sleep record data, and storing the sleep record data in the local storage module; after the wake-up process is completed, the sleep record data is synchronized to the control terminal, and the control terminal organizes the sleep record data and generates a wake-up process report.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. This invention proposes a method for mild sleep awakening based on the user's sleep curve. By integrating multidimensional data analysis of sleep stages and mood index, it can intelligently select the best time to start waking up when the user is in a light sleep stage and in a relatively pleasant mood within a preset time range. This breaks through the limitations of the traditional alarm clock's fixed time and rough wake-up, reduces physiological discomfort and psychological resistance during the wake-up process, and greatly improves the comfort and awakening experience when waking up.

[0043] 2. This invention proposes a method for mild sleep arousal based on the user's sleep curve. By constructing a complete system that includes real-time monitoring, intelligent decision-making, multimodal stimulation, and closed-loop feedback regulation, this invention can dynamically respond to changes in the user's sleep state during the arousal process. When the system detects that the user has a tendency to fall asleep again, it can automatically adjust the stimulation intensity to effectively maintain the arousal effect, thereby ensuring the final success of the arousal process and achieving a highly intelligent, personalized, gentle, and reliable natural arousal. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the sleep light awakening method based on the user's sleep curve according to the present invention;

[0045] Figure 2 This is a flowchart illustrating the principle of the sleep-light awakening method based on the user's sleep curve in this invention.

[0046] Figure 3 This is a flowchart illustrating the generation of the sleep-mood alignment feature set in the sleep-light awakening method based on user sleep curves of this invention. Detailed Implementation

[0047] Example 1:

[0048] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a method for mild sleep arousal based on a user's sleep curve, the method comprising S1 to S4, including the following steps:

[0049] S1: Physiological signals are continuously collected through the sensor array in the sleeping mat and preprocessed to obtain a set of preprocessed physiological signals;

[0050] Furthermore, the sensor array in the sleeping mat includes a piezoelectric thin film sensor, a millimeter-wave radar sensor, and a temperature sensor. The piezoelectric thin film sensor is used to collect the user's body movement signals and respiratory vibration signals. The millimeter-wave radar sensor is used to collect the user's heart rate signals and chest rise and fall signals. The temperature sensor is used to collect the sleep environment temperature signal. The sensor array uses the collected body movement signals, respiratory vibration signals, heart rate signals, chest rise and fall signals, and sleep environment temperature signals together as the physiological signals.

[0051] S2: Based on the preprocessed physiological signal set, a sleep stage curve is generated using a pre-trained sleep staging model. At the same time, heart rate variability and respiratory coordination are analyzed in parallel to generate a mood index curve. The sleep curve data and the mood index curve data are superimposed to form a sleep-mood alignment feature set.

[0052] S3: Within the preset wake-up time range, determine the optimal wake-up initiation time based on the sleep-mood alignment feature set. When the optimal wake-up initiation time is reached, send a control command to the linked alarm clock to trigger multimodal stimulation in a preset order.

[0053] In the process of sending control commands to the linked alarm clock, the control unit of the sleep mat establishes a communication connection with the control module of the linked alarm clock via Bluetooth protocol; when the optimal wake-up start time is reached, the control unit generates a control command; the control command is sent to the control module of the linked alarm clock via Bluetooth communication, and after receiving the control command, the control module of the linked alarm clock parses the parameter information in the command and starts multimodal stimulation in a preset order.

[0054] Furthermore, the preset wake-up time range is 30-60 minutes before the user sets the wake-up time. This time range is set by the user through the interactive interface of the control terminal. The control terminal stores the set wake-up time range in the local storage module and sends it to the control unit of the sleep mat simultaneously.

[0055] S4: During the wake-up process, the sleep pad continuously monitors the user's sleep status and obtains the real-time sleep depth level. If the user's sleep depth level is detected to rise above the critical threshold, the control terminal immediately adjusts the alarm clock intensity parameters until the sleep depth level is detected to fall back to the critical threshold again.

[0056] Furthermore, if it is detected that the user's sleep depth level has risen above a critical threshold, the method includes: during the execution of the multimodal stimulation, comparing the real-time sleep depth level with the critical threshold in real time; if the real-time sleep depth level changes from below the critical threshold to above the critical threshold, it is determined that the user's sleep depth level has risen above the critical threshold.

[0057] Furthermore, the control terminal immediately adjusts the alarm intensity parameters, including:

[0058] (1) The control terminal receives the judgment result of the sleep depth level rebound and immediately generates a new control command according to the preset adjustment strategy;

[0059] The preset adjustment strategy refers to a set of predefined rules, algorithms, or logic used by the control terminal to calculate and generate new stimulation parameters after determining that the user's sleep depth has rebounded. This mainly includes:

[0060] Progressive increase rule: Increase the intensity parameter of the currently triggered multimodal stimulus by a preset fixed step value;

[0061] Control algorithm rules: Based on the difference between the real-time sleep depth level and the critical threshold, a new intensity parameter is calculated through a proportional-integral-derivative control algorithm. The proportional-integral-derivative control algorithm is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0062] Modal coordination rule: While increasing the intensity of the current modality stimulus, activate or enhance the stimulus of another modality to carry out coordinated intervention;

[0063] Personalized rules: Based on the user's historical response data, select the most effective stimulus modality for priority adjustment and adopt personalized adjustment range.

[0064] (2) The new control command is used to increase the intensity parameter of at least one of the currently triggered multimodal stimuli; the intensity parameter is volume for auditory stimuli, brightness for visual stimuli, and vibration amplitude for tactile stimuli.

[0065] Furthermore, when the real-time sleep depth level drops from above the critical threshold to below or equal to the critical threshold, it is determined to have fallen back to the critical threshold.

[0066] The sleep-based mild wake-up method based on the user's sleep curve also includes recording data of the wake-up process. The control unit of the sleep pad records the wake-up initiation time, the real-time sleep depth level at each time point, the alarm intensity adjustment time, the intensity parameters before and after the adjustment, and the changes in sleep depth level in real time, forming sleep record data. The sleep record data is stored in the local storage module. After the wake-up process is completed, the sleep record data is synchronized to the control terminal. The control terminal organizes the sleep record data and generates a wake-up process report.

