Intelligent sleep environment adjusting system and method

Abstract

The invention relates to the technical field of sleep regulation, and provides an intelligent sleep environment regulation system and method. The system comprises a main control module, a sleep sensing module, a sleep stage inference module, a lighting effect strategy generation module and a lighting effect execution module, wherein the sleep sensing module senses vital sign signals, temperature distribution information and space contour information in a monitoring area in a non-contact manner; the sleep stage inference module is used for performing fusion analysis on the multi-source heterogeneous data acquired by the sleep perception module and inferring the sleep stage of the user; the main control module generates different light control instructions based on different sleep stages, the lighting effect strategy generation module responds to the light control instructions and outputs lighting effect driving signals corresponding to the sleep stages, and the lighting effect execution module generates sleep adjustment lighting effects according to the lighting effect driving signals. According to the scheme, wearing discomfort and sleep interference are eliminated, and the defects that a single sensor is low in precision, poor in interference resistance and weak in scene understanding are overcome.

Description

Technical Field

[0001] This invention relates to the field of sleep regulation technology, and in particular to an intelligent sleep environment regulation system and method. Background Technology

[0002] Sleep is an essential life activity for humans, and good sleep is crucial to human health. Sleep problems such as sleep disorders and sleep diseases can affect a person's work and life, and can become a trigger for many diseases. Therefore, it is necessary to monitor a person's sleep quality.

[0003] Chinese patent CN110585551A discloses a light-based wake-up control system based on sleep stages. The system includes a basic data acquisition unit, a processor, and a wake-up device. It calculates the user's sleep stage based on basic data and controls the wake-up device to generate differentiated light stimuli that gradually increase in intensity throughout the wake-up process. However, the basic data acquisition unit obtains the user's basic data through a portable wearable device, and this wearable monitoring can cause discomfort and disrupt sleep.

[0004] Therefore, existing sleep regulation solutions suffer from technical problems such as user discomfort and high misjudgment rates. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent sleep environment regulation system and method, which aims to solve the technical problems of user discomfort and high misjudgment rate in existing sleep regulation solutions.

[0006] In a first aspect, this application provides an intelligent sleep environment regulation system, including a main control module, a sleep sensing module, a sleep stage inference module, a light effect strategy generation module, and a light effect execution module. The sleep sensing module uses non-contact sensing to monitor vital signs, temperature distribution information, and spatial contour information within the monitoring area. The vital signs include the user's respiratory rate, heart rate, and body movement signals. The sleep stage inference module fuses and analyzes the multi-source heterogeneous data sensed by the sleep sensing module and infers the user's sleep stage; the main control module generates different light control commands based on different sleep stages; the light effect strategy generation module responds to the light control commands and outputs a light effect driving signal corresponding to the sleep stage; and the light effect execution module generates sleep adjustment light effects according to the light effect driving signal.

[0007] Secondly, this application provides an intelligent sleep environment regulation method, applied to any of the intelligent sleep environment regulation systems described above, comprising the following steps: S100: The sleep sensing module continuously acquires multi-source heterogeneous data synchronously in time, and the main control module performs filtering, noise reduction, non-uniformity correction and anti-shake preprocessing on the original multi-source heterogeneous data. S200: The sleep stage inference module performs feature extraction and deep fusion on preprocessed multi-source heterogeneous data, and infers the user's sleep stage through a neural network model; S300: The main control module receives the sleep stage and dynamically generates light control commands by combining the user's historical preference model; S400: The light effect strategy generation module responds to the light control command and outputs a light effect driving signal that matches the current sleep stage; S500: The system encrypts and stores complete closed-loop data, including raw sensor data, recognition results, light control commands, and status changes after feedback; S600: Return to step S100 to continue real-time monitoring. During idle periods, use the closed-loop data recorded in step S500 to perform local incremental learning on the neural network model, optimize the recognition accuracy and the adaptability of the feedback strategy, and form an intelligent closed loop.

[0008] The beneficial effects of the intelligent sleep environment regulation system and method provided by this invention are as follows: The sleep sensing module uses non-contact sensing to monitor vital signs, temperature distribution information, and spatial contour information within the monitoring area, eliminating wearing discomfort and sleep interference, and supporting long-term imperceptible monitoring; The sleep stage inference module deeply integrates multi-source heterogeneous data to achieve accurate sleep stage identification, overcoming the shortcomings of low accuracy, poor anti-interference, and weak scene understanding of single sensors, significantly reducing the false judgment rate and improving system robustness; Closed-loop regulation is completed through sleep regulation light effect, realizing the upgrade from passive monitoring to active sleep regulation, making the interaction more intelligent, and solving the technical problems of user discomfort and high false judgment rate in existing sleep regulation solutions. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the intelligent sleep environment regulation system provided in an embodiment of the present invention; Figure 2 Another structural schematic diagram of the intelligent sleep environment regulation system provided in the embodiment; Figure 3 A schematic diagram showing the connection between the blue light sensing unit and the main control module provided in this embodiment; Figure 4 A schematic diagram of the workflow of the intelligent sleep environment regulation system provided in this embodiment; Figure 5 This is a flowchart illustrating the intelligent sleep environment adjustment method provided in this embodiment.

[0011] The following are the labeling elements in the figure: 10. Main control module; 20. Sleep sensing module; 21. Radar module; 22. Thermal imaging module; 23. Passive infrared sensor; 24. Blue light sensing unit; 241. PEN flexible substrate; 242. Indium tin oxide conductive layer; 243. Nano-tin oxide and boron-nitrogen hybrid semiconductor heterojunction layer; 244. Molybdenum oxide electron blocking layer; 245. Silver electrode layer; 30. Sleep stage inference module; 40. Light effect strategy generation module; 50. Light effect execution module; 60. Audio module; 61. Digital audio decoding chip; 62. Miniature speaker; 71. Touch sensing module; 72. Housing; 73. Power management module. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0013] Throughout this specification, references to "an embodiment" or "an embodiment" mean that a particular feature, structure, or characteristic described in connection with an embodiment is included in at least one embodiment of this application. Therefore, the phrases "in one embodiment" or "in some embodiments" appear in various places throughout the specification, and not all refer to the same embodiment. Furthermore, in one or more embodiments, particular features, structures, or characteristics may be combined in any suitable manner.

