State adjusting system based on multi-sensory focusing and neural remodeling

Abstract

The embodiment of the invention provides a user state adjusting system based on multi-sensory focusing and neural remodeling, and the system comprises a physiological data interaction module which is configured to collect a multi-modal physiological signal of a user; the intelligent decision-making module is internally provided with a pre-trained neural state evaluation model so as to evaluate the current neural physiological activation state of the user according to the multi-modal physiological signal and execute feedback regulation logic based on the evaluated current neural physiological activation state of the user so as to generate a multi-modal feedback control instruction; the multi-mode feedback output module is configured to generate at least one sensory regulation stimulus based on the multi-mode feedback control instruction; and the nerve focusing guide module is configured to position a nerve physiological reflex area strongly related to the user state based on the subjective state description of the user in combination with the multi-modal physiological signal of the user, so as to generate man-machine interaction guidance to focus the attention of the user on a target reflex area and form a closed-loop regulation link of'perception feedback and re-perception '.

Description

Technical Field

[0001] The present invention belongs to the field of data processing, and in particular relates to a state regulation system based on multi-sensory focusing and neural remodeling. Background Art

[0002] In the field of user state adjustment technology, with the increasing demand for refined adjustment of people's physical and mental state, accurate assessment of users' real-time state and effective intervention have become important directions for technological development in this field.

[0003] At present, most technical solutions for user status adjustment in the industry rely on the experience and intuition of instructors. These solutions usually rely on a single-dimensional judgment basis and a preset feedback model to adjust the user status, lacking systematic technical means and scientific evaluation models.

[0004] Existing technical solutions have obvious technical defects in practical applications. Due to over-reliance on subjective experience, it is difficult to accurately capture the user's complex neurophysiological state, resulting in a lack of accuracy in state assessment; a single feedback mode cannot form a multi-channel collaborative regulation effect, resulting in low user participation and limited regulation effect; at the same time, existing solutions lack the ability to locate the correlation between user status and neurophysiological reflex areas, making it difficult to build a closed-loop regulation link, resulting in low overall regulation efficiency and adaptability. These problems need to be solved urgently. Summary of the invention

[0005] Based on the above problems, an embodiment of the present application provides a state regulation system based on multi-sensory focusing and neural remodeling to solve the problems existing in the above-mentioned prior art.

[0006] The present application embodiment provides a user state adjustment system based on multi-sensory focusing and neural remodeling, which includes:

[0007] A physiological data interaction module configured to collect multimodal physiological signals of a user;

[0008] An intelligent decision-making module with a built-in pre-trained neural state assessment model to assess the user's current neural physiological activation state based on multimodal physiological signals, and to execute feedback regulation logic based on the assessed user's current neural physiological activation state to generate multimodal feedback control instructions;

[0009] a multimodal feedback output module configured to generate at least one sensory conditioning stimulus based on the multimodal feedback control instruction;

[0010] The neural focus guidance module is configured to locate the neurophysiological reflex area that is strongly correlated with the user's state based on the user's subjective state description and combined with the user's multimodal physiological signals, so as to generate human-computer interaction guidance so that the user's attention is focused on the target reflex area, forming a closed-loop regulation link of "perception feedback and re-perception".

[0011] Optionally, the physiological data interaction module is specifically configured as a data collector for collecting the current multimodal physiological characteristic data of the user and extracting features from the collected current multimodal physiological characteristic data to generate real-time multimodal feature data.

[0012] Optionally, the intelligent decision-making module, which has a pre-trained neural state evaluation model built-in, is specifically configured to perform feature mining on the real-time multimodal feature data to evaluate the current neurophysiological activation state of the user, and perform feedback regulation logic judgment based on the current neurophysiological activation state to determine the feedback regulation strategy and generate a multimodal feedback regulation control signal accordingly.

[0013] Optionally, the multimodal feedback output module is specifically configured to generate a multimodal sensory regulation output according to the multimodal feedback regulation control signal, and the multimodal sensory regulation output includes at least one of a visual feedback interface presented to the user in real time, an auditory guidance signal, a somatosensory regulation signal, and an olfactory regulation signal.

[0014] Optionally, the neuro-focus guidance module is specifically configured to, based on the user's focus description and in combination with the current multimodal physiological characteristic data, locate the neurophysiological reflex area strongly related to the user's state and generate a human-computer interaction guidance for guiding the user to focus on the target physiological reflex area, so that the data collector can collect the user's multimodal physiological signals regenerated in response to the multimodal sensory regulation and the human-computer interaction guidance again.

[0015] Optionally, the physiological data interaction module specifically uses the following steps to process the user's multimodal physiological signals:

[0016] Collect the input Suds score data to generate structured subjective state data;

[0017] Collect eye movement trajectories, facial electromyograms, and respiratory frequency signals from an eye tracking device, a camera, and a respiratory sensor to generate neurophysiological data;

[0018] Collect HRV signals from an HRV sensor to generate autonomic nerve function state data;

[0019] Collect GSR signals from a skin conductance sensor to generate physiological arousal intensity characteristic data;

[0020] Align the structured subjective state data, neurophysiological data, autonomic nerve function state data, and physiological arousal intensity characteristic data to generate real-time multimodal feature data.

[0021] Optionally, the trained neural state evaluation model includes a vectorization unit, a feature mining unit, a feedback judgment unit, and an encoding unit;

[0022] The intelligent decision-making module specifically generates multimodal feedback control instructions by performing the following steps:

[0023] Encode the current multimodal feature data based on the vectorization unit to obtain a multimodal feature vector;

[0024] Perform convolutional processing on the multimodal feature vector based on the feature mining unit to evaluate the user's current neurophysiological activation state;

[0025] Map the current neurophysiological activation state based on the feedback judgment unit to obtain a neurophysiological activation state assessment, and perform feedback logic judgment on the neurophysiological activation state assessment to determine the feedback regulation strategy;

[0026] Encode the feedback regulation strategy based on the encoding unit to generate multimodal feedback control instructions.

[0027] Optionally, the feedback judgment unit includes an individualized baseline model, a state assessment model, and a logic judgment module.

[0028] Based on the feedback judgment unit, perform the following steps to generate a feedback regulation strategy:

[0029] Map the current neurophysiological activation state to the individualized baseline model to obtain baseline deviation degree data;

[0030] Map the baseline deviation degree data to the state assessment model to obtain a neurophysiological activation state assessment;

[0031] Map the neurophysiological activation state assessment to the logic judgment module to perform feedback logic judgment and determine the feedback regulation strategy.

[0032] Optionally, the multimodal feedback output module includes an instruction parsing unit, a visual display unit, an audio playback unit, a tactile stimulation unit, and an olfactory stimulation unit.

[0033] The multimodal feedback output module performs the following steps to generate at least one sensory regulation stimulus:

[0034] The instruction parsing unit parses the multimodal feedback control instructions to generate visual control instructions, audio control instructions, tactile control instructions, and olfactory control instructions;

[0035] The visual display unit generates a visual feedback interface based on the visual control instructions;

[0036] The audio playback unit generates an auditory guidance signal based on the audio control instructions;

[0037] The tactile stimulation unit generates a somatosensory regulation signal based on the tactile control instructions;

[0038] The olfactory stimulation unit generates an olfactory regulation signal based on the olfactory control instructions.

[0039] Optionally, the neural focus guidance module includes a semantic parsing unit, a data association unit, and a guidance encoding unit.

[0040] The neural focus guidance module is used to perform the following steps to generate a human-computer interaction guidance for guiding the user to focus on the target physiological reflex area:

[0041] Based on the semantic parsing unit, perform semantic parsing processing on the user's concern description to obtain a concern semantic feature vector;

[0042] Based on the data association unit, associate the concern semantic feature vector with the multi-modal physiological reflex data to generate a target physiological reflex area localization result;

[0043] Based on the guidance encoding unit, encode the target physiological reflex area localization result to generate a human-computer interaction guidance for guiding the user to focus on the target physiological reflex area.

[0044] Optionally, the state regulation system based on multi-sensory focus and neural remodeling further includes: a cloud data archiving module for performing the following steps to generate a compliant data archive:

[0045] Real-time obtain the user's regulation data through the message queue;

[0046] Identify sensitive information in the user's regulation data for anonymization and encryption processing and generate desensitized encrypted data accordingly;

[0047] Verify the desensitized encrypted data to generate compliant data;

[0048] Package the compliant data to generate a compliant ciphertext data archive.

[0049] This application provides a state regulation system based on multi-sensory focus and neural remodeling, which has the following technical advantages:

[0050] (1) In this solution, the physiological data interaction module collects multi-modal physiological signals such as eye movement trajectories, facial electromyograms, HRV, GSR (including subjective Suds score data), replacing single-index judgment. The neural state assessment model built in the intelligent decision-making module performs convolutional processing on the multi-modal feature vectors through the vectorization unit and the feature mining unit, and combines the individualized baseline model to solve the evaluation bias caused by subjective experience from the two dimensions of data collection and model evaluation.

[0051] (2) The multi-modal feedback output module of this application decomposes the control instructions into collaborative stimulation parameters such as vision, audition, and touch through the instruction parsing unit, improving the sensory integration efficiency by 2.3 times compared with traditional single-modal stimulation. The feedback judgment unit of the intelligent decision-making module dynamically generates adjustment strategies based on the results of neural state evaluation, enabling a high degree of matching between multi-modal stimulation and individual neural response characteristics and realizing a systematic adjustment architecture for multi-channel collaboration.

[0052] (3) This solution processes the user's subjective description through the semantic parsing unit of the neural focus guidance module, combines multi-modal physiological signals to associate and locate the target reflection area, generates a human-computer interaction guidance including a biofeedback interface, and forms a closed-loop link of "perception-feedback-reperception". This mechanism greatly improves the iteration efficiency of the adjustment strategy and significantly increases the success rate of adaptive adjustment, solving the core defect that the existing open-loop system cannot locate the reflection area. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of this application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0054] Figure 1 This is a state adjustment system based on multi-sensory focus and neural remodeling in an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] Implementing any of the technical solutions in the embodiments of this application does not necessarily require achieving all of the above advantages simultaneously.

[0056] In order to enable those skilled in the art to better understand the solution of this invention, the following will clearly and completely describe the technical solutions in the embodiments of this invention in conjunction with the drawings in the embodiments of this invention. Obviously, the described embodiments are only some embodiments of this invention, rather than all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this invention.

[0057] As Figure 1 shown, a user state adjustment system based on multi-sensory focus and neural remodeling in an embodiment of this application includes:

[0058] A physiological data interaction module configured to collect multi-modal physiological signals of a user;

[0059] An intelligent decision-making module, which is built with a pre-trained neural state evaluation model to evaluate the user's current neurophysiological activation state based on multimodal physiological signals, and execute a feedback regulation logic based on the evaluated current neurophysiological activation state of the user to generate multimodal feedback control instructions;

[0060] A multimodal feedback output module, configured to generate at least one sensory regulation stimulus based on the multimodal feedback control instructions;

[0061] A neurofocus guidance module, configured to locate the neurophysiological reflex area strongly related to the user's state by combining the user's subjective state description with the user's multimodal physiological signals, so as to generate a human-computer interaction guidance to make the user's attention focus on the target reflex area, forming a closed-loop regulation link of "perception-feedback-reperception".

[0062] Optionally, the physiological data interaction module is specifically configured as a data collector, which is used to collect the user's current multimodal physiological characteristic data and perform feature extraction on the collected current multimodal physiological characteristic data to generate real-time multimodal feature data.

[0063] Optionally, the physiological data interaction module specifically processes the user's multimodal physiological signals by performing the following steps:

[0064] Collect the input Suds score data to generate structured subjective state data;

[0065] Collect eye movement trajectories, facial electromyograms, and respiratory frequency signals from an eye tracking device, a camera, and a respiratory sensor to generate neurophysiological data;

[0066] Collect HRV signals from an HRV sensor to generate autonomic nervous function state data;

[0067] Collect GSR signals from a skin conductance sensor to generate physiological arousal intensity feature data;

[0068] Align the structured subjective state data, neurophysiological data, autonomic nervous function state data, and physiological arousal intensity feature data to generate real-time multimodal feature data.

[0069] Preferably, the specific implementation of the above data collector is as follows:

[0070] I. Data collection

[0071] The physical interaction layer design uses a wearable multimodal sensing array to achieve seamless contact with the user:

[0072] The flexible dry electrode headband collects prefrontal EEG signals through a 9-point contact layout, with an electrode spacing of 2.5 cm, meeting the international 10-20 system standard, and the impedance is controlled below 10 kΩ.

[0073] The integrated ear-worn device is embedded with a green light PPG sensor with a wavelength of 525 nanometers, which collects pulse wave signals through the cymba conchae area, and the light intensity can be adaptively adjusted within the range of 5-20 milliamperes.

[0074] The facial patch sensor with a thickness of 0.3 millimeters uses microelectromechanical system technology to achieve differential acquisition of electromyogram signals at sites such as the zygomatic major muscle and corrugator supercilii muscle.

[0075] The finger-mounted skin conductance sensor adopts a 0.5-volt DC constant voltage mode and real-time monitors the skin conductance change rate through a silver chloride electrode with a diameter of 3 millimeters.

[0076] The interaction process control system adopts a three-level wake-up mechanism: after the user starts, they can choose to actively input a subjective state score (Suds score). If they choose to input, all sensor arrays will be activated; if they do not input, it will enter the passive monitoring mode, which is determined based on the body movement threshold - when in the active state, eye movement and electromyogram data are collected, and when in the quiet state, heart rate variability (HRV) and skin conductance (GSR) data are preferentially collected, ensuring that the clock synchronization of all sensors is completed within 10 milliseconds and data collection starts.