[0067] Furthermore, the sleep-based mild wake-up method based on the user's sleep curve also includes a wake-up success determination step: if, after the wake-up process is completed, the sensor array continuously detects that the user's body movement signal exceeds the activity threshold and continues for a certain period of time within the preset bed-leaving detection time, then the wake-up is determined to be successful; if no movement signal is detected, then the wake-up is determined to be unsuccessful, and a backup wake-up scheme is activated.

[0068] Furthermore, the sleep-based mild awakening method based on user sleep curves also includes data feedback and model optimization steps: the complete data of each awakening process, including the final sleep-mood alignment feature set, awakening timing decision, stimulus parameter adjustment record, and awakening success or failure result, is uploaded to the cloud server; this is used to incrementally learn and optimize the pre-trained sleep staging model, first mood regression model, second mood regression model, and real-time sleep depth assessment model.

[0069] Example 2:

[0070] Please see Figure 3 In this embodiment, the pre-trained sleep staging model is a hybrid model based on convolutional neural network-long short-term memory network; the training of the hybrid model is based on multi-channel sleep graph data labeled with N1, N2, N3, REM and wakefulness stages, and the training process uses cross-entropy loss function and Adam optimizer;

[0071] Furthermore, N1, N2, and N3 are non-rapid eye movement (NREM) sleep, while REM sleep is rapid eye movement (REM) sleep. Together, they constitute a complete human sleep cycle, which lasts approximately 90-120 minutes and typically occurs 4-5 times per night.

[0072] Furthermore, N1 is the first stage of non-rapid eye movement sleep, which is the initial stage of light sleep. It is the introductory stage of the transition from wakefulness to sleep, and it is the lightest sleep state with a relatively short duration.

[0073] Furthermore, N2 is the second stage of non-rapid eye movement sleep, also known as the main stage of light sleep. It is the stage with the highest proportion of nighttime sleep, accounting for about 45%-55% of total sleep. It follows N1 and is the core part of light sleep, as well as the longest stage in the sleep cycle.

[0074] Furthermore, N3 is the third stage of non-rapid eye movement sleep, also known as deep sleep or slow-wave sleep, which is the deepest sleep state and usually accounts for 15%-25% of total sleep at night, mostly concentrated in the first half of the night.

[0075] Furthermore, REM sleep, also known as dream sleep, is the most unique stage in the sleep cycle. It is characterized by rapid and irregular left-right movements of the eyes and typically accounts for 20%-25% of total nighttime sleep. As the sleep cycle progresses, the duration of REM sleep gradually increases.

[0076] The hybrid model takes as input the preprocessed physiological signal set the body pressure signal, heart vibration signal and respiratory vibration signal, and outputs the sleep stage label at 30-second intervals. The sleep stage labels at consecutive time points are arranged in chronological order to generate a sleep stage curve. The sleep stages include wakefulness, REM sleep, light sleep and deep sleep.

[0077] Furthermore, the hybrid model takes as input the preprocessed physiological signal set the body motion pressure signal, heart rate vibration signal, and respiratory vibration signal, and outputs sleep stage labels at 30-second intervals, including:

[0078] (1) The preprocessed body pressure signal, heart vibration signal and respiratory vibration signal are first sent into the hybrid model. Then, the three signals are standardized respectively, that is, the value of each signal is subtracted from its mean and then divided by its standard deviation, so that each signal is converted into standardized data with a mean of zero and a standard deviation of one, ensuring that the hybrid model will not assign an unreasonable weight to any signal because the value of any signal is large.

[0079] (2) According to the recognized standards in the field of sleep staging, a time window is set. This time window is usually set to thirty seconds and is called a period. The system uses a fixed sampling rate to slide and extract a data segment from the continuous data stream in units of thirty seconds without overlap. Each thirty-second segment contains three synchronized and standardized signal data. This data segment serves as the input for one forward propagation of the hybrid model.

[0080] (3) The deep convolutional neural network in the hybrid model performs convolution operations on the three input signals in parallel. The convolution kernel is like a miniature sliding window that moves on the time axis. Its core function is to automatically and efficiently extract the local patterns of the signal. For example, shorter convolution kernels are good at capturing subtle changes such as the interval between heartbeats and the single cycle of the respiratory wave, while longer convolution kernels are good at recognizing macroscopic trends such as long-term signal drift caused by body movement and turning over, and the overall acceleration or deceleration of the respiratory rhythm. At the same time, the convolution operation generates feature maps in each layer. These feature maps are essentially high-level representations of the original signal after filtering and compression, highlighting the parts of the signal that are useful for the classification task and suppressing irrelevant noise. Finally, each 30-second signal segment outputs a highly abstract and highly condensed feature vector. This feature vector comprehensively represents the complex patterns of the user's body movement, heartbeat and breathing during that time period.

[0081] (4) The system arranges the feature vectors corresponding to multiple consecutive 30-second segments in chronological order to form a feature vector sequence. The feature vector sequence is input into the Long Short-Term Memory Network in sequence. The Long Short-Term Memory Network has a unique internal control mechanism that can autonomously decide which long-term information to remember, which irrelevant information to forget, and how to combine historical information with the current input. Through this mechanism, when the hybrid model judges the current 30-second sleep stage, it will comprehensively consider the evolution trend of the sleep state over the previous 10 minutes or even longer. For example, if the user was in a deep sleep period for a period of time, the possibility of suddenly switching to the awake period at the current moment is very low. It is more likely to be measurement noise or a slight movement of the body. The output of the Long Short-Term Memory Network is another feature vector sequence containing historical context information. Each feature vector corresponds to a 30-second time point, but its meaning already includes the historical background of the previous period.

[0082] (5) The feature vector after time-series modeling is passed to the final classifier layer, which is usually a fully connected layer with a softmax activation function. The classifier layer maps the high-dimensional feature vector to a four-dimensional probability distribution vector. These four dimensions correspond to four sleep stages: wakefulness, REM sleep, light sleep, and deep sleep. The value at each position in the probability distribution vector is a probability value between zero and one. The sum of all four probability values ​​is one, and this probability distribution represents the confidence of the model in that the current 30-second signal segment belongs to each sleep stage.

[0083] (6) Generate the final sleep stage label. Usually, the winner-takes-all rule is adopted, that is, the sleep stage corresponding to the largest of the four probability values ​​is selected as the final judgment label of the 30-second segment. For example, if the probability distribution output by the hybrid model is 5% for the awake period, 10% for the REM period, 80% for the light sleep period and 5% for the deep sleep period, the system will judge the time period as the light sleep period.

[0084] (7) The system repeats the process of (1)-(6) to process the continuous signal stream of the whole night in 30-second increments, thereby obtaining a series of sleep stage labels arranged in chronological order. After encoding these sleep stage labels with their first letter in English, they are arranged in chronological order on the time axis to generate the final sleep stage curve.

[0085] The parallel analysis of heart rate variability and respiratory coordination generates a mood index curve, including:

[0086] A1: Peak detection is performed on the heartbeat vibration signal in the preprocessed physiological signal set, the time difference between two adjacent heartbeat intervals is calculated to obtain the heartbeat interval sequence, and the heart rate variability index is calculated based on the heartbeat interval sequence; the heart rate variability index includes time domain index and frequency domain index; the frequency domain index includes low-frequency component, high-frequency component, and the ratio of low-frequency component to high-frequency component;

[0087] Furthermore, time-domain metrics are directly calculated from the statistical characteristics of the heart rate interval sequence, reflecting the overall amplitude and short-term changes in heart rate variability. The most commonly used time-domain metrics include: the standard deviation of all normal heart rate intervals, reflecting the overall amplitude of heart rate variation throughout the recording period; a larger value indicates higher heart rate variability, usually signifying better cardiac function and adaptability; the root mean square (RMS) value of the difference between adjacent heart rate intervals, calculated by first squaring the differences between all adjacent intervals, then averaging these squares, and finally taking the square root of this average; this metric is particularly sensitive to parasympathetic activity and effectively reflects rapid, transient changes in heart rate; and the percentage of times the difference between adjacent heart rate intervals exceeds a specific millisecond value out of all intervals, which is also used to quantify rapid, short-term changes in heart rate.

[0088] Furthermore, frequency domain indices, through spectral analysis, decompose the rhythmic variations inherent in the intercardiac sequence into different frequency components, thereby revealing the balance between the sympathetic and parasympathetic nervous systems in the autonomic nervous system. Since the intercardiac sequence is non-uniformly sampled in time, direct spectral analysis is difficult. Therefore, it is first necessary to resample it into a uniform time series using cubic spline interpolation at a frequency of 4 Hz. Subsequently, a Fast Fourier Transform (FFT) algorithm is applied to this uniform time series, transforming the signal from the time domain to the frequency domain and outputting a power spectral density map. This power spectral density map shows the distribution of signal power at different frequencies. Frequency bands are identified on the obtained power spectral density map, including:

[0089] Low-frequency components: Their frequency range is usually between 0.04 Hz and 0.15 Hz. The power in this frequency band mainly reflects the joint regulatory activities of the sympathetic and parasympathetic nervous systems, but is also affected by other factors such as vasomotor tension and the renin-angiotensin system.

[0090] High-frequency components: Their frequency range is usually between 0.15 Hz and 0.4 Hz. The power in this frequency band is highly synchronized with the respiratory cycle, and therefore mainly reflects the activity of the parasympathetic nervous system.

[0091] The ratio of the power of the low-frequency component to the power of the high-frequency component is calculated. This ratio provides a simple indicator of the balance between sympathetic and parasympathetic activity. An increase in the ratio is usually interpreted as a relative increase in sympathetic activity or a relative decrease in parasympathetic activity.

[0092] Furthermore, the process of obtaining the intercardiac interval sequence includes:

[0093] (1) Receive the preprocessed heartbeat vibration signal; the preprocessed heartbeat vibration signal has been denoised and filtered to present a clear heartbeat waveform. It should be noted that each heartbeat cycle usually produces a most prominent and sharp peak point on the signal. This peak point corresponds to the beginning of the heart ventricular contraction and can be compared with the R wave peak in the electrocardiogram.

[0094] (2) A peak detection algorithm is used to automatically identify peak points. Specifically, the peak detection algorithm uses a sliding window to traverse the entire preprocessed heartbeat vibration signal. Within the window, the peak detection algorithm finds points that meet the following three conditions: First, the amplitude value of the point must be the maximum value among the multiple points on its left and right sides, i.e., a local maximum value; Second, the amplitude value of the point must exceed an adaptive dynamic threshold. This adaptive dynamic threshold is dynamically adjusted according to the overall amplitude level of the signal to exclude small and meaningless fluctuations; Finally, the minimum time interval requirement must be met between two adjacent identified peak points. The minimum time interval requirement is set based on the fastest possible heart rate physiologically to prevent multiple fluctuations in one heartbeat from being misjudged as multiple heartbeats, thereby outputting an accurate timestamp. Each timestamp corresponds to the moment when a heartbeat peak occurs.

[0095] (3) After receiving the timestamp sequence of the heartbeat peak, the system calculates the time difference between two adjacent timestamps. This time difference is the time interval between two adjacent heartbeats, i.e., the heartbeat interval. All heartbeat interval values ​​are calculated and arranged in chronological order to obtain the initial heartbeat interval sequence.

[0096] (4) Calculate the median and interquartile range of the entire initial heartbeat interval sequence. Then, set a range, such as the upper and lower limits being the median plus N times and minus N times the dispersion statistic, respectively. Any heartbeat interval falling outside this range will be marked as an outlier. For outliers, use linear interpolation to calculate the replacement value using the normal interval values ​​before and after it, thereby generating a cleaned and smoothed heartbeat interval sequence.

[0097] A2: Input the time-domain index and frequency-domain index into the pre-trained first mood regression model and output the first mood index sub-score;

[0098] Furthermore, the specific steps of A2 include:

[0099] (1) Collect a set of heart rate variability indicators for different individuals under different emotional states, including all time-domain and frequency-domain indicators calculated. At the same time, collect labels, which are the actual mood scores corresponding to these heart rate variability data.