[0014] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0015] Existing sleep monitoring solutions are limited to "monitoring" and "data reporting," forming an open-loop system. Even those few solutions that attempt to provide feedback (such as light reminders) often employ fixed, preset, and non-personalized feedback strategies. They cannot dynamically adjust and adaptively optimize based on the user's real-time physiological state, sleep stage, and historical habits, resulting in low levels of intelligence and limited improvement in user experience.

[0016] Example 1 Combination Figure 1 The intelligent sleep environment regulation system provided in this application includes a main control module 10, a sleep sensing module 20, a sleep stage inference module 30, a light effect strategy generation module 40, and a light effect execution module 50. The sleep sensing module 20 uses non-contact sensing to monitor vital signs, temperature distribution information, and spatial contour information within the monitoring area. Vital signs include the user's respiratory rate, heart rate, and body movement signals.

[0017] The sleep stage inference module 30 fuses and analyzes the multi-source heterogeneous data sensed by the sleep sensing module 20 to infer the user's sleep stage. The multi-source heterogeneous data includes vital signs, temperature distribution information, and spatial contour information. The main control module 10 generates different light control commands based on different sleep stages. The light effect strategy generation module 40 responds to the light control commands, outputting light effect driving signals corresponding to the sleep stage. The light effect execution module 50 generates sleep-regulating light effects based on the light effect driving signals. All modules work collaboratively to achieve a complete closed loop of "non-contact sensing - AI analysis - dynamic adjustment - data feedback."

[0018] Based on this, the intelligent sleep environment adjustment system has the following advantages: First, the sleep sensing module 20 completely eliminates user wear and contact, achieving truly imperceptible sleep monitoring. This addresses the root causes of discomfort and sleep disturbance caused by wearable monitoring, supporting long-term, continuous, and burden-free home sleep monitoring. It does not use optical imaging modules such as cameras, but only monitors through radar micro-motion signals and thermal imaging temperature / contour data. There is no facial or human image acquisition, avoiding the dual defects of privacy leaks and failure in low-light environments associated with camera-based solutions. Second, the sleep stage inference module 30 performs feature-level fusion and intelligent inference on heterogeneous data, achieving accurate identification of sleep stages. This solves the common industry problem of high misjudgment rates and poor robustness in complex home scenarios (obstruction, environmental interference, changes in body position). Third, after the sleep stage inference module 30 identifies the sleep stage, the main control module 10 directly and dynamically generates light control commands, and the light effect execution module 50 outputs sleep adjustment light effects matching the sleep stage, achieving active sleep adjustment monitoring.

[0019] In this embodiment, the light effect execution module 50 uses RGBW full-color LED beads with a color temperature adjustment range of 2700K~6500K and a brightness adjustment range of 0~100%. The light effect strategy generation module 40 uses a high-performance lighting control driver chip, such as the TPIC6B595 chip, which supports 8-channel PWM output, is connected to the main control module 10 through the SPI interface, has a drive current ≤500mA, and can control RGBW four-channel LED beads.

[0020] In this embodiment, the sleep-regulating light effect specifically includes: (1) Intelligent sleep aid light effect: The rhythm and brightness change curve of the "breathing light" are dynamically adjusted according to the user's difficulty in falling asleep. It can be paired with the sleep-aid white noise of the audio module 60 below to help the user fall asleep quickly. (2) Unobstructed night light effect, accurately recognizes the intention to "sit up" or "get out of bed", and automatically triggers low brightness (brightness ≤10%) guide light to ensure safety without interrupting sleep; (3) Health intervention light effect: When a possible sleep apnea event is detected, a very soft light (brightness ≤5%, color temperature 2700K) is used to provide a warning and non-invasive intervention to avoid affecting the user's sleep; (4) Personalized wake-up light effects, combined with sleep cycle (light sleep period) and user history preferences, simulate the gradual change of light temperature at sunrise in the morning to achieve painless wake-up and improve the user's morning state.

[0021] In some embodiments, the sleep sensing module 20 includes a radar module 21 and a thermal imaging module 22, both capable of non-contact sensing. The radar module 21 is used to non-contactly sense and monitor the user's vital signs signals such as breathing, heart rate, and body movement within the monitoring area. The thermal imaging module 22 is used to non-contactly sense and monitor temperature distribution and spatial contour information within the monitoring area. The radar module 21 has the advantages of penetrating bedding and resisting obstruction, while the thermal imaging module 22 is responsible for spatial contour, temperature distribution, and human position determination, compensating for the radar module 21's inability to intuitively distinguish human contours and scene environments. The data from the two modules complement each other, solving the problems of single infrared being susceptible to temperature interference, single radar being unable to determine whether the user is in bed or out of bed, and weak scene understanding.

[0022] In one embodiment, combined Figure 1The sleep sensing module 20 also includes a passive infrared sensor 23 (PIR sensor). The PIR sensor 23 is used to sense human movement signals within the monitoring area, assisting the sleep stage inference module 30 in identifying scene states and compensating for the perception deficiencies of the radar module 21 in static scenes. The PIR sensor 23 outputs a digital switch signal, which is connected to the main control module 10 via a GPIO pin, providing low-power, wide-range presence sensing. At this time, multi-source heterogeneous data includes vital signs, temperature distribution information, spatial contour information, and human movement signals. The sleep stage inference module 30 identifies scene states based on human movement signals, including whether the user is in bed, out of bed, or distinguishes between humans and pets. That is, the system can not only analyze physiological signals but also combine spatial information obtained from thermal imaging with activity information obtained from the PIR sensor 23 to accurately determine whether the user is "lying still in bed," "active out of bed," or "the room is empty," and can distinguish between humans and pets, thereby achieving more realistic intelligent decision-making. The light effect strategy generation module 40 combines the scene state and sleep stage to output the corresponding light effect driving signal to adjust the brightness, color, illumination range and beam angle of the light effect execution module 50.

[0023] Based on this, the system deeply integrates the radar module 21, thermal imaging module 22, passive infrared sensor 23, and lightweight sleep stage inference module 30 to form an edge intelligent agent that integrates perception, cognition, and interaction. This system not only achieves high-precision non-contact vital sign (respiration, heart rate, body movement) monitoring through multimodal fusion, but also performs real-time analysis and state inference (such as sleep stage and abnormal events) of the monitoring data through the built-in sleep stage inference module 30. Based on this, it dynamically generates personalized light effects, sound, and interactive feedback strategies that match the user's current state and long-term habits, thus forming a next-generation intelligent terminal with environmental perception, health monitoring, intelligent companionship, and adaptive adjustment capabilities.

[0024] It should be noted that this is combined with the frequency of radar micro-motion signals (human breathing rate 12-20 breaths / minute, pet breathing rate 20-40 breaths / minute) and the area threshold of the thermal imaging contour (human body ≥ 0.5m). 2 Pets ≤ 0.3m 2 The duration of the PIR motion signal is used to form a triple verification, which improves the accuracy of human-pet differentiation to over 98%.

[0025] In one embodiment, combined Figure 2 and Figure 3The sleep sensing module 20 also includes a blue light sensing unit 24, which comprises, from bottom to top, a PEN flexible substrate 241, an indium tin oxide conductive layer 242, a nano-tin oxide and boron-nitrogen hybrid semiconductor heterojunction layer 243, a molybdenum oxide electron blocking layer 244, and a silver electrode layer 245. The PEN flexible substrate 241 is flexible and bendable, adaptable to the inner wall of the ambient light housing 72 or mounted near the light effect execution module 50. Specifically, the PEN flexible substrate 241 is attached to the curved surface of the LED beads, ensuring that the blue light detection angle matches the emission angle of the light effect execution module 50, thus improving detection accuracy. The indium tin oxide conductive layer 242 serves as the bottom electrode. The nano-tin oxide and boron-nitrogen hybrid semiconductor heterojunction layer 243 is the core photosensitive layer, absorbing wavelengths of 400-470nm (harmful blue light). The molybdenum oxide electron blocking layer 244 suppresses dark current and improves detection rate. The silver electrode layer 245 serves as the top electrode, outputting a photocurrent signal.

[0026] Among them, the blue light sensing unit 24, which uses nano-tin oxide and boron-nitrogen hybrid semiconductor heterojunction layer 243, has good response capability in the 400nm to 468nm spectral band, has a narrow spectral response half-width (FWHM) of less than 36nm, and the external quantum efficiency of the blue light sensing unit 24 is greater than 14%, which helps to realize the light detection of blue light that is harmful to the eyes.

[0027] Specifically, the indium tin oxide conductive layer 242 and the silver electrode layer 245 of the blue light sensing unit 24 are connected to the main control module 10 via GPIO, outputting an analog photocurrent signal, which is converted into a digital blue light intensity signal by the built-in ADC. The blue light sensing unit 24 faces the monitoring area, simultaneously monitoring the ambient blue light and the blue light component of the light effect execution module 50 itself. The blue light sensing unit 24 is uniformly supplied with 3.3V DC power by the power management module 73, with a static power consumption ≤5mW, without adding additional system load.

[0028] Specifically, in the nano-tin oxide and boron-nitrogen hybrid semiconductor heterojunction layer 243, the mass ratio of nano-tin oxide to boron-nitrogen is 8:2 to 9:1, which improves the external quantum efficiency of blue light detection to more than 18%, reduces the full width at half maximum (FWHM) of the spectral response to less than 30 nm, and reduces the static power consumption to less than 3 mW.

[0029] When the blue light sensing unit 24 detects that the blue light intensity is greater than the nighttime threshold (e.g., 10 μW / cm²), 2 If the sleep stage inference module 30 detects that the user is awake or has difficulty falling asleep, the main control module 10 determines it to be blue light interference-induced insomnia. The light effect strategy generation module 40 immediately reduces the 400-470nm blue light component in the light effect execution module 50, locking the color temperature at 2700-3000K (warm yellow light). The nighttime threshold can be selected as 10μW / cm². 2 .

[0030] When the blue light sensing unit 24 detects that the blue light intensity is less than or equal to the nighttime threshold, and the sleep stage inference module 30 identifies that the user is having difficulty falling asleep, the main control module 10 determines that it is physiological insomnia and executes a sleep aid strategy, namely intelligent sleep aid light effect, which dynamically adjusts the rhythm of "breathing light" and the type of sleep-accompanied white noise according to the user's difficulty in falling asleep.

[0031] When the blue light sensing unit 24 detects that the blue light intensity is greater than the nighttime threshold, and the sleep stage inference module 30 identifies that the user is in deep sleep or REM sleep, the main control module 10 determines that it is an ambient light interference and triggers the extremely low brightness warm light protection.

[0032] In this embodiment, the sleep sensing module 20 consists of a radar module 21, a thermal imaging module 22, a passive infrared sensor 23, and a blue light sensing unit 24. It has a simple structure and does not include other modules.

[0033] In one embodiment, the main control module 10 employs a high-performance, low-power wireless communication microcontroller (MCU), such as the ESP32-S3. This MCU integrates a Wi-Fi wireless communication unit and / or a Bluetooth wireless communication unit, and possesses basic AI instruction acceleration capabilities (such as vector operations), and is responsible for overall system control, task scheduling, network communication, and some lightweight AI inference tasks.

