Aromatic sleep intervention feedback system based on lightweight electroencephalogram sleep staging model

Abstract

The invention discloses an aromatic sleep intervention feedback system based on a lightweight electroencephalogram sleep staging model, which relates to the technical field of intelligent sleep monitoring and intervention and comprises an electroencephalogram sensor module, a data processing module and an atomizer module. The electroencephalogram sensor module is used for collecting an original electroencephalogram signal of a prefrontal lobe Fpz point location in a single-point mode, and the original electroencephalogram signal is transmitted to the data processing module through Bluetooth after being preprocessed and subjected to digital-to-analog conversion. The data processing module is used for preprocessing electroencephalogram data and then inputting the electroencephalogram data into the lightweight model for sleep staging, and a dynamic aroma intervention instruction is generated. The atomizer module controls essential oil atomization parameters according to the instruction to realize closed-loop intervention; according to the method, the forehead lobe Fpz point position electroencephalogram signals are collected in a single-point mode, real-time processing is achieved in combination with a lightweight sleep staging model, aroma intervention parameters are dynamically adjusted based on different sleep stages, and the problems that in the prior art, equipment wearing is complex, algorithm delay is large, and the intervention mode is single are solved; the method has the advantages of simplifying the wearing structure, improving the processing efficiency and optimizing the intervention effect.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent sleep monitoring and intervention, and in particular to an aroma sleep intervention feedback system based on a lightweight EEG sleep staging model. Background Art

[0002] Existing sleep monitoring technology primarily relies on the collection of multimodal physiological signals, such as EEG, EOG, EMG, and blood oxygenation signals. While this multi-sensor combination can improve monitoring accuracy, it complicates the wearing of the device and may interfere with the user's sleep. The technical solution disclosed in Chinese patent CN202310128060.0 requires the placement of electrodes in multiple locations, such as the forehead, eyes, and temples. This multi-point contact monitoring can cause significant discomfort to the user.

[0003] At the algorithmic level, existing sleep staging models generally suffer from high computational complexity. While deep learning models can achieve highly accurate sleep staging, their large number of parameters requires high hardware computing power, making real-time processing difficult on embedded devices. While the closed-loop management system proposed in Chinese patent CN202510386560.3 implements sleep intervention, its deep learning algorithm requires processing multi-channel physiological signals, resulting in significant system response delays and failing to meet the needs of real-time intervention.

[0004] Existing technologies for aromatherapy interventions often rely on a fixed-concentration, continuous-release control strategy. This single-mode approach can easily lead to olfactory fatigue and fails to tailor the response to different sleep stages. Especially during deep sleep, the continuous release of essential oils can disrupt a user's natural sleep rhythm. Summary of the Invention

[0005] In view of this, the present invention provides an aroma sleep intervention feedback system based on a lightweight EEG sleep staging model, which has the advantages of simplifying the complexity of device wearing, realizing real-time sleep staging and dynamic aroma intervention, and avoiding olfactory fatigue and sleep interference.

[0006] The present invention provides an aroma sleep intervention feedback system based on a lightweight EEG sleep staging model, comprising:

[0007] The EEG sensor module is connected to the brain electrodes and is used to collect the original EEG signal of the human frontal lobe Fpz point at a single point; and pre-process and digital-to-analog convert the original EEG signal to obtain the corresponding EEG data, and transmit the EEG data to the data processing module via Bluetooth;

[0008] a data processing module connected to the EEG sensor module, configured to pre-process received EEG data; and input the pre-processed EEG data into a pre-trained lightweight EEG sleep staging model for sleep staging, and generate dynamic aroma intervention instructions based on the sleep staging results; the sleep staging results include wakefulness, non-rapid eye movement (NREM) 1, non-rapid eye movement (NREM) 2, non-rapid eye movement (NREM) 3, and rapid eye movement (REM); wherein non-REM 1 and non-REM 2 are collectively referred to as light sleep, and non-REM 3 is referred to as deep sleep;

[0009] The atomizer module is connected to the data processing module and is used to control the atomization concentration, continuous intervention time and interval time of the aromatic essential oil according to the received aromatic intervention instructions, so as to realize closed-loop intervention associated with the sleep staging results.

[0010] In an optional embodiment, the EEG sensor module includes:

[0011] The electrostatic protection circuit is connected to the brain electrodes to prevent the electrostatic effect generated by direct contact between the input interface of the EEG sensor module and the human body from damaging the subsequent circuit;

[0012] The impedance detection circuit is connected to the electrostatic protection circuit and is used to detect whether the contact between the brain electrodes and the human body is good;

[0013] The front-end filtering circuit is connected to the electrostatic protection circuit and is used to filter out high-frequency signals in the EEG signal by low-pass filtering;

[0014] The instrument amplifier circuit is connected to the front-end filter circuit and consists of two common-input operational amplifiers and a differential operational amplifier. It is used to perform differential amplification on the EEG signal and suppress common-mode noise.

[0015] The right leg drive circuit is connected to the electrostatic protection circuit and the instrument amplifier circuit respectively, and is used to amplify the common mode voltage collected from the human body through the instrument amplifier circuit and then feed it back to the human body to offset the common mode interference signal;

[0016] A DC correction circuit is connected to the instrument amplifier circuit and is used to superimpose the amplified EEG signal with the reference voltage to eliminate the DC offset of the EEG signal;

[0017] A bandpass filter circuit is connected to the DC correction circuit and is used to extract the EEG signal in the target signal frequency band by using a bandpass filtering method;

[0018] A voltage follower circuit is connected to the bandpass filter circuit and is used to reduce the output impedance of the previous stage circuit;

[0019] an analog-to-digital conversion circuit, connected to the voltage follower circuit, for converting the analog EEG signal output by the voltage follower circuit into digital EEG data;

[0020] The EEG sensor main control circuit is connected to the impedance detection circuit, the analog-to-digital conversion circuit, and the Bluetooth transmission circuit, respectively, and is used for data exchange and command control between multiple circuits to support the functions of the EEG sensor module;

[0021] The Bluetooth transmission circuit is connected to the EEG sensor main control circuit and is used to transmit the converted EEG data to the data processing module via Bluetooth.

[0022] In an optional embodiment, the lightweight EEG sleep staging model includes a first convolution activation layer, a second convolution activation layer, a third convolution activation layer, an LSTM layer, a fully connected layer and a Softmax layer that are cascaded in sequence.

[0023] In an optional implementation, the preprocessing of the received EEG data includes:

[0024] Band-pass filtering is performed on the EEG data to retain the EEG data within the target frequency range;

[0025] Perform wavelet denoising on the EEG data within the target frequency range to filter out noise and artifacts including other physiological signals;

[0026] Downsample the denoised EEG data.

[0027] In an optional embodiment, the data processing module generates dynamic aroma intervention instructions using a four-stage dynamic control strategy:

[0028] Normal awake mode: When the sleep stage result is awake, the atomization concentration of the essential oil is controlled to 2%, the continuous intervention time is 10 seconds, and the interval time is 30 seconds;

[0029] Light sleep activation mode: When the sleep staging result is non-rapid eye movement stage 1, the atomization concentration of the essential oil is controlled to 2%, the continuous intervention time is 5 seconds, and the interval time is 25 seconds; when the sleep staging result is non-rapid eye movement stage 2, the atomization concentration of the essential oil is controlled to 1.5%, the continuous intervention time is 5 seconds, and the interval time is 25 seconds;

[0030] Deep sleep protection mode: When the sleep stage result is non-rapid eye movement stage 3 or rapid eye movement stage, the essential oil release is stopped and only the basic diffusion airflow is maintained.

[0031] In an optional embodiment, the atomizer module includes:

[0032] The serial port transmission circuit is connected to the data processing module and the atomizer main control circuit respectively, and is used to receive the aroma intervention instruction from the data processing module and transmit it to the atomizer main control circuit;

[0033] The atomizer main control circuit is connected to the serial port transmission circuit and is used to generate a corresponding low-voltage PWM signal according to the aroma intervention instruction;

[0034] The atomizer drive circuit is connected to the atomizer main control circuit and is used to boost the low-voltage PWM signal output by the atomizer main control circuit to the operating voltage required by the atomizer plate, thereby driving the atomizer plate to a resonant state;

[0035] The atomizer is connected to the atomizer drive circuit and is used to generate high-frequency vibration in a resonant state to vibrate and atomize the essential oil liquid into essential oil gas;

[0036] In an optional embodiment, the atomizer module further includes:

[0037] The liquid level detection circuit is connected to the atomizer main control circuit and is used to detect the liquid level of the essential oil liquid in real time;

[0038] The atomizer main control circuit is connected to the liquid level detection circuit and is also used to immediately cut off the PWM signal output to the atomization drive circuit when the liquid level height is lower than the safety threshold, stop the atomization plate from vibrating, and avoid no-load operation.

[0039] In an optional embodiment, the atomizer module further includes:

[0040] The isolation circuit is connected to the atomizer main control circuit and is used to isolate the atomizer main control circuit, serial port transmission circuit, liquid level detection circuit, atomization drive circuit, atomizer plate and fan to prevent the high power supply voltage of the fan from affecting other circuits.

[0041] The fan is connected to the isolation circuit and is used to accelerate the diffusion of the atomized essential oil gas through airflow, so that the essential oil gas is quickly inhaled by the human nasal cavity.

[0042] In an optional embodiment, the data processing module is also used to display the time domain and frequency domain graphs of the EEG data and the sleep staging results in real time, and save the EEG data, the time domain and frequency domain graphs of the EEG data, the sleep staging results and the aroma intervention instruction information in a built-in memory.

[0043] The present invention collects EEG signals at the Fpz point of the frontal lobe at a single point, combines it with a lightweight sleep staging model to achieve real-time processing, and dynamically adjusts the aroma intervention parameters based on different sleep stages. It solves the problems of complex equipment wearing, large algorithm delay, and single intervention mode in the existing technology, and has the advantages of simplifying the wearing structure, improving processing efficiency, and optimizing the intervention effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 2. It is a structural diagram of an aroma sleep intervention feedback system based on a lightweight EEG sleep staging model according to an embodiment of the present invention;

[0046] Figure 2 is a circuit schematic diagram of an EEG sensor module according to an embodiment of the present invention;

[0047] Figure 3 is a flowchart of processing EEG data according to an embodiment of the present invention;

[0048] Figure 4 4 is a circuit diagram of an atomizer module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0050] In existing technologies, sleep monitoring usually relies on multimodal physiological signal acquisition, such as the simultaneous use of EEG, EOG, EMG, and blood oxygen sensors, which results in high data processing complexity and cumbersome wearing equipment. Traditional sleep staging algorithms require complex calculations and are difficult to run in real time on low-power hardware, which limits practical application scenarios. In addition, existing aroma intervention solutions usually adopt a fixed release mode and cannot be dynamically adjusted according to the sleep stage. Long-term use can easily lead to olfactory fatigue, and the intervention effect gradually weakens. For example, existing systems require sensors to be arranged in multiple locations such as the forehead, eyes, and temples, which may interfere with the user's sleep comfort. At the same time, multi-channel signal processing increases system power consumption and latency.

[0051] To address these issues, we first considered reducing the dimensionality of physiological signal acquisition. Analysis revealed that EEG signals at the Fpz point in the prefrontal cortex contain sufficient information about sleep stages, paving the way for multimodal acquisition. Secondly, to address algorithmic complexity, we investigated lightweight model structures and found that combining a convolutional neural network with an LSTM could reduce computational complexity while maintaining accuracy. Finally, to address the issue of a single intervention strategy, we proposed dynamically adjusting atomization parameters based on sleep stage. Experiments verified the differences in sensitivity to fragrance concentration across sleep stages, leading to a hierarchical control strategy.

[0052] Therefore, if Figure 1 As shown, this application proposes an aromatherapy sleep intervention feedback system based on a lightweight EEG sleep staging model, including an EEG sensor module, a data processing module, and an atomizer module. The EEG sensor module is connected to the brain electrodes and is used to collect the raw EEG signal at the Fpz point of the prefrontal lobe. After preprocessing and digital-to-analog conversion, it is transmitted to the data processing module via Bluetooth. The data processing module preprocesses the EEG data and inputs it into the lightweight model for sleep staging, generating dynamic aromatherapy intervention instructions. The atomizer module controls the essential oil atomization parameters according to the instructions to achieve closed-loop intervention.