[0100] (2) Load a multi-layer fully connected neural network model architecture. The number of neurons in the input layer of the multi-layer fully connected neural network model must be equal to the number of input heart rate variability indicators. Then connect one or more hidden layers. Each hidden layer contains multiple neurons, and the last output layer contains only one neuron, which is responsible for outputting the final prediction score. After the design is completed, all internal parameters of the multi-layer fully connected neural network model architecture, such as the connection weights and bias terms between neurons, will be initialized to random small values. The multi-layer fully connected neural network model is the prior art in this field and is not an inventive solution of this application. It will not be described in detail here.

[0101] (3) The heart rate variability index of different individuals under different emotional states is collected and input into the multilayer fully connected neural network model. The data is passed from the input layer to the next layer. Each neuron receives the output values ​​of all neurons in the previous layer, sums them by weight and adds a bias value, and then produces the output value of the neuron through a non-linear activation function, such as the ReLU function. Finally, the neuron in the output layer produces a predicted mood score.

[0102] (4) The predicted mood score output by the multilayer fully connected neural network model is compared with the actual mood state score of the sample using the mean squared error loss function. The squared value obtained is the loss value of this prediction. The calculation formula of the mean squared error loss function is the prior art in this field and is not an inventive solution of this application. It will not be elaborated here.

[0103] (5) Calculate the gradient of each model parameter, such as each weight and bias, with respect to the final loss value through the backpropagation algorithm. Then, use the Adam optimizer to adjust all model parameters proportionally according to the calculated gradient direction. The direction of adjustment is to reduce the loss value. The backpropagation algorithm and the Adam optimizer are existing technologies in this field and are not the inventive solutions of this application. They will not be described in detail here.

[0104] (6) Based on the collection of heart rate variability indexes of different individuals under different emotional states, iterative training is carried out to obtain a well-trained first mood regression model;

[0105] (7) Based on the time domain and frequency domain indices calculated by the current user in real time, a feature vector to be predicted is constructed and standardized.

[0106] (8) The standardized feature vector is fed into the pre-trained first mood regression model and flows through the network along the same forward propagation path as during training. The neurons in each layer perform their fixed calculations: weighted summation, biasing, and activation function. Finally, the signal is transmitted to the output layer, outputting a unique neuron. The calculation result of this neuron is a continuous real value. This value is the inference score of the first mood regression model for the user's current emotional state, i.e., the first mood index sub-score. The first mood index sub-score is a relative value. Its absolute size may not be meaningful, but its trend of change effectively reflects the emotional fluctuations represented by the state of the autonomic nervous system. The higher the score, the more the mood state inferred from the heart rate variability tends to be positive and relaxed.

[0107] A3: Envelope extraction and peak detection are performed on the respiratory vibration signals in the preprocessed physiological signal set to generate a continuous respiratory cycle sequence, and the variance and disorder index of the respiratory cycle sequence are calculated.

[0108] Furthermore, the envelope is a smooth curve that outlines the trend of the highest point of signal amplitude.

[0109] Furthermore, the specific steps in A3 include:

[0110] (1) Receive the pre-processed respiratory vibration signal; the pre-processed respiratory vibration signal exhibits a sine wave-like fluctuation that is synchronized with the respiratory cycle, with the signal amplitude increasing during inhalation and decreasing during exhalation;

[0111] (2) The Hilbert transform method is used to extract the envelope, including: first, applying the Hilbert transform to the preprocessed respiratory signal, then summing the squares of the preprocessed respiratory signal and its Hilbert transform result, and then calculating the square root to finally obtain a smooth envelope line that outlines the amplitude profile of each breath. The Hilbert transform is a prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0112] (3) The envelope obtained by traversing the peak detection algorithm is used to output accurate timestamps. Each timestamp corresponds to the peak moment when inhalation is completed in a respiratory cycle, and the timestamp sequence of respiratory peaks is obtained. The peak detection algorithm traverses the envelope with a sliding window. Within the window, the peak detection algorithm finds points that meet specific conditions: First, the amplitude value of the point must be the maximum value among the multiple points on its left and right sides, that is, it is a local maximum. Second, the amplitude value of the point must exceed an adaptive dynamic threshold. The dynamic threshold is dynamically calculated based on the median amplitude of the envelope over a period of time to exclude false peaks that are too small and may be caused by noise. Finally, the minimum and maximum time interval requirements set based on the fastest and slowest possible physiological respiratory rates must be met between two adjacent identified peak points to prevent false detection.

[0113] (4) After receiving the timestamp sequence of respiratory peaks, the system calculates the time difference between two adjacent peaks. The time difference is the time taken for a complete respiratory cycle, or the cycle value. All these time differences are calculated and arranged in chronological order to obtain the initial respiratory cycle sequence.

[0114] (5) The initial respiratory cycle sequence is preprocessed using a statistical outlier detection method to generate a respiratory cycle sequence;

[0115] (6) Calculate the arithmetic mean of all values ​​in the entire respiratory cycle sequence, and calculate the difference between each individual cycle value in the respiratory cycle sequence and this mean. Then, square each difference and sum all the squares. Finally, divide this sum by the total number of time differences in the respiratory cycle sequence. The result is the variance. The larger the variance, the more violently the respiratory cycle length fluctuates around the mean and the more unstable the respiratory rhythm. The smaller the variance, the more stable the breathing and the stronger the rhythm.

[0116] (7) Calculate the absolute value of the difference between all two adjacent respiratory cycles in the respiratory cycle sequence, and calculate the arithmetic mean of all absolute differences to obtain the first mean. Then calculate the arithmetic mean of the entire respiratory cycle sequence itself to obtain the respiratory cycle average. Finally, divide the first mean by the respiratory cycle average to obtain the disorder index.

[0117] A4: Analyze the phase synchronicity between respiration and heartbeat to generate a coordination index for respiratory sinus arrhythmia.

[0118] Furthermore, the specific steps for A4 include:

[0119] (1) Obtain respiratory cycle sequence and inter-cardiac interval sequence;

[0120] (2) A linear phase interpolation model is used to perform phase allocation for each respiratory cycle in the respiratory cycle sequence, including: defining the starting phase of a respiratory cycle as zero degrees and the ending phase as 360 degrees. For each time point in this cycle, its phase value is linearly increased according to the relative position of the time point in the whole cycle. The linear phase interpolation is the prior art in this field and is not an inventive solution of this application. It will not be described in detail here.

[0121] It needs to be explained that by using a linear phase interpolation model, the system transforms the continuous breathing signal into a breathing phase time series that continuously cycles and monotonically increases from zero to 360 degrees. The breathing phase time series clearly indicates at any given moment where the breathing motion is in its own cycle, for example, whether it is at the beginning of inhalation or in the middle of exhalation.

[0122] (3) For each heartbeat timestamp, query the generated respiratory phase time sequence, and calculate the continuous respiratory phase value corresponding to the precise heartbeat time through interpolation, thereby obtaining the respiratory phase value sequence triggered by the heartbeat event;

[0123] It needs to be explained that this process assigns a respiratory phase label to each heartbeat. For example, one heartbeat may occur when the respiratory phase is 90 degrees, and another heartbeat may occur when the respiratory phase is 270 degrees.

[0124] (4) Using the method of cyclic statistics, the respiratory phase value corresponding to each heartbeat is regarded as a point on a unit circle. That is, each phase value is converted into its two-dimensional coordinates on the unit circle. Then, the average abscissa and average ordinate of all points are calculated to obtain the average coordinate point. The distance from the center of the circle to the average coordinate point is the final coordination index.

[0125] Furthermore, the range of the coordination index is [0,1]. If all heartbeat points fall on the same position on the unit circle, their average coordinate point is also at this position, and its length is 1. If all heartbeat points are evenly distributed on the entire circumference, their average coordinates will converge at the center of the circle, and its length is 0.

[0126] Furthermore, the closer the coordination index is to 1, the more significant the respiratory sinus arrhythmia is, and the stronger the regulatory effect of respiration on heart rate. This usually means that parasympathetic nerve activity is dominant, and the body and mind are in a more relaxed and coordinated state. The closer the coordination index is to zero, the more it indicates that the rhythms of the two are decoupled, the autonomic nervous system regulation function may be impaired, or the patient is in a state of stress or tension.

[0127] A5: Input the variance, disorder index, and respiratory sinus arrhythmia coordination index of the respiratory cycle sequence into the pre-trained second mood regression model and output the second mood index sub-score;

[0128] Furthermore, the construction and training process of the second mood regression model includes:

[0129] (1) Obtain a set of breathing pattern indicators under different emotional states; the set of breathing pattern indicators includes the variance of the respiratory cycle sequence, the disorder index, and the coordination index of respiratory sinus arrhythmia.

[0130] (2) Obtain the true mood score recorded synchronously with the breathing pattern index; the true mood score is obtained by having the subject complete a standardized psychological questionnaire while collecting physiological signals;

[0131] (3) Loading a shallow, fully connected neural network model architecture;

[0132] It should be noted that since there are only three input features, the input layer only needs to be set with three neurons. Then, one or more hidden layers are connected. Each hidden layer contains multiple neurons and is responsible for learning the non-linear interaction relationship between features. The last output layer contains only one neuron and is used to output the final prediction score. All internal parameters of the shallow fully connected neural network model are initialized to random small values. This is a starting point.

[0133] (4) The obtained breathing pattern index set under different emotional states is input into a shallow fully connected neural network model. The three feature values ​​of the sample in each breathing pattern index set enter from the input layer and are passed through the network layer by layer. Each neuron in each layer receives the output values ​​of all neurons in the previous layer, performs a weighted summation and adds a bias value. Then the result is passed through a non-linear activation function, such as a rectified linear unit function, to generate the output of the neuron. Finally, the neurons in the output layer generate a second predicted mood score.

[0134] (5) The second predicted mood score is compared with the actual mood state score using the mean square error method to obtain the final loss value. The mean square error is the prior art in this field and is not an inventive solution of this application. It will not be described in detail here.

[0135] (6) The gradient of each model parameter with respect to the final loss value is calculated by backpropagation algorithm. Then, the Adam optimizer is used to adaptively adjust all model parameters according to the calculated gradient direction. The adjustment direction is to make the overall loss value continue to decrease.

[0136] (7) Repeat the iteration until the performance of the shallow fully connected neural network model reaches its optimal level and stabilizes on the reserved validation dataset. At this point, the model training is complete, its parameters are fixed and saved, and the second mood regression model is obtained.

[0137] A6: The first mood index sub-score and the second mood index sub-score are weighted and merged according to a preset weight to generate a mood index value at a time node of 30 seconds. The mood index values ​​at consecutive time nodes are arranged in chronological order to generate a mood index curve.

[0138] The process of overlaying sleep curve data and mood index curve data to form a sleep-mood alignment feature set includes:

[0139] B1: Align the sleep stage curve and the mood index curve on the time axis;

[0140] B2: For each identical timestamp, construct a feature vector; the feature vector shall at least include the sleep stage code, mood index value, and the sleep stage change trend and mood index change trend extracted from the previous preset time window, and the interval of the timestamps shall be consistent with the time node interval of the sleep stage curve and the mood index curve.

[0141] B3: Arrange the feature vectors of all timestamps in chronological order to form the sleep-mood alignment feature set.

[0142] The step of determining the optimal wake-up initiation time based on the sleep-mood alignment feature set includes setting wake-up initiation criteria: the sleep stage label value in the feature pair is greater than or equal to 3, and the mood index value is... satisfy ;

[0143] Within the preset wake-up time range, the feature pairs in the sleep-mood alignment feature set are traversed. When the feature pairs of three consecutive time nodes meet the wake-up start judgment conditions for the first time, the first time node is determined as the best wake-up start time.