[0034] In one embodiment, the sleep stage inference module 30 is integrated inside the main control module 10, sharing a signal transmission channel with the main control module 10 to achieve real-time data interaction. The main control module 10 has AI hardware acceleration capabilities, utilizing its dedicated computing unit to accelerate neural network inference, and complete multimodal data fusion, sleep stage classification, and anomaly detection.

[0035] In one embodiment, the sleep stage inference module 30 is separately configured from the main control module 10. The sleep stage inference module 30 is connected to the main control module 10 via a communication bus to realize the transmission of multi-source heterogeneous data and the feedback of recognition results. Specifically, the sleep stage inference module 30 is an independent low-power AI acceleration chip that works separately and collaboratively with the main control module 10 via a high-speed SPI or I2C communication bus. The sleep stage inference module 30 runs a lightweight neural network model, receives preprocessed multimodal data, performs multimodal feature fusion, temporal pattern recognition, sleep stage classification, and abnormal event prediction, and returns the inference results (sleep stage label, apnea probability, and scene state) to the main control module 10 for decision-making.

[0036] In one embodiment, radar module 21 is a millimeter-wave radar, specifically employing a 60GHz or 77GHz FMCW radar chip, such as the Infineon BGT60LTR11AIP. Radar module 21 communicates with the main control module 10 or the sleep phase inference module 30 via SPI or I2C interfaces for configuration, and outputs intermediate frequency (IF) signals through analog or digital interfaces. The radar IF output signal is input to the main control module 10 for preprocessing via an analog-to-digital converter (ADC) (which can be built into the radar chip or externally placed on the MCU).

[0037] The radar module 21 emits a linear frequency modulated continuous wave (frequency changes linearly with time). Upon encountering a human body, the emitted wave is reflected back, and the transmitted wave and the reflected echo are mixed to generate a difference frequency signal (output to the main control module 10 via an intermediate frequency). The difference frequency is proportional to the target distance and can be used to locate the human body. Since human breathing, heartbeat, and body movement cause slight displacements in the chest and abdomen, the radar module 21 detects the phase changes and Doppler frequency shifts caused by these displacements to invert vital signs signals.

[0038] In one embodiment, the thermal imaging module 22 adopts an uncooled microthermometer module, which outputs digital temperature matrix data. It communicates directly with the main control module 10 or the sleep stage inference module 30 through the I2C or SPI interface, providing scene temperature distribution and spatial contour information, and enabling auxiliary judgment of human body position and existence status. This compensates for the shortcomings of the radar module 21 in not being able to intuitively distinguish human body contours and scene environment, and assists in judging the user's status in bed / out of bed.

[0039] Among them, the thermal imaging module 22 adopts an uncooled microthermometer to receive infrared thermal radiation radiated by objects in the monitoring area. It converts the infrared signal into an electrical signal through thermoelectric conversion, and outputs a two-dimensional digital temperature matrix after calibration processing, thereby obtaining the temperature distribution information of the monitoring area. At the same time, based on the temperature difference between the human body and the environment, it identifies and outputs spatial contour information such as the human body's spatial contour, position, and area size through temperature segmentation and edge extraction.

[0040] In one embodiment, the light effect strategy generation module 40 receives the PWM (pulse width modulation) signal from the main control module 10 and drives the multiple (such as RGBW) high-brightness LED beads of the light effect execution module 50 to achieve precise control of color and brightness.

[0041] In one embodiment, the light effect execution module 50 generates sleep regulation light effects that match the sleep stage based on the light effect driving signal, including adaptive pulsating light, color temperature gradient light, low brightness guiding light, sunrise simulated wake-up light, abnormal indication light, etc.

[0042] In the above embodiments, the system adopts a modular design. The hardware utilizes industry-standard chipsets (such as Infineon radar chips, ESP32 main control, and uncooled thermal imaging module 22), while the software algorithms can be built based on the rich development kits provided by the chip manufacturers. This modular design significantly reduces R&D difficulty and production costs, ensuring rapid product deployment and stable reliability.

[0043] In one embodiment, the sleep stage inference module 30 runs a pre-trained lightweight neural network model (hereinafter referred to as the "model"). The model is built based on the Transformer architecture and consists of an input layer, a multi-head attention layer, a feedforward neural network layer, and an output layer from bottom to top. The input layer receives multimodal fusion features, the multi-head attention layer sets up four attention heads to process the input data, the feedforward neural network layer contains two fully connected layers for further data processing, and the output layer provides the probability distribution of sleep stages and sleep events. The model is pre-trained using publicly available sleep monitoring datasets such as CitySleep and SleepEDF, and then fine-tuned and optimized based on the measured data collected by this system. Specifically, it performs at least one of the following functions: Function 1 involves time-aligning and feature-level fusion of multi-source heterogeneous data, and using an adaptive filtering algorithm to suppress interference from ambient temperature and light, generating joint features resistant to environmental interference. The process is as follows: Step 1: The sleep stage inference module 30 collects and extracts structured features from each perception module to form a 10-dimensional basic feature set, with the following dimensions: 1) Radar module 21 outputs 4-dimensional features: respiratory rate (f_breath), heart rate (f_heart), body motion amplitude (A_motion), and body motion event flag (Flag_motion, 0=no body motion, 1=body motion). 2) The thermal imaging module 22 outputs 3D features: human core temperature (T_core), human contour area (S_contour), and contour movement distance (D_contour); 3) Passive infrared sensor 23 outputs 2D features: motion signal intensity (I_pir) and motion event flag (Flag_pir, 0 = no motion, 1 = motion). 4) The blue light sensing unit 24 outputs a 1-dimensional feature: blue light intensity (I_blue).

[0044] Step 2: The Transformer model of the sleep stage inference module 30 preprocesses the data.

[0045] 1) Time alignment: The model unifies all features to a 1Hz sampling frequency through linear interpolation and downsampling; a time series data sequence D={d1,d2,...,d...} is constructed with a time window length of 30s. 30}, where each time step d t It is a 10-dimensional feature vector (t∈[1,30]), and the sequence dimension is 30×10.

