A two-way emotional interaction head ring system based on a multi-modal brain-computer interface

Abstract

The application provides a kind of two-way emotional interaction head ring system based on multi-modal brain-computer interface.The two-way emotional interaction head ring system based on multi-modal brain-computer interface includes: exclusive emotion recognition algorithm module, two-way emotional interaction output module, intimate relationship synchronization and emotion interaction module, system control module and head ring structure;The system control module and head ring structure include master control MCU, multi-modal fusion processor, Bluetooth / ultra-low power communication module, electrode adaptive pressure structure, ultra-light flexible head ring structure.The two-way emotional interaction head ring system based on multi-modal brain-computer interface provided by the application has multiple advantages such as high monitoring accuracy, timely and effective intervention, high family participation and continuous wearing, which changes emotion management from passive recording to active adjustment, provides scientific, stable and sustainable emotion management support for users at each stage, improves the overall mental health level of users, and has clear practical value and industrial application prospect.

Description

Technical Field

[0001] This invention belongs to the field of emotional interaction technology, and in particular relates to a bidirectional emotional interaction headband system based on a multimodal brain-computer interface. Background Technology

[0002] Emotional state monitoring and intervention have become research hotspots in the fields of neural engineering, wearable devices, and smart healthcare in recent years. Human emotional states are closely related to brain neural activity, autonomic nervous system signals, and cardiovascular activity. Especially during specific physiological cycles (such as pregnancy and menopause) or high-pressure life stages, the risk of emotional fluctuations, anxiety, and depression increases significantly due to factors such as hormonal fluctuations, work pressure, sleep deprivation, and anxiety expectations. Therefore, non-drug-based smart devices that can continuously and objectively monitor users' emotions and provide timely intervention have significant medical value.

[0003] Existing emotion monitoring devices mainly include wearable devices based on single-modal or dual-modal methods such as electroencephalography (EEG), photoplethysmography (PPG), and electrical skin activity (EDA). However, existing technologies generally suffer from the following problems:

[0004] 1. The emotion acquisition signal is singular, and the noise resistance is insufficient:

[0005] Currently available meditation headsets (such as Muse and Flowtime) mostly rely on single- or dual-lead frontal EEG for simple meditation feedback. Their EEG signal acquisition channels are limited, and they lack effective extraction of more clinically significant emotional biomarkers such as the theta / alpha ratio, frontal asymmetry, and low-frequency brain wave changes. They are also susceptible to motion artifacts and cannot provide reliable indicators for daily emotion monitoring.

[0006] 2. Lack of dedicated emotion recognition algorithms:

[0007] Existing emotion classification algorithms are mostly based on datasets of the general population and do not consider the impact of specific physiological states (such as hormonal fluctuations, changes in autonomic nervous patterns, and sleep deprivation) on EEG and heart rate variability (HRV). Therefore, existing systems are not accurate enough when facing scenarios such as anxiety, stress, and depression, and are difficult to use for medical-grade risk screening.

[0008] 3. The existing system lacks two-way interaction capabilities:

[0009] Most brain-computer interface devices operate in a "one-way monitoring" mode, meaning they only collect EEG / PPG signals and provide analysis results. They cannot output "stimuli that actually regulate the human brain" based on the user's emotional feedback in real time. Examples include breathing rhythm guidance, rhythmic resonance light stimulation, transcranial microcurrent stimulation (CES), and neurofeedback gamification training. These bidirectional regulatory functions are crucial in emotional intervention, but they are generally lacking in existing commercial devices.

[0010] 4. Lack of an emotional linkage system within close relationships:

[0011] Existing emotion monitoring devices only provide individual monitoring results and do not support features such as synchronized mother-infant EEG analysis, real-time emotional alerts for partners, or family support systems. In everyday scenarios, immediate support from partners or family members and the quality of mother-infant interaction are of great significance for emotional recovery and postpartum depression recovery, but no mature products have yet been found to achieve this function.

[0012] 5. The device's form factor and wearing comfort are insufficient, making it difficult to adapt to the needs of prolonged wear:

[0013] Existing mood monitoring devices are mostly geared towards scientific research or sports scenarios. Their structures are rigid and heavy, making them unsuitable for pregnant and postpartum users to wear for extended periods (such as during sleep or in multiple scenarios). Insufficient electrode quality and unstable contact will further reduce the quality of EEG data.

[0014] Current technologies cannot provide a brain-computer headband system that is based on multimodal high-precision emotion acquisition, has a dedicated emotion model, and can achieve bidirectional emotional interaction and support for intimate relationships. In particular, there is a lack of a multimodal emotion acquisition module, emotion recognition algorithms that are not adapted to physiological characteristics, a lack of a truly meaningful "brain-machine-human" closed-loop emotion interaction mechanism, and a lack of "bidirectional emotional connection" capabilities at the level of intimate relationships. Therefore, it is necessary to propose a brain-computer headband that has multimodal emotion acquisition, physiological specialization, bidirectional emotion intervention, and an emotional interaction mechanism for intimate relationships, so as to solve the above-mentioned shortcomings of existing technologies. Summary of the Invention

[0015] The technical problem solved by this invention is to provide a bidirectional emotional interaction headband system based on a multimodal brain-computer interface, which has multiple advantages such as high monitoring accuracy, timely and effective intervention, high family participation, and continuous wearability. This system enables emotion management to shift from passive recording to active regulation, and can provide users with scientific, stable, and sustainable emotion management support at all stages, thereby improving the overall mental health level of users. It has clear practical value and industrial application prospects.

[0016] To address the aforementioned technical problems, the present invention provides a bidirectional emotional interaction headband system based on a multimodal brain-computer interface, comprising: a dedicated emotion recognition algorithm module, a bidirectional emotional interaction output module, an intimate relationship synchronization and emotional interaction module, a system control module, and a headband structure;

[0017] The system control module and headband structure include a main control MCU, a multimodal fusion processor, a Bluetooth / ultra-low power communication module, an electrode adaptive pressure structure, and an ultra-light flexible headband structure. The ultra-light flexible headband structure integrates a multimodal emotion acquisition module, and the headband is worn in scenarios such as sleep, commuting, waking up at night, working, and meditation.

[0018] The multimodal emotion acquisition module consists of several parts: an EEG electroencephalogram acquisition unit, a PPG photoelectric sensing unit, an EDA skin conductance response unit, an IMU inertial measurement unit, and a multimodal emotion fusion algorithm module.

[0019] The multimodal emotion acquisition module can acquire and achieve higher accuracy in emotion recognition and stronger anti-interference capabilities;

[0020] The dedicated emotion recognition algorithm module can more accurately identify unique emotional fluctuation characteristics.

[0021] The bidirectional emotional interaction output module significantly improves the efficiency of emotion regulation based on the brain-computer feedback bidirectional emotional interaction mechanism, realizing a real-time emotional intervention closed loop.

[0022] As a further embodiment of the present invention, the EEG acquisition unit is located in the frontal region inside the head ring, and is used to acquire α, β, θ, and δ band signals of the user's frontal cortex in a multi-channel, low-noise, continuous manner. It adopts a differential acquisition structure composed of at least four flexible electrodes, two of which are respectively arranged at the functional key points of the user's left and right frontal lobes, at positions AF7 and AF8, one electrode is arranged in the middle frontal lobe region at position Fz, and the other is a reference electrode located at the posterior edge of the head ring. The electrodes have a flexible fit structure, automatically sense contact impedance and automatically adjust sampling gain to reduce motion artifacts.