[0144] Multimodal stimuli are triggered in a pre-defined sequence, including:

[0145] C1: The control command includes at least the stimulus type, intensity parameter, and triggering sequence;

[0146] C2: The linked alarm clock first triggers the first modal stimulus according to the control command; the first modal stimulus is a progressively stronger white noise auditory stimulus;

[0147] C3: After the white noise auditory stimulus has lasted for a first preset duration, a second modal stimulus is triggered by superimposing a control command; the second modal stimulus is a gradually increasing warm-colored light visual stimulus;

[0148] Furthermore, the steps in C3 include:

[0149] (1) When the first modal stimulus is initiated, that is, while the progressively stronger white noise auditory stimulus is being performed, the timer is activated and starts running. The sole task of the timer is to continuously and in real time accumulate the time that the first modal stimulus has been played.

[0150] (2) During the entire first modal stimulus playback, the timer will continuously compare the current cumulative duration with the stored first preset duration. If and only if the value read by the timer reaches or exceeds the first preset duration, an internal trigger signal will be automatically generated. The trigger signal is an event marker, which means that the first preset duration condition has been met and a request is made to start the second modal stimulus. The first preset duration is a fixed value preset by humans. In this invention, the first preset duration is five minutes.

[0151] (3) After the trigger signal is generated, the system immediately parses the intensity parameters specific to the second modality visual stimulus from the original control command sent by the control terminal; the intensity parameters include, but are not limited to, light color value, initial brightness value, maximum brightness target value, and the slope or total duration of brightness gradual increase;

[0152] (4) The system loads the intensity parameters into the driver program of the hardware that controls the visual stimulus output, such as the LED light board, to complete the initialization preparation for the output.

[0153] (5) The driver calculates the brightness value that should be output at the current moment based on the loaded gradual increase rule.

[0154] It should be noted that the implementation of the gradual increase rule is a linear gradual increase, which includes: there is an independent and more refined timer inside the driver program, which starts timing from the moment the visual stimulus is activated, and sets the current brightness output value as a linear function of this timing duration. Specifically, the brightness value is equal to the initial brightness value, plus a first coefficient, and then multiplied by the duration that has been sustained. The first coefficient is pre-calculated based on the maximum brightness target value and the expected total gradual increase duration, which determines the speed at which the brightness increases.

[0155] Furthermore, in this way, the actual output brightness of the light source will smoothly, continuously, and linearly increase over time from a very dark starting point to a bright but not dazzling level, perfectly achieving a gradual increase effect.

[0156] Furthermore, through C3, the sensory stimulation presented to the user by the system is no longer singular. The user's auditory system is still receiving white noise that may also be gradually increasing in intensity, while at the same time, the user's visual system begins to receive a warm-colored light stimulus with gradually increasing brightness.

[0157] Furthermore, the two different modalities of stimulation are seamlessly superimposed in time: the auditory stimulus begins first, followed by the visual stimulus after a period of time. The two work together for a period of time to form a synergistic effect. This superposition method is gentler than triggering two stimuli at the same time, giving the nervous system a gradual adaptation process and greatly reducing the risk of being startled, which is in line with the core principle of mild arousal.

[0158] C4: If the user is not detected leaving the bed after the warm light visual stimulation has lasted for a second preset time, the third modal stimulation is triggered by superimposing the control command; the third modal stimulation is the tactile stimulation of the micro-vibration unit built into the sleep pad.

[0159] Furthermore, when executing C4, the first step the system performs is to ensure that the second modal stimulus, namely the gradually increasing warm-colored light visual stimulus, has been successfully triggered and continuously output. At the same time, an independent high-precision timer dedicated to monitoring the visual stimulus is activated and starts running. The task of this timer is to continuously accumulate the time that the visual stimulus has been output.

[0160] Furthermore, the second preset duration represents the time window that the auditory and visual bimodal synergistic stimulation is considered to require to produce an effective arousal effect, as determined by the system design. In this invention, it is set to three minutes.

[0161] Furthermore, the system's internal timing condition is met only when the duration of the visual stimulus reaches or exceeds the second preset duration. At this point, the system does not immediately terminate the first two types of stimulation, but instead initiates an exit-from-bed status verification procedure. This procedure uses a sensor array built into the sleep mat, such as a distributed pressure sensor, to determine the user's in-bed status in real time. It analyzes the current pressure distribution map: if a wide and stable distribution of pressure centers is detected, the user is determined to still be in bed; if the pressure centers disappear or only scattered at the edges are detected, such as when the user sits up, the user is determined to have left the bed. Based on the result of the exit-from-bed status verification, a branch decision is made. This branch decision includes: if the verification result indicates the user has left the bed, the awakening is deemed successful, and the system generates an instruction to gradually stop all ongoing stimuli. The system records the successful completion of the wake-up process. If the verification result shows that the user has not yet left the bed, it means that the first two rounds of gentle stimulation have failed to achieve the wake-up goal. At this time, if the internal logic conditions of the system have not yet detected that the user has left the bed, the system automatically generates an internal trigger signal. After generating the trigger signal, the system parses the third intensity parameter specific to the third modality tactile stimulation from the original control command sent by the control terminal. The third intensity parameter includes, but is not limited to, the initial intensity of the vibration, the maximum intensity target value, the vibration frequency, and the total duration of intensity increase. The third intensity parameter is loaded into the drive circuit of the micro-vibration unit built into the sleep pad, such as a linear resonant actuator, to complete the initialization preparation of the output. The drive circuit calculates the vibration intensity that should be output at the current moment according to the loading gradient rule.

[0162] Furthermore, the gradual increase rule is implemented in a linear manner: the driving circuit has an independent precision timer that starts timing from the moment the tactile stimulus is activated, and sets the current vibration intensity output value as a linear function of this timing duration. Specifically, the vibration intensity is equal to the initial intensity value, plus a pre-calculated second coefficient, and then multiplied by the duration of continuous vibration. The second coefficient determines the rate of increase in intensity.

[0163] It is important to emphasize that the stimulation of these three different modalities is precisely sequentially superimposed in time and gradually increases in intensity, working together to act on the user's nervous system and maximizing the reliability of arousal. During the entire third modal stimulation output, the system continuously monitors the user's out-of-bed status and sleep depth, forming a closed-loop control. Once the system detects that the user has left the bed or that the sleep depth has significantly decreased, it will adjust or stop the stimulation.