[0053] Single-point acquisition involves placing a single electrode at the Fpz point in the prefrontal cortex to collect EEG signals. This selection of this area allows for the acquisition of low-frequency rhythmic signals that characterize sleep stages, reducing the number of sensors required. Preprocessing and digital-to-analog conversion, including signal amplification, filtering, and analog-to-digital conversion, eliminates ambient noise and adapts to digital processing requirements. Bluetooth transmission utilizes a low-power protocol for wireless data transmission, eliminating the impact of cable constraints on sleeping posture. A lightweight EEG sleep staging model, comprised of convolutional layers and a recurrent neural network, enables real-time sleep stage classification on embedded devices. Dynamic aroma intervention commands generate different atomization parameters based on wakefulness, light sleep, and deep sleep, matching intervention intensity to sleep depth. Closed-loop intervention analyzes sleep status in real time and adjusts atomization parameters, forming a perception-decision-execution feedback control loop.

[0054] Specifically, the raw EEG signals collected by the EEG sensor module are amplified and filtered, then converted into digital signals and transmitted via Bluetooth to the data processing module. Bandpass filtering and wavelet denoising are performed during the preprocessing phase to eliminate myoelectric interference and power-frequency noise. A lightweight model performs feature extraction and time series analysis on the processed EEG data, outputting the current sleep stage. When wakefulness is detected, the system increases the atomization concentration and shortens the interval to promote sleep onset; reduces the concentration during light sleep to maintain a sleep state; and stops atomization after entering deep sleep to avoid interference. The atomizer module adjusts the drive voltage and duty cycle according to instructions to precisely control the amount of essential oil released.

[0055] Compared with existing technologies, existing systems require the simultaneous acquisition of EEG, EOG, and EMG signals. However, the present invention only requires single-point EEG acquisition to complete staging, reducing equipment complexity and data processing burden. Traditional methods use complex deep learning models, resulting in high latency. The present invention optimizes the model structure to enable real-time execution on low-computing hardware. Existing aromatherapy interventions use a fixed release pattern, while the present invention achieves dynamic regulation through sleep stage recognition, avoiding olfactory fatigue while improving intervention efficiency.

[0056] Through the above technical solutions, the present invention solves the equipment complexity issues associated with multimodal signal acquisition and simplifies the system architecture through single-point EEG acquisition. The lightweight model design reduces computing resource requirements, making real-time sleep staging feasible on embedded devices. The dynamic aromatherapy intervention strategy adjusts release parameters based on sleep stages, minimizing essential oil consumption and user adaptability while ensuring intervention effectiveness.

[0057] In an optional embodiment, as Figure 2 As shown, the EEG sensor module includes: electrostatic protection circuit, impedance detection circuit, front-end filtering circuit, instrument amplifier circuit, right leg drive circuit, DC correction circuit, bandpass filtering circuit, voltage follower circuit, analog-to-digital conversion circuit, EEG sensor main control circuit and Bluetooth transmission circuit.

[0058] The electrostatic protection circuit is a combination of a transient voltage suppressor diode (TVS) and a gas discharge tube (GDT). Specifically, it can be implemented using a bidirectional TVS diode in parallel with a varistor, used to eliminate electrostatic pulses generated by human contact. The impedance detection circuit is a contact impedance measurement circuit based on the AC excitation method. Specifically, it can inject a 50Hz square wave signal into the electrode and determine the contact status by measuring the feedback voltage amplitude. The front-end filter circuit is a second-order passive low-pass filter with a cutoff frequency of 100Hz. Specifically, it can be implemented using an RC ladder network structure to filter out high-frequency interference caused by electromyographic signals. The instrumentation amplifier circuit is a three-op-amp differential amplifier structure composed of INA128 chips. Specifically, the gain factor can be adjusted up to 1000x using gain resistors, used to extract microvolt-level EEG signals. The right leg drive circuit is a closed-loop compensation circuit that feeds common-mode voltage back to the human body. Specifically, it can be implemented using an inverting amplifier and a current-limiting resistor to form a common-mode cancellation loop. The DC correction circuit is an adder circuit composed of operational amplifiers. Specifically, it can offset the DC offset component in the signal by adjusting the reference voltage. The bandpass filter circuit is an eighth-order Butterworth active filter with a passband range of 0.5-35Hz. It can be implemented using a cascaded multi-stage op amp to extract the delta and theta wave frequency bands. The voltage follower circuit is a unity-gain buffer based on the LM358 chip. It reduces the output impedance to below 50Ω, enhancing signal driving capability. The analog-to-digital conversion circuit is a 24-bit high-precision Σ-Δ ADC chip, specifically implemented using the ADS1299 analog front end, which digitizes the signal at a 512Hz sampling rate.

[0059] Specifically, during the signal acquisition process, the EEG signal of the human frontal lobe first passes through the electrostatic protection circuit to eliminate electrostatic interference, and then the impedance detection circuit monitors the electrode contact impedance in real time. When the impedance value exceeds 200kΩ, an alarm prompt is triggered. The effective signal passes through the front-end filter circuit to filter out high-frequency noise and then enters the instrument amplifier circuit for differential amplification. At the same time, the right leg drive circuit injects the amplified common-mode interference back into the human body to form negative feedback. The EEG signal, which has been eliminated from baseline drift by the DC correction circuit, enters the bandpass filter circuit to extract the target frequency band (0.5-35Hz). The voltage follower circuit reduces the signal output impedance and sends it to the analog-to-digital conversion circuit for digitization. The EEG sensor main control circuit controls the ADC sampling parameters through the SPI interface, and encapsulates the digital signal into a data packet and transmits it to the data processing module via Bluetooth.

[0060] Compared with the prior art, the CN202310128060.0 patent disclosed in the prior art only uses single-stage amplification and simple filtering processing. The right leg drive circuit is not set, resulting in a common-mode rejection ratio of less than 80dB, and the lack of impedance detection function easily causes signal acquisition failure. The present invention forms a dual common-mode rejection mechanism through instrument amplification and right leg drive, which increases the common-mode rejection ratio to more than 120dB, and combines bandpass filtering and DC correction technology to increase the signal-to-noise ratio within the effective bandwidth of the signal by 40%. In addition, the full-link signal optimization from electrostatic protection to wireless transmission is achieved through multi-stage circuit sequential processing, which solves the signal distortion problem caused by circuit design defects in the prior art.

[0061] Through the above technical solutions, this application effectively solves the problems of electrostatic breakdown risk, electrode contact failure, high-frequency noise interference, common-mode interference introduction and signal baseline drift during the EEG signal acquisition process. Among them, the electrostatic protection circuit increases the electrostatic withstand voltage value to more than 8kV, the impedance detection circuit makes the detection rate of poor electrode contact reach 99%, the instrument amplification and right leg drive combination attenuates the common-mode interference by 60dB, and the bandpass filter circuit suppresses the myoelectric interference amplitude to less than 5% of the original signal. Ultimately, the input reference noise of the single-channel EEG signal acquisition system is achieved to be less than 2μVpp, which meets the signal quality requirements of the sleep staging algorithm.

[0062] In an optional embodiment, as Figure 3 As shown in the figure, the lightweight EEG sleep staging model includes a first convolution activation layer, a second convolution activation layer, a third convolution activation layer, an LSTM layer, a fully connected layer and a Softmax layer, which are cascaded in sequence.

[0063] The first convolutional activation layer is a neural network layer that performs primary feature extraction on the input EEG data. It extracts low-frequency features from the EEG data for dimensionality reduction. Specifically, it uses a one-dimensional convolution kernel combined with an activation function to extract local time-domain features and reduce parameter redundancy.

[0064] The second convolutional activation layer is a neural network layer that performs intermediate-level abstraction on primary features, extracting high-frequency features from EEG data for dimensionality reduction. This can be achieved using larger convolution kernels to enhance feature representation and suppress noise.

[0065] Among them, the third convolution activation layer refers to a neural network layer that deeply represents intermediate features and can extract high-frequency features in EEG data. Specifically, it can be implemented using a multi-channel convolution kernel to capture EEG signal features in different frequency ranges.

[0066] The LSTM layer is a recurrent neural network layer with memory cells that extracts time-related features from EEG data and reduces its dimensionality. This is achieved using a gating mechanism to model the temporal dependencies of sleep stages.

[0067] Among them, the fully connected layer refers to the neural network layer that maps spatiotemporal features to the sleep stage dimension. It can be implemented using a linear transformation matrix to establish the association between features and classification results.

[0068] The Softmax layer is a normalization layer that outputs a probability distribution between 0 and 1. It can be implemented using an exponential function and normalization calculation to generate the classification confidence of each sleep stage.

[0069] Specifically, EEG data is sequentially processed through three convolutional activation layers for multi-level feature extraction: the first layer extracts local waveform features using small convolution kernels, the second layer captures correlated features between adjacent time points by expanding the receptive field, and the third layer separates features across different frequency bands using multi-channel convolution. The resulting feature sequence is then fed into an LSTM layer, whose memory cells track the temporal evolution of sleep stages, such as the transition from wakefulness to light sleep. A fully connected layer maps the temporal features output by the LSTM into a five-dimensional vector corresponding to wakefulness, non-REM 1, non-REM 2, non-REM 3, and REM. The softmax layer converts the five-dimensional vector into a probability distribution, selecting the sleep stage corresponding to the maximum probability value as the final classification result. By combining hierarchical feature extraction with spatiotemporal modeling, the number of model parameters is kept within the operational range of embedded devices while ensuring classification accuracy.

[0070] Existing sleep staging methods rely on multimodal signals such as EEG, EOG, heart rate, and blood oxygen, and use deep learning algorithms to process multidimensional data, resulting in high model complexity and high computing power requirements. However, the present invention uses only a single-channel EEG signal, achieves efficient feature compression through a three-level convolutional layer, and combines it with an LSTM layer to replace traditional multimodal signal fusion, completing time series modeling on a single data source. The parameters of the existing multimodal sleep staging model usually exceed 10MB, while the parameters of the lightweight model of the present invention can be compressed to less than 2MB, making it suitable for real-time reasoning on low-power embedded processors.

[0071] It should be noted that the lightweight EEG sleep staging model can perform sleep staging in real time, but it also requires corresponding training in advance and saving the trained weight data. When the lightweight EEG sleep staging model is running, the weight data can be called to complete the sleep staging.

[0072] Through the above-mentioned technical solution, the present invention solves the high processing complexity problem caused by the reliance on multimodal data in existing sleep staging algorithms. By reducing feature redundancy through a hierarchical convolutional structure and replacing additional physiological signal acquisition equipment with LSTM layers, time series modeling can be achieved using only single-channel EEG signals. The reduced number of model parameters makes real-time inference possible on embedded devices, avoiding the processing delays caused by insufficient computing power in traditional solutions. Furthermore, the single data source processing mechanism eliminates the interference of multiple sensors on sleep quality, meeting the lightweight requirements of practical application scenarios.

[0073] In an optional embodiment, preprocessing the received EEG data includes: bandpass filtering the EEG data to retain the EEG data within the target frequency range (0.5-30Hz); performing wavelet denoising on the EEG data within the target frequency range to filter out noise and artifacts including other physiological signals; and downsampling the denoised EEG data.

[0074] Bandpass filtering refers to retaining only EEG signals in the target frequency band by setting upper and lower limits of the frequency. This can be achieved by using a Butterworth filter, for example, setting the cutoff frequency to 0.5Hz to 30Hz to eliminate high-frequency myoelectric interference and low-frequency baseline drift. Wavelet denoising refers to decomposing the signal using the time-frequency localization characteristics of the wavelet transform and suppressing noise through threshold processing. This can be achieved by using the Daubechies wavelet basis function to separate physiological artifacts such as electrooculogram and electrocardiogram from EEG signals. Downsampling refers to lowering the signal sampling rate to reduce the amount of data. This can be achieved by decimation, for example, reducing the original sampling rate from 200Hz to 100Hz, thereby reducing the input dimension of the subsequent model while retaining valid information.

[0075] Specifically, bandpass filtering first extracts the target frequency band from the raw EEG data, such as the effective components in the range of 0.5Hz to 30Hz, to eliminate the interference of irrelevant frequency bands on subsequent processing. Subsequently, wavelet denoising further eliminates artifacts generated by physiological activities such as eye movements and heartbeats within the retained frequency band, removes high-frequency noise through multi-scale decomposition and threshold shrinkage, and retains low-frequency features reflecting sleep stages. Finally, the downsampling operation proportionally reduces the data sampling rate, for example, from 200Hz to 100Hz, compressing the data size in the time domain, while avoiding signal distortion through anti-aliasing filters. These three steps form a progressive processing chain. Bandpass filtering limits the effective frequency band for wavelet denoising and reduces the noise processing range; wavelet denoising improves signal quality and avoids the need for multimodal signal processing; downsampling reduces the data dimension to adapt to the low computing power requirements of lightweight models.

[0076] Compared with existing technologies such as CN202310128060.0, which require the simultaneous acquisition of multimodal signals such as EEG, EOG, and EMG, the present invention achieves effective denoising with a single-channel EEG signal through a combination of bandpass filtering and wavelet denoising, thus avoiding the complexity of multi-sensor data fusion. Furthermore, existing technologies such as CN202510386560.3, which use deep learning algorithms to process raw signals, require high-sampling rate data. However, the present invention reduces the amount of input data through downsampling, enabling lightweight models to achieve real-time inference on low-power hardware.

[0077] Through the above-mentioned technical solution, the present invention solves the problem of traditional sleep staging algorithms relying on multimodal signals and complex processing procedures. Through a three-stage process consisting of bandpass filtering to limit the effective frequency band, wavelet denoising to eliminate cross-modal noise, and downsampling to compress the data size, the amount of single-channel EEG data is reduced to 25% of the original size while maintaining sleep staging accuracy, significantly reducing the model's computational complexity. Furthermore, this preprocessing process eliminates the need for auxiliary signals such as electrooculography and electromyography, simplifying the design complexity of the hardware acquisition module and providing adaptive data input conditions for the real-time deployment of lightweight sleep staging models.

[0078] In an optional embodiment, the data processing module uses a four-stage dynamic control strategy to generate dynamic aroma intervention instructions, including a wakeful normal mode, a light sleep activation mode, and a deep sleep protection mode. In the wakeful normal mode, when the sleep staging result is the wakeful stage (W stage), the atomization concentration of the aromatic essential oil is controlled to be 2%, the continuous intervention time is 10 seconds, and the interval time is 30 seconds; the light sleep activation mode maintains a 2% concentration in the non-rapid eye movement stage 1 (N1 stage) but shortens the continuous intervention time to 5 seconds and adjusts the interval time to 25 seconds, and reduces the concentration to 1.5% in the non-rapid eye movement stage 2 (N2 stage) and maintains a continuous intervention of 5 seconds and an interval time of 25 seconds; the deep sleep protection mode stops the release of essential oil in the non-rapid eye movement stage 3 (N3 stage) or the rapid eye movement (REM stage) and only maintains the basic diffusion airflow.

[0079] Among them, the dynamic control strategy refers to a closed-loop control method that dynamically adjusts the aroma intervention parameters according to the real-time sleep staging results. Specifically, it can be implemented by combining preset concentration gradients with time parameters, and precise control can be achieved by establishing intervention modes corresponding to different sleep stages. Atomization concentration refers to the volume ratio of essential oil in the atomized gas, which can be achieved by adjusting the duty cycle of the atomizer drive signal. Different concentrations correspond to different sleep-promoting effects. Continuous intervention time refers to the duration of a single atomization release, which can be achieved by using a timer to control the conduction time of the atomization drive circuit, which directly affects the intensity of a single intervention. Interval time refers to the interval period between two adjacent atomization releases, which can be achieved by using a cycle counter for time management to control the frequency of intervention. Basic diffusion airflow refers to the minimum wind speed to maintain the flow of gas inside the atomizer, which can be achieved by controlling the fan to run at the lowest speed to avoid the accumulation of essential oil residues.

[0080] Specifically, during the wakefulness stage, a higher atomization concentration is used in combination with a longer single intervention time to help users quickly enter a sleep state by enhancing olfactory stimulation. When it is detected that the user has entered a light sleep stage, the intervention frequency is increased by shortening the interval time, while the atomization concentration is simultaneously reduced in the non-rapid eye movement stage 2, forming a gradient intervention strategy to promote the depth of sleep. When entering the deep sleep stage or the rapid eye movement stage, the release of essential oils is completely cut off, and only the basic airflow is retained to keep the device clean, so as to avoid olfactory stimulation interfering with physiological activities in the deep sleep stage. The parameter combination of each stage forms a dynamic closed loop through the three-dimensional regulation of atomization concentration, duration and interval time. The awake normal mode adopts 10 seconds of continuous intervention and 30 seconds of periodic release. The light sleep activation mode achieves high-frequency low-dose intervention through 5 seconds of continuous intervention and 25 seconds of interval. The deep sleep protection mode eliminates external interference through zero concentration release.

[0081] Compared to existing technologies, traditional aromatherapy interventions employ a continuous release method with a fixed concentration and fixed intervals, which can easily lead to decreased sensitivity of olfactory receptors and fail to adapt to the physiological needs of different sleep stages. This invention establishes a dynamic parameter adjustment mechanism linked to sleep staging, proactively reducing intervention intensity during light sleep and completely stopping intervention during deep sleep. This avoids olfactory fatigue caused by continuous stimulation and reduces the interference of ineffective intervention on deep sleep.

[0082] Through this technical solution, the present invention achieves precise matching of aromatherapy intervention parameters with sleep stages. High-intensity intervention accelerates sleep onset during wakefulness, gradient adjustment promotes sleep stage transitions during light sleep, and completely stops intervention during deep sleep to ensure sleep quality. This solution effectively addresses the olfactory fatigue caused by continuous intervention in traditional methods. Furthermore, through phased regulation, it reduces ineffective essential oil consumption and lowers the operating load of the hardware system.

[0083] In an optional embodiment, as Figure 4 As shown, the atomizer module includes a serial port transmission circuit, an atomizer main control circuit, an atomizer drive circuit and an atomizer plate. The serial port transmission circuit is connected to the data processing module and the atomizer main control circuit respectively, and is used to receive the aroma intervention instruction of the data processing module and transmit it to the atomizer main control circuit. The atomizer main control circuit is connected to the serial port transmission circuit, and is used to generate a corresponding low-voltage PWM signal according to the aroma intervention instruction. The atomizer drive circuit is connected to the atomizer main control circuit, and is used to boost the low-voltage PWM signal output by the atomizer main control circuit to the working voltage required by the atomizer plate to drive the atomizer plate to reach a resonant state. The atomizer plate is connected to the atomization drive circuit, and is used to generate high-frequency vibrations in a resonant state to vibrate the essential oil liquid and atomize it into essential oil gas.

[0084] The serial port transmission circuit refers to a communication interface circuit that enables reliable transmission of command signals. This can be implemented using a UART interface chip, which receives and interprets command data via an asynchronous serial communication protocol. The atomizer main control circuit refers to a microprocessor unit that executes the control logic. This can be implemented using a low-power single-chip microcomputer, generating a PWM waveform with an adjustable duty cycle via a built-in timer module. The atomization drive circuit refers to a power amplifier circuit that performs voltage conversion. This can be implemented using a combination of a boost converter and a MOS transistor, converting a 3.3V logic level into a high-frequency drive signal with a peak-to-peak value of 24V. The atomizer plate refers to a piezoelectric ceramic element that atomizes liquid. This can be implemented using a ring-shaped piezoelectric transducer, which generates longitudinal vibrations at a resonant frequency of 1.7MHz to achieve liquid dropletization.

[0085] Specifically, when the aroma intervention command is received via the serial port transmission circuit, the atomizer main control circuit generates a PWM control signal with a corresponding duty cycle based on the command parameters. This low-voltage PWM signal is fed into the atomizer drive circuit for voltage amplification, and the signal amplitude is increased to the 24V drive level required for resonance of the atomizer plate via a boost converter. The boosted high-frequency AC signal is applied to the piezoelectric ceramic layer of the atomizer plate, causing it to produce axial mechanical vibration. When the vibration frequency reaches the natural resonant frequency of the piezoelectric ceramic, a stable standing wave field forms on the surface of the atomizer plate, causing the contacting essential oil liquid to break into micron-sized atomized particles under the action of surface tension. This process controls the vibration intensity of the atomizer plate by adjusting the duty cycle of the PWM signal, thereby precisely adjusting the amount of essential oil atomized per unit time.

[0086] Compared with the existing technology, the traditional atomization device directly adopts a fixed voltage drive, which causes the atomization efficiency to be affected by power supply fluctuations and lacks dynamic adjustment capabilities. The present invention combines a low-voltage PWM signal with a boost drive to ensure the safety of the control circuit and achieve precise adjustment of the driving voltage. The resistive heating atomization method used in the existing technology is prone to high-temperature decomposition of essential oil components, while the present invention adopts piezoelectric atomization technology to avoid the influence of thermal effects. In addition, the existing atomizer control interface mostly adopts analog voltage signal transmission, which is susceptible to interference. The present invention adopts digital serial communication to significantly improve the reliability of command transmission.

[0087] Through the above technical solution, the present invention realizes reliable signal interaction between the aroma intervention instruction and the atomization actuator, and ensures the complete transmission of the control instruction through the digital communication interface. The architecture combining low-voltage PWM and boost drive is adopted to meet the high-voltage drive requirements of the atomizer while ensuring circuit safety. The resonant working mode of the piezoelectric atomizer realizes efficient atomization of the essential oil liquid, and the atomization concentration can be accurately controlled by adjusting the PWM duty cycle. This technical solution effectively solves the problem of low atomization efficiency caused by insufficient driving voltage of traditional atomizers, and avoids the decomposition of effective ingredients caused by heated atomization.

[0088] In an optional embodiment, the atomizer module further includes a liquid level detection circuit and an atomizer main control circuit. The liquid level detection circuit is connected to the atomizer main control circuit for real-time detection of the liquid level of the essential oil. The atomizer main control circuit is also connected to the liquid level detection circuit for immediately shutting off the PWM signal output to the atomization drive circuit when the liquid level falls below a safety threshold, thereby stopping the vibration of the atomizer plate and preventing no-load operation.

[0089] Among them, the liquid level detection circuit refers to a detection device that uses a sensor to obtain the liquid height in the essential oil container in real time. Specifically, it can be implemented using a capacitive or photoelectric liquid level sensor. Its function is to convert the physical liquid level into an electrical signal to provide real-time status feedback for the system.

[0090] Among them, the safety threshold refers to a pre-set liquid level critical value, which can be determined through experimental calibration or container structure parameters. Its function is to trigger the equipment protection mechanism to prevent the atomizer from continuing to work when there is no liquid supply.

[0091] Among them, PWM signal cutting refers to directly interrupting the output of the pulse width modulation signal through a hardware circuit. Specifically, it can be implemented using a transistor or relay switching circuit. Its function is to cut off the power supply path of the driving circuit from a physical level to ensure that the atomizer stops vibrating immediately.

[0092] Specifically, the liquid level detection circuit continuously monitors the liquid level in the essential oil container and transmits the detection signal to the atomizer's main control circuit. When the liquid level drops below a safe threshold, the atomizer's main control circuit directly cuts off the PWM signal output channel through a hardware switch circuit, causing the atomizer drive circuit to lose operating voltage. At this point, the atomizer plate ceases high-frequency vibration due to the loss of the drive signal, avoiding continuous no-load operation in the absence of liquid. This process completes state switching within milliseconds through the hardware circuit's rapid response mechanism, effectively preventing device damage caused by delays that may exist in traditional software control.

[0093] Compared to existing technologies, existing sleep intervention devices often lack liquid level monitoring capabilities. For example, the closed-loop management system proposed by Pan Tingrui focuses solely on physiological signal monitoring and fails to address atomizer protection mechanisms. This can easily lead to overheating and damage to the atomizer when essential oil depletion occurs. However, this invention, by implementing an independent liquid level detection hardware circuit and combining it with a rapid power-off mechanism within the main control circuit, creates a physical device protection barrier, resolving the existing risk of the device running idle due to insufficient liquid level.

[0094] Through the above technical solution, the present invention implements automatic shutdown protection for the atomizer module in the event of low liquid level, preventing mechanical wear of the atomizer disc due to no-load vibration. It also avoids power loss in the drive circuit under abnormal operating conditions, thereby extending the device's service life. This technical solution utilizes a fast hardware-level response mechanism to ensure that the device immediately stops operating upon detecting an abnormal liquid level, effectively eliminating the risk of device damage caused by software delays in traditional control methods.

[0095] In an optional embodiment, the atomizer module further includes an isolation circuit and a fan. The isolation circuit is connected to the atomizer main control circuit and is used to isolate the atomizer main control circuit, serial port transmission circuit, liquid level detection circuit, atomization drive circuit, and atomizing plate from the fan. The fan is connected to the isolation circuit and is used to accelerate the diffusion of atomized essential oil vapor through airflow.

[0096] An isolation circuit, specifically an optocoupler or magnetic isolator, blocks the direct connection between high-voltage and low-voltage circuits through electrical isolation. This circuit eliminates voltage fluctuations from interfering with the main control circuit through physical separation and signal conversion. A fan, specifically a device that generates directional airflow through mechanical rotation and employs a miniature DC motor with a fan blade structure, rapidly delivers atomized gas to the human respiratory zone through forced convection.

[0097] Specifically, the isolation circuit creates an electrical barrier between the atomizer's main control circuit and the fan's power supply line, preventing the high-voltage power supply required for fan operation from affecting the main control circuit's low-voltage signal transmission through the common ground path. The fan activates immediately after the atomizer produces essential oil vapor, creating a directional airflow from the atomizer to the person's head, shortening the gas's residence time in the air. This solution, through the synergistic effect of circuit isolation and airflow control, ensures circuit stability while improving the timeliness of intervention.

[0098] Compared to existing technologies, traditional atomizers share a common power supply system between the fan circuit and the main control circuit, leading to voltage interference that affects signal acquisition accuracy. While multimodal signal processing is mentioned, it does not address the issue of internal circuit interference within the atomizer, and reliance on natural diffusion leads to delayed intervention. This invention uses an isolation circuit to block the interference propagation path and combines it with active airflow acceleration to achieve precise control.

[0099] Through this technical solution, the present invention effectively prevents high-voltage circuit interference with sensors and control signals, ensuring the measurement accuracy of the EEG signal processing module. Furthermore, the forced airflow diffusion mechanism enables the atomized essential oil to quickly reach the human nasal cavity, improving the response speed and effectiveness of sleep intervention.

[0100] In an optional embodiment, the data processing module is also used to display the time domain and frequency domain graphs of the EEG data and the sleep staging results in real time, and save the EEG data, the time domain and frequency domain graphs of the EEG data, the sleep staging results and the aroma intervention instruction information in a built-in memory.

[0101] Among them, the time domain frequency domain diagram refers to a visual chart that jointly displays the time series waveform and spectral characteristics of the EEG signal. Specifically, the fast Fourier transform algorithm can be used to convert the time domain signal into frequency domain characteristics, and the dynamic refresh display can be achieved through the graphics processor to synchronously reflect the rhythm changes and energy distribution of the EEG signal.

[0102] Among them, the sleep staging result refers to the sleep stage label obtained by classifying the pre-processed EEG data through a lightweight EEG sleep staging model. Specifically, a model architecture based on a combination of convolutional neural networks and long short-term memory networks can be used to achieve real-time inference, which is used to dynamically identify the current state of wakefulness, light sleep, deep sleep or rapid eye movement.

[0103] Among them, the aroma intervention instruction information refers to a set of control parameters including atomization concentration, duration and interval time. Specifically, the instruction type and execution timestamp can be stored in a structured data format to record the triggering conditions and execution details of each intervention operation.

[0104] Among them, built-in memory refers to the non-volatile storage medium integrated in the data processing module. Specifically, flash memory chips or solid-state drives can be used to achieve persistent storage of local data to avoid the risk of data loss due to interruption of connection to external storage devices.

[0105] Specifically, the data processing module generates a composite view containing time domain waveforms and spectrum graphs through a graphical user interface, updating the display content once per second, allowing the operator to synchronously observe the original EEG signals and the processed frequency domain features. The sleep staging results are superimposed on the top of the view in the form of text labels and are dynamically updated with the model inference results. All raw EEG data are stored in binary format in chronological order, the time domain and frequency domain graphs are saved in image format by frame, and the sleep staging results and aromatherapy intervention instruction information are associated and stored in the form of a structured database table. The memory uses a circular overwrite mechanism to manage storage space, and automatically deletes the oldest historical data when the storage capacity reaches the threshold.

[0106] Compared to existing technologies, existing sleep monitoring systems only store raw EEG signals and staging results, without synchronously saving time-frequency domain analysis data and intervention instruction information. This makes it impossible to retrospectively analyze the correlation between intervention strategies and changes in EEG characteristics. This invention uses built-in memory to achieve multi-dimensional data association storage, forming a complete data chain of EEG signal characteristics, staging results, and instruction parameters corresponding to each intervention operation, providing data support for subsequent optimization of intervention strategies.

[0107] Through the above technical solution, the present invention solves the problem of the lack of real-time visualization and full-cycle data recording in the existing system, realizes the synchronous monitoring and associated storage of EEG signal characteristics, sleep stages and intervention parameters during the intervention process, provides a complete data basis for analyzing the relationship between aromatherapy intervention effects and sleep quality, and avoids the risk of data loss caused by relying on external storage devices.

[0108] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. An aroma sleep intervention feedback system based on a lightweight EEG sleep staging model, characterized by: include: The EEG sensor module is connected to the brain electrodes and is used to collect the original EEG signal of the Fpz point of the human frontal lobe; and performing pre-processing and digital-to-analog conversion on the raw EEG signal to obtain corresponding EEG data, and transmitting the EEG data to a data processing module via Bluetooth; a data processing module connected to the EEG sensor module, configured to pre-process received EEG data; and input the pre-processed EEG data into a pre-trained lightweight EEG sleep staging model for sleep staging, and generate dynamic aroma intervention instructions based on the sleep staging results; the sleep staging results include wakefulness, non-rapid eye movement (NREM) 1, non-rapid eye movement (NREM) 2, non-rapid eye movement (NREM) 3, and rapid eye movement (REM); wherein non-REM 1 and non-REM 2 are collectively referred to as light sleep, and non-REM 3 is referred to as deep sleep; The atomizer module is connected to the data processing module and is used to control the atomization concentration, continuous intervention time and interval time of the aromatic essential oil according to the received aromatic intervention instructions, so as to realize closed-loop intervention associated with the sleep staging results.

2. The system according to claim 1, wherein: The EEG sensor module includes: The electrostatic protection circuit is connected to the brain electrodes to prevent the electrostatic effect generated by direct contact between the input interface of the EEG sensor module and the human body from damaging the subsequent circuit; The impedance detection circuit is connected to the electrostatic protection circuit and is used to detect whether the contact between the brain electrodes and the human body is good; The front-end filtering circuit is connected to the electrostatic protection circuit and is used to filter out high-frequency signals in the EEG signal by low-pass filtering; The instrument amplifier circuit is connected to the front-end filter circuit and consists of two common-input operational amplifiers and a differential operational amplifier. It is used to perform differential amplification on the EEG signal and suppress common-mode noise. The right leg drive circuit is connected to the electrostatic protection circuit and the instrument amplifier circuit respectively, and is used to amplify the common mode voltage collected from the human body through the instrument amplifier circuit and then feed it back to the human body to offset the common mode interference signal; A DC correction circuit is connected to the instrument amplifier circuit and is used to superimpose the amplified EEG signal with the reference voltage to eliminate the DC offset of the EEG signal; A bandpass filter circuit is connected to the DC correction circuit and is used to extract the EEG signal in the target signal frequency band by using a bandpass filtering method; A voltage follower circuit is connected to the bandpass filter circuit and is used to reduce the output impedance of the previous stage circuit; an analog-to-digital conversion circuit, connected to the voltage follower circuit, for converting the analog EEG signal output by the voltage follower circuit into digital EEG data; The EEG sensor main control circuit is connected to the impedance detection circuit, the analog-to-digital conversion circuit, and the Bluetooth transmission circuit, respectively, and is used for data exchange and command control between multiple circuits to support the functions of the EEG sensor module; The Bluetooth transmission circuit is connected to the EEG sensor main control circuit and is used to transmit the converted EEG data to the data processing module via Bluetooth.

3. The system according to claim 1, wherein: The lightweight EEG sleep staging model includes a first convolution activation layer, a second convolution activation layer, a third convolution activation layer, an LSTM layer, a fully connected layer and a Softmax layer, which are cascaded in sequence.

4. The system according to claim 1, wherein: The preprocessing of the received EEG data includes: Band-pass filtering is performed on the EEG data to retain the EEG data within the target frequency range; Perform wavelet denoising on the EEG data within the target frequency range to filter out noise and artifacts including other physiological signals; Downsample the denoised EEG data.

5. The system according to claim 1, wherein: The data processing module uses a four-stage dynamic control strategy to generate dynamic aroma intervention instructions: Normal awake mode: When the sleep stage result is awake, the atomization concentration of the essential oil is controlled to 2%, the continuous intervention time is 10 seconds, and the interval time is 30 seconds; Light sleep activation mode: When the sleep staging result is non-rapid eye movement stage 1, the atomization concentration of the essential oil is controlled to 2%, the continuous intervention time is 5 seconds, and the interval time is 25 seconds; when the sleep staging result is non-rapid eye movement stage 2, the atomization concentration of the essential oil is controlled to 1.5%, the continuous intervention time is 5 seconds, and the interval time is 25 seconds; Deep sleep protection mode: When the sleep stage result is non-rapid eye movement stage 3 or rapid eye movement stage, the essential oil release is stopped and only the basic diffusion airflow is maintained.

6. The system according to claim 1, wherein: The atomizer module comprises: The serial port transmission circuit is connected to the data processing module and the atomizer main control circuit respectively, and is used to receive the aroma intervention instruction from the data processing module and transmit it to the atomizer main control circuit; The atomizer main control circuit is connected to the serial port transmission circuit and is used to generate a corresponding low-voltage PWM signal according to the aroma intervention instruction; The atomizer drive circuit is connected to the atomizer main control circuit and is used to boost the low-voltage PWM signal output by the atomizer main control circuit to the operating voltage required by the atomizer plate, thereby driving the atomizer plate to a resonant state; The atomizer is connected to the atomizer drive circuit and is used to generate high-frequency vibration under a resonant state to vibrate and atomize the essential oil liquid into essential oil gas.

7. The system according to claim 1, wherein: The atomizer module further comprises: The liquid level detection circuit is connected to the atomizer main control circuit and is used to detect the liquid level of the essential oil liquid in real time; The atomizer main control circuit is connected to the liquid level detection circuit and is also used to immediately cut off the PWM signal output to the atomization drive circuit when the liquid level height is lower than the safety threshold, stop the atomization plate from vibrating, and avoid no-load operation.

8. The system according to claim 7, characterized in that The atomizer module further comprises: The isolation circuit is connected to the atomizer main control circuit and is used to isolate the atomizer main control circuit, serial port transmission circuit, liquid level detection circuit, atomization drive circuit, atomizer plate and fan to prevent the high power supply voltage of the fan from affecting other circuits. The fan is connected to the isolation circuit and is used to accelerate the diffusion of the atomized essential oil gas through airflow, so that the essential oil gas is quickly inhaled by the human nasal cavity.

9. The system according to claim 1, wherein: The data processing module is also used to display the time domain and frequency domain diagrams of the EEG data and the sleep staging results in real time, and save the EEG data, the time domain and frequency domain diagrams of the EEG data, the sleep staging results and the aroma intervention instruction information in the built-in memory.

Images