[0046] 2) Z-score normalization: x′=(x-μ) / σ, where μ is the mean of the feature on the training set and σ is the standard deviation of the feature on the training set. After normalization, the feature range is mapped to [-1,1], eliminating the difference in dimensions.

[0047] Step 3: The Transformer model performs feature-level fusion. 1) Input layer encoding. The 30×10-dimensional temporal feature sequence is converted into a vector that can be processed by the Transformer. First, a position vector is added for each time step (1~30s) to preserve the temporal information; second, the 10-dimensional features are mapped to a 64-dimensional embedding space through a fully connected layer (lightweight design to reduce computation), and the output is a 30×64-dimensional embedding sequence.

[0048] 2) The four attention heads focus on the feature associations of different dimensions.

[0049] Attention Head 1: Captures the physiological correlation between respiration and heart rate (e.g., the pattern of heart rate changes during apnea); Attention Head 2: Captures the behavioral correlation between body movement and contour movement (e.g., the synchronicity of body movement amplitude and contour movement when turning over); Attention Head 3: Captures the cross-modal correlation between physiology and behavior (e.g., the correspondence between respiratory rate and body movement during light sleep); Attention Head 4: Captures temporal correlations (e.g., the trend of respiratory rate changes within 30 seconds). Each attention head outputs 30×16-dimensional features (64 total dimensions ÷ 4 heads), which are then concatenated to obtain 30×64-dimensional fused features, achieving multi-dimensional feature extraction of "physiology + behavior + temporal correlation".

[0050] 3) Extraction of high-order features from feedforward neural network layers.

[0051] First layer: 64-dimensional → 128-dimensional fully connected layer, with ReLU activation function, to explore higher-order non-linear relationships between features; The second layer is a 128-dimensional to 64-dimensional fully connected layer, which reduces the dimensionality and retains the core features, outputting a fused 30×64-dimensional final feature sequence.

[0052] Step four: Infer the fused features using the classification head. The 30×64-dimensional final feature sequence is compressed into a 64-dimensional feature vector through global average pooling, input into the Softmax layer, and outputs the probability distribution of various sleep events (such as "apnea", "normal sleep", "body movement disturbance", "environmental disturbance", etc.). The event category with the highest probability is taken as the final recognition result.

[0053] Function 2: Classify users' sleep stages, which include at least the waking state, light sleep state, deep sleep state, and REM sleep state; identify sleep sub-states, including difficulty falling asleep in light sleep state, frequent turning over in light sleep state, and mild breathing disorders in deep sleep state.

[0054] For example, see the table below for specific classifications.

[0055] ; Function 3: Detects abnormal physiological events, including mild respiratory disturbances and sleep apnea.

[0056] In one embodiment, the neural network model possesses local incremental learning capabilities. Specifically, it utilizes idle periods to adaptively fine-tune model parameters based on historical monitoring data stored in the system, feedback effects from user input via terminal or local touch, and sleep preference information, thereby optimizing state recognition accuracy and feedback strategy adaptability. Idle periods are defined as the time after the user leaves the bed and during which the system does not perform real-time sleep monitoring.

[0057] In one embodiment, the sleep regulation lighting effect is dynamically generated by the main control module 10 based on the recognition results and user preferences. It prioritizes matching the sleep stage and then fine-tunes it according to user preferences. Specifically, the sleep regulation lighting effect includes: adaptive pulsating light synchronized with the user's breathing rhythm to guide breathing and assist in falling asleep; color temperature gradient light corresponding to sleep stages such as light sleep / deep sleep to maintain the corresponding sleep stage; low-brightness guiding light corresponding to the out-of-bed scenario to avoid strong light stimulation; wake-up light simulating sunrise to gently wake the user; and soft alert light for abnormal states to indicate abnormalities without interfering with sleep.

[0058] In one embodiment, the intelligent sleep environment adjustment system further includes a photoelectric module 50, which is electrically connected to the main control module 10. The photoelectric module is installed on the light effect execution module 50 so that the pulse light emitted by the photoelectric module 50 is consistent with the light emission angle of the light effect execution module 50, and can accurately reach the user's fingertips to obtain the user's blood oxygen data.

[0059] In one embodiment, combined Figure 1 and Figure 2 The intelligent sleep environment adjustment system also includes an audio module 60 that is communicatively connected to the main control module 10. The audio module 60 includes a digital audio decoding chip 61 and a miniature speaker 62. Based on the sleep stage recognition results, the main control module 10 controls the audio module 60 to play audio content that coordinates with the sleep adjustment light effect. The coordination method is that the volume and rhythm of the audio are synchronized with the brightness and pulse frequency of the sleep adjustment light effect. The audio content includes sleep-aiding white noise and natural sounds to guide sleep, and gradually increasing music for gentle wake-up.

[0060] Specifically, the audio module 60 is connected to the main control module 10 via an I2S interface and is used to play sleep-aid or wake-up audio content selected by the AI ​​strategy engine.

[0061] In one embodiment, combined Figure 1 and Figure 2 The intelligent sleep environment adjustment system also includes a touch sensing module 71 that is communicatively connected to the main control module 10. The touch sensing module 71 uses a capacitive touch chip to receive local user control input. Local control input includes commands to adjust sleep adjustment light effect parameters, audio playback mode adjustments, and sleep mode switching. The main control module 10 responds to user input, adjusting the parameters of the sleep adjustment light effect and the audio playback mode of the audio module 60 in real time. Specifically, the touch sensing module 71 uses a capacitive touch chip (such as the QT series) to detect user touch, providing an intuitive local input interface.

[0062] In one embodiment, the main control module 10 integrates a wireless communication unit, which is used to encrypt and transmit the anonymized sleep report and recognition results to the user terminal. The anonymization method involves removing the user's personal identification information and retaining only sleep-related data. The user terminal includes mobile terminals such as smartphones, tablets, and smartwatches. The main control module 10 integrates a wireless communication unit to encrypt and transmit the recognition results, vital sign trends, and system feedback records to the user terminal APP for local storage, analysis, and visualization.

[0063] This solution employs millimeter-wave radar and thermal imaging technology, which do not involve optical imaging, fundamentally eliminating the risk of privacy leaks. All raw data and AI models are processed and analyzed locally on the device, with only anonymized sleep reports being encrypted and synchronized to the user's mobile app, ensuring the absolute privacy of personal health data.

[0064] In one embodiment, combined Figure 1 and Figure 2 The system also includes a housing 72, in which the main control module 10, sleep sensing module 20, sleep stage inference module 30, light effect strategy generation module 40, and light effect execution module 50 are all integrated. No additional equipment needs to be placed in the bedroom, maintaining the aesthetics and simplicity of the home environment. The monitoring process is completely contactless and wearable, relieving users of any burden or psychological resistance.

[0065] In one embodiment, combined Figure 1 and Figure 2 The system also includes a power management module 73, which provides a stable and efficient DC voltage for all the above modules and supports power supply through common interfaces such as USB-C. In particular, it provides sufficient current to ensure the stable operation of the sleep phase inference module 30 under high load.

[0066] In summary, this system, based on embedded AI-powered multimodal sleep perception and closed-loop adjustment, deeply integrates millimeter-wave radar, thermal imaging, and PIR sensors into a smart ambient light, forming a multimodal perception array. It also innovatively introduces an embedded sleep stage inference module 30 as the system's intelligent hub. This sleep stage inference module 30 performs real-time fusion analysis and deep understanding of multi-source heterogeneous sensor data, accurately identifying the user's sleep stage, physiological state, and abnormal events. Based on the AI's recognition results, the main control module 10 dynamically generates and executes highly personalized lighting effects (such as adaptive breathing rhythm lights and intelligent wake-up gradient lights) and audio (such as context-matched white noise) feedback strategies through an autonomous decision engine, achieving a leap from "imperceptible monitoring" to "intelligent cognition and proactive personalized adjustment." This system provides a highly integrated, privacy-protecting, and continuously learning-capable bedroom health sleep companion solution.

[0067] Combination Figure 4 The specific working process of this system is as follows: The first step is system power-on initialization. In other words, the system powers on and initializes all hardware modules. A pre-trained and user-tuned lightweight neural network model is loaded from the device's local storage into the sleep stage inference module 30. The lighting enters the default warm scene mode (the color temperature of the lighting effect execution module 50 is 2700K, and the brightness is 30%).

[0068] In the second step, the sleep sensing module 20 acquires multi-source heterogeneous data. The radar module 21, thermal imaging module 22, passive infrared sensor 23, and blue light sensing unit 24 begin continuous data acquisition in synchronized time. The main control module 10 performs first-level preprocessing on the raw data to improve data quality. Specifically, this includes: filtering and denoising the radar intermediate frequency signal; correcting the non-uniformity of the thermal imaging temperature matrix; performing image stabilization on the PIR signal; and enhancing the current signal sensed by the blue light sensing unit 24 using a moving average method.

[0069] Specifically, the main control module 10 integrates a filter, function generator, lock-in amplifier, and data acquisition card, specifically for filtering and denoising radar intermediate frequency signals. First, a 50kHz high-pass filter is used to filter the vital signs signal, focusing on removing low-frequency interference noise. The pre-processed signal is then acquired and initially buffered by the data acquisition card, providing a foundation for further denoising. Subsequently, the high-pass filtered vital signs signal is fed into the lock-in amplifier. Simultaneously, the function generator generates a reference signal with the same modulation frequency as the signal, which is synchronously input into the lock-in amplifier. By identifying the frequency characteristics of the reference signal, the lock-in amplifier selectively suppresses various types of clutter interference unrelated to the reference signal frequency, thereby achieving high signal-to-noise ratio detection and accurate measurement of the vital signs signal, effectively improving signal purity.

[0070] In the entire radar signal processing flow, the signal-to-noise ratio (SNR) of the processed vital signs signals needs to be calculated to determine the signal quality. This SNR is denoted as . The specific calculation formula is as follows:

[0071] Among them, V radar V represents the effective value of the vital sign signal voltage output by radar module 21. noise The SNR represents the effective value of the noise signal voltage generated by the radar module 21 itself. radar This refers to the signal-to-noise ratio (SNR) calculated using both methods. After the calculation is complete, the resulting SNR is... radar The signal-to-noise ratio (SNR) is compared with a preset threshold for determination: if the SNR... radar If the SNR is greater than the preset threshold, it indicates that the vital sign signal quality meets the standard and the signal acquisition is successfully completed; if the SNR is greater than the preset threshold, it indicates that the vital sign signal quality meets the standard and the signal acquisition is successfully completed. radar If the preset threshold is not reached, the control radar module 21 will restart the signal detection process until the signal-to-noise ratio (SNR) of the acquired signal is reached. radar The system meets the set requirements. As a preferred implementation, the preset signal-to-noise ratio threshold in this system can be selected as 30dB, which balances signal quality and detection efficiency.

[0072] Specifically, during the calibration phase, before leaving the factory, the thermal imaging module 22 is placed in two known calibration temperature environments (low temperature T_low, such as 25℃; high temperature T_high, such as 50℃) and two sets of temperature matrices are acquired: R awL [H][W] (Original pixel values ​​at low temperature), R awH [H][W] (Original pixel values ​​at high temperature); then, calculate the gain coefficient Gain[i][j] and offset coefficient Offset[i][j] for each pixel (i,j):

[0073] ; Finally, the Gain and Offse matrices are stored in the Flash memory of the main control module 10.

[0074] During the real-time correction phase, the main control module 10 reads the raw temperature matrix Raw[H][W] output by the thermal imaging module 22; performs correction calculations for each pixel (i,j) to obtain the homogenized temperature value: ; The corrected matrix is ​​then subjected to boundary value clipping (e.g., limited to -20℃~80℃ to adapt to human body monitoring scenarios), and the final temperature matrix is ​​output.

[0075] Specifically, PIR sensors are susceptible to false signals triggered by minor environmental disturbances (such as airflow and changes in light). The image stabilization process filters out short-pulse interference through "continuous sampling confirmation within a time window," retaining only valid human motion signals, thus meeting the low false triggering requirements of sleep monitoring scenarios.

[0076] In a specific example, first, the parameter settings are as follows: Anti-shake confirmation window: 200ms (balancing response speed and anti-interference); Sampling frequency: 10Hz (sampling PIR level once every 100ms); Valid judgment condition: two consecutive high-level samples within the 200ms window (trigger state) are judged as valid human motion, and two consecutive low-level samples (idle state) are judged as no motion.

[0077] Secondly, the main control module 10 reads the PIR signal level (0 = no motion, 1 = motion detected) through GPIO interrupt or timed sampling; maintains a sampling buffer of length 2 to store the two most recent sampling values; if two consecutive values ​​in the buffer are 1, outputs a "valid motion" signal; if two consecutive values ​​are 0, outputs a "no motion" signal; otherwise, it maintains the previous state to avoid frequent switching.

[0078] Specifically, the analog current signal output by the blue light sensing unit 24 contains random noise after being converted by the ADC. The moving average method smooths out noise fluctuations by calculating the average value of the most recent N sampling points, thereby improving the stability of blue light intensity detection and ensuring the real-time performance of the signal.

[0079] In a specific example, first, the parameter settings: Sliding window size N=16 (balancing smoothness and real-time performance, with controllable computational load in embedded scenarios); ADC sampling frequency: 10Hz (sampled once every 100ms to adapt to the slow changes in blue light intensity); Circular buffer: stores the 16 most recent ADC sample values, with new values ​​overwriting the oldest values.

[0080] Secondly, the main control module 10 reads the current signal of the blue light sensing unit 24 through the built-in ADC and converts it into a digital value adc_val; writes adc_val into a circular buffer and updates the buffer pointer; calculates the average value of 16 values ​​in the buffer as the enhanced blue light intensity value.

[0081] The third step involves feature fusion driven by the sleep stage inference module 30. Preprocessed multi-source time-series data (radar micro-motion signals, thermal imaging profiles / temperature fields, PIR trigger events, and blue light intensity) are simultaneously input into the neural network model running in the sleep stage inference module 30. The neural network model automatically learns and extracts the most effective features from each sensor's data, and performs deep fusion at the feature layer to generate a joint feature representation that is more robust to environmental interference.

[0082] The fourth step is the sleep stage inference module 30, which drives sleep stage classification. Based on the fused features, the model directly outputs a high-dimensional probability distribution of the user's current state, which includes not only the basic sleep stages (awake, light sleep, deep sleep, REM), but also more refined sub-states (such as "difficulty falling asleep", "frequent turning over", "mild breathing disorder") and contextual scenarios ("user gets out of bed", "pet interference").

[0083] Furthermore, by leveraging time-series modeling capabilities, the model can detect trends of abnormal events such as sleep apnea earlier and more accurately, rather than simply relying on threshold triggers.

[0084] Step 5: Generation of personalized feedback strategies based on the AI ​​decision engine. The main control module 10 receives the recognition results (state labels and confidence levels) from the sleep stage inference module 30. The personalized strategy engine (which can be based on rule-based and model-based collaboration) is activated, dynamically generating the optimal combination of feedback instructions, including: (1) Light effect strategy: Dynamically calculate the brightness change curve and color temperature transition function of the "breathing light" (instead of a fixed preset), or when the night light is triggered, adaptively adjust the beam angle and brightness according to the user's direction of leaving the bed and thermal imaging information. By detecting the actions of sitting up and getting out of bed, a small range of low-brightness guide light is automatically lit to avoid glare and prevent falls.

[0085] (2) Sound effect strategy: intelligently select the most suitable sleep aid sound or wake-up music from the audio library, and dynamically adjust its volume envelope, accompanied by sleep white noise, and matched with light-on / off changes.

[0086] (3) Interaction strategy: Decide whether and when to light up the light to alert the user of possible abnormalities, or provide contextualized shortcut options for touch interaction.

[0087] Finally, data recording and presentation. Detailed sleep stages identified by AI, vital sign trends, abnormal event reports, and system feedback actions are encrypted and stored locally. These can be used to generate intuitive visual reports via the app, providing users with in-depth sleep insights.

[0088] Example 2 Combination Figure 5 This application provides an intelligent sleep environment regulation method, applied to the intelligent sleep environment regulation system of any one of the embodiments, comprising the following steps: S100: The sleep sensing module 20 continuously acquires multi-source heterogeneous data synchronously. The main control module 10 performs filtering, noise reduction, non-uniformity correction, and anti-shake preprocessing on the raw multi-source heterogeneous data. Specifically, the clock unit of the main control module 10 achieves time synchronization between the radar module 21 and the thermal imaging module 22, ensuring consistent timing of multi-source heterogeneous data acquisition. Specifically, the vital sign signals of the radar module 21 are filtered and denoised, and the image data of the thermal imaging module 22 undergoes non-uniformity correction and anti-shake preprocessing.

[0089] S200: The sleep stage inference module 30 performs feature extraction and deep fusion on the preprocessed multi-source heterogeneous data, and infers the user's sleep stage identification result through a neural network model.

[0090] S300: The main control module 10 receives the recognition results and, in conjunction with the user's historical preference model trained based on the user's long-term preference input and system feedback data, dynamically generates light control commands.

[0091] S400: The light effect strategy generation module 40 responds to the light control command and outputs a light effect driving signal that matches the current sleep stage.

[0092] S500: The system stores complete closed-loop data, including raw sensor data, recognition results, light control commands, and subsequent state changes, through local storage units or encrypted cloud storage. This closed-loop data provides data support for subsequent incremental model learning.