[0164] The control terminal immediately adjusts the alarm clock intensity parameters, including:

[0165] D1: After the control unit of the sleep mat detects that the sleep depth level has risen above the critical threshold, it sends an adjustment request signal to the control terminal through the communication protocol. The adjustment request signal includes the current sleep depth level, the type of multimodal stimulation that has been triggered, and the current intensity parameters. The intensity parameters include the intensity of light, sound, and vibration. The critical threshold is a preset sleep depth level value, which corresponds to the critical state in which the user is easily awakened and difficult to fall asleep again.

[0166] D2: After receiving the adjustment request signal, the control terminal calculates the new alarm clock intensity parameters according to the preset intensity adjustment rules;

[0167] Furthermore, the preset intensity adjustment rules refer to a set of pre-programmed, automated decision-making logic and calculation methods followed by the control terminal after confirming that the stimulation intensity needs to be adjusted. The core objective of these rules is to quickly suppress the user's sleep depth back below the critical threshold in the most effective and user-friendly way, thereby ensuring the ultimate success of the wake-up process.

[0168] Furthermore, the intensity adjustment rules include:

[0169] Adjustment trigger condition: When the real-time sleep depth level continuously monitored by the system changes from below the critical threshold to above the critical threshold, the adjustment trigger condition is immediately met;

[0170] Adjustment mode selection: 1) Single-modal enhancement rule: Only increase the intensity of the most dominant stimulus modality that is currently being output. For example, if it is currently in the stage of only playing white noise, only increase the volume. If it has entered the vibration stage, prioritize increasing the vibration intensity. 2) Multimodal synergy rule: On the basis of the currently dominant stimulus modality, simultaneously activate or enhance another unused stimulus modality for synergistic intervention. For example, while adjusting the volume of white noise, the rule decides to immediately turn on the gradually increasing light stimulation to enhance the arousal effect by using the multi-sensory superposition effect. 3) Mode switching rule: When the intensity of any modality is adjusted more than a certain number of times or after a certain intensity is still ineffective, switch to another stronger stimulus mode. For example, after the combination of sound and light is ineffective, the rule decides to stop them and switch to the highest intensity vibration mode.

[0171] Constraints and Protection Conditions: Rules must include boundary restrictions to prevent unlimited enhancement of stimulation and protect users from fright or injury, including: 1) Intensity Limit Constraint: Set an absolute maximum intensity value for each stimulus modality. Regardless of the calculation result, the adjusted intensity parameter must not exceed this limit. For example, the rule stipulates that the maximum volume must not exceed 70 decibels and the maximum brightness must not exceed 500 lux; 2) Time Decay Constraint: In the final stage of the preset wake-up time range, such as the last 2 minutes, adopt a more aggressive adjustment strategy, such as directly using the maximum allowed intensity, to ensure that the user wakes up on time; 3) Fatigue Protection Rule: If the number of consecutive adjustments exceeds a limit, such as 5 times, it is judged as abnormally difficult to wake up, triggering a backup process, such as stopping wake-up and notifying the user.

[0172] D3: The control terminal encapsulates the new intensity parameters into an adjustment command and sends it to the control module of the linked alarm clock;

[0173] D4: After receiving the adjustment command, the control module of the linked alarm clock updates the intensity parameters and executes the adjusted stimulus output.

[0174] During the wake-up process, the sleep mat continuously monitors the user's sleep state and obtains real-time sleep depth levels, including:

[0175] E1: After being triggered by a wake-up stimulus, the sensor array continues to collect real-time physiological signals;

[0176] E2: The real-time physiological signals are preprocessed in real time and input into a preset lightweight real-time sleep depth assessment model; the lightweight real-time sleep depth assessment model is constructed based on a gradient boosting tree, and the gradient boosting tree is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0177] It's important to note that the system maintains a sliding window buffer in memory. This window continuously adds the latest signal data and discards the oldest, always maintaining a fixed time interval. The length of this window is carefully chosen, typically slightly shorter than the standard 30-second sleep phase cycle, such as 20 or 25 seconds, to achieve an optimal balance between evaluation latency and data sufficiency. Whenever the system needs to perform a new evaluation, for example, triggered once per second, it extracts a segment of signal data from this sliding window, calculated backwards by one window length from the current moment, as the signal data block for this real-time processing.

[0178] E3: The real-time sleep depth assessment model outputs a quantified real-time sleep depth level; the real-time sleep depth level is a continuous value or discrete score from deep sleep to wakefulness.

[0179] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A method for mild sleep arousal based on user sleep curves, characterized in that, include: Physiological signals are continuously collected by a sensor array in the sleeping mat and preprocessed to obtain a set of preprocessed physiological signals. Based on the preprocessed set of physiological signals, a sleep stage curve is generated using a pre-trained sleep staging model. Simultaneously, heart rate variability and respiratory coordination are analyzed in parallel to generate a mood index curve. The sleep curve data and the mood index curve data are superimposed to form a sleep-mood alignment feature set. Within the preset wake-up time range, the optimal wake-up initiation time is determined based on the sleep-mood alignment feature set. When the optimal wake-up initiation time is reached, a control command is sent to the linked alarm clock to trigger multimodal stimulation in a preset order. During the wake-up process, the sleep pad continuously monitors the user's sleep status and obtains the real-time sleep depth level. If the user's sleep depth level is detected to rise above the critical threshold, the control terminal immediately adjusts the alarm clock intensity parameters until the sleep depth level is detected to fall back to the critical threshold again.

2. The method for mild sleep arousal based on user sleep curves as described in claim 1, characterized in that, The pre-trained sleep staging model is a hybrid model based on convolutional neural network-long short-term memory network; the training of the hybrid model is based on multi-channel sleep data labeled with N1, N2, N3, REM and wakefulness stages, and the training process uses cross-entropy loss function and Adam optimizer; The hybrid model takes as input the preprocessed physiological signal set the body pressure signal, heart vibration signal and respiratory vibration signal, and outputs the sleep stage label at 30-second intervals. The sleep stage labels at consecutive time points are arranged in chronological order to generate a sleep stage curve. The sleep stages include wakefulness, REM sleep, light sleep and deep sleep.

3. The method for mild sleep arousal based on user sleep curves as described in claim 2, characterized in that, The parallel analysis of heart rate variability and respiratory coordination generates a mood index curve, including: Peak detection is performed on the heartbeat vibration signals in the preprocessed physiological signal set, the time difference between two adjacent heartbeat intervals is calculated to obtain the heartbeat interval sequence, and the heart rate variability index is calculated based on the heartbeat interval sequence; the heart rate variability index includes time domain index and frequency domain index; the frequency domain index includes low-frequency component, high-frequency component, and the ratio of low-frequency component to high-frequency component; The time-domain and frequency-domain indicators are input into the pre-trained first mood regression model, and the first mood index sub-score is output. Envelope extraction and peak detection are performed on the respiratory vibration signals in the preprocessed physiological signal set to generate a continuous respiratory cycle sequence, and the variance and disorder index of the respiratory cycle sequence are calculated. Analyze the phase synchronicity between respiration and heartbeat to generate a coordination index for respiratory sinus arrhythmia; The variance, disorder index, and respiratory sinus arrhythmia coordination index of the respiratory cycle sequence are input into the pre-trained second mood regression model, and the second mood index sub-score is output. The first and second mood index sub-scores are weighted and merged according to preset weights to generate a mood index value at a time node of 30 seconds. The mood index values ​​at consecutive time nodes are arranged in chronological order to generate a mood index curve.

4. The method for mild sleep arousal based on user sleep curves as described in claim 3, characterized in that, The process of overlaying sleep curve data and mood index curve data to form a sleep-mood alignment feature set includes: Align the sleep stage curve and the mood index curve on the time axis; For each identical timestamp, a feature vector is constructed; the feature vector includes at least the sleep stage code, mood index value, and the sleep stage change trend and mood index change trend extracted from the previous preset time window for that timestamp, and the interval of the timestamp is consistent with the time node interval of the sleep stage curve and the mood index curve. The feature vectors of all timestamps are arranged in chronological order to form the sleep-mood alignment feature set.

5. The method for mild sleep arousal based on user sleep curves as described in claim 4, characterized in that, The process of determining the optimal wake-up initiation time based on the sleep-mood alignment feature set includes: setting wake-up initiation criteria: the sleep stage label value in the feature pair is greater than or equal to 3, and the mood index value is... satisfy ; Within the preset wake-up time range, the feature pairs in the sleep-mood alignment feature set are traversed. When the feature pairs of three consecutive time nodes meet the wake-up start judgment conditions for the first time, the first time node is determined as the best wake-up start time.

6. The method for mild sleep arousal based on user sleep curves as described in claim 5, characterized in that, In the process of sending control commands to the linked alarm clock, the control unit of the sleep mat establishes a communication connection with the control module of the linked alarm clock via Bluetooth protocol; when the optimal wake-up start time is reached, the control unit generates a control command; the control command is sent to the control module of the linked alarm clock via Bluetooth communication, and after receiving the control command, the control module of the linked alarm clock parses the parameter information in the command and starts multimodal stimulation in a preset order.

7. The method for mild sleep arousal based on user sleep curves as described in claim 6, characterized in that, Multimodal stimuli are triggered in a pre-defined sequence, including: The control instructions include at least the stimulus type, intensity parameter, and triggering sequence; The linked alarm clock first triggers the first modal stimulus according to the control command; the first modal stimulus is a gradually increasing white noise auditory stimulus; After the white noise auditory stimulus has lasted for a first preset duration, a second modal stimulus is triggered by superimposing a control command; the second modal stimulus is a gradually increasing warm-colored light visual stimulus. If the user is not detected leaving the bed after the warm light visual stimulation has lasted for a second preset time, the third modal stimulation is triggered by superimposing the control command; the third modal stimulation is the tactile stimulation of the micro-vibration unit built into the sleep pad.

8. The method for mild sleep arousal based on user sleep curves as described in claim 7, characterized in that, The control terminal immediately adjusts the alarm clock intensity parameters, including: After the control unit of the sleep mat detects that the sleep depth level has risen above the critical threshold, it sends an adjustment request signal to the control terminal through the communication protocol. The adjustment request signal includes the current sleep depth level, the type of multimodal stimulation that has been triggered, and the current intensity parameters. The intensity parameters include the intensity of light, sound, and vibration. The critical threshold is a preset sleep depth level value, which corresponds to the critical state in which the user is easily awakened and difficult to fall asleep again. After receiving the adjustment request signal, the control terminal calculates the new alarm clock intensity parameters according to the preset intensity adjustment rules. The control terminal encapsulates the new intensity parameters into adjustment instructions and sends them to the control module of the linked alarm clock; After receiving the adjustment command, the control module of the alarm clock immediately updates the intensity parameters and executes the adjusted stimulus output.

9. The method for mild sleep arousal based on user sleep curves as described in claim 8, characterized in that, During the wake-up process, the sleep mat continuously monitors the user's sleep state and obtains real-time sleep depth levels, including: After the arousal stimulus is triggered, the sensor array continues to collect real-time physiological signals; The real-time physiological signals are preprocessed in real time and input into a preset lightweight real-time sleep depth assessment model. The real-time sleep depth assessment model outputs a quantified real-time sleep depth level; the real-time sleep depth level is a continuous value or discrete score from deep sleep to wakefulness.

10. The method for mild sleep arousal based on user sleep curves as described in claim 9, characterized in that, The method further includes: recording data of the wake-up process; the control unit of the sleep mat records the wake-up start time, the real-time sleep depth level at each time point, the alarm intensity adjustment time, the intensity parameters before and after adjustment, and the changes in sleep depth level in real time, forming sleep record data, and storing the sleep record data in the local storage module; after the wake-up process is completed, the sleep record data is synchronized to the control terminal, and the control terminal organizes the sleep record data and generates a wake-up process report.

Images