Multi-model closed-loop dream sleep intervention system and method

Abstract

The invention relates to the technical field of intelligent sleep, and provides a multi-model closed-loop dream sleep intervention system and method.The system is composed of a sensing layer, a decision-making layer, an execution layer and an application layer, the sensing layer collects multi-modal physiological and environmental signals, and the application layer obtains dream texts and mental stress scores; the decision-making layer utilizes a world model, a time sequence model, a multi-mode representation model and a language model to construct a prediction body and mind model, maps the sleep stage, the dream emotion and the pressure level into implicit states, and performs rolling prediction on the nightmare risk of a future time window. An intervention decision engine selects an intervention combination from sound, illumination, smell, touch sense, temperature control and nerve stimulation according to the health prescription, and generates a multi-channel instruction to drive an execution layer to intervene in advance; the system quantitatively evaluates the effect through the physiological and psychological steady-state index, and realizes individualized closed-loop adjustment of the dream state and the mental stress on the basis of a sensing-prediction-intervention-feedback data self-adaptive updating model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent sleep, in particular to a multi-model closed-loop dream sleep intervention system and method. BACKGROUND

[0002] With the acceleration of work rhythm and the increase of mental stress, more and more people have problems such as difficulty falling asleep, frequent nightmares, and daytime fatigue. Traditional sleep monitoring devices mainly focus on heart rate, body movement or snoring, and concentrate on the statistics of sleep duration, sleep staging and wake-up times. The relationship between dream content and mental stress is not analyzed in depth, and the intervention of nightmares and negative dreams is mostly limited to simple white noise playing, environmental noise reduction or single light adjustment.

[0003] In the prior art, one type of scheme uses wearable devices to collect heart rate and body movement information, and identifies sleep stages and wake-up events through traditional machine learning or deep learning models. Another type of scheme introduces sound or light devices to assist with fixed music, white noise or weak light during the sleep stage or light sleep stage to improve subjective comfort. This type of scheme mainly solves the problems of falling asleep and maintaining sleep, and has less consideration for dream content, nightmare frequency and mental stress adjustment. Moreover, the intervention is mostly passive response, lacking the ability of forward-looking adjustment based on prediction.

[0004] Some studies attempt to use electroencephalogram and eye movement signals to identify rapid eye movement periods, thereby inferring that they are in a dream state. However, these studies still focus on label classification and lack detailed description of dream types and emotional tones. They do not fully utilize subjective information such as user dream text and mental stress scales. On the other hand, existing systems mostly use single model structures, which have limited ability to learn complex relationships and are difficult to balance short-term early warning and long-term trend judgment.

[0005] In terms of intervention strategies, traditional schemes mainly focus on sleep stability and wake-up risk, and adjust temperature, noise or light to reduce external disturbances. However, they do not design multi-channel and scenario-based interventions from the perspective of "actively guiding dream direction, buffering nightmare emotions, and reducing mental stress."

[0006] Therefore, there is an urgent need in the art for a new technical solution that can prospectively predict dream state, mental stress and nightmare risk during sleep and provide individualized intervention to address the shortcomings of existing technology. SUMMARY

[0007] To overcome the above-mentioned defects of the prior art, the present application provides a multi-model closed-loop dream sleep intervention system and method, which builds a multi-model voting prediction psychosomatic model and drives individualized multi-channel dream mental intervention to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purposes, the present application provides the following technical solutions: A multi-model closed-loop nightmare sleep intervention system, comprising a perception layer, an execution layer, a decision layer and an application layer, the decision layer integrates a world model, a timing model, a multi-modal representation model and a language model to form a prediction psychosomatic model, for receiving multi-modal physiological signals and environmental signals collected by the perception layer, and combining dream content and subjective evaluation of mental stress obtained by the application layer, mapping sleep stage, dream emotion and mental stress level into hidden state, rolling prediction of psychosomatic state in future multiple time windows, and giving nightmare occurrence and mental stress change risk; when the prediction result shows that there is nightmare or high stress risk in the forward-looking time window, the intervention decision engine passes through the multi-model integration and voting module, according to the risk assessment, dream semantic vector, health prescription target and scene script in the historical data, the physiological and psychological stability index of the scene script is improved, the target script is selected by weighted voting of the preset scene script, and the intervention instruction including sound, light, smell, touch, partition temperature control and neural stimulation control parameters is generated, driving the execution layer to gradually apply multi-channel stimulation before the high-risk state is formed; after the intervention, the psychosomatic state before and after the intervention is quantitatively evaluated by the physiological and psychological stability index calculation module, and the evaluation result is generated and stored, which is used for subsequent model training, and the "perception-prediction-intervention-response" data of multiple nights is used as an incremental sample to update the prediction psychosomatic model and the intervention decision engine parameters, so that the system forms individualized nightmare and mental intervention mode for different users in long-term operation.

[0009] As a further scheme of the present application, the perception layer comprises a physiological signal acquisition unit and an environmental signal acquisition unit, the physiological signal acquisition unit is composed of a wristband wearable device, a mattress sensing array, an electroencephalogram acquisition device and a skin electricity and body surface temperature sensor, the wristband wearable device acquires heart rate, heart rate variability and body movement information through photoplethysmography and accelerometer, the mattress sensing array identifies turning over and kicking behavior through multi-point pressure detection, the electroencephalogram acquisition device extracts multiple frequency band electroencephalogram power to support sleep stage and dream activity level discrimination, and the skin electricity and body surface temperature sensor is used for monitoring sympathetic excitation level and peripheral temperature change; the environmental signal acquisition unit integrates temperature and humidity, noise intensity and frequency spectrum, light intensity and color temperature, and air quality sensors.