[0093] S600: Return to step S100 to continue real-time monitoring. During idle periods when the user is out of bed and the system is not performing real-time sleep monitoring, use the closed-loop data recorded in step S500 to perform local incremental learning on the neural network model, optimize the recognition accuracy and the adaptability of the feedback strategy, and form an intelligent closed loop.

[0094] This method constructs a sleep stage classification network based on the Transformer architecture. Using multimodal fusion features as input, it captures long-sequence data dependencies through a self-attention mechanism, outputting the probability distributions of sleep stages (wakefulness, light sleep, deep sleep, REM sleep) and sleep events (apnea, motor arousal). The model is pre-trained on publicly available datasets such as CitySleep and SleepEDF, and then fine-tuned and optimized using data collected by this system.

[0095] In step S300, dynamic generation includes: selecting from a preset strategy mapping library or calculating in real time the brightness change curve of the light effect, the color temperature transition function, and the type and volume envelope of the audio content based on the sleep stage in the recognition result.

[0096] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent sleep environment regulation system, characterized in that, It includes a main control module, a sleep perception module, a sleep stage inference module, a light effect strategy generation module, and a light effect execution module. The sleep perception module uses non-contact sensing to monitor vital signs, temperature distribution information, and spatial contour information within the monitoring area. The vital signs include the user's respiratory rate, heart rate, and body movement signals. The sleep stage inference module fuses and analyzes the multi-source heterogeneous data sensed by the sleep sensing module and infers the user's sleep stage; the main control module generates different light control commands based on different sleep stages; the light effect strategy generation module responds to the light control commands and outputs a light effect driving signal corresponding to the sleep stage; and the light effect execution module generates sleep regulation light effects according to the light effect driving signal.

2. The intelligent sleep environment regulation system according to claim 1, characterized in that: The sleep sensing module includes a radar module and a thermal imaging module. The radar module is used to sense vital signs signals within the monitoring area, and the thermal imaging module is used to sense temperature distribution information and spatial contour information within the monitoring area.

3. The intelligent sleep environment regulation system according to claim 2, characterized in that: The sleep sensing module includes a passive infrared sensor, which is used to sense human movement signals within the monitoring area. The sleep stage inference module identifies the scene state based on the human movement signals. The scene state includes user in bed, out of bed, and human / pet distinction. The light effect strategy generation module combines the scene state with the sleep stage identification results to output a corresponding light effect driving signal to adjust the brightness, color, illumination range, and beam angle of the light effect execution module.

4. The intelligent sleep environment regulation system according to claim 1, characterized in that: The intelligent sleep environment regulation system also includes a photoelectric module, which is installed in the light effect execution module and electrically connected to the main control module. The photoelectric module emits pulsed light to the user's fingertip along with the light effect execution module to obtain the user's blood oxygen data.

5. The intelligent sleep environment regulation system according to claim 1, characterized in that: The sleep stage inference module is integrated inside the main control module and shares a signal transmission channel with the main control module to achieve real-time data interaction; Alternatively, the sleep stage inference module can be set separately from the main control module, and the sleep stage inference module can communicate with the main control module through a communication bus to realize the transmission of multi-source heterogeneous data and feedback of sleep stages.

6. The intelligent sleep environment regulation system according to claim 1, characterized in that: The sleep stage inference module runs a pre-trained lightweight neural network model, which performs the following functions: The multi-source heterogeneous data is time-aligned and feature-level fused to generate joint features resistant to environmental interference. The user's sleep stages are classified, and the sleep stages include at least the waking state, light sleep state, deep sleep state, and REM sleep stage; Detect abnormal physiological events, including mild respiratory disturbances and sleep apnea.

7. The intelligent sleep environment regulation system according to claim 1, characterized in that: The intelligent sleep environment regulation system also includes an audio module that is communicatively connected to the main control module. The audio module includes a digital audio decoding chip and a miniature speaker. The main control module controls the audio module to play audio content that coordinates with the sleep regulation light effect according to the sleep stage. The audio content includes sleep-inducing white noise and natural sounds, and gradually increasing music for gentle wake-up.

8. The intelligent sleep environment regulation system according to claim 7, characterized in that: The intelligent sleep environment adjustment system also includes a touch sensing module that is communicatively connected to the main control module. The touch sensing module uses a capacitive touch chip to receive local user control input. The main control module responds to user input and adjusts the parameters of the sleep adjustment light effect and the audio playback mode of the audio module in real time.

9. The intelligent sleep environment regulation system according to any one of claims 1 to 8, characterized in that: The main control module integrates a wireless communication unit, which is used to encrypt and transmit the desensitized sleep report and recognition results to the user terminal. And / or, the system further includes a housing, in which the main control module, the sleep sensing module, the sleep stage inference module, the light effect strategy generation module, and the light effect execution module are all integrated.

10. A method for regulating an intelligent sleep environment, applied to the intelligent sleep environment regulation system according to any one of claims 1 to 9, characterized in that, Includes the following steps: S100: The sleep sensing module continuously acquires multi-source heterogeneous data synchronously in time, and the main control module performs filtering, noise reduction, non-uniformity correction and anti-shake preprocessing on the original multi-source heterogeneous data. S2 00: The sleep stage inference module performs feature extraction and deep fusion on preprocessed multi-source heterogeneous data, and infers the user's sleep stage through a neural network model; S300: The main control module receives the sleep stage and, in conjunction with the user's historical preference model, generates corresponding light control commands; S400: The light effect strategy generation module responds to the light control command and outputs a light effect driving signal that matches the current sleep stage; S500: The system encrypts and stores complete closed-loop data, including raw sensor data, recognition results, light control commands, and status changes after feedback; S600: Return to step S100 to continue real-time monitoring. During idle periods, use the closed-loop data recorded in step S500 to perform local incremental learning on the neural network model, optimize the recognition accuracy and the adaptability of the feedback strategy, and form an intelligent closed loop.

Images