[0023] As a further aspect of the present invention, the PPG photoelectric sensing unit is used to collect heart rate (HR) and heart rate variability (HRV). The LF / HF ratio of HRV is used to assist in determining the pressure level. The PPG photoelectric sensing unit is used to non-invasively and continuously collect volumetric pulsation signals related to blood perfusion to obtain heart rate (HR), heart rate variability (HRV), and pulse waveform characteristics. The PPG sensor is integrated into the forehead or temporal region of the headband, an area rich in blood vessels and with thin skin, to improve signal amplitude and reduce the required light penetration depth. It is designed to work in parallel with the EEG, EDA, and IMU multimodal units, forming a time-synchronized physiological data stream. The PPG channel is located in the headband. In some scenarios, a combination of green light 520–560nm and near-infrared or red light 660nm or 850–940nm is used to take into account both HR and HRV extraction. The PPG photoelectric sensing unit consists of a light emitter (LED), a light receiver, an optical isolation window, an analog front-end (AFE), and an ADC synchronized with the MCU, forming a complete photoelectric acquisition channel. The photoelectric signal enters the analog front-end after passing through the photodetector, including a transimpedance amplifier, variable gain amplification, low-pass anti-aliasing filter, and DC / AC coupling switching. The ADC equivalent sampling rate is set to 125–200Hz with a resolution of 16bit to ensure dynamic range under low amplitude signals.

[0024] As a further aspect of the present invention, the EDA skin conductance unit collects sympathetic nerve activity indicators to identify anxiety and stress states. The EDA skin conductance unit is placed on the upper edge of the forehead where the skin is thin and sweat glands are active. It adopts a differential electrode pair measurement method and shares a time reference with EEG, PPG and IMU for multimodal synchronization. The EDA analog front end includes a high input impedance differential amplifier, a programmable gain amplifier (PGA), an anti-aliasing low-pass filter and an ADC synchronous sampling interface, with a sampling rate of 50-100Hz, a resolution of 16-bit ADC, and a measurement dynamic range covering 0-20μS to cope with skin conductance changes from quiet to highly aroused states.

[0025] As a further aspect of the present invention, the IMU (Inertial Measurement Unit) is used to identify wearing stability and motion artifacts, and is used in conjunction with EEG for artifact removal and sleep stage determination. The multimodal emotion fusion algorithm module extracts relevant features of frontal lobe α asymmetry, θ / β ratio, HRV index, and EDASCR index by fusing EEG, HRV, and EDA through a neural network, as detailed below:

[0026] Each modality is time-aligned and noise-labeled. Then, intermediate representations are extracted using their respective modal encoders. Finally, the modal representations are jointly modeled in the fusion layer to obtain a unified fusion vector. The process is as follows:

[0027]

[0028] Timestamp all channels using a high-precision master clock as a reference; perform isochronous window statistics on low sampling rate modes (HRV intervals); perform short-window spectral feature analysis on EEG and aggregate it over a longer window; perform baseline removal and SCR detection on EDA and statistically analyze the average amplitude; and after alignment, form the original feature vector over a time window T, with the time window uniformly set to 30-60 seconds.

[0029]

[0030] Each mode uses a dedicated encoder. Mapping to low-dimensional representation The EEG encoder uses a lightweight convolutional network with short-time spectrogram input and can output a fixed-dimensional vector; the PPG encoder uses an MLP based on statistical features RMSSD and LF / HF; and the EDA encoder uses a convolutional neural network with peak sequence plus wavelet energy input.

[0031]

[0032] Attention-weighted fusion is used to calculate the relative importance of each modality to the final representation:

[0033]

[0034] in, , , These are trainable parameters.

[0035] As a further aspect of the present invention, the composition and principle of the dedicated emotion recognition algorithm module are as follows:

[0036] (1). Emotional fluctuation model: Establish the correspondence between physiological signs and EEG / HRV to identify anxiety and stress;

[0037] (2). Depression prediction model: Using nighttime sleep fragments EEG and HRV, the nighttime awakening pattern and the proportion of deep sleep were analyzed, and the high θ and low α of the mood band were added for risk prediction;

[0038] (3). Periodic mood fluctuation prediction model: detect periodic EEG rhythm changes and integrate sleep quality and EDA fluctuations to form periodic prediction;

[0039] The final algorithm output includes emotion level, risk classification, intervention priority, and intimate relationship support reminders;

[0040] The frequency domain energy to spectrum ratio extracted from frontal lobe multi-channel EEG was used as a core physiological indicator of emotion. The raw EEG signals were bandpass filtered at 0.5–45 Hz and the power spectrum was reconstructed by sampling at 250–500 Hz. The power of each frequency band was calculated at each time window T using the following formula:

[0041]

[0042] The frequency band values ​​are δ: 0.5-4Hz, θ: 4-8Hz, α: 8-13Hz, β: 13-30Hz. Based on the above power values, a mood-related index is constructed: Frontal α-asymmetry (FAA) is defined as:

[0043]

[0044] in, and To determine the alpha band power of the corresponding channel within the window, the FAA is z-score normalized relative to the user's sliding baseline over the past 7 days:

[0045]

[0046] in , The individual's baseline mean and standard deviation. ≤−1.0: Significant right-side dominance, associated with negative tendency; -1.0 < <1.0, no obvious bias. ≥1.0: Significant left-side dominance, associated with a positive tendency;

[0047] The θ / β ratio is defined as:

[0048]

[0049] TBR < 0.5: Stronger β, higher attention; 0.5 ≤ TBR ≤ 1.5: Neutral range; TBR > 1.5: Relatively enhanced θ, reflecting fatigue or stress.

[0050] Heartbeat interval sequence obtained from PPG Using the raw data, time-domain and frequency-domain HRV indices were calculated to reflect autonomic neural regulation. Short-term RMSSD was selected as the time-domain indicator.

[0051]

[0052] In the frequency domain, the power spectrum is estimated by interpolating the RR sequence to calculate the power at low frequency (LF) 0.04–0.15 Hz and high frequency (HF) 0.15–0.4 Hz. The LF / HF ratio is used as an indicator of the balance between the sympathetic and non-sympathetic inductance.

[0053]

[0054] By combining IMU annotations and employing stronger artifact suppression or discarding strategies during high-activity periods, increased LF / HF or decreased RMSSD are often associated with stress and anxiety and can serve as univariate indicators of stress levels in the model. If RMSSD < 25ms and LF / HF > 2, it is judged as "sympathetic dominance"; if RMSSD ≥ 35ms and LF / HF < 1, it is judged as "parasympathetic dominance".

[0055] EDA reflects short-term sympathetic arousal and is used to detect transient emotional arousal. The EDA signal is decomposed into a slow baseline tonic and a rapid synaptic response (SCR). The implementation steps include filtering 0.05–5Hz, baseline trend removal, and peak detection. The metrics used include SCR count and the average peak value of a single SCR. SCR duration distribution, defining SCR counts on window T. Compared with average peak amplitude:

[0056]

[0057] in For the first Secondary SCR peak amplitude, low activity: / min≤1, moderately active: 1< / min≤4, high activity: / min>4;

[0058] Sleep-related indicators were extracted from continuous monitoring to determine periods of emotional vulnerability: total sleep time, deep sleep ratio, number of nighttime awakenings, and micro-awakening density. Stage labels could be obtained based on multimodal EEG+IMU+PPG. And calculate the statistics, the average deep sleep ratio over the most recent 24 hours and 7 days. with standard deviation These long-term indicators serve as background input to the emotion recognition model, thus distinguishing short-term fluctuations from long-term fatigue / sleep deprivation. Short-term deviations are represented in the model using a difference term.