[0010] As a further scheme of the present application, the perception layer is provided with a preprocessing module for filtering, outlier rejection and time alignment of the physiological signals and environmental signals, and extracting the average heart rate and heart rate variability, respiratory rate and variability, body movement energy, turning event, brain electrical frequency band power, skin electrical fluctuation amplitude, body surface temperature and environmental temperature, humidity, noise and illumination features in a fixed length sliding time window to generate a multi-modal feature vector for inputting into the decision layer in time sequence, and the types of physiological signals are not limited to heart rate, respiration, body movement, brain electrical activity, skin electrical activity and body surface temperature, and can be extended to blood oxygen saturation, blood pressure, muscle electrical activity, eye movement, mandibular muscle electrical activity or other physiological and behavioral indicators related to sleep and emotional state according to the development of sensing technology.

[0011] As a further scheme of the present application, the world model in the prediction psychosomatic model represents the sleep stage, dream emotion, mental stress level and nightmare risk level by introducing hidden state variables, learns the hidden state transition relationship and observation mapping relationship in the training stage by using multi-night multi-modal feature sequences and labeled sleep stages, nightmare segments and wake-up events, and updates the current hidden state according to the current multi-modal feature vector and the hidden state of the last time window in the online running stage, and performs multi-step rolling prediction in the hidden state space to obtain the hidden state sequence and corresponding nightmare risk and mental stress change curve of future multiple time windows.

[0012] As a further scheme of the present application, the time series model in the prediction psychosomatic model constructs a short-term prediction structure for heart rate variability, skin electrical fluctuation and body movement energy, which is used to identify the pre-nightmare physiological pattern of gradual increase in heart rate accompanied by decrease in variability, superimposed continuous fluctuation of skin electrical activity and increase in body movement in a limited time window, and outputs the sympathetic excitation risk and wake-up risk evaluation results of the next few time windows.

[0013] As a further scheme of the present application, the multi-modal representation model in the prediction psychosomatic model is used for normalizing and fusion mapping of the preprocessed heart rate, respiration, body movement, brain electrical activity, skin electrical activity, body surface temperature and environmental features, compressing signals of different sources into a unified dimensional feature vector, which is used as the input of the world model and the time series model, and as the state representation when the scene script scoring is performed by the intervention decision engine.

[0014] As a further scheme of the present application, the language model in the prediction psychosomatic model is used for semantic analysis of the dream text and the mental stress scale collected by the application layer, the dream content is divided into a combination of tags related to work stress, academic stress, family conflicts, traumatic memories or childhood safe memories, the dream emotional tendency, control and safety level are judged, and the above-mentioned tags and scale information are encoded into a dream semantic vector to input the world model and the intervention decision engine, so that the nightmare risk prediction and the intervention strategy selection consider the dream theme and the emotional background at the same time, and the language model is preferably a domain language model constructed based on a large-scale pre-trained language model.

[0015] As a further scheme of the present application, the intervention decision engine is connected to an intervention strategy library, and the intervention strategy library has multiple scene scripts of a seaside safety scene, a forest quiet scene, a neutral breathing guidance scene and an environmental steady state optimization scene, each scene script specifies the sound source type, the volume range and the gradual change rhythm of the sound channel, the brightness, the color temperature and the change curve of the light channel, the fragrance type and the release rate of the olfactory channel, the wrapping feeling and the vibration mode of the tactile channel, the target temperature interval of the partition temperature control channel, and the stimulation amplitude and the frequency range of the neural stimulation channel, and the intervention decision engine is internally provided with a multi-model integration and voting module, the risk curve, the dream semantic vector, the health prescription target output by the world model and the time sequence model, and the physiological and psychological steady state index improvement performance of each scene script in historical data are used to perform weighted voting scoring and sorting on the candidate scene scripts and select a target scene script.

[0016] As a further scheme of the present application, the sound intervention unit in the execution layer plays environmental sound, music or voice guidance content through a bedside sound box or earphones, the light intervention unit controls the light intensity and color temperature through adjustable light fixtures, the olfactory intervention unit releases fragrance through atomization devices, the tactile and pressure intervention unit adjusts the local pressure and micro-vibration mode through the internal actuator of the mattress, the partition temperature control unit cooperates with air conditioners, floor heating or mattress temperature control layers to adjust the temperature of the trunk and limb areas, and the neural stimulation unit applies controlled electric stimulation through skin surface electrodes, and the execution layer starts the corresponding channel in a gradual manner before the predicted high-risk time window according to the intervention instructions generated by the intervention decision engine.

[0017] A multi-model closed-loop dream sleep intervention method, comprising the following steps: Step one, baseline modeling, the perception layer is used to collect multi-modal physiological signals and environmental signals in several nights, and the application layer is used to collect dream content and mental state subjective evaluation, the above-mentioned data is input into the prediction psychosomatic model and the feedback self-adaptive learning unit for training, and the individualized implicit state space and the scene script intervention effect statistics of the user are established. Step two, online prediction, after the user falls asleep, the perception layer continuously outputs multi-modal features in a sliding time window, the physical and mental model updates the current hidden state at each time window, and the world model and the time series model are used to make rolling predictions for multiple future time windows, outputting nightmare risk curves and stress change curves; Step three, strategy decision, the intervention decision engine generates intervention instructions containing sound, light, olfactory, tactile, zoned temperature control and neural stimulation control parameters based on the curves, dream semantic vectors and sleep health prescription targets, and the multi-model integration and voting module weights and votes multiple scene scripts in the intervention strategy library to generate intervention instructions containing sound, light, olfactory, tactile, zoned temperature control and neural stimulation control parameters; Step four, multi-channel intervention, the execution layer controls each intervention channel according to the intervention instructions, gradually applies multi-modal stimulation in the predicted high-risk time window, and guides the dream emotion and physiological indicators to regress to the safe area; Step five, closed-loop feedback, the feedback adaptive learning unit calculates the physiological and psychological stability index before and after intervention, statistically analyzes the actual improvement effect of each scene script under different hidden state and dream label combinations, and updates the prediction physical and mental model and intervention decision engine parameters with the "perception-prediction-intervention-response" data as incremental samples, so that the system forms individualized dream and mental intervention patterns for different users in the long-term operation.

[0018] The technical effects and advantages of the multi-model closed-loop dream sleep intervention system and method of the present application are as follows: First, by combining the world model and the time series model, the sleep stage evolution, dream emotion change, nightmare occurrence risk and stress trend of multiple future time windows are predicted in the hidden state space, so that the system can identify the risk in advance before the nightmare and high stress state is fully formed, realizing the transition from post-event recording to pre-warning.

[0019] Second, by the cooperation of the multi-modal representation model and the language model, the heart rate, respiration, body movement, electroencephalogram, electrodermal and environmental signals are combined with the user's subjective dream text, dream comfort and stress scale to build a prediction physical and mental model considering physiological indicators and dream semantic information, so that the nightmare risk assessment can make differentiated judgments for different dream themes and emotional backgrounds, and improve the recognition ability of complex dream state and stress change.

[0020] Third, by intervening the decision engine and its internal multi-model integration and voting module, under the target and safety limit conditions of health prescription agreement, the preset seaside safety scene, forest quiet scene, neutral respiratory guidance scene and environmental steady state optimization scene and other scene scripts are comprehensively scored to generate sound, lighting, olfactory, tactile, partition temperature control and neural stimulation combination for specific users and specific night state, which can provide more personalized adjustment that fits individual dream characteristics and spiritual needs, especially in buffering and working pressure, nightmare related to traumatic memories.

[0021] Fourth, through the physiological and psychological steady state index calculation module and the feedback adaptive learning unit, the whole process of "perception-prediction-intervention-response" is quantitatively evaluated and incrementally learned, and the prediction physical and mental model and the intervention strategy library are continuously corrected on the multi-night time scale, so that the system gradually forms the dream and spiritual intervention preferences for different users in long-term use, realizes closed-loop control and self-optimization, and has stable comprehensive improvement effect in reducing the number of nightmare nights, reducing the number of night awakenings, improving dream comfort and relieving spiritual stress. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is the overall structure schematic diagram of a multi-model closed-loop dream sleep intervention system of the present application; Figure 2 is the structure schematic diagram of the prediction physical and mental model and the multi-model integration and voting module in the decision layer of the present application; Figure 3 is the multi-channel intervention execution layer structure schematic diagram of the present application; Figure 4 is the physiological and psychological steady state index calculation module schematic diagram of the present application; Figure 5 is the multi-model closed-loop dream sleep intervention method flowchart schematic diagram of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0024] Embodiment 1 In this embodiment, a multi-model closed-loop dream sleep intervention system adopts a hierarchical architecture and is deployed in a family bedroom environment, which is used for long-term monitoring of the physiological and environmental state of a sleeper, predicting the dream state, nightmare risk and spiritual stress, and realizing closed-loop adjustment through multi-channel intervention.

[0025] like Figure 1 As shown, the system consists of a perception layer, a decision-making layer, an execution layer, and an application layer. The interaction of data and control commands between the layers is achieved through wired or wireless networks.

[0026] The perception layer is mainly responsible for data acquisition and preprocessing. The physiological signal acquisition unit consists of a wristband wearable device, a mattress sensor array, an EEG acquisition device, and skin conductance and body surface temperature sensors. The wristband wearable device is worn on the sleeper's wrist and has a built-in photoplethysmography sensor and a triaxial accelerometer. The former continuously acquires pulse waveforms at a preset sampling frequency, and the latter acquires hand movement information. The system calculates heart rate and heart rate variability by detecting pulse waves and analyzing rhythms of the pulse waveform. It also estimates body movement intensity and turning events based on acceleration signals. A mattress sensor array, embedded within the mattress, consists of several piezoresistive or capacitive sensing units deployed in key areas such as the torso, waist, and legs. By detecting changes in pressure magnitude and distribution over time in different areas, it identifies changes in sleeping posture, turning over, and leg kicking. The EEG acquisition device uses a comfortable headband with electrodes distributed on the forehead or top of the head. It acquires EEG signals through front-end amplification and filtering circuits and extracts energy in frequency bands such as δ, θ, α, σ, and β, providing a basis for determining sleep stages and dream activity levels. Electrodermal sensors are located on the fingers or palms to measure skin conductance or skin response, reflecting the level of sympathetic nerve excitation. Body surface temperature sensors are positioned on the wrist or forearm to monitor changes in peripheral body surface temperature.

[0027] The environmental signal acquisition unit is installed on the bedroom wall or near the bedside, integrating a temperature and humidity sensor, a noise sensor, a light sensor, and an air quality sensor. The temperature and humidity sensor outputs indoor temperature and relative humidity. The noise sensor not only measures sound pressure level but also records the sound spectrum distribution to distinguish between continuous noise, intermittent noise, and sudden sound sources. The light sensor measures illuminance intensity and color temperature to understand the current room lighting environment. The air quality sensor detects indicators such as volatile organic compound concentration or carbon dioxide concentration to determine the freshness of indoor air.

[0028] In this embodiment, the system perception layer is configured with a unified time reference module to correct the timestamps of data from various sensors, ensuring synchronization of multimodal signals on the time axis. The perception layer includes a preprocessing module that uses conventional bandpass filtering, wavelet denoising, or adaptive filtering algorithms to remove artifacts from ECG, EEG, and ESK signals, and performs outlier detection and interpolation correction on acceleration, pressure, and environmental signals. The preprocessing module extracts the average heart rate, heart rate variability index, respiratory rate and variability (which can be estimated from chest movement or pulse wave), body energy, number of turns, relative power of each EEG band, ESK fluctuation amplitude, body surface temperature, and environmental temperature, humidity, noise intensity, noise spectrum characteristics, and light intensity and color temperature from the original signal according to a preset sliding time window (e.g., every 30 seconds or every minute). It then normalizes each feature to form a fixed-dimensional multimodal feature vector, which is buffered in chronological order and sent to the decision layer. It should be noted that the types of physiological signals are not limited to the examples above. Blood oxygen saturation, blood pressure, electromyography, eye movement, mandibular electromyography, or other physiological and behavioral indicators related to sleep and emotional state can also be introduced. This embodiment does not limit the types of signals.

[0029] The internal structure of the decision-making level, such as Figure 2 As shown, the system includes a predictive mind-body model unit, an intervention decision engine unit, and a feedback adaptive learning unit. The predictive mind-body model unit consists of a multimodal representation model, a world model, a temporal model, and a language model. The multimodal representation model receives multimodal feature vectors from the perceptual layer, normalizes and linearly or nonlinearly maps features from different sources, and compresses them into a unified-dimensional state vector space, so that the subsequent world model and temporal model do not need to perform independent input processing for each specific signal. The world model introduces latent state variables to comprehensively represent the current sleep stage, dream emotion level, mental stress background, and nightmare risk level. During the training phase, the world model uses multimodal feature sequences collected over multiple nights, as well as post-processed sleep stage labels, suspected nightmare fragment labels, and awakening event labels, to establish a probabilistic mapping relationship between latent states and observable features, and learn the transition rules between latent states. During the online operation phase, the world model updates the current latent state estimate at each time window using the current multimodal state vector and the latent state of the previous time window, and makes rolling predictions for multiple future time windows in the latent state space, outputting the latent state sequence of several future time windows as well as the corresponding nightmare risk curve and mental stress change curve.

[0030] The time-series model focuses on analyzing the changing trends of heart rate variability, skin conductance fluctuations, and kinetic energy over a shorter time scale. During the training phase, the time-series model uses labeled pre-nightmare time windows and ordinary dream time windows to extract pattern features of characteristic sequences, such as typical pre-nightmare physiological combinations like slow but continuously rising heart rate, decreasing heart rate variability, and rapid skin conductance fluctuations coupled with increased kinetic energy, to construct a short-term prediction structure. During the online operation phase, the time-series model continuously analyzes the features of the latest time windows and outputs estimates of changes in sympathetic excitation and arousal risk within the next one to three time windows, providing more sensitive early warning signals for intervention decisions.

[0031] The language model is built upon a large-scale pre-trained language model and fine-tuned using corpora from sleep medicine and psychology to improve the understanding of scenes, characters, emotions, conflicts, and psychological stress descriptions in dream texts. The application layer guides users to describe their previous night's dreams in natural language through a dream recording interface after waking, and scores dream comfort, nightmare intensity, and psychological stress on a scale. The language model receives dream text and scale information, segments and encodes the text, identifies typical scenes (e.g., office, school, home, public places), event types (e.g., failing an exam, making a mistake on a task, conflict, escape, fall), emotional tendencies (e.g., tension, fear, shame, security), and the subject and dialogue content. Combined with scale scores, dreams are categorized into label combinations corresponding to work stress, academic stress, family conflicts, traumatic memories, or childhood safety memories, and an overall assessment of emotional tendency, sense of control, and level of security is provided. These labels and scale information are further encoded into fixed-dimensional dream semantic vectors, stored in the user profile, and used as long-term background input in subsequent nighttime predictions.

[0032] The intervention decision engine unit is used to select appropriate intervention strategies based on the output of the predictive mind-body model. It includes a multi-model integration and voting module, such as... Figure 2 As shown, this module receives, on the one hand, the implicit state sequence of multiple future time windows and the nightmare risk curve and mental stress change curve output by the world model; on the other hand, it receives the short-term sympathetic arousal and arousal risk assessment output by the time series model, and the current state vector output by the multimodal representation model. Simultaneously, it obtains the dream semantic vector output by the language model and the sleep health prescription goals issued by the application layer. In this embodiment, the sleep health prescription is typically configured by the user or a professional to set the relative weights of goals such as reducing the frequency of nightmares, alleviating long-term mental stress, and maintaining a high proportion of deep sleep.

[0033] The multi-model integration and voting module synthesizes the above information into several indicators, such as the peak risk of nightmares in the future time window, the rate of risk increase, the short-term arousal risk, and the type of dream theme and emotional background that the current implicit state is close to. It also calculates the expected improvement in the physiological and psychological homeostasis index for each scenario script in the intervention strategy library. During the calculation, the module references historical data on the actual changes in the physiological and psychological homeostasis index after the user applied the script under similar implicit states and dream tag combinations, using this historical improvement performance as an important reference feature. Subsequently, the multi-model integration and voting module generates a comprehensive score for each scenario script, sorts the candidate scripts, selects the target scenario script with the highest comprehensive score that meets the current health prescription constraints and safety limitations, and converts it into specific intervention instructions. These instructions include the sound source type, target volume, and gradation time for the sound channel; the target brightness, color temperature, and change curve for the light channel; the fragrance type and release rate for the olfactory channel; the pressure adjustment amplitude and vibration mode for the tactile channel; the target temperature and gradient for each area of ​​the zoned temperature control channel; and whether to activate the neural stimulation channel and its stimulation amplitude and frequency parameters.

[0034] The structure of the execution layer is as follows Figure 3 As shown, the sound intervention unit connects to the bedside speaker or bone conduction headphones, supporting the playback of ambient sounds, music, or voice guidance content, while also outputting interactive audio information. It transmits intervention data related to psychological intervention, relaxation training, or positive suggestion to the user in audio form, through the user's dreams, without interrupting sleep. This is used in conjunction with multi-channel intervention to regulate the user's dream emotions and mental stress. The light intervention unit connects to dimmable lamps, allowing adjustment of illuminance and color temperature within a certain range. The olfactory intervention unit releases specific fragrances, such as lavender, forest scents, sea breeze, or baking aromas, through a misting device and fragrance box. The tactile and pressure intervention unit consists of airbags and vibrators inside the mattress, which can change local support strength or generate slight rhythmic vibrations within a safe range. The zoned temperature control unit is linked to air conditioning, underfloor heating, or the mattress temperature control layer to achieve independent temperature regulation of the trunk and limb areas. The nerve stimulation unit applies low-intensity transcutaneous electrical stimulation to the back of the ear or forehead through electrodes on the skin surface, according to pre-set safety parameters, to assist in regulating sympathetic and parasympathetic activity. In this embodiment, each intervention unit receives intervention instructions from the decision-making layer from the local controller and executes them step by step according to the start time, initial intensity and change rhythm in the instructions, thereby forming a smooth scene transition in time.

[0035] To achieve a unified quantification of intervention effects, this embodiment includes a physiological and psychological homeostasis index calculation module in the decision-making layer, such as... Figure 4As shown, this module receives physiological characteristics such as average and variability of heart rate, respiratory rate and variability, kinetic energy, power of various EEG bands, and amplitude of skin conductance fluctuations within each preset assessment window, as well as subjective characteristics such as dream comfort score, nightmare intensity score, and mental stress score. After normalization, each feature is weighted and fused according to weight coefficients to obtain a physiological and psychological homeostasis index between zero and one. The weight coefficients can be set by professionals based on general sleep medicine experience, or fine-tuned through data-driven methods over long-term operation. The physiological and psychological homeostasis index serves as a comprehensive indicator for evaluating current sleep and mental state; a higher index value indicates a more stable current physiological rhythm, a more comfortable dream experience, and greater relief from mental stress. The feedback adaptive learning unit statistically analyzes the changes in the index within adjacent assessment windows, comparing the difference in the index before and after intervention with the world model's expectation of the intervention effect. If a script for a certain scenario consistently improves the index better than predicted in multiple applications, the priority of that script in similar states is increased; conversely, its priority is reduced or the stimulus intensity range is tightened.

[0036] To visually demonstrate the effectiveness of the system in actual operation, we take a user with work-related nightmares as an example and statistically analyze the main indicators during the baseline and intervention periods, resulting in the results shown in Table 1.

[0037] Table 1 Comparison of key indicators of participants before and after intervention.

[0038] As shown in Table 1, during the baseline period, when only monitoring was conducted without prospective intervention, the user had a higher number of nightmare nights per week, more nighttime awakenings, a lower average level of physiological and psychological homeostasis index, a low dream comfort score, and a high mental stress score. After enabling the multi-model prediction and multi-channel intervention functions of the system in this embodiment and running it continuously for several weeks, the number of nightmare nights decreased significantly, the number of nighttime awakenings decreased, the physiological and psychological homeostasis index increased from 0.50 to 0.73, the dream comfort score increased from 3.5 to 7.1, and the mental stress score decreased from 8.2 to 4.9. This indicates that the system in this embodiment can achieve prospective identification and effective intervention of dream state, nightmare risk, and mental stress in engineering.

[0039] like Figure 5As shown, after a user falls asleep at night, the perception layer continuously collects multimodal signals and generates feature vectors; the predictive mind-body model updates the latent state in each time window and performs rolling predictions for multiple future time windows, generating nightmare risk and mental stress change curves; when the multi-model integration and voting module determines that there is a high risk of nightmares or high stress in the future time window, the intervention decision engine selects the most suitable scenario script from the intervention strategy library for the combination of the current latent state and dream label and converts it into intervention instructions; the execution layer gradually initiates multi-channel interventions before the arrival of high-risk time windows, guiding dreams and physiological states back to the safe zone; subsequently, the physiological and psychological homeostasis index calculation module quantitatively evaluates the state before and after intervention, and the feedback adaptive learning unit uses multi-night "perception-prediction-intervention-response" data to incrementally update the predictive mind-body model and intervention strategy library, completing a closed loop. Through long-term operation, the system in this embodiment can gradually form an individualized intervention mode for each specific user, playing a stable role in dream regulation and mental health assistance.

[0040] Example 2 In this embodiment, we take a user who has been under high-intensity work pressure for a long time as an example to illustrate the application process of the system of the present invention in a real-world usage scenario. This user has frequently experienced work-related nightmares over the past year, with the main dream content being "failing exams," "project errors," and "being reprimanded by superiors." The user experiences nightmares multiple nights per week and reports significant mental stress.

[0041] according to Figure 5 The illustrated process flow begins with the system entering the baseline modeling phase. Over two consecutive weeks, the perception layer collects data on heart rate, heart rate variability, body movement, number of turns, EEG frequency band power, skin conductance, body surface temperature, and ambient temperature, humidity, noise, and light characteristics within the user's normal living environment, such as their bedroom. Each morning upon waking, the application layer reminds the user to complete a dream log and a mental state questionnaire. The user provides a brief text description of the main plot of their dream and scores dream comfort, nightmare intensity, and mental stress on the scale. The decision layer inputs this data into a predictive psychosomatic model for training. The multimodal representation model learns the distribution of the user's physiological signal characteristics, the world model learns the implicit state evolution patterns under different dream and stress backgrounds, the temporal model identifies short-term patterns of increased heart rate, decreased heart rate variability, and increased skin conductance fluctuations before nightmares, and the language model extracts high-frequency tags such as "work stress," "exam failure," and "blame scenarios" from the dream text to form corresponding dream semantic vectors. The feedback adaptive learning unit only records data during the baseline period and does not output any intervention.

[0042] After baseline modeling was completed, the user and professionals configured a sleep health prescription at the application layer. The primary goal was to reduce the frequency of work-related nightmares and alleviate nighttime fluctuations in mental stress, with an additional goal of maximizing the proportion of deep sleep. The prescription translated these goals into weights and safety constraints, which were then sent to the decision-making layer for use by the intervention decision engine.

[0043] On a typical workday evening, the user prepares to sleep at their usual time. The perceptual layer automatically begins recording signals after the user goes to bed, and the predictive mind-body model updates the latent state as the user transitions from wakefulness to light sleep and then to deep sleep. Entering the middle of the night, EEG characteristics show the user has entered REM sleep, and the feature vector output by the multimodal representation model indicates increased dream activity. The world model classifies the current latent state as "active dreams but moderate mood," with a low nightmare risk curve. As time progresses, the perceptual layer observes a slow increase in heart rate, a slight decrease in heart rate variability, increased skin conductance fluctuations, and slight leg movements recorded by the mattress sensor array. Based on these changes, the temporal model determines that the user has developed a short-term physiological pattern approaching nightmares, outputting an assessment of increased sympathetic arousal and awakening risk for the next few time windows. The world model performs rolling predictions in the latent state space for several future time windows, showing a significant increase in the probability curve of the state transitioning from "mildly negative dreams" to "nightmare threshold."

[0044] After receiving the above prediction results, the intervention decision engine integrates information from the world model, time series model, multimodal representation model, and language model through the multi-model integration and voting module. Given the user's recent high workload, the language model infers that the current dream is more likely a "work stress dream" based on dream history tags and daytime events, giving the corresponding target a higher weight in the sleep health prescription. The intervention strategy library already contains scripts for a safe beach scenario, a neutral breathing guidance scenario, and a tranquil forest scenario, all tested during the baseline period. The safe beach scenario demonstrated a better improvement in the user's physiological and psychological homeostasis index during the latter half of the baseline period and a limited number of trials. The multi-model integration and voting module weights the expected improvement in homeostasis index for each scenario script under the current implicit state and dream tag combination, ranks the candidate scripts, and ultimately selects the safe beach scenario script as the intervention plan, generating intervention instructions that include control parameters for sound, light, smell, touch, and temperature channels. Considering the user's low subjective acceptance of transcutaneous electrical stimulation, the neural stimulation channels are kept closed in this intervention instruction, only to be used as a backup in case of future high-risk situations and no improvement in physiological signals.

[0045] The execution layer initiates a gradual intervention approximately one minute before the anticipated high-risk time window, following the intervention instructions. The sound intervention unit plays ocean wave sounds at a low volume, followed by distant birdsong. It also periodically overlays short voice messages at extremely low volume, such as "It's safe now," "You can relax," and "The task is complete," based on information interaction control parameters. This allows these messages to enter the user's dream experience along with the background sounds, thus transmitting intervention data to the user through the dream and enhancing the regulatory effect on negative dreams and mental stress. The light intervention unit slowly adjusts the room light from a weak, cool color to a warm, low-brightness yellow. The olfactory intervention unit releases a low concentration of sea breeze and citrus-based aromas. The tactile and pressure intervention unit slightly enhances back support through mattress airbags, making the user feel supported by a soft medium. The zoned temperature control unit fine-tunes the temperature of the torso area to the user's comfort range recorded at baseline, while maintaining stable temperatures in the limb areas. Throughout the process, the intensity of each channel changes smoothly according to a preset curve, avoiding sudden changes that could cause awakening or discomfort.

[0046] After the intervention, the sensory layer continuously collected physiological signals. After approximately one time window, the heart rate trend showed a decline, and heart rate variability returned to the user's normal level during REM sleep. Skin conductance fluctuations weakened, and body movements changed from continuous, short movements to sporadic, slight adjustments. The world model revised the latent state from "nightmare critical trend" to "mild dream buffer," and the predicted nightmare risk curve subsequently decreased. The short-term risk assessment given by the time-series model returned to a moderate level. Based on the latest prediction results and health prescriptions, the intervention decision engine maintained the current seaside safety scenario script for several time windows and gradually reduced the sound volume and olfactory release rate, allowing the intervention intensity to converge slowly as the risk decreased.

[0047] Upon waking, the user reviewed their dream through the application interface, describing it as follows: "I dreamt I was taking an exam at work. I was very nervous at first, but then the scene changed to a relaxing beach scene. I didn't feel reprimanded, and I didn't wake up startled." The dream comfort score was 7, the nightmare severity score was 2, and the mental stress score was slightly lower than the recent average. The decision-making layer input these subjective scores along with nighttime physiological signals into the physiological and psychological homeostasis index calculation module to calculate the index for multiple evaluation windows. The feedback adaptive learning unit recorded the index improvement brought about by using the beach safety scenario script under the "work stress dream + increased nightmare risk" tag and incorporated this result into the update of scenario script priority and model parameters.

[0048] In the following weeks, the system performed prospective interventions on multiple nights following a similar pattern. Statistics showed that, compared to the baseline period, the user experienced a significant reduction in the number of nights with nightmares per week, a decrease in the average number of awakenings per night, an increase in the average physiological and psychological homeostasis index during REM sleep, an overall increase in dream comfort scores, and a downward trend in mental stress scores. This indicates that the system and method of this invention can continuously and individually regulate dream states and mental stress in real-world environments through multi-model prediction and multi-channel intervention.

[0049] The system of this invention can be deployed in a real home environment. It utilizes multi-modal perception, world model and temporal model combined with language model for multi-model prediction. With the help of multi-model integration and voting module, it selects individualized scene scripts and achieves adaptive learning through multi-channel intervention and physiological and psychological homeostasis index feedback.

[0050] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0051] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-model closed-loop dream sleep intervention system, comprising a perception layer, an execution layer, a decision-making layer, and an application layer, characterized in that, The decision-making layer integrates a world model, a time series model, a multimodal representation model, and a language model to form a predictive mind-body model. This model receives multimodal physiological and environmental signals collected by the perception layer and combines them with dream content and subjective evaluations of mental stress obtained from the application layer. It maps sleep stages, dream emotions, and mental stress levels as latent states, performing rolling predictions of mind-body states over multiple future time windows, providing information on the risk of nightmares and changes in mental stress. When the prediction results indicate a risk of nightmares or high stress within the prospective time window, the intervention decision engine, through a multi-model integration and voting module, considers the risk assessment, dream semantic vectors, health prescription goals, and scenario scripts output by each model in historical data. The system improves the physiological and psychological homeostasis index, performs weighted voting on preset scenario scripts to select target scripts, and generates intervention instructions containing parameters for sound, light, smell, touch, zoned temperature control, and neural stimulation control. This drives the execution layer to gradually apply multi-channel stimulation before a high-risk state is formed. After the intervention, the physiological and psychological homeostasis index calculation module quantitatively evaluates the physical and mental state before and after the intervention and generates and stores the evaluation results for subsequent model training. Multi-night "perception-prediction-intervention-response" data is used as incremental samples to update the parameters of the predictive physical and mental model and the intervention decision engine, enabling the system to form individualized dream and mental intervention patterns for different users in long-term operation.

2. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The perception layer includes a physiological signal acquisition unit and an environmental signal acquisition unit. The physiological signal acquisition unit consists of a wristband-type wearable device, a mattress sensor array, an EEG acquisition device, and skin conductance and body surface temperature sensors. The wristband-type wearable device acquires heart rate, heart rate variability, and body movement information through photoelectric volume recording and accelerometer. The mattress sensor array identifies turning over and kicking behaviors through multi-point pressure detection. The EEG acquisition device extracts multiple frequency bands of EEG power to support the discrimination of sleep stages and dream activity levels. The skin conductance and body surface temperature sensors are used to monitor sympathetic excitation levels and changes in peripheral temperature. The environmental signal acquisition unit integrates temperature and humidity, noise intensity and spectrum, light intensity and color temperature, and air quality sensors.

3. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The perception layer is equipped with a preprocessing module for filtering, outlier removal, and time alignment of physiological and environmental signals. Within a fixed-length sliding time window, it extracts the average heart rate and heart rate variability, respiratory rate and variability, body energy, turning events, EEG frequency band power, skin conductance fluctuation amplitude, body surface temperature, and environmental temperature, humidity, noise, and light characteristics. These are combined to generate a multimodal feature vector, which is then input into the decision layer in chronological order. The types of physiological signals are not limited to heart rate, respiration, body movement, EEG, skin conductance, and body surface temperature. They can also be expanded to include blood oxygen saturation, blood pressure, electromyography, eye movement, mandibular electromyography, or other physiological and behavioral indicators related to sleep and emotional state, depending on the development of sensing technology.

4. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The world model in the predictive mind-body model introduces latent state variables to represent sleep stages, dream emotions, mental stress levels, and nightmare risk levels. During the training phase, it uses multi-night multimodal feature sequences and labeled sleep stages, nightmare fragments, and awakening events to learn latent state transition relationships and observation mapping relationships. During the online operation phase, it updates the current latent state based on the current multimodal feature vector and the latent state of the previous time window, and performs multi-step rolling prediction within the latent state space to obtain the latent state sequences of multiple future time windows and the corresponding nightmare risk and mental stress change curves.

5. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The temporal model in the predictive mind-body model constructs a short-term predictive structure for heart rate variability, skin conductance fluctuations, and body energy. It is used to identify a pre-nightmare physiological pattern that occurs within a limited time window, characterized by a gradual increase in heart rate accompanied by a decrease in variability, continuous fluctuations in skin conductance, and an increase in body energy. It also outputs the assessment results of sympathetic excitation risk and arousal risk for the most recent time windows.

6. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The multimodal representation model in the predictive mind-body model is used to normalize and fuse preprocessed heart rate, respiration, body movement, EEG, skin conductance, body surface temperature and environmental features, compressing signals from different sources into feature vectors of a unified dimension, which are used as inputs to the world model and time series model, and as state representations when the intervention decision engine scores scene scripts.

7. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The language model in the predictive psychosomatic model is used to perform semantic analysis on dream texts and mental stress scales collected from the application layer. It divides dream content into tag combinations related to work stress, academic stress, family conflicts, traumatic memories, or childhood safety memories, and judges the emotional tendency, sense of control, and sense of security of dreams. The above-mentioned tag and scale information are encoded into dream semantic vectors and input into the world model and intervention decision engine, so that nightmare risk prediction and intervention strategy selection take into account both dream theme and emotional background. The language model is preferably a domain-specific language model built on a large-scale pre-trained language model.

8. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The intervention decision engine is connected to an intervention strategy library, which contains multiple scenario scripts: a safe seaside scenario, a tranquil forest scenario, a neutral breathing guidance scenario, and an environmental homeostasis optimization scenario. Each scenario script specifies the sound source type, volume range, and gradation rhythm for the sound channel; the brightness, color temperature, and change curve for the light channel; the fragrance type and release rate for the olfactory channel; the enveloping sensation and vibration mode for the tactile channel; the target temperature range for the zoned temperature control channel; and the stimulation amplitude and frequency range for the neural stimulation channel. The intervention decision engine has a multi-model integration and voting module. Based on the risk curves output by the world model and the time series model, the dream semantic vector, the health prescription target, and the physiological and psychological homeostasis index improvement performance of each scenario script in historical data, the engine performs weighted voting scoring and ranking of candidate scenario scripts and selects the target scenario script.

9. The multi-model closed-loop dream sleep intervention system according to claim 1, characterized in that, The sound intervention unit in the execution layer plays ambient sounds, music, or voice guidance content through bedside speakers or headphones, and outputs interactive audio information. It presents intervention data related to psychological intervention, relaxation training, or positive suggestion to the user in the form of audio information without interrupting the user's sleep, allowing the intervention data to be transmitted to the user through the user's dreams. The light intervention unit controls the light intensity and color temperature through dimmable lamps, the olfactory intervention unit releases fragrance through atomizing devices, the tactile and pressure intervention unit adjusts local pressure and micro-vibration patterns through actuators inside the mattress, the zoned temperature control unit works with air conditioning, floor heating, or the mattress temperature control layer to regulate the temperature of the torso and limb areas, and the nerve stimulation unit applies controlled electrical stimulation through electrodes on the skin surface. The execution layer gradually activates the corresponding channels before the predicted high-risk time window according to the intervention instructions generated by the intervention decision engine.

10. A multi-model closed-loop dream sleep intervention method, applied to the system according to any one of claims 1 to 9, characterized in that, Includes the following steps: Step 1: Baseline modeling. The perception layer collects multimodal physiological and environmental signals over several nights, and the application layer collects dream content and subjective evaluation of mental state. The above data is input into the predictive mind-body model and the feedback adaptive learning unit for training, and a user-individualized implicit state space and scene script intervention effect statistics are established. Step 2, online prediction: After the user falls asleep, the perception layer continuously outputs multimodal features by sliding time windows, predicts the mind-body model to update the current latent state in each time window, and makes rolling predictions for multiple future time windows based on the world model and time series model, outputting nightmare risk curves and mental stress change curves. Step 3, Strategy Decision-Making: The intervention decision engine, based on the nightmare risk curve, mental stress change curve, dream semantic vector, and sleep health prescription goals, uses a multi-model integration and voting module to perform weighted voting scoring and sorting of multiple scenario scripts in the intervention strategy library, generating intervention instructions that include sound, light, smell, touch, zoned temperature control, and neural stimulation control parameters. Step four, multi-channel intervention: The execution layer, following the intervention instructions output by the intervention decision engine, jointly controls channels such as sound, light, smell, touch, zoned temperature control, and nerve stimulation. Before the predicted high-risk time window, multimodal stimulation is applied in a progressive manner to guide the user's dream emotions and physiological indicators back to the safe zone. Among them, the sound intervention channel outputs information interaction audio while playing ambient sounds or music, and transmits intervention data to the user through the user's dream to work with the multi-channel intervention to regulate the user's dream experience and mental stress. Step 5: Closed-loop feedback. The adaptive learning unit calculates the physiological and psychological homeostasis index before and after the intervention, and statistically analyzes the actual improvement effect of each scenario script under different implicit states and dream label combinations. The "perception-prediction-intervention-response" data is used as incremental samples to update the parameters of the predictive mind-body model and the intervention decision engine, so that the system can form an individualized dream and mental intervention model for different users in the long run.

Images