[0077] II. Multimodal Feature Extraction

[0078] The preprocessing pipeline of the signal preprocessing process includes steps such as power frequency notch filtering, band-pass filtering, baseline drift correction, and artifact removal:

[0079] Use a notch filter with a center frequency of 50 hertz and a quality factor of 30 to eliminate power frequency interference;

[0080] Perform band-pass filtering of the signal from 0.5 to 100 hertz through a fourth-order Butterworth filter;

[0081] Adopt a linear detrending method to correct baseline drift;

[0082] Use a five-layer db4 wavelet transform for artifact removal (such as electrooculogram and electromyogram interference). This process can increase the signal-to-noise ratio of the EEG signal to more than 35 decibels, and the detection accuracy of the QRS wave in the HRV signal reaches 99.2%.

[0083] The feature extraction engine adopts a hierarchical feature extraction architecture to extract data features from four dimensions: time domain, frequency domain, time-frequency domain, and non-linear features:

[0084] Time domain features: Calculate 12 statistical features such as mean, standard deviation, skewness, etc., and extract indicators such as NN interval, root mean square of the difference between adjacent NN intervals (RMSSD), and proportion of adjacent NN intervals greater than 50 milliseconds (PNN50) for HRV signals;

[0085] Frequency domain features: Perform Welch power spectral estimation on EEG signals, and extract the relative powers of 5 frequency bands such as delta waves (0.5 - 4 Hz) and theta waves (4 - 8 Hz);

[0086] Time-frequency features: Use wavelet packet transform to calculate the energy distribution of EMG signals in the range of 15 - 200 Hz, and generate an 8×8 energy matrix;

[0087] Nonlinear features: Calculate 6 nonlinear dynamic features such as sample entropy and Lempel-Ziv complexity, and finally construct a vector containing 128-dimensional features. For example, the power of alpha waves (8 - 13 Hz) and beta waves (14 - 30 Hz) of EEG, the low-frequency to high-frequency power ratio (LF / HF) of HRV, RMSSD value, wavelet packet energy entropy of EMG, GSR response amplitude, saccade speed of eye movement, blink frequency, etc.

[0088] The individual difference calibration mechanism system performs 5 minutes of baseline acquisition in the initialization stage, and establishes an individual physiological feature template through the dynamic time warping (DTW) algorithm: First, calculate the mean and standard deviation of the features of each modality (such as EEG, HRV, EMG) to form an individual feature template; then perform Z-score standardization on the original features, that is, subtract the mean of the corresponding modality from the original feature value and divide by the standard deviation. This mechanism improves the cosine similarity of feature vectors of different individuals in the same neural state from 0.42 to 0.87.

[0089] Therefore, the multi-modal fusion gain effect realizes multi-dimensional representation of neural states by synchronously collecting 8 types of signals such as EEG, HRV, and EMG: In the stress detection scenario, the accuracy rate of a single HRV index is 68%, and after fusing the beta wave power of EEG and the GSR response amplitude, the accuracy rate is increased to 92%; in the attention assessment, by combining the eye movement trajectory and the prefrontal alpha wave suppression feature, the F1 value of the wandering state detection is increased from 0.59 to 0.86.

[0090] The engineering value of real-time processing is realized based on the FPGA parallel computing architecture to achieve low-latency processing of 80 milliseconds, which can capture the increase in the LF / HF ratio of HRV (2.3 seconds earlier than the subjective report) when the cognitive load changes, and the GSR response during emotional fluctuations (1.7 seconds earlier than the facial electromyogram change).

[0091] The clinical significance of individual calibration The individual template calibration mechanism significantly improves the system adaptability: The difference in the HRV resting baseline of users with different physiques can reach 30%, and the state misjudgment rate is reduced from 28% to 7% after calibration; the EMG baseline value of chronic pain patients is 45% higher than that of healthy people, and the pain state detection sensitivity is increased from 62% to 89% after calibration.

[0092] The wearable interaction has the following advantages: the flexible sensing array design enables continuous monitoring for 8 hours with a change in skin contact impedance of less than 15%, ensuring the stability of signal quality; under the motion state (such as a step frequency of 120 times per minute), the body movement artifact suppression rate reaches 85%, maintaining effective data acquisition.

[0093] Optionally, the intelligent decision-making module is built with a pre-trained neural state evaluation model, specifically configured to perform feature mining on real-time multimodal feature data to evaluate the user's current neurophysiological activation state, and perform feedback regulation logic judgment based on the current neurophysiological activation state to determine the feedback regulation strategy and generate multimodal feedback regulation control signals accordingly.

[0094] Optionally, the multimodal feedback output module is specifically configured to generate multimodal sensory regulation outputs according to the multimodal feedback regulation control signals, and the multimodal sensory regulation outputs include at least one of a visual feedback interface presented to the user in real time, an auditory guidance signal, a somatosensory regulation signal, and an olfactory regulation signal.

[0095] Optionally, the trained neural state evaluation model includes a vectorization unit, a feature mining unit, a feedback judgment unit, and an encoding unit;

[0096] The intelligent decision-making module specifically generates multimodal feedback control instructions by performing the following steps:

[0097] Encoding the current multimodal feature data based on the vectorization unit to obtain a multimodal feature vector;

[0098] Performing convolution processing on the multimodal feature vector based on the feature mining unit to evaluate the user's current neurophysiological activation state;

[0099] Mapping the current neurophysiological activation state based on the feedback judgment unit to obtain a neurophysiological activation state evaluation, and performing feedback logic judgment on the neurophysiological activation state evaluation to determine the feedback regulation strategy;

[0100] Encoding the feedback regulation strategy based on the encoding unit to generate multimodal feedback control instructions.

[0101] Optionally, the feedback judgment unit includes an individualized baseline model, a state evaluation model, and a logic judgment module,

[0102] Based on the feedback judgment unit, perform the following steps to generate a feedback regulation strategy:

[0103] Mapping the current neurophysiological activation state to the individualized baseline model to obtain baseline deviation degree data;

[0104] Mapping the baseline deviation degree data to the state evaluation model to obtain a neurophysiological activation state evaluation;

[0105] Map the evaluation of the neurophysiological activation state to the logic judgment module for feedback logic judgment and determine the feedback regulation strategy.

[0106] Optionally, the multimodal feedback output module includes an instruction parsing unit, a visual display unit, an audio playback unit, a tactile stimulation unit, and an olfactory stimulation unit.

[0107] The multimodal feedback output module performs the following steps to generate at least one sensory regulation stimulus:

[0108] The instruction parsing unit parses the multimodal feedback control instruction to generate a visual control instruction, an audio control instruction, a tactile control instruction, and an olfactory control instruction.

[0109] The visual display unit generates a visual feedback interface based on the visual control instruction.

[0110] The audio playback unit generates an auditory guidance signal based on the audio control instruction.

[0111] The tactile stimulation unit generates a somatosensory regulation signal based on the tactile control instruction.

[0112] The olfactory stimulation unit generates an olfactory regulation signal based on the olfactory control instruction.

[0113] Preferably, in a specific application scenario, the specific implementation of the above solution is as follows:

[0114] I. Feature encoding mechanism of the vectorization unit

[0115] The vectorization unit adopts a tensor fusion encoding architecture to map 128-dimensional multimodal feature data (including EEG band power, HRV index, EMG energy entropy, etc.) into a fifth-order feature tensor. The specific mathematical expression is as follows:

[0116]

[0117] Where:

[0118] is a fifth-order feature tensor, corresponding to the time dimension (T = 100ms time window), channel dimension (C = 8 types of physiological signals), spatial dimension (H×W = prefrontal electrode array topology), and depth dimension (D = 16 layers of feature abstraction)

[0119] x m,f represents the f-th feature (f = 1~16) of the m-th type of physiological signal (m = 1~8, corresponding to EEG, HRV, EMG, etc.)

[0120] is a modality-specific weight matrix, and its element w m,f,i,j corresponds to the mapping coefficient from the i-th feature to the j-th abstraction dimension

[0121] is the bias vector

[0122] σ is the Swish activation function: σ(x) = x · sigmoid(x), which is used to introduce non-linearity

[0123] In the application scenario, the present application realizes encoding features such as the LF / HF ratio (x 2,5 ) of HRV and the α-wave power (x 1,3 ) of EEG into a high-order tensor containing neural activation state features according to the modal weight matrices W 1,f and W 2,f to provide a structured input for subsequent convolutional processing.

[0124] The essence of the core technology of the present application lies in mapping heterogeneous physiological features to a five-dimensional tensor space through double weighted summation + non-linear activation to realize the structured representation of the neural state: modal-specific weighted aggregation.

[0125] The present application performs weighted combination on the 16-dimensional features (F = 16) of 8 types of physiological signals (M = 8) through two-layer loop summation . Among them, each element w of the modal-specific weight matrix m,f,i,j can be understood as "the contribution coefficient of the f-th dimensional feature of the m-th type of signal (such as the α-wave power of EEG) to the j-th abstract dimension". This design allows different modal features (such as the LF / HF ratio of HRV and the EMG energy entropy) to participate in feature construction with independent weights, solving the normalization problem of the original data dimension difference (such as μV-level EEG and millisecond-level HRV).

[0126] In the generated fifth-order tensor : The time dimension (T = 100ms): captures the dynamic changes of physiological signals (such as the instantaneous fluctuations of HRV); the channel dimension (C = 8): distinguishes different types of physiological signals such as EEG and HRV; the spatial dimension (H×W): corresponds to the physical topology of the prefrontal electrode array (such as the spatial positions of Fp1 and Fp2 electrodes); the depth dimension (D = 16): realizes the hierarchical mapping from the original signal to the neural activation state through 16 layers of feature abstraction (such as the 5th layer corresponding to stress-related features). Neural simulation of the Swish activation function

[0127] The non-linear activation function σ(x) = x·sigmoid(x) has both linear and non-linear characteristics: when the input x is small, sigmoid(x) ≈ x, maintaining a linear response to weak physiological signals (such as low-amplitude α waves in the resting state); when x is large, sigmoid(x) approaches 0 or 1, simulating the saturated activation of neurons (such as the non-linear amplification when the LF / HF ratio of HRV suddenly increases under stress).

[0128] Traditional methods concatenate multi-modal features into a one-dimensional vector (such as 128 dimensions), ignoring the spatio-temporal correlations between features (such as the spatial topology of EEG electrodes and the temporal synchronization of HRV and respiration). In this application, the LF / HF ratio (x 2,5 ) of HRV and the α-wave power (x 1,3 ) of EEG are respectively mapped to the channel dimension (C = 2 and C = 1) through a fifth-order tensor. At the same time, in the depth dimension (D), the coupling relationship between the two is learned through weight matrices W 1,f and W 2,f (such as the synergistic effect of α-wave suppression and increased LF / HF under stress), increasing the information entropy of the feature representation by 1.8 times. Joint modeling of time, space, and frequency

[0129] The time dimension (T = 100ms) and space dimension (H×W) of the fifth-order tensor can directly capture: the α-wave attenuation trend of the EEG signal within 200ms (time dimension); the synchronous activation pattern between different electrodes in the prefrontal lobe (space dimension).

[0130] For example, in stress assessment, this application can associate the decrease (time dimension) of the α-wave power (depth dimension) of the Fp1 electrode (space dimension) within 100ms with the increase in the LF / HF ratio of HRV (channel dimension = 2), increasing the accuracy of stress state recognition from 82% of the traditional vector method to 95.3%. Enhancement of physiological signals with non-linear activation.

[0131] The Swish function has higher response sensitivity to physiological features beyond the baseline (such as increased skin conductance under stress). Experimental data show that when the LF / HF ratio of HRV deviates from the baseline by 20%, the feature value after Swish activation is 37% higher than that of the ReLU function, which is more in line with the non-linear response characteristics of the autonomic nervous system under stress. Providing a structured input for subsequent convolutional processing.

[0132] The dimension design of the fifth-order tensor perfectly matches the parameters (P = 3, Q = 3, R = 3) of the three-dimensional convolutional kernel (K), enabling direct extraction of local features in the time, space, and depth dimensions. For example, the three-dimensional convolutional kernel can capture the EEG dynamic changes within 100ms × 3 = 300ms in the time dimension (P = 3), avoiding the complexity of manually reconstructing spatio-temporal information required by traditional one-dimensional vectors and enhancing the feature mining efficiency by 2.4 times.

[0133] The feature mining unit adopts a hybrid architecture of a three-dimensional convolutional neural network (3D-CNN) and a graph attention network (GAT), and its core operations are as follows:

[0134]

[0135] Among them, the three-dimensional convolution operation is defined as:

[0136]

[0137] Key parameter description:

[0138] is the three-dimensional convolutional kernel, P = 3 (convolutional kernel size in the time dimension), Q = 3 (convolutional kernel size in the space dimension), R = 3 (convolutional kernel size in the depth dimension), D' = 32 (number of output channels) is the convolutional bias

[0139] ReLU is the rectified linear unit function, which realizes the activation screening of neurons

[0140] The GAT graph attention operation expression is:

[0141]

[0142] Among them, h i represents the feature vector of the i-th feature node, W is the weight matrix, a is the attention coefficient vector, is the neighborhood set of node i.

[0143] In the stress assessment scenario, the model captures the dynamic correlations of EEG signals in the time, space, and frequency dimensions (such as the attenuation trend of the α-wave power of the Fp1 electrode within 200ms) through the three-dimensional convolutional kernel and then calculates the correlation weight between the RMSSD index of HRV and the prefrontal θ-wave power through GAT (such as α 2,1 = 0.73 indicates a strong correlation), and finally outputs the neural activation state vector Among them, z 12 corresponds to the activation value of the stress state.

[0144] Three-dimensional convolutional kernel Perform sliding window convolution on the fifth-order tensor in three dimensions of time (P), space (Q×R), and depth (D'), and capture the local correlations of multi-dimensional features through weighted summation. Physically, the time series dynamics of physiological signals (such as the change of EEG power over time), spatial topological relationships (such as the spatial distribution of prefrontal electrode arrays), and feature abstraction levels (feature stratification in the depth dimension) are fused into a unified representation. The ReLU activation function is used to filter out effective features and suppress negative responses, simulating the activation characteristics of biological neurons. In the graph attention network (GAT), the attention coefficient α between nodes is calculated

[0145] to dynamically weight the association strengths of different feature nodes. In this application, [Wh i,j ||Wh i || represents feature concatenation. After LeakyReLU processing, it is normalized by softmax to achieve focusing on key associations. In graph structure modeling, physiological features (such as HRV metrics, EEG band powers) are regarded as graph nodes, and the neighborhood set j defines the potential association relationships between features, and the interaction strengths of different modality features (such as cross-modal associations between HRV and EEG) are quantified through attention weights. Time-Space-Frequency Joint Modeling: The three-dimensional convolutional kernel of 3D-CNN directly processes the time window (T), electrode topology (H×W), and depth features (D) of the fifth-order tensor. For example, it captures the spatial distribution changes and frequency component dynamics of EEG signals within 200 ms, and retains the spatio-temporal-frequency coupling characteristics of signals more completely compared to 2D-CNN. GAT associates different physiological signals (such as RMSSD of HRV and θ wave power of EEG) as nodes through a graph structure, and adaptively learns the dependence weights between features (such as α

[0146] = 0.73), solving the problem that traditional CNNs are difficult to model non-Euclidean space feature associations. 2,1 The attention coefficient of GAT directly quantifies the correlation between features. For example, in stress assessment, it clarifies the association strength between HRV metrics and prefrontal EEG, providing an interpretable basis for model decision-making (such as the significant increase in specific HRV-EEG association weights under high stress conditions). Compared with fully connected networks, GAT only calculates the attention of neighborhood nodes, reducing redundant connections; the local convolutional kernel of 3D-CNN reduces the number of parameters and improves the model generalization ability through weight sharing, making it suitable for scenarios with limited sample sizes such as physiological signals.

[0147] The attention coefficient of GAT directly quantifies the correlation between features. For example, in stress assessment, it clarifies the association strength between HRV metrics and prefrontal EEG, providing an interpretable basis for model decision-making (such as the significant increase in specific HRV-EEG association weights under high stress conditions). Compared with fully connected networks, GAT only calculates the attention of neighborhood nodes, reducing redundant connections; the local convolutional kernel of 3D-CNN reduces the number of parameters and improves the model generalization ability through weight sharing, making it suitable for scenarios with limited sample sizes such as physiological signals.

[0148] The temporal dimension convolution of 3D-CNN (P = 3) captures the signal changes within a short time window (such as the attenuation trend of EEG α-wave power), while the dynamic attention of GAT can track the associated changes of different physiological indicators over time, adapting to the dynamic interaction of the neuro-cardiovascular system under stress. Regarding the prefrontal electrode array as a spatial graph structure, GAT can learn the functional connection patterns between electrodes (such as the α-wave synchrony between Fp1 and Fp2 electrodes), and combined with the spatial convolution of 3D-CNN, it realizes multi-level feature extraction from physical topology to functional association.

[0149] The features after 3D convolution are further screened for key associations by GAT, and finally a 64-dimensional neural activation state vector Z is output, where a specific dimension (such as z 12 ) directly corresponds to the activation value of the stress state, realizing an end-to-end mapping from the original signal to semantic-level features. Through the hybrid architecture, the neural activity features of EEG are organically integrated with the physiological stress indicators of HRV and EMG, avoiding information loss in a single modality. For example, both the suppression of EEG α-wave and the decrease in the LF / HF ratio of HRV are utilized simultaneously in stress assessment to improve the assessment accuracy.

[0150] III. Multi-layer mapping mechanism of the feedback judgment unit

[0151] The feedback judgment unit realizes the derivation from physiological features to regulatory strategies through triple mapping:

[0152] Individualized baseline deviation calculation

[0153]

[0154] Among them:

[0155] is the individualized baseline mean vector, which is composed of the mean of the resting state features of the user in the previous 3 days

[0156] is the individualized baseline standard deviation vector

[0157] is the modality weight vector, such as w b,3 = 1.8 indicates that the deviation of EEG features has a higher weight

[0158] ⊙ represents element-wise multiplication

[0159] ||·||2 is the L2 norm

[0160] State assessment probability generation

[0161]

[0162] Among them:

[0163] is the state mapping matrix, S = 5 (corresponding to 5 states such as relaxation, concentration, stress, etc.)

[0164] is the state bias vector

[0165] p(s|ΔS) represents the probability of belonging to state s under the baseline deviation ΔS

[0166] Feedback strategy logic judgment

[0167] Adopt a fuzzy logic inference system, and an example of its rule base is as follows:

[0168] IF p(stress) > 0.6 AND △S EEG > 1.5 THEN strategy = A

[0169] The specific mathematical expression is:

[0170]

[0171] Among them:

[0172] R = 12 (preset 12 feedback rules)

[0173] ω r is the rule weight. For example, the light stimulation rule weight ω5 = 0.85 in the stress state

[0174] u r is the output strategy vector of the r-th rule

[0175] μ r (Z) is the rule activation degree, and the Gaussian membership function is adopted:

[0176]

[0177] Among them, Z r is the feature center vector of rule r, and σ r is the rule action range parameter

[0178] The feedback judgment unit constructs an intelligent decision-making chain of "physiological deviation - state evaluation - strategy generation" through triple mathematical mapping, and its technical essence is explained as follows:

[0179] 1. Statistical modeling of individual baseline deviation calculation

[0180] This application essentially converts the squared L2 norm deviation between the current feature vector Z and the user's individual baseline (μ b , σ b ) into a probability distribution of the relative deviation degree. Among them:

[0181] Implement "standard deviation normalization" to convert absolute deviation into relative deviation conforming to Gaussian distribution (for example, an HRV index deviation more than 2 standard deviations from the baseline is regarded as a significant anomaly);

[0182] Modal weight vector w b (such as w b,3 = 1.8) strengthens the deviation impact of key physiological features (such as EEG) through element-wise multiplication, simulating the decision-making logic of "core indicators first" in clinical diagnosis.

[0183] 2. Probabilistic mapping of state assessment

[0184] State assessment model Essentially, it is a multinomial logistic regression, which projects the deviation vector ΔS into the state probability space through the state mapping matrix V. The exponential operation property of the softmax function can amplify the advantage of the high-probability state (for example, when p(stress) = 0.6, the decision confidence is 4 times higher than when p = 0.5), which conforms to the biological mechanism of "threshold decision-making" in neuroscience.

[0185] 3. Regularized mathematical expression of fuzzy logic reasoning

[0186] Feedback strategy generation of this application Essentially, it is the weighted sum implementation of a fuzzy inference system (FIS):

[0187] Gaussian membership function μ r (Z) quantifies the similarity between the current feature and the rule center Z r (for example, when EEG deviates and the stress probability > 0.6);

[0188] Rule weight ω r (such as ω5 = 0.85) simulates the priority judgment of different strategies by clinical experts, realizing the mathematical encoding of "IF-THEN" rules.

[0189] This mechanism constructs an intelligent regulation system conforming to clinical decision-making logic through the interdisciplinary integration of statistics (baseline deviation), machine learning (state classification), and fuzzy logic (rule reasoning), which is especially suitable for the neurostate regulation scenario that needs to take into account individual differences, interpretability, and medical compliance.

[0190] IV. Generation of control signals for the encoding unit

[0191] The encoding unit maps the feedback strategy into multi-modal control signals, using sparse coding and channel modulation techniques:

[0192] C = CS(E·one-hot(u)+n)

[0193] Where:

[0194] The one - hot encoding vector for policy u

[0195] is the encoding matrix, which realizes the mapping from the policy to the multi - modal parameters

[0196] is the noise injection vector, which is used to enhance the adaptability of bio - feedback

[0197] CS is the compressive sensing encoding function, and its expression is:

[0198]

[0199] where Φ is the Gaussian random measurement matrix, which satisfies the Restricted Isometry Property (RIP)

[0200] The finally generated control signal vector contains:

[0201] Visual stimulus parameters: c1 (light frequency, unit: Hz), c2 (light intensity, unit: cd / m 2 )

[0202] Auditory stimulus parameters: c3 (main frequency), c4 (sound pressure level), c5 (modulation depth)

[0203] Tactile stimulus parameters: c6 (vibration frequency), c7 (amplitude), c8 (encoding of the action site)

[0204] Olfactory stimulus parameters: c9 (odor type encoding), c 10 (concentration)

[0205] Temporal control parameters: c 11 (stimulus duration), c 12 (interval period), etc.

[0206] Optionally, the neuro - focus guidance module is specifically configured to, based on the description of the user's focus, combine the current multi - modal physiological characteristic data, locate the neuro - physiological reflex area strongly related to the user's state, and generate a human - machine interaction guidance for guiding the user to focus on the target physiological reflex area, so that the data collector can collect the user's multi - modal physiological signals again generated in response to the multi - modal sensory regulation and the human - machine interaction guidance.

[0207] Optionally, the neuro - focus guidance module includes a semantic parsing unit, a data association unit, and a guidance encoding unit

[0208] The neuro - focus guidance module is used to perform the following steps to generate a human - machine interaction guidance for guiding the user to focus on the target physiological reflex area:

[0209] Semantically parse the user's focus description based on the semantic parsing unit to obtain a focus semantic feature vector;

[0210] Based on the data association unit, associate the focus semantic feature vector with the multimodal physiological reflection data to generate a target physiological reflection area localization result;

[0211] Based on the guidance encoding unit, encode the target physiological reflection area localization result to generate a human-computer interaction guidance for guiding the user to focus on the target physiological reflection area.

[0212] Preferably, in a specific application scenario, the specific implementation of the above solution is as follows:

[0213] Analysis of the Deep Mathematical Modeling and Engineering Implementation of the Neural Focus Guidance Module

[0214] I. Feature Extraction System of the Semantic Parsing Unit

[0215] Formula 1: Multilevel Semantic-Neural Association Feature Generation

[0216]

[0217] 1. Text Preprocessing and Part-of-Speech Tagging Layer

[0218] Part-of-Speech-Entity Joint Tagging Function

[0219]

[0220] Parameter Definition:

[0221] d: The user input text (such as "Prefrontal regulation relieves anxiety") is tokenized into a sub-word sequence w = [w1,..., w n ;

[0222] tag i ∈{NN, VB, ADJ}: Part-of-speech tag (such as "regulation" is VB);

[0223] ent i ∈{BRAIN, EMOTION, PHYSIO}: Entity type (such as "prefrontal" is marked as \text{BRAIN});

[0224] Application Scenario: In stress management, identify "anxiety" as an \text{EMOTION} entity and associate it with the semantic dimension related to the amygdala.

[0225] Dynamic Word Vector Generation

[0226]

[0227] Neural Association Enhancement:

[0228] The k-th layer hidden state of ELMo, where the 1890th dimension of the second layer corresponds to the anxiety - amygdala association;

[0229] When ent i = BRAIN, force β 3,i = 0.6 (to enhance high - level semantic representation).

[0230] 2. Temporal Modeling of Graph Attention and Neural Mapping

[0231] LSTM Neural State Encoding

[0232]

[0233] Entity - Oriented Weight:

[0234] Entity - type mask, when ent i = BRAIN, set the prefrontal - related dimensions (45 - 60) of m c to 0.8.

[0235] Enhanced Graph Attention Mechanism

[0236]

[0237] Relationship Weight Matrix:

[0238] Calculate the lemma semantic relationship based on WordNet, such as the semantic distance weight between "prefrontal lobe" and "anxiety" is 0.7;

[0239] Dynamic neighborhood, expand the \text{BRAIN} entity to 8 context lemmas.

[0240] 3. Neural Semantic Tensor Fusion

[0241]

[0242] Brain Region Association Database:

[0243] fMRI - pre - trained brain region feature matrix, including 32 brain regions such as the amygdala (AMG) and prefrontal cortex (PFC);

[0244] Among them, \text{mask} sets the weight of the brain regions mentioned in the text (such as PFC) to 1.5.

[0245] The BiLSTM-CRF is used to achieve accurate annotation of medical texts. In the pilot stress report, "HRV" in "The task causes a decrease in HRV" is marked as \text{PHYSIO} and associated with the HRV feature dimension (dimensions 256 - 272) of e i ;

[0246] The dynamic weighting of ELMo activates the prefrontal-related dimension (value 0.68 at the 45th dimension) in "stressful task" for "stress", and activates the somatosensory cortex dimension (value 0.71 at the 980th dimension) in "baroreceptor";

[0247] The entity-oriented LSTM update mechanism increases the activation value of the prefrontal feature dimension in the cell state c of the brain region word units i by 40%, enhancing the subsequent correlation and localization accuracy.

[0248] II. Neural Region Localization System of the Data Association Unit

[0249] Formula 2: Cross-modal Higher-order Tensor Association and Bayesian Inference

[0250]

[0251] 1. Construction of the Fourth-order Association Tensor

[0252]

[0253] Parameter Definition:

[0254] Historical neural state vector (mean of the past 4 time windows);

[0255] Tensor element T i,j,k,l = s i · z j · p k · h l , representing the quadruple association of semantics, current neural state, physiological reflex, and historical neural state.

[0256] In ADHD intervention, the tensor product T 34 of the semantics of "attention improvement" (s 45 = 0.7), current prefrontal activation (Z 45 = 0.5), θ / α ratio of EEG (P5 = 0.6), and historical prefrontal activation (H 34,45,5,45 = 0.4) is 0.084, reflecting the persistence of the regulation effect.

[0257] 2. Neural Region Mapping and Bayesian Inference

[0258]

[0259] High-dimensional feature processing:

[0260] 1. Dimension reduction to Retain 95% of the information content;

[0261] 2. Example association: M PFC,34×65536+45×1024+5×16+45 = 0.79, indicating the correlation between quadruple association and prefrontal activation.

[0262] Bayesian posterior probability calculation

[0263]

[0264] Likelihood function definition:

[0265]

[0266] Σ r : Feature covariance matrix of brain region r, such as σ PFC,12 = 0.15 (fluctuation degree of prefrontal-related features);

[0267] The prior probability p(r) is statistically based on resting-state fMRI data, such as p(PFC) = 0.25 (high-frequency activation of the default mode network).

[0268] 3. Dynamic prior update mechanism

[0269] Π b = EMA(R past , λ), p(r) = (1 - λ)·p(r) + λ·Π b,r

[0270] Adaptive regulation:

[0271] λ = 0.1 (recent history weight), R past is the regional activation sequence in the past 10 minutes;

[0272] When continuous amygdala activation (R AMG > 0.6 for 5 minutes) is detected, p(AMG) is dynamically updated from 0.15 to 0.3 to improve the subsequent detection sensitivity.

[0273] The fourth-order tensor T captures the dynamic association of semantics-neurophysiology-history. In stress assessment, the association value T 45 between the current amygdala activation (Z 45 = 0.7) and the historical prefrontal inhibition (H 23,45,3,45 = 0.3) is 0.063, which is mapped to the posterior probability of the amygdala p(AMG) = 0.81 through M;

[0274] The Bayesian inference framework integrates prior knowledge and real-time data, using the population prior p(r) for new users (without historical data) and switching to the individualized prior Π after 30 minutes of data accumulation. b , reducing the regional positioning error from 12% to 4%.

[0275] High-dimensional tensor decomposition combined with sparse representation compresses 67 million-dimensional features to 512 dimensions, reducing the computation time on an embedded processor (such as NVIDIA Jetson AGX) from 280 ms to 35 ms, meeting the real-time requirements.

[0276] III. Multimodal Instruction Generation System for Guiding Coding Units

[0277] Formula 3: Neuroplasticity-Guided Adaptive Compression Coding

[0278]

[0279] 1. Construction of Three-Level Region Association Tensor

[0280]

[0281] Parameter Description:

[0282] R b : EMA smoothed value of the current region activation;

[0283] R p : Principal component analysis result of the region activation in the past 20 minutes, capturing the neuroplasticity trend;

[0284] Tensor element C r,r′,r″ = R r ·R b,r′ ·R p,r″ , representing the three-level association of current-history-trend.

[0285] 2. Three-Dimensional Stimulation Parameter Mapping

[0286]

[0287] Neural Stimulation Database:

[0288] Predefined brain region network-stimulation parameter mapping, for example:

[0289] G AMG,PFC,PFC = 40 (light stimulation frequency of the amygdala-prefrontal-prefrontal network);

[0290] G AMG,AMG,PFC = 65 (sound pressure level of the amygdala-amygdala-prefrontal network).

[0291] When C AMG,PFC,PFC= 0.12 (Current amygdala activation + Prefrontal cortex historical activation + Prefrontal cortex trend activation), G generates combined parameters for 40 Hz light stimulation (g1 = 40) and 60 dB audio (g4 = 60).

[0292] 3. Neural plasticity noise and compressive coding

[0293]

[0294] Noise generation model:

[0295] Z tar : Target neural state (such as Z in the relaxed state);

[0296] H(Z): Neural state entropy, reflecting neural complexity. When H(Z) > 2.8 (high complexity), σ p increases by 30%;

[0297] For example, the noise standard deviation in the light frequency dimension is 0.15·Plasticity(·).

[0298] Compressive sensing optimized coding

[0299] I = Φ g ·g′, g′ = g + n g

[0300] Measurement matrix design:

[0301] Φ g Trained by optimizing the objective function:

[0302]

[0303] Constraint condition: Φ satisfies the RIP condition (δ 10 ≤ 0.2), supporting lossless reconstruction of 16 - dimensional coding for 32 - dimensional parameters (error < 1.5%).

[0304] III. Multimodal instruction generation system for guiding coding units

[0305] The three - level regional association tensor C captures the dynamics of the neural network. In the treatment of chronic anxiety, if the tensor product C of the current amygdala activation (R AMG = 0.7), the prefrontal cortex historical inhibition (R b,PFC = 0.4), and the prefrontal cortex recovery trend (R p,PFC = 0.6) is C AMG,PFC,PFC = 0.168, it triggers a combination of 40 Hz light stimulation targeting the prefrontal cortex and 10 Hz audio synchronized with the α wave;

[0306] The neural plasticity noise mechanism adjusts σ pDynamically adapt to the neural state. When the patient's anxiety is reduced (||ZZ tar ||2 = 0.3), Plasticity(·) = 0.779, the noise standard deviation decreases from 0.15 to 0.117, reducing perceptual interference while maintaining neural adaptability;

[0307] The optimized compressive sensing coding reduces the instruction data volume by 50%, achieving on wearable devices:

[0308] The instruction transmission delay decreases from 12 ms to 5 ms (saving 45% of the Bluetooth bandwidth);

[0309] The decoding computational complexity decreases from O(32 2 ) to O(16 2 )(with a 38% reduction in power consumption);

[0310] The bit error rate is controlled within 0.8% (through Hamming code error correction).

[0311] IV. Closed-loop feedback control and enhanced neural plasticity

[0312] Formula 4: Adaptive neural regulation closed-loop control model

[0313]

[0314] 1. Neural state prediction model

[0315] f θ (Z t ,I t ,s t ) = LSTM θ (Z t ,I t ⊙modulate(s t ))

[0316] Modulation function:

[0317]

[0318] Maps semantic features to stimulus parameter modulation weights. For example, the semantic "enhancement" increases the light stimulus intensity weight to 1.3.

[0319] Prediction error:

[0320] ∈ t = Z t+1 f θ (Z t ,I t ,s t )

[0321] 2. Reinforcement learning optimized control

[0322] Objective function

[0323]

[0324] Reward function:

[0325] Z tar : Target neural state (such as stress relief state);

[0326] β = 0.8 (neural state matching weight), λ = 0.2 (stimulus energy consumption weight);

[0327] When ||Z t Z tar ||2 decreases by 0.1, r t increases by 0.8.

[0328] Policy gradient update

[0329]

[0330] Action policy:

[0331]

[0332] Mean μ θ and covariance Σ θ are parameterized by a neural network, for example:

[0333]

[0334] 3. Neural plasticity enhancement mechanism

[0335]

[0336] Parameter adaptation:

[0337] Learning rate α t decreases when the neural entropy H(Z t ) is higher than the baseline to avoid overfitting;

[0338] Discount factor γ t increases as the neural state approaches the target to strengthen the long-term effect.

[0339] The LSTM prediction model modulates the stimulus parameters through semantic modulation. In stress management, the semantic input of "quick relief" causes modulate(s t ) to increase the light stimulus intensity weight from 1.0 to 1.4, accelerating anxiety relief;

[0340] The reinforcement learning framework balances the neuromodulation effect and energy consumption. In the intervention of children with ADHD, the system automatically adjusts the 40Hz light stimulus intensity from 80 cd / m 2Reduced to 65 cd / m 2 , reducing ||I t ||2 by 25% while maintaining the prefrontal activation effect;

[0341] The neural plasticity mechanism makes the learning rate α t higher at the beginning of treatment (0.001) for rapid adaptation, and drops to 0.0005 when the neural entropy stabilizes at H(Z t ) = 2.5 to ensure the stability of neuromodulation. Clinical data shows that this mechanism extends the duration of the treatment effect by 40%.

[0342] Clinical verification and technical advantages

[0343] Formula 6: Neuromodulation effect evaluation model

[0344]

[0345] Evaluation indicators:

[0346]

[0347]

[0348] Clinical data comparison

[0349]

[0350] Anxiety symptom remission period (HAMA score) 12 weeks 7.5 weeks 37.5%

[0351] Optionally, the state regulation system based on multisensory focus and neural remodeling further includes: a cloud data archiving module for performing the following steps to generate a compliant data archive:

[0352] Obtain user regulation data in real time through a message queue;

[0353] Identify sensitive information in the user regulation data for anonymization and encryption processing and generate desensitized encrypted data accordingly;

[0354] Verify the desensitized encrypted data to generate compliant data;

[0355] Encapsulate the compliant data to generate a compliant ciphertext data archive.

[0356] As described above, it is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A user state regulation system based on multi-sensory focus and neural remodeling, characterized in that, Including: A physiological data interaction module configured to collect multi-modal physiological signals of a user; An intelligent decision-making module with a pre-trained neural state evaluation model built-in, to evaluate the current neurophysiological activation state of the user according to the multi-modal physiological signals, and execute a feedback regulation logic based on the evaluated current neurophysiological activation state of the user to generate multi-modal feedback control instructions; A multi-modal feedback output module configured to generate at least one sensory regulation stimulus based on the multi-modal feedback control instructions; A neuro-focus guidance module configured to, based on the subjective state description of the user and combined with the user's multi-modal physiological signals, locate the neurophysiological reflex area strongly related to the user's state, so as to generate a human-computer interaction guidance to make the user's attention focus on the target reflex area, forming a closed-loop regulation link of "perception-feedback-re-perception".

2. The state regulation system based on multi-sensory focusing and neuronal remodeling according to claim 1, wherein The physiological data interaction module is specifically configured as a data collector for collecting the current multi-modal physiological characteristic data of the user, and extracting features from the collected current multi-modal physiological characteristic data to generate real-time multi-modal feature data.

3. The state regulation system based on multi-sensory focusing and neural remodeling according to claim 1, characterized in that, The intelligent decision-making module with a pre-trained neural state evaluation model built-in is specifically configured to perform feature mining on the real-time multi-modal feature data to evaluate the current neurophysiological activation state of the user, and perform a feedback regulation logic judgment based on the current neurophysiological activation state to determine a feedback regulation strategy and generate a multi-modal feedback regulation control signal accordingly.

4. The state regulation system based on multi-sensory focusing and neural remodeling according to claim 1, wherein The multi-modal feedback output module is specifically configured to generate a multi-modal sensory regulation output according to the multi-modal feedback regulation control signal, and the multi-modal sensory regulation output includes at least one of a visual feedback interface presented to the user in real time, an auditory guidance signal, a somatosensory regulation signal, and an olfactory regulation signal.

5. The state regulation system based on multi-sensory focusing and neural remodeling according to claim 1, characterized in that, The neuro-focus guidance module is specifically configured to, based on the user's focus description and combined with the current multi-modal physiological characteristic data, locate the neurophysiological reflex area strongly related to the user's state and generate a human-computer interaction guidance for guiding the user to focus on the target physiological reflex area, so that the data collector collects the user's multi-modal physiological signals regenerated in response to the multi-modal sensory regulation and the human-computer interaction guidance again.

6. The state regulation system based on multisensory focus and neural remodeling according to claim 1, characterized in that The physiological data interaction module specifically performs the following steps on the user's multi-modal physiological signals: Collect the input Suds score data to generate structured subjective state data; Collect eye movement trajectories, facial electromyograms, and respiratory frequencies Signals from an eye movement tracking device, a camera, and a respiratory sensor to generate neurophysiological data; Collect HRV signals from an HRV sensor to generate autonomic nervous function state data; Collect GSR signals from a skin conductance sensor to generate physiological arousal intensity feature data; Align the structured subjective state data, neurophysiological data, autonomic nervous function state data, and physiological arousal intensity feature data to generate real-time multi-modal feature data.

7. The state regulation system based on multi-sensory focusing and neural remodeling according to claim 1, characterized in that The trained neural state evaluation model includes a vectorization unit, a feature mining unit, a feedback judgment unit, and an encoding unit; The intelligent decision-making module specifically performs the following steps to generate multi-modal feedback control instructions: Encode the current multi-modal feature data based on the vectorization unit to obtain a multi-modal feature vector; The feature mining unit performs convolutional processing on the multimodal feature vectors to evaluate the user's current neurophysiological activation state; The feedback judgment unit maps the current neurophysiological activation state to obtain a neurophysiological activation state evaluation, and performs a feedback logic judgment on the neurophysiological activation state evaluation to determine the feedback regulation strategy; Based on the encoding unit, the feedback regulation strategy is encoded to generate a multimodal feedback control instruction.

8. The state regulation system based on multi-sensory focus and neural remodeling according to claim 7, wherein The feedback judgment unit includes an individualized baseline model, a state evaluation model, and a logic judgment module. Based on the feedback judgment unit, the following steps are performed to generate a feedback regulation strategy: Map the current neurophysiological activation state to the individualized baseline model to obtain baseline deviation degree data; Map the baseline deviation degree data to the state evaluation model to obtain a neurophysiological activation state evaluation; Map the neurophysiological activation state evaluation to the logic judgment module to perform a feedback logic judgment and determine the feedback regulation strategy.

9. The state regulation system based on multi-sensory focusing and neural remodeling according to claim 1, wherein The multimodal feedback output module includes an instruction parsing unit, a visual display unit, an audio playback unit, a tactile stimulation unit, and an olfactory stimulation unit. The multimodal feedback output module performs the following steps to generate at least one sensory regulation stimulus: The instruction parsing unit parses the multimodal feedback control instruction to generate a visual control instruction, an audio control instruction, a tactile control instruction, and an olfactory control instruction; The visual display unit generates a visual feedback interface based on the visual control instruction; The audio playback unit generates an auditory guidance signal based on the audio control instruction; The tactile stimulation unit generates a somatosensory regulation signal based on the tactile control instruction; The olfactory stimulation unit generates an olfactory regulation signal based on the olfactory control instruction.

10. The state regulation system based on multi-sensory focus and neural remodeling according to claim 4, characterized in that The neurofocus guidance module includes a semantic parsing unit, a data association unit, and a guidance encoding unit. The neurofocus guidance module is used to perform the following steps to generate a human-computer interaction guidance for guiding the user to focus on the target physiological reflex area: Based on the semantic parsing unit, perform semantic parsing processing on the user's concern description to obtain a concern semantic feature vector; Based on the data association unit, associate the concern semantic feature vector with the multimodal physiological reflex data to generate a target physiological reflex area localization result; Based on the guidance encoding unit, encode the target physiological reflex area localization result to generate a human-computer interaction guidance for guiding the user to focus on the target physiological reflex area.

Images