[0059]

[0060] Emotion levels are fused from end-to-end neural network features The multi-class softmax output is obtained, which outputs the probability distribution for each emotion category. The category with the highest probability is ultimately selected as the current emotion level, while the probability is retained to represent confidence for subsequent decisions. Let the category set be... End-to-end neural networks Output unnormalized logits vector The probability vector is obtained by softmax. :

[0061]

[0062] Emotional level:

[0063]

[0064] Static risk scoring uses the same fusion vector Risk score is generated as input. Used for threshold alarms and generating interpretable weights For user viewing, the training process involves first training a neural network to obtain high-quality data. Then use Fine-tuning logistic regression makes This becomes a local linear approximation of the neural model, thereby achieving a dual-track output that is both high-performance and interpretable;

[0065]

[0066] in ∈ (0,1) represents a risk probability estimate that can be interpreted based on weights:

[0067] Risk grading is based on the interpretable static score of the fusion vector. Continuous risk scores are obtained by synthesizing the relevant anomaly probabilities with neural networks. Then, by mapping to discrete levels through a preset threshold, the anomaly coefficient of the neural network is defined as the aggregation of the probabilities of high-risk classes. ={Anxiety, Depressive Tendency}, Definition:

[0068]

[0069] Weights are combined with static and neural outputs, with coefficients λ∈[0,1] adjusted during deployment: ;

[0070] Discrete risk level mapping:

[0071] The initial thresholds are set as follows: T1=0.2, T2=0.5, T3=0.8;

[0072] Emotions are continuous and time-dependent states. A simplified linear Gaussian state-space model is used to track latent emotional states. With observation :

[0073]

[0074] in For known external interventions, , For process and observation noise, Kalman filtering can provide the optimal online estimate. And based on the prediction error, the intervention strategy is adaptively adjusted when the predicted sentiment trend deteriorates and The intervention intensity is automatically increased when the threshold is approached.

[0075] To mitigate the impact of cross-individual variability, an individualized baseline and an online adaptive mechanism are introduced into the model. The baseline is estimated using exponential smoothing.

[0076]

[0077] in It represents the original feature value at a certain moment, and α∈(0,1) controls the adaptation rate. At the model training level, it is first pre-trained with large-scale general data, and then transferred learning is achieved by using a small amount of labeled data or unlabeled self-supervised fine-tuning for new users, thereby improving the accuracy of individual recognition and reducing misjudgments during the cold start period.

[0078] For the "Intimate Relationship Synchronization" module, the coherence of signals from two or more individuals is used as a quantitative indicator, specifically frequency domain coherence.

[0079]

[0080] in For mutual spectrum, , For each power spectrum, select the emotion-related frequency band to calculate the intra-band coherence, and combine the confidence interval to judge significant synchronization enhancement. If the synchronization of the two intimate parties decreases significantly in the short term or one party shows high risk, this system can trigger communication suggestions and joint practice intervention prompts at the relationship level.

[0081] Intimate relationship support alerts are determined by three parts: ontological risk, intimate partner risk, and relevance indicators, with ontological risk being the most significant factor. The risk of intimate partners is obtained Synchronization metrics:

[0082]

[0083] Where [F1, F2] are the selected emotion-related frequency bands mentioned above. If one party has a high risk and the synchronicity decreases or the asymmetry increases, it indicates that intervention is needed at the relationship level. Asymmetry is defined as follows:

[0084]

[0085] Intimate support trigger logic:

[0086]

[0087] The thresholds are defined as 0.6, 0.3, and 0.25 respectively. When the trigger is true, the system executes the following sequence of actions: pushes instant companionship suggestions, arranges synchronous practice, reminds the close partner to pay attention to and participate in comforting tasks, or suggests referring to clinical consultation.

[0088] As a further aspect of the present invention, the principle and effect of the two-way emotional interaction output module are as follows:

[0089] When the system detects a risk of emotional fluctuation, it will automatically select the best intervention plan based on the intensity of the emotion to achieve "personalized emotion regulation," including the following intervention methods:

[0090] (1) Rhythm-guided audio breathing feedback: induces enhancement of alpha and theta waves, improving relaxation;

[0091] (2) Breathing rhythm guidance algorithm: The headband plays rhythm prompts to synchronize the breathing rate with the ideal HRV resonance frequency of 0.1Hz;

[0092] (3) Neurofeedback training: EEG emotion indicators are converted into audiovisual feedback in real time, and users can return their emotions to a steady state through self-regulation;

[0093] (4) CES: Improve anxiety and sleep quality through low-intensity transcranial micro-electric stimulation, and avoid overstimulation by synchronous monitoring with EEG.

[0094] As a further aspect of the present invention, the specific principle of the intimate relationship synchronization and emotion interaction module is as follows:

[0095] (1) Intimate Relationship EEG Synchronization Analysis: Used to detect the emotional synchronicity between the user and their partner or family member;

[0096] (2) Support system: When a user experiences a high risk of emotional distress, the system will notify the partner or family support to provide comfort and companionship suggestions, thereby achieving two-way emotional linkage between the user and the intimate relationship;

[0097] (3) AI Emotion Coach: Based on the collected emotion data, it provides communication suggestions and emotional support strategies to the user's close relationships.

[0098] As a further aspect of the present invention, the ultralight flexible headband structure is constructed with a lightweight flexible material to form the headband body, and combined with adaptively adjustable flexible electrodes, so that EEG acquisition maintains a stable skin contact state during long-term wear or nighttime sleep. The real-time monitoring of electrode contact impedance and automatic gain adjustment function reduce signal loss caused by position changes, thus improving the continuous efficiency of EEG data.

[0099] As a further aspect of the present invention, the CES uses low-amplitude current stimulation, which can gently regulate the activity of neurotransmitters in the brain.

[0100] Compared with related technologies, the bidirectional emotional interaction headband system based on a multimodal brain-computer interface provided by this invention has the following beneficial effects:

[0101] 1. This invention can identify specific emotional states such as anxiety, increased stress, mood instability after waking up at night, and postpartum depressive tendencies, making the system's output more consistent with the actual emotional expressions of the population. This proprietary model overcomes the shortcomings of general emotion algorithms in identifying insensitive and high false positive rates in specific scenarios, thus improving the applicability and scientific value of this invention in the medical and health field.

[0102] 2. This invention can more accurately adjust emotional states in an individualized manner, achieving the immediacy, continuity and controllability of emotion improvement. Unlike traditional meditation devices or one-way emotion prompting systems, the closed-loop feedback constructed by this invention improves the efficiency of emotion regulation and has obvious technological progress.

[0103] 3. This invention can intervene in a timely manner during the peak of emotional fluctuations, improving the user's sense of stability and the speed of emotional recovery. The non-drug intervention method adopted by this invention is feasible, safe and controllable, and has an effect that is difficult to achieve with existing equipment.

[0104] 4. This invention integrates multimodal physiological data acquisition technology, a dedicated emotion recognition model, bidirectional brain-computer interface intervention, and intimate relationship support algorithms into a unified emotion management platform. This forms a complete closed-loop process from data acquisition to intervention feedback and relationship linkage. Compared with traditional devices that can only provide one-way emotion monitoring, this invention has multiple advantages such as high monitoring accuracy, timely and effective intervention, high family participation, and continuous wearability. It transforms emotion management from passive recording to active regulation. The technical solution of this invention can provide users with scientific, stable, and sustainable emotion management support at all stages, improve the overall mental health level of users, and has clear practical value and industrial application prospects. Attached Figure Description

[0105] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0106] Figure 1 This is a flowchart of the two-way emotional interaction method of the present invention;

[0107] Figure 2 This is a block diagram of the overall composition of the present invention. Detailed Implementation

[0108] Please refer to the following: Figures 1 to 2 ,in, Figure 1 This is a flowchart of the two-way emotional interaction method of the present invention; Figure 2 This is a block diagram of the overall composition of the present invention. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface includes: a dedicated emotion recognition algorithm module, a bidirectional emotional interaction output module, an intimate relationship synchronization and emotional interaction module, a system control module, and a headband structure;

[0109] The system control module and headband structure include a main control MCU, a multimodal fusion processor, a Bluetooth / ultra-low power communication module, an electrode adaptive pressure structure, and an ultra-light flexible headband structure. The ultra-light flexible headband structure integrates a multimodal emotion acquisition module, and the headband is worn in scenarios such as sleep, commuting, waking up at night, working, and meditation.

[0110] Flexible wearing structure and adaptive electrode design significantly improve long-term wearing comfort and data continuity:

[0111] This invention uses lightweight, flexible materials to construct the headband body and incorporates adaptively adjustable flexible electrodes. This ensures stable skin contact during EEG acquisition, whether worn for extended periods or during sleep. Real-time monitoring of electrode contact impedance and automatic gain adjustment reduce signal loss due to positional changes, improving the continuous effectiveness of EEG data. The flexible headband structure can autonomously adjust pressure distribution according to head shape, reducing pressure during prolonged wear and enhancing the user experience. Furthermore, the adaptive contact mechanism reduces the probability of electrode detachment and artifact interference, enabling the invention to achieve higher data stability in real-life scenarios (such as sleep, rest, and movement), meeting the needs of people for long-term continuous monitoring. Compared with traditional rigid EEG devices, this invention has significant advantages.

[0112] The multimodal emotion acquisition module consists of several parts: an EEG electroencephalogram acquisition unit, a PPG photoelectric sensing unit, an EDA skin conductance response unit, an IMU inertial measurement unit, and a multimodal emotion fusion algorithm module.

[0113] The multimodal emotion acquisition module can acquire and achieve higher accuracy in emotion recognition and stronger anti-interference capabilities;

[0114] Multimodal emotion acquisition achieves higher accuracy in emotion recognition and stronger resistance to interference:

[0115] By integrating multimodal physiological signal acquisition units such as EEG, PPG, EDA, and IMU into the headband structure, the system can simultaneously acquire emotion-related data from three dimensions: brain electrical activity, autonomic nervous system response, and motor state. The four-channel frontal lobe signals acquired by EEG can provide highly emotion-correlated α, θ, and β band features. The PPG module can extract HR and HRV in real time to reflect the degree of autonomic nervous tension. The EDA module is used to capture the skin conductance response of the sympathetic nervous system, and the IMU is used to identify motion artifacts and compensate for them in the fusion algorithm. After the above multi-source data are jointly trained by the fusion model, the system can simultaneously rely on changes in brain electrical rhythm, heart rate rhythm regulation, and skin conductance changes to make emotion judgments, significantly reducing the error caused by noise in a single physiological signal.

[0116] Since the responses of each modality signal to emotional changes are complementary, the multimodal structure can improve the robustness and stability of the recognition model, enabling the system to maintain stable emotion recognition performance in daily environments (such as walking, resting at night, and light activity). Compared with traditional single-modal EEG or heart rate detection, the multimodal fusion model can achieve higher classification accuracy and lower false recognition rate in laboratory labeled data, demonstrating the substantial improvement of the reliability of emotion recognition in this invention.

[0117] The dedicated emotion recognition algorithm module can more accurately identify unique emotional fluctuation characteristics.

[0118] Dedicated emotion recognition models can more accurately identify unique emotional fluctuation characteristics:

[0119] Based on traditional emotion recognition algorithms, this invention constructs a specific emotion recognition model targeting factors such as endocrine fluctuations, changes in autonomic nervous system regulation, and changes in sleep structure. This model focuses on extracting emotion-sensitive indicators such as asymmetric patterns in the frontal lobe α band and changes in the θ / β ratio from EEG signals, and combines them with characteristic parameters reflecting psychological stress, such as HRV LF / HF and RMSSD. At the same time, it introduces indicators reflecting sympathetic arousal, such as EDA skin conductance peak value and skin conductance level, to adapt to the user's physiological state.

[0120] By introducing a model training method based on specific samples, this invention can identify specific emotional states such as anxiety, increased stress, emotional instability after waking up at night, and postpartum depression tendencies, making the system's output more consistent with the actual emotional expressions of the population. This dedicated model overcomes the shortcomings of general emotion algorithms in identifying insensitive and having a high false alarm rate in specific scenarios, thereby improving the applicability and scientific value of this invention in the field of medical and health care.

[0121] The bidirectional emotional interaction output module significantly improves the efficiency of emotion regulation based on the brain-computer feedback bidirectional emotional interaction mechanism, realizing a real-time emotional intervention closed loop.

[0122] The EEG acquisition unit is located in the frontal region inside the headband and is used to acquire α, β, θ, and δ band signals of the user's frontal cortex in a multi-channel, low-noise, continuous manner. It uses at least four flexible electrodes to form a differential acquisition structure. Two electrodes are respectively placed at the functional key points on the user's left and right frontal lobes, at positions AF7 and AF8. One electrode is placed in the middle frontal lobe region at position Fz. The other is a reference electrode located at the posterior edge of the headband. The electrodes have a flexible fit structure, automatically sense contact impedance and automatically adjust the sampling gain to reduce motion artifacts.

[0123] The PPG photoelectric sensing unit is used to collect heart rate (HR) and heart rate variability (HRV). The LF / HF ratio of HRV is used to assist in determining pressure levels. This unit is used for non-invasive, continuous acquisition of volumetric pulsatility signals related to blood perfusion to obtain heart rate (HR), heart rate variability (HRV), and pulse waveform characteristics. The PPG sensor is integrated into the forehead or temporal region of the headband, an area rich in blood vessels and with thin skin, to improve signal amplitude and reduce the required light penetration depth. It is designed to work in parallel with the EEG, EDA, and IMU multimodal units, forming a time-synchronized physiological data stream. The PPG channel uses green light in head-mounted applications. The combination of 520–560nm light with near-infrared or red light at 660nm or 850–940nm is used to balance HR and HRV extraction. The PPG photoelectric sensing unit consists of a light emitter (LED), a light receiver, an optical isolation window, an analog front-end (AFE), and an ADC synchronized with the MCU, forming a complete photoelectric acquisition channel. The photoelectric signal enters the analog front-end after passing through the photodetector, which includes a transimpedance amplifier, variable gain amplifier, low-pass anti-aliasing filter, and DC / AC coupling switching. The ADC equivalent sampling rate is set to 125–200Hz with a resolution of 16bit to ensure dynamic range under low amplitude signals.

[0124] The EDA (Electrodermal Analog) skin activity unit collects sympathetic nerve activity indicators to identify anxiety and stress states. The EDA skin activity unit is placed on the upper edge of the forehead where the skin is thin and sweat glands are active. It uses a differential electrode pair measurement method and shares a time reference with EEG, PPG, and IMU for multimodal synchronization. The EDA analog front end includes a high input impedance differential amplifier, a programmable gain amplifier (PGA), an anti-aliasing low-pass filter, and an ADC synchronous sampling interface with a sampling rate of 50–100Hz, a resolution of 16-bit ADC, and a measurement dynamic range covering 0–20μS to cope with skin activity changes from quiet to highly aroused states.

[0125] The IMU (Inertial Measurement Unit) is used to identify wearing stability and motion artifacts, and is used in conjunction with EEG for artifact removal and sleep stage determination. The multimodal emotion fusion algorithm module uses a neural network to fuse EEG, HRV, and EDA to extract relevant features of emotion, such as frontal α asymmetry, θ / β ratio, HRV index, and EDASCR index, as detailed below:

[0126] Each modality is time-aligned and noise-labeled. Then, intermediate representations are extracted using their respective modal encoders. Finally, the modal representations are jointly modeled in the fusion layer to obtain a unified fusion vector. The process is as follows:

[0127]

[0128] Timestamp all channels using a high-precision master clock as a reference; perform isochronous window statistics on low sampling rate modes (HRV intervals); perform short-window spectral feature analysis on EEG and aggregate it over a longer window; perform baseline removal and SCR detection on EDA and statistically analyze the average amplitude; and after alignment, form the original feature vector over a time window T, with the time window uniformly set to 30-60 seconds.

[0129]

[0130] Each mode uses a dedicated encoder. Mapping to low-dimensional representation The EEG encoder uses a lightweight convolutional network with short-time spectrogram input and can output a fixed-dimensional vector; the PPG encoder uses an MLP based on statistical features RMSSD and LF / HF; and the EDA encoder uses a convolutional neural network with peak sequence plus wavelet energy input.

[0131]

[0132] Attention-weighted fusion is used to calculate the relative importance of each modality to the final representation:

[0133]

[0134] in, , , These are trainable parameters.

[0135] A two-way emotion interaction mechanism based on brain-computer feedback significantly improves the efficiency of emotion regulation and achieves a closed loop for real-time emotion intervention:

[0136] This invention constructs a closed-loop "collection-recognition-feedback" system based on emotion recognition, realizing two-way emotional interaction. The system judges the user's emotional state by monitoring the user's EEG rhythm, HRV level, and EDA changes in real time. When increased stress or anxiety is detected, it can automatically trigger corresponding intervention modes, including rhythmic guided audio, breathing rhythm cues, audiovisual feedback, or neurofeedback training. Rhythmic guided audio can induce an increase in alpha waves or stabilize theta waves, thereby promoting relaxation. Breathing rhythm cues can help users achieve a stable resonant breathing frequency of 0.08–0.12 Hz, improving the balance of the autonomic nervous system. Neurofeedback training allows users to actively adjust their own EEG patterns through visual feedback.

[0137] Since the feedback stimulus parameters are dynamically adjusted based entirely on real-time collected physiological signals, this invention can more accurately adjust emotional states individually, achieving immediacy, continuity, and controllability in emotion improvement. Unlike traditional meditation devices or one-way emotion prompting systems, the closed-loop feedback constructed by this invention improves the efficiency of emotion regulation and has significant technological advancements.

[0138] The composition and principle of the dedicated emotion recognition algorithm module are as follows:

[0139] (1). Emotional fluctuation model: Establish the correspondence between physiological signs and EEG / HRV to identify anxiety and stress;

[0140] (2). Depression prediction model: Using nighttime sleep fragments EEG and HRV, the nighttime awakening pattern and the proportion of deep sleep were analyzed, and the high θ and low α of the mood band were added for risk prediction;

[0141] (3). Periodic mood fluctuation prediction model: detect periodic EEG rhythm changes and integrate sleep quality and EDA fluctuations to form periodic prediction;

[0142] The final algorithm output includes emotion level, risk classification, intervention priority, and intimate relationship support reminders;

[0143] The frequency domain energy to spectrum ratio extracted from frontal lobe multi-channel EEG was used as a core physiological indicator of emotion. The raw EEG signals were bandpass filtered at 0.5–45 Hz and the power spectrum was reconstructed by sampling at 250–500 Hz. The power of each frequency band was calculated at each time window T using the following formula:

[0144]

[0145] The frequency band values ​​are δ: 0.5-4Hz, θ: 4-8Hz, α: 8-13Hz, β: 13-30Hz. Based on the above power values, a mood-related index is constructed: Frontal α-asymmetry (FAA) is defined as:

[0146]

[0147] in, and To determine the alpha band power of the corresponding channel within the window, the FAA is z-score normalized relative to the user's sliding baseline over the past 7 days:

[0148]

[0149] in , The individual's baseline mean and standard deviation. ≤−1.0: Significant right-side dominance, associated with negative tendency; −1.0 < <1.0, no obvious bias. ≥1.0: Significant left-side dominance, associated with a positive tendency;

[0150] The θ / β ratio is defined as:

[0151]

[0152] TBR < 0.5: Stronger β, higher attention; 0.5 ≤ TBR ≤ 1.5: Neutral range; TBR > 1.5: Relatively enhanced θ, reflecting fatigue or stress.

[0153] Heartbeat interval sequence obtained from PPG Using the raw data, time-domain and frequency-domain HRV indices were calculated to reflect autonomic neural regulation. Short-term RMSSD was selected as the time-domain indicator.

[0154]

[0155] In the frequency domain, the power spectrum is estimated by interpolating the RR sequence to calculate the power at low frequency (LF) 0.04–0.15 Hz and high frequency (HF) 0.15–0.4 Hz. The LF / HF ratio is used as an indicator of the balance between the sympathetic and non-sympathetic inductance.

[0156]

[0157] By combining IMU annotations and employing stronger artifact suppression or discarding strategies during high-activity periods, increased LF / HF or decreased RMSSD are often associated with stress and anxiety and can serve as univariate indicators of stress levels in the model. If RMSSD < 25ms and LF / HF > 2, it is judged as "sympathetic dominance"; if RMSSD ≥ 35ms and LF / HF < 1, it is judged as "parasympathetic dominance".

[0158] EDA reflects short-term sympathetic arousal and is used to detect transient emotional arousal. The EDA signal is decomposed into a slow baseline tonic and a rapid synaptic response (SCR). The implementation steps include filtering 0.05–5Hz, baseline trend removal, and peak detection. The metrics used include SCR count and the average peak value of a single SCR. SCR duration distribution, defining SCR counts on window T. Compared with average peak amplitude:

[0159]

[0160] in For the first Secondary SCR peak amplitude, low activity: / min≤1, moderately active: 1< / min≤4, high activity: / min>4;

[0161] Sleep-related indicators were extracted from continuous monitoring to determine periods of emotional vulnerability: total sleep time, deep sleep ratio, number of nighttime awakenings, and micro-awakening density. Stage labels could be obtained based on multimodal EEG+IMU+PPG. And calculate the statistics, the average deep sleep ratio over the most recent 24 hours and 7 days. with standard deviation These long-term indicators serve as background input to the emotion recognition model, thus distinguishing short-term fluctuations from long-term fatigue / sleep deprivation. Short-term deviations are represented in the model using a difference term.

[0162]

[0163] Emotion levels are fused from end-to-end neural network features The multi-class softmax output is obtained, which outputs the probability distribution for each emotion category. The category with the highest probability is ultimately selected as the current emotion level, while the probability is retained to represent confidence for subsequent decisions. Let the category set be... End-to-end neural networks Output unnormalized logits vector The probability vector is obtained by softmax. :

[0164]

[0165] Emotional level:

[0166]

[0167] Static risk scoring uses the same fusion vector Risk score is generated as input. Used for threshold alarms and generating interpretable weights For user viewing, the training process involves first training a neural network to obtain high-quality data. Then use Fine-tuning logistic regression makes This becomes a local linear approximation of the neural model, thereby achieving a dual-track output that is both high-performance and interpretable;

[0168]

[0169] in ∈ (0,1) represents a risk probability estimate that can be interpreted based on weights:

[0170] Risk grading is based on the interpretable static score of the fusion vector. Continuous risk scores are obtained by synthesizing the relevant anomaly probabilities with neural networks. Then, by mapping to discrete levels through a preset threshold, the anomaly coefficient of the neural network is defined as the aggregation of the probabilities of high-risk classes. ={Anxiety, Depressive Tendency}, Definition:

[0171]

[0172] Weights are combined with static and neural outputs, with coefficients λ∈[0,1] adjusted during deployment: ;

[0173] Discrete risk level mapping:

[0174] The initial thresholds are set as follows: T1=0.2, T2=0.5, T3=0.8;

[0175] Emotions are continuous and time-dependent states. A simplified linear Gaussian state-space model is used to track latent emotional states. With observation :

[0176]

[0177] in For known external interventions, , For process and observation noise, Kalman filtering can provide the optimal online estimate. And based on the prediction error, the intervention strategy is adaptively adjusted when the predicted sentiment trend deteriorates and The intervention intensity is automatically increased when the threshold is approached.

[0178] To mitigate the impact of cross-individual variability, an individualized baseline and an online adaptive mechanism are introduced into the model. The baseline is estimated using exponential smoothing.

[0179]

[0180] in It represents the original feature value at a certain moment, and α∈(0,1) controls the adaptation rate. At the model training level, it is first pre-trained with large-scale general data, and then transferred learning is achieved by using a small amount of labeled data or unlabeled self-supervised fine-tuning for new users, thereby improving the accuracy of individual recognition and reducing misjudgments during the cold start period.

[0181] For the "Intimate Relationship Synchronization" module, the coherence of signals from two or more individuals is used as a quantitative indicator, specifically frequency domain coherence.

[0182]

[0183] in For mutual spectrum, , For each power spectrum, select the emotion-related frequency band to calculate the intra-band coherence, and combine the confidence interval to judge significant synchronization enhancement. If the synchronization of the two intimate parties decreases significantly in the short term or one party shows high risk, this system can trigger communication suggestions and joint practice intervention prompts at the relationship level.

[0184] Intimate relationship support alerts are determined by three parts: ontological risk, intimate partner risk, and relevance indicators, with ontological risk being the most significant factor. The risk of intimate partners is obtained Synchronization metrics:

[0185]

[0186] Where [F1, F2] are the selected emotion-related frequency bands mentioned above. If one party has a high risk and the synchronicity decreases or the asymmetry increases, it indicates that intervention is needed at the relationship level. Asymmetry is defined as follows:

[0187]

[0188] Intimate support trigger logic:

[0189]

[0190] Where the threshold Defined as 0.6, 0.3, and 0.25 respectively, when the trigger is true, the system executes the following sequence of actions: pushes instant companionship suggestions, arranges synchronous exercises, reminds the close partner to pay attention to and participate in comforting tasks, or suggests referring to clinical consultation.

[0191] The system's intervention output method improves emotional stability without altering medication treatment, and is suitable for safe emotion regulation.

[0192] The emotional intervention output designed in this invention includes non-contact audio stimulation, breathing regulation cues, visual feedback, and a CES module. All of the above intervention methods are physical, non-pharmacological regulation methods, which are particularly suitable for drug-sensitive user groups. Rhythmic audio can guide users into a more stable brainwave state through binaural beats or natural sounds; breathing cues increase vagal nerve tone by reducing breathing frequency; visual feedback can provide a positive reinforcement mechanism based on neural feedback; and CES uses low-amplitude current stimulation to gently regulate neurotransmitter activity in the brain.

[0193] Since all stimulation methods are selected based on real-time emotional state, this invention can intervene in a timely manner during the peak of emotional fluctuations, improving the user's sense of stability and the speed of emotional recovery. The non-drug intervention method adopted by this invention is feasible, safe and controllable, and has an effect that is difficult to achieve with existing equipment.

[0194] The system enables emotional linkage among members of close relationships, which helps improve the efficiency of mutual support interventions.

[0195] The emotional linkage mechanism constructed by this invention can push a user's emotional state, stress index, or risk level to the terminals of close relationship members, and generate communication suggestions, companionship prompts, or intervention plans based on an emotional model. For specific groups (such as newborns and the elderly), this invention can be equipped with a lightweight heart rate patch or sleep monitoring device to detect indicators such as sleep synchronicity and emotional rhythm consistency between users. The system provides interactive feedback to users by analyzing the degree of rhythm coupling. This process enables close relationship members to understand the user's psychological state in a timely manner and provide practical support when the system issues high-risk warnings, such as providing comfort, sharing tasks, or cooperating in adjustment exercises. This function effectively improves the effectiveness of support interventions between close relationships, realizing multi-subject participatory emotional management that extends from individual emotional regulation to the relationship level, an effect that traditional emotional monitoring devices cannot achieve.

[0196] The principle and effect of the two-way emotional interaction output module are as follows:

[0197] When the system detects a risk of emotional fluctuation, it will automatically select the best intervention plan based on the intensity of the emotion to achieve "personalized emotion regulation," including the following intervention methods:

[0198] (1) Rhythm-guided audio breathing feedback: induces enhancement of alpha and theta waves, improving relaxation;

[0199] (2) Breathing rhythm guidance algorithm: The headband plays rhythm prompts to synchronize the breathing rate with the ideal HRV resonance frequency of 0.1Hz;

[0200] (3) Neurofeedback training: EEG emotion indicators are converted into audiovisual feedback in real time, and users can return their emotions to a steady state through self-regulation;

[0201] (4) CES: Improve anxiety and sleep quality through low-intensity transcranial micro-electric stimulation, and avoid overstimulation by synchronous monitoring with EEG.

[0202] The specific principle behind the intimate relationship synchronization and emotional interaction module is as follows:

[0203] (1) Intimate Relationship EEG Synchronization Analysis: Used to detect the emotional synchronicity between the user and their partner or family member;

[0204] (2) Support system: When a user experiences a high risk of emotional distress, the system will notify the partner or family support to provide comfort and companionship suggestions, thereby achieving two-way emotional linkage between the user and the intimate relationship;

[0205] (3) AI Emotion Coach: Based on the collected emotion data, it provides communication suggestions and emotional support strategies to the user's close relationships.

[0206] The ultralight flexible headband structure uses lightweight flexible materials to construct the headband body and combines it with adaptively adjustable flexible electrodes to maintain a stable skin contact state during EEG acquisition during long-term wear or overnight sleep. The real-time monitoring of electrode contact impedance and automatic gain adjustment function reduce signal loss caused by position changes, thus improving the continuous efficiency of EEG data.

[0207] The CES uses low-amplitude electrical stimulation, which can gently regulate neurotransmitter activity in the brain.

[0208] The overall system forms a complete closed loop of "collection-identification-intervention-synchronization-support", improving the overall efficiency of mental health management.

[0209] This invention integrates multimodal physiological data acquisition technology, a dedicated emotion recognition model, a two-way brain-computer interface system, and an intimate relationship support algorithm into a unified emotion management platform. This forms a complete closed-loop process from data acquisition to intervention feedback and relationship linkage. Compared with traditional devices that can only provide one-way emotion monitoring, this invention has multiple advantages such as high monitoring accuracy, timely and effective intervention, high family participation, and continuous wearability. It transforms emotion management from passive recording to active regulation. The technical solution of this invention can provide users with scientific, stable, and sustainable emotion management support at all stages, improve users' overall mental health, and has clear practical value and industrial application prospects.

[0210] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments, or they can be used directly or indirectly, without departing from the principles and spirit of the invention. In other related technical fields, the scope of the invention is defined by the appended claims and their equivalents, and they are similarly included within the scope of patent protection of the invention.

Claims

1. A bidirectional emotional interaction headband system based on a multimodal brain-computer interface, characterized in that, include: The system includes a dedicated emotion recognition algorithm module, a two-way emotion interaction output module, an intimate relationship synchronization and emotion interaction module, a system control module, and a headband structure. The system control module and headband structure include a main control MCU, a multimodal fusion processor, a Bluetooth / ultra-low power communication module, an electrode adaptive pressure structure, and an ultra-light flexible headband structure. The ultra-light flexible headband structure integrates a multimodal emotion acquisition module, and the headband is worn in scenarios such as sleep, commuting, waking up at night, working, and meditation. The multimodal emotion acquisition module consists of several parts: an EEG electroencephalogram acquisition unit, a PPG photoelectric sensing unit, an EDA skin conductance response unit, an IMU inertial measurement unit, and a multimodal emotion fusion algorithm module. The multimodal emotion acquisition module can acquire and achieve higher accuracy in emotion recognition and stronger anti-interference capabilities; The dedicated emotion recognition algorithm module can more accurately identify unique emotional fluctuation characteristics. The bidirectional emotional interaction output module significantly improves the efficiency of emotion regulation based on the bidirectional emotional interaction mechanism of brain-computer feedback, and realizes a real-time emotional intervention closed loop. The IMU (Inertial Measurement Unit) is used to identify wearing stability and motion artifacts, and is used in conjunction with EEG for artifact removal and sleep stage determination. The multimodal emotion fusion algorithm module uses a neural network to fuse EEG, HRV, and EDA to extract relevant features of emotion, such as frontal α asymmetry, θ / β ratio, HRV index, and EDASCR index, as detailed below: Each modality is time-aligned and noise-labeled. Then, intermediate representations are extracted using their respective modal encoders. Finally, the modal representations are jointly modeled in the fusion layer to obtain a unified fusion vector. The process is as follows: ； Timestamp all channels using a high-precision master clock as a reference; perform isochronous window statistics on low sampling rate modes (HRV intervals); perform short-window spectral feature analysis on EEG and aggregate it over a longer window; perform baseline removal and SCR detection on EDA and statistically analyze the average amplitude; and after alignment, form the original feature vector over a time window T, with the time window uniformly set to 30-60 seconds. ； Each mode uses a dedicated encoder. Mapping to low-dimensional representation The EEG encoder uses a lightweight convolutional network with short-time spectrogram input and can output a fixed-dimensional vector; the PPG encoder uses an MLP based on statistical features RMSSD and LF / HF; and the EDA encoder uses a convolutional neural network with peak sequence plus wavelet energy input. ； Attention-weighted fusion is used to calculate the relative importance of each modality to the final representation: ； in, , , These are trainable parameters; The composition and principle of the dedicated emotion recognition algorithm module are as follows: (1). Emotional fluctuation model: Establish the correspondence between physiological signs and EEG / HRV to identify anxiety and stress; (2). Depression prediction model: Using nighttime sleep fragments EEG and HRV, the nighttime awakening pattern and the proportion of deep sleep were analyzed, and the high θ and low α of the mood band were added for risk prediction; (3). Periodic mood fluctuation prediction model: detect periodic EEG rhythm changes and integrate sleep quality and EDA fluctuations to form periodic prediction; The algorithm's final output includes emotion level, risk rating, intervention priority, and reminders for close relationship support.

2. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 1, characterized in that: The EEG acquisition unit is located in the frontal region inside the headband and is used to acquire α, β, θ, and δ band signals of the user's frontal cortex in a multi-channel, low-noise, continuous manner. It uses at least four flexible electrodes to form a differential acquisition structure. Two electrodes are respectively placed at the functional key points on the user's left and right frontal lobes, at positions AF7 and AF8. One electrode is placed in the middle frontal lobe region at position Fz. The other is a reference electrode located at the posterior edge of the headband. The electrodes have a flexible fit structure, automatically sense contact impedance and automatically adjust the sampling gain to reduce motion artifacts.

3. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 1, characterized in that: The PPG photoelectric sensing unit is used to collect heart rate (HR) and heart rate variability (HRV). The LF / HF ratio of HRV is used to assist in determining pressure levels. This unit is used for non-invasive, continuous acquisition of volumetric pulsatility signals related to blood perfusion to obtain heart rate (HR), heart rate variability (HRV), and pulse waveform characteristics. The PPG sensor is integrated into the forehead or temporal region of the headband, an area rich in blood vessels and with thin skin, to improve signal amplitude and reduce the required light penetration depth. It is designed to work in parallel with the EEG, EDA, and IMU multimodal units, forming a time-synchronized physiological data stream. The PPG channel uses green light in head-mounted applications. The combination of 520–560nm light with near-infrared or red light at 660nm or 850–940nm is used to balance HR and HRV extraction. The PPG photoelectric sensing unit consists of a light emitter (LED), a light receiver, an optical isolation window, an analog front-end (AFE), and an ADC synchronized with the MCU, forming a complete photoelectric acquisition channel. The photoelectric signal enters the analog front-end after passing through the photodetector, which includes a transimpedance amplifier, variable gain amplifier, low-pass anti-aliasing filter, and DC / AC coupling switching. The ADC equivalent sampling rate is set to 125–200Hz with a resolution of 16bit to ensure dynamic range under low amplitude signals.

4. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 1, characterized in that: The EDA (Electrodermal Analog) skin activity unit collects sympathetic nerve activity indicators to identify anxiety and stress states. The EDA skin activity unit is placed on the upper edge of the forehead where the skin is thin and sweat glands are active. It uses a differential electrode pair measurement method and shares a time reference with EEG, PPG, and IMU for multimodal synchronization. The EDA analog front end includes a high input impedance differential amplifier, a programmable gain amplifier (PGA), an anti-aliasing low-pass filter, and an ADC synchronous sampling interface with a sampling rate of 50–100Hz, a resolution of 16-bit ADC, and a measurement dynamic range covering 0–20μS to cope with skin activity changes from quiet to highly aroused states.

5. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 1, characterized in that: The frequency domain energy to spectrum ratio extracted from frontal lobe multi-channel EEG was used as a core physiological indicator of emotion. The raw EEG signals were bandpass filtered at 0.5–45 Hz and the power spectrum was reconstructed by sampling at 250–500 Hz. The power of each frequency band was calculated at each time window T using the following formula: ； The frequency band values ​​are δ: 0.5-4Hz, θ: 4-8Hz, α: 8-13Hz, β: 13-30Hz. Based on the above power values, a mood-related index is constructed: Frontal α-asymmetry (FAA) is defined as: ； in, and To determine the alpha band power of the corresponding channel within the window, the FAA is z-score normalized relative to the user's sliding baseline over the past 7 days: ； in , The individual's baseline mean and standard deviation. ≤-1.0: Significant right-side dominance, associated with negative tendency; -1.0 < <1.0, no obvious bias. ≥1.0: Significant left-side dominance, associated with a positive tendency; The θ / β ratio is defined as: ； TBR < 0.5: Stronger β, higher attention; 0.5 ≤ TBR ≤ 1.5: Neutral range; TBR > 1.5: Relatively enhanced θ, reflecting fatigue or stress. Heartbeat interval sequence obtained from PPG Using the raw data, time-domain and frequency-domain HRV indices were calculated to reflect autonomic neural regulation. Short-term RMSSD was selected as the time-domain indicator. ； In the frequency domain, the power spectrum is estimated by interpolating the RR sequence to calculate the power at low frequency (LF) 0.04–0.15 Hz and high frequency (HF) 0.15–0.4 Hz. The LF / HF ratio is used as an indicator of the balance between the sympathetic and non-sympathetic inductance. ； By combining IMU annotations and employing stronger artifact suppression or discarding strategies during high-activity periods, increased LF / HF or decreased RMSSD are often associated with stress and anxiety and can serve as univariate indicators of stress levels in the model. If RMSSD < 25ms and LF / HF > 2, it is judged as "sympathetic dominance"; if RMSSD ≥ 35ms and LF / HF < 1, it is judged as "parasympathetic dominance". EDA reflects short-term sympathetic arousal and is used to detect transient emotional arousal. The EDA signal is decomposed into a slow baseline tonic and a rapid synaptic response (SCR). The implementation steps include filtering 0.05–5Hz, baseline trend removal, and peak detection. The metrics used include SCR count and the average peak value of a single SCR. SCR duration distribution, defining SCR counts on window T. Compared with average peak amplitude: ； in For the first Secondary SCR peak amplitude, low activity: / min≤1, moderately active: 1< / min≤4, high activity: / min>4; Sleep-related indicators were extracted from continuous monitoring to determine periods of emotional vulnerability: total sleep time, deep sleep ratio, number of nighttime awakenings, and micro-awakening density. Stage labels could be obtained based on multimodal EEG+IMU+PPG. And calculate the statistics, the average deep sleep ratio over the most recent 24 hours and 7 days. with standard deviation These long-term indicators serve as background input to the emotion recognition model, thus distinguishing short-term fluctuations from long-term fatigue / sleep deprivation. Short-term deviations are represented in the model using a difference term. ； Emotion levels are fused from end-to-end neural network features The multi-class softmax output is obtained, which outputs the probability distribution for each emotion category. The category with the highest probability is selected as the current emotion level, while the probability is retained to represent confidence for subsequent decisions. Let the category set be... End-to-end neural networks Output unnormalized logits vector The probability vector is obtained by softmax. : ； Emotional level: ； Static risk scoring uses the same fusion vector Risk score is generated as input. Used for threshold alarms and generating interpretable weights For user viewing, the training process involves first training a neural network to obtain high-quality data. Then use Fine-tuning logistic regression makes This becomes a local linear approximation of the neural model, thereby achieving a dual-track output that is both high-performance and interpretable; ； in ∈ (0,1) represents a risk probability estimate that can be interpreted based on weights: Risk grading is based on the interpretable static score of the fusion vector. Continuous risk scores are obtained by synthesizing the relevant anomaly probabilities with neural networks. Then, by mapping to discrete levels through a preset threshold, the anomaly coefficient of the neural network is defined as the aggregation of the probabilities of high-risk classes. ={Anxiety, Depressive Tendency}, Definition: ； Weights are combined with static and neural outputs, with coefficients λ∈[0,1] adjusted during deployment: ; Discrete risk level mapping: ; The initial thresholds are set as follows: T1=0.2, T2=0.5, T3=0.8; Emotions are continuous and time-dependent states. A simplified linear Gaussian state-space model is used to track latent emotional states. With observation : ； in For known external interventions, , For process and observation noise, Kalman filtering can provide the optimal online estimate. And based on the prediction error, the intervention strategy is adaptively adjusted when the predicted sentiment trend deteriorates and The intervention intensity is automatically increased when the threshold is approached. To mitigate the impact of cross-individual variability, an individualized baseline and an online adaptive mechanism are introduced into the model. The baseline is estimated using exponential smoothing. ； in It represents the original feature value at a certain moment, and α∈(0,1) controls the adaptation rate. At the model training level, it is first pre-trained with large-scale general data, and then transferred learning is achieved by using a small amount of labeled data or unlabeled self-supervised fine-tuning for new users, thereby improving the accuracy of individual recognition and reducing misjudgments during the cold start period. For the "Intimate Relationship Synchronization" module, the coherence of signals from two or more individuals is used as a quantitative indicator, specifically frequency domain coherence. ； in For mutual spectrum, , For each power spectrum, select the emotion-related frequency band to calculate the intra-band coherence, and combine the confidence interval to judge significant synchronization enhancement. If the synchronization of the two intimate parties decreases significantly in the short term or one party shows high risk, this system can trigger communication suggestions and joint practice intervention prompts at the relationship level. Intimate relationship support alerts are determined by three parts: ontological risk, intimate partner risk, and relevance indicators, with ontological risk being the most significant factor. The risk of intimate partners is obtained Synchronization metrics: ； Where [F1, F2] are the selected emotion-related frequency bands mentioned above. If one party has a high risk and the synchronicity decreases or the asymmetry increases, it indicates that intervention is needed at the relationship level. Asymmetry is defined as follows: ； Intimate support trigger logic: ； Where the threshold Defined as 0.6, 0.3, and 0.25 respectively, when the trigger is true, the system executes the following sequence of actions: pushes instant companionship suggestions, arranges synchronous exercises, reminds the close partner to pay attention to and participate in comforting tasks, or suggests referring to clinical consultation.

6. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 1, characterized in that: The principle and effect of the two-way emotional interaction output module are as follows: When the system detects a risk of emotional fluctuation, it will automatically select the best intervention plan based on the intensity of the emotion to achieve "personalized emotion regulation," including the following intervention methods: (1) Rhythm-guided audio breathing feedback: induces enhancement of alpha and theta waves, improving relaxation; (2) Breathing rhythm guidance algorithm: The headband plays rhythm prompts to synchronize the breathing rate with the ideal HRV resonance frequency of 0.1Hz; (3) Neurofeedback training: EEG emotion indicators are converted into audiovisual feedback in real time, and users can return their emotions to a steady state through self-regulation; (4) CES: Improve anxiety and sleep quality through low-intensity transcranial micro-electric stimulation, and avoid overstimulation by synchronous monitoring with EEG.

7. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 1, characterized in that: The specific principle behind the intimate relationship synchronization and emotional interaction module is as follows: (1) Intimate Relationship EEG Synchronization Analysis: Used to detect the emotional synchronicity between the user and their partner or family member; (2) Support system: When a user experiences a high risk of emotional distress, the system will notify the partner or family support to provide comfort and companionship suggestions, thereby achieving two-way emotional linkage between the user and the intimate relationship; (3) AI Emotion Coach: Based on the collected emotion data, it provides communication suggestions and emotional support strategies to the user's close relationships.

8. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 1, characterized in that: The ultralight flexible headband structure uses lightweight flexible materials to construct the headband body and combines it with adaptively adjustable flexible electrodes to maintain a stable skin contact state during EEG acquisition during long-term wear or overnight sleep. The real-time monitoring of electrode contact impedance and automatic gain adjustment function reduce signal loss caused by position changes, thus improving the continuous efficiency of EEG data.

9. The bidirectional emotional interaction headband system based on a multimodal brain-computer interface according to claim 6, characterized in that: The CES uses low-amplitude electrical stimulation, which can gently regulate neurotransmitter activity in the brain.

Images