Oral cavity detection method based on intelligent tooth socket

Abstract

The invention relates to the technical field of oral health monitoring, and discloses an oral detection method based on an intelligent tooth socket. The method comprises the following steps: acquiring an original oral cavity pressure time sequence and frequency domain characteristics through an intelligent tooth socket, determining an optimized decomposition parameter, and performing multi-scale decomposition on an original signal to generate a multi-scale oral cavity signal component; nonlinear dynamic features are extracted, a component complexity index is obtained, and key signal components related to oral health are screened out; generating a personalized physiological response frequency in combination with the user oral cavity baseline features, the real-time occlusion state and the component complexity index; reconstructing an oral cavity physiological feature sequence based on the key signal component and the personalized physiological response frequency, and separating a low-frequency component and a high-frequency component through modal decomposition; identifying oral cavity abnormity categories according to the two components, establishing a mapping relation with a diagnosis result, and generating an oral cavity health detection instruction. According to the method, convenient and comprehensive oral cavity detection can be realized, the detection adaptability and precision are improved, and support is provided for oral cavity health management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oral health monitoring, in particular to an oral detection method based on an intelligent mouthpiece. BACKGROUND

[0002] With the improvement of oral health awareness, early screening and dynamic monitoring of oral diseases have become an important topic in modern preventive medicine. Traditional oral detection relies on the equipment of professional medical institutions and the experience of doctors, which has problems such as long detection period, high cost, and dependence on subjective judgment, and is difficult to meet the needs of the public for real-time, convenient and personalized oral health management. At present, the oral detection technology on the market is mainly divided into two categories: invasive detection such as oral endoscope and periodontal probe, which can obtain high-precision data, but the operation process is easy to cause discomfort to patients, and long-term continuous monitoring cannot be realized; non-invasive detection such as oral CT and ultrasonic scanning, which can provide three-dimensional structural information, but the equipment is bulky and expensive, and is only suitable for medical institutions, and is difficult to popularize to the family scene. In recent years, the application of wearable devices in the field of health monitoring has promoted the innovation of oral detection technology. As a wearable device that fits the oral structure, the intelligent mouthpiece has the potential to continuously collect oral physiological signals, but there are still significant technical bottlenecks in signal processing and health status recognition. The existing detection methods based on intelligent mouthpieces mostly rely on single-dimensional signal features (such as pressure peak value, occlusion frequency), ignoring the nonlinear and multi-scale characteristics of oral physiological signals - oral pressure signals contain not only high-frequency components generated by dynamic actions such as chewing and swallowing, but also low-frequency information of the steady-state changes of periodontal tissue. Single feature extraction method cannot fully reflect the oral health status. Individual oral physiological characteristics have significant differences, such as biting force, tooth arrangement, periodontal tissue condition, etc., which have personalized characteristics. Universal signal analysis model often leads to insufficient detection accuracy, making it difficult to accurately identify early oral diseases (such as gingivitis, periodontitis, occlusion abnormalities, etc.). At the same time, the existing technology lacks adaptive optimization mechanism in signal decomposition and feature selection, which makes it difficult to effectively separate noise and effective signals, resulting in insufficient extraction of key physiological features and affecting the accuracy of subsequent health status evaluation.

[0003] It is of great significance to develop an intelligent detection method that can adaptively process oral pressure signals, extract personalized physiological features, and accurately identify abnormal oral conditions, to improve the convenience, real-time and accuracy of oral health monitoring. It is also an important research direction in the field of wearable medical devices. SUMMARY

[0004] The purpose of the present application is to provide an oral detection method based on an intelligent mouthpiece to solve the problems raised in the background.

[0005] To achieve the above object, the present application provides a mouth detection method based on intelligent mouthpiece, which comprises: obtaining original mouth pressure time series and corresponding frequency domain features collected by a user wearing an intelligent mouthpiece, determining optimization decomposition parameters corresponding to the frequency domain features, performing multi-scale signal decomposition on the original mouth pressure time series based on the optimization decomposition parameters, and generating multi-scale mouth signal components; extracting nonlinear dynamic features of the multi-scale mouth signal components, and obtaining component complexity indicators; screening key signal components related to the oral health status according to the multi-scale mouth signal components and the component complexity indicators; obtaining oral baseline features of the user, real-time occlusion state, and combining the component complexity indicators to generate individual physiological response frequencies; reconstructing oral physiological feature sequences based on the key signal components and the individual physiological response frequencies, performing modal decomposition on the oral physiological feature sequences, and separating low-frequency components reflecting steady-state physiological characteristics and high-frequency components reflecting dynamic changes; identifying oral abnormal categories of the user according to the low-frequency components and the high-frequency components, establishing a mapping relationship between the abnormal categories and diagnostic results, and generating oral health detection instructions.

[0006] Preferably, the obtaining of the oral baseline features of the user, the real-time occlusion state, and the component complexity indicators comprises: collecting multi-stage oral baseline data, wherein the multi-stage includes a resting state, a light occlusion state, and a chewing task state; extracting occlusion rhythm features from the multi-stage oral baseline data, and generating real-time occlusion state labels and component complexity indicators based on the occlusion rhythm features; calculating matching degree scores of preset candidate frequencies according to the frequency band energy distribution of the oral baseline features, the real-time occlusion state labels, and the component complexity indicators, wherein the matching degree scores include physiological rhythm consistency scores, state adaptability scores, and complexity correlation scores; dynamically adjusting weight coefficients of the matching degree scores using an optimization algorithm to determine individual physiological response frequencies.

[0007] Preferably, the multi-scale signal decomposition on the original mouth pressure time series to generate multi-scale mouth signal components comprises: processing the original mouth pressure time series using an adaptive filtering algorithm combined with the optimization decomposition parameters; and suppressing end effect using a period extension method at the sequence boundary; extracting a multi-scale coefficient set after decomposition, and constructing the multi-scale mouth signal components according to the multi-scale coefficient set.

[0008] Preferably, reconstructing the oral physiological feature sequence based on the key signal component and the personalized physiological response frequency comprises: constructing a time-frequency feature matrix, elements of the time-frequency feature matrix being generated by nonlinear mapping of the amplitude of the key signal component, the frequency band weight corresponding to the personalized physiological response frequency, and the preset physiological modulation coefficient; performing multi-scale time series analysis on the time-frequency feature matrix, including short-time scale analysis to capture occlusion transient features, medium-time scale analysis to track physiological state evolution, and long-time scale analysis to monitor health trends; generating an oral physiological feature sequence containing multi-mode physiological features by dynamically weighting and fusing multi-scale time series analysis results.

[0009] Preferably, the modal decomposition of the oral physiological feature sequence separates low-frequency components reflecting steady-state physiological characteristics and high-frequency components reflecting dynamic changes, comprising: detecting extreme points of the oral physiological feature sequence and performing mirror extension processing, and extracting multi-order modal components satisfying the intrinsic condition by iterative screening; According to the average period and energy concentration degree of the multi-order modal components, the front-order short-period components are classified as high-frequency components, and the back-order long-period components and residual terms are classified as low-frequency components.

[0010] Preferably, the identification of the user's oral abnormality category according to the low-frequency component and the high-frequency component comprises: extracting time-domain statistical features, energy distribution features, and component correlation features from the low-frequency component to construct a steady-state feature vector; extracting instantaneous change rate, peak feature, and waveform distortion feature from the high-frequency component to construct a transient feature vector; adopting a hierarchical classification strategy to classify the steady-state feature vector into a health state, and identifying abnormal events corresponding to the transient feature vector through an adaptive detector; Fusing health state classification results and abnormal events, combining spatiotemporal continuity constraints for confidence verification, and outputting an oral abnormality category.

[0011] Preferably, the acquisition of the original oral pressure time sequence collected by the user wearing the smart mouthpiece comprises: real-time capture of multi-point pressure data and temperature auxiliary data in the oral cavity; classifying occlusion layer, tongue pressure layer, and buccal mucosa layer signals to analyze multi-source physiological parameters; integrating and standardizing the processed data to generate a structured original oral pressure time sequence.

[0012] Preferably, the dynamic adjustment of the weight coefficient of the matching degree score by the optimization algorithm comprises: Initialize the position and velocity parameters of the particle swarm, and define the target function as the weighted harmonic mean of the matching degree score; Iteratively update the particle position and calculate the global optimal solution, and output the weight coefficient optimization result when the target function converges.

[0013] Preferably, the fusion health state classification result and abnormal event comprises: A dynamic probability graph model is constructed, taking the health state classification result as the node prior probability and the abnormal event as the observation evidence. The node probability distribution is updated through Bayesian inference, and the joint probability output is generated in combination with the state transition constraint.

[0014] Preferably, after the oral health detection instruction is generated, it comprises: Real-time acquisition of instruction execution feedback data, the feedback data comprising abnormal area pressure recovery rate, temperature fluctuation deviation value and health index change amount; According to the feedback data, the oral baseline feature and the component complexity index are dynamically updated.

[0015] Compared with the prior art, the present application has the following advantages: The oral cavity detection method based on the intelligent dental tray can realize real-time and dynamic monitoring of the oral cavity signal by collecting the original oral cavity pressure time sequence and the corresponding frequency domain feature through the intelligent dental tray, thereby eliminating the dependence on traditional large medical equipment and improving the convenience of detection. In the signal processing link, the optimization decomposition parameters are determined based on the frequency domain feature, the original signal is subjected to multi-scale decomposition to generate multi-scale oral cavity signal components, and this multi-scale decomposition method can decompose the complex oral cavity signal into components of different scales, more meticulously capture various feature information in the oral cavity signal, and help in-depth analysis of the signal. The nonlinear dynamic features of the multi-scale oral cavity signal components are extracted, the component complexity index is obtained, the key signal components are screened in combination with the multi-scale oral cavity signal components and the component complexity index, the important information related to the oral health state can be effectively focused, the interference of irrelevant signals is reduced, and the subsequent analysis is more targeted. By obtaining the oral baseline feature of the user, the real-time occlusion state and combining the component complexity index, a personalized physiological response frequency is generated, which fully considers the differences in individual physiological characteristics, makes the detection more suitable for the actual situation of different users, and helps to improve the adaptability and accuracy of the detection. The oral physiological feature sequence is reconstructed based on the key signal components and the personalized physiological response frequency, and modal decomposition is performed thereon to separate the low-frequency components reflecting the steady-state physiological characteristics and the high-frequency components reflecting the dynamic changes. This separation method can analyze the steady-state and dynamic characteristics of the oral cavity respectively, and comprehensively grasp the health status of the oral cavity. According to the low-frequency component and the high-frequency component, the oral cavity abnormality category is identified, and a mapping relationship with a diagnosis result is established, an oral cavity health detection instruction is generated, the detection result can be directly associated with the diagnosis result, a more intuitive and effective basis is provided for oral cavity health evaluation, potential oral cavity problems can be found in time, and oral cavity health management is assisted. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A working principle diagram of the oral cavity detection method based on the intelligent mouthpiece is shown. Figure 2 A flowchart of multi-scale signal decomposition is shown. Figure 3 A flowchart of modal decomposition is shown. Figure 4 A flowchart of obtaining an original oral cavity pressure time sequence is shown. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0018] Please refer to Figure 1 The present application provides an oral cavity detection method based on an intelligent mouthpiece, which comprises: An original oral cavity pressure time sequence and frequency domain features in a user's oral cavity are collected by the intelligent mouthpiece. First, the optimal decomposition parameters corresponding to the frequency domain features are determined, and the original oral cavity pressure time sequence is subjected to multi-scale signal decomposition based on the parameters, to generate multi-scale oral cavity signal components. Then, the nonlinear dynamic features of the multi-scale oral cavity signal components are extracted, and the component complexity indexes are calculated. According to the multi-scale oral cavity signal components and the component complexity indexes, the key signal components related to the oral cavity health status are screened. The personalized physiological response frequency is generated in combination with the user's oral cavity baseline features, the real-time occlusion state and the component complexity indexes. The oral cavity physiological feature sequence is reconstructed based on the key signal components and the personalized physiological response frequency, and the sequence is subjected to modal decomposition to separate the low-frequency component and the high-frequency component. Finally, the user's oral cavity abnormality category is identified according to the low-frequency component and the high-frequency component, a mapping relationship between the abnormality category and a diagnosis result is established, and an oral cavity health detection instruction is generated.

[0019] Embodiment 1: refer to Figure 2The multi-stage oral baseline data acquisition process is achieved through the pressure sensor array built-in the smart mouthguard, covering three typical physiological scenarios of resting state, light bite state and chewing task state. In the resting state, the user keeps the mouth naturally closed, and the sensor records the baseline pressure distribution under no bite load. The light bite state requires the user to close the teeth with constant force, simulating the daily unconscious bite behavior. The chewing task state is activated through standardized chewing action, and the test material with specific hardness and size is used to ensure data comparability. The sensor sampling frequency is set in the range of 200Hz to 1000Hz, to balance the high-frequency dynamic capture and data storage efficiency.

[0020] The extraction of occlusal rhythm features uses a sliding window analysis method, and the window length is adaptively adjusted according to the user's average occlusal period. The time domain statistics of the pressure signal are calculated in each window, including peak interval, amplitude variation coefficient and rising slope. The resting state data is used to establish an individualized noise base, the light bite state data reveals the basic bite pattern, and the chewing task data contains more rich time-frequency features. By comparing the power spectral density differences of the three groups of data, the feature frequency band related to functional movement is identified. The real-time occlusal state label uses a finite state machine model to generate, which classifies the continuous pressure signal into three discrete states of resting, light bite or chewing. The component complexity index is calculated based on the multi-scale entropy algorithm, reflecting the irregularity of the signal in time and frequency domains.

[0021] The frequency band energy distribution of the oral baseline features is obtained through short-time Fourier transform, which is divided into four physiological frequency bands of delta, theta, alpha and beta. The delta band corresponds to the slow wave of 0.5-4Hz, the theta band covers the basic rhythm of 4-8Hz, the alpha band of 8-13Hz is related to autonomic nervous regulation, and the beta band of 13-30Hz reflects fast muscle regulation. The energy proportion of each frequency band and the real-time occlusal state label constitute a three-dimensional feature space, and the Mahalanobis distance is used to measure the matching degree of the candidate frequency and the feature space. The physiological rhythm consistency score measures the degree of consistency between the candidate frequency and the user's historical dominant rhythm, the state adaptability score evaluates the ability of the frequency to distinguish different bite states, and the complexity correlation score represents the correlation between the frequency component and the complexity of oral movement.

[0022] The optimization algorithm adopts an improved particle swarm optimization framework, and the particle dimension corresponds to the weight coefficient of the three matching degree scores. The initial population size is set to 50, and the particle position is randomly initialized in the [0, 1] interval. The speed parameter is limited in the range of [-0.1, 0.1] to prevent premature convergence. The objective function is defined as the weighted harmonic mean of the three scores, and a nonlinear decay factor is introduced to balance the exploration and development stages. In each iteration, the particle updates the velocity vector according to the individual historical optimal solution and the global optimal solution, and performs boundary reflection processing after position update to avoid falling into local extremum. The convergence condition is set to the global optimal solution with a change amplitude less than 1e-5 for 20 consecutive generations, and the final output weight coefficient combination makes the objective function reach the Pareto optimality.

[0023] The preprocessing of the original oral pressure time series includes baseline drift correction and motion artifact elimination. The adaptive filter design adopts the LMS algorithm, and the step size parameter is dynamically adjusted according to the signal signal-to-noise ratio. The decomposition layer is determined by the optimized decomposition parameter, and the over-decomposition is automatically terminated by the spectrum flatness detection. The boundary processing adopts the symmetric periodic extension strategy, and the endpoint extreme value is mirrored and copied to the both ends of the sequence to form a virtual extension segment. The multi-scale coefficient set is constructed by discrete wavelet transform, and the Daubechies wavelet basis function is selected to ensure the time-frequency localization property. The detail coefficients and approximation coefficients are recombined after threshold denoising to form 6-8 components with clear physiological interpretation, corresponding to different sources such as occlusal force main frequency, harmonic component, and electromyographic interference.

[0024] The physical significance verification of multi-scale oral signal components is realized through synchronous electromyographic signals. The surface electromyographic electrode array of the temporalis and masseter muscles provides independent reference, and the coherence analysis is performed with the pressure signal components. The components with coherence coefficients exceeding 0.7 are marked as effective physiological components, and the rest are considered as environmental noise or device artifacts. The effective components are further classified by cluster analysis, and a component dictionary is established based on their time domain waveform, frequency band energy, and complexity characteristics. This dictionary serves as a priori knowledge base in subsequent signal screening, improving the specificity of key signal component identification.

[0025] The update mechanism of the real-time occlusal state label adopts an event-driven mode, which triggers reclassification when the pressure signal amplitude exceeds 3 times the standard deviation of the resting state. The state duration statistical characteristics are included in the observation variables of the hidden Markov model, and the Viterbi algorithm is used to decode the most likely state sequence. The sliding calculation window of the component complexity index is set to 5 seconds, and it is updated every 1 second to balance real-time and stability. The dynamic Z-score method is used for index normalization, which is standardized with reference to the moving mean and standard deviation of the user's historical data.

[0026] The final determination of the personalized physiological response frequency is completed through a multi-criteria decision. The candidate frequency list is ranked in descending order of total matching score, and the top 5% frequencies are subjected to expert system for physiological reasonableness verification. The verification rules include whether the frequency falls within the typical range of human jawbone resonance (2-8 Hz), whether it is associated with the user's past medical history, etc. The frequencies that pass the verification are output as the personalized physiological response frequency, which is used to guide the subsequent feature reconstruction. The entire process is implemented in an embedded system, with a calculation delay controlled within 50 ms to meet real-time requirements.

[0027] Residual analysis of signal decomposition is used to monitor the system health status. When the residual energy ratio exceeds 15%, a sensor calibration process is triggered, and a self-checking sequence is performed through a built-in vibration motor. Adaptive adjustment of decomposition parameters uses a reinforcement learning framework, with the mutual information between the reconstructed signal and the original signal as the reward signal. The length optimization of periodic continuation is achieved through cross-validation, selecting the continuation ratio that minimizes the energy of the endpoint components. The storage of multi-scale coefficient sets uses sparse coding technology, retaining only the top 10% of the amplitude coefficients to reduce the transmission bandwidth requirement.

[0028] Dynamic maintenance of oral baseline features uses an incremental learning mechanism. One minute of standard occlusion data is automatically collected every week, and the distribution drift is detected through KL divergence. When the difference between the old and new data distributions exceeds the threshold, a baseline feature retraining process is triggered. The training data is weighted to combine recent samples and historical samples, with different weights given to new and old data using exponential decay. Long-term trend analysis of component complexity indicators uses a seasonal decomposition method to separate the effects of slow-changing factors such as age growth and wearing habits. This mechanism enables the system to continuously adapt to the natural evolution of the user's physiological state.

[0029] Embodiment 2: refer to Figure 3 The construction process of the time-frequency feature matrix is based on the amplitude characteristics of the key signal components, the frequency band weight of the personalized physiological response frequency, and the preset physiological modulation coefficient. The amplitude of the key signal components is extracted through Hilbert transform to obtain the instantaneous envelope line, eliminating the interference of high-frequency oscillation on amplitude estimation. The frequency band weight of the personalized physiological response frequency is obtained based on user historical data analysis, reflecting the representation ability of different frequency bands for individual oral status. The physiological modulation coefficient is derived from large-scale clinical research data, describing the statistical association between typical oral abnormalities and energy changes in specific frequency bands. These three elements generate matrix elements through a nonlinear mapping function, which uses a combination of piecewise linear approximation and Sigmoid activation to preserve the physical meaning of the features while enhancing numerical stability.

[0030] The row dimension of the matrix corresponds to the discrete sampling points of the time series, and the column dimension covers the physiological frequency band from 0.5 Hz to 30 Hz. The generation process of each matrix element includes a dynamic normalization step to convert the original feature value to the [0, 1] interval to eliminate the dimensional difference. The time axis adopts a non-uniform sampling strategy, increasing the sampling density in areas with rapid signal changes and reducing the sampling rate in stable areas to optimize storage efficiency. The frequency band division adopts an overlapping sub-band design, with a 15%-20% frequency overlap between adjacent frequency bands to avoid the feature truncation effect caused by strict frequency band division. The application of physiological modulation coefficient introduces adaptive gain control, dynamically adjusting the contribution weight of different frequency bands according to the user's real-time bite force.

[0031] The implementation of multi-scale time series analysis is divided into short, medium and long time scales. Short-time scale analysis focuses on a 50-200 ms time window, capturing transient features of bite actions through differential operators, including micro-parameters such as pressure rise rate, peak holding time, etc. The analysis process uses a sliding window strategy, with a window step size of 10 ms to ensure feature continuity. The extraction of transient features combines morphological filtering techniques to eliminate false transients caused by sensor noise. Medium-time scale analysis covers a 5-30 second time range, modeling the evolution trajectory of physiological state using a recurrent neural network. The network hidden layer units store the time series dependence relationship, and through the gating mechanism, they control the information transmission path, distinguishing between normal physiological fluctuations and abnormal state transitions. Long-time scale analysis processes data blocks longer than 5 minutes, using a trend decomposition algorithm to separate periodic patterns and baseline drift. Health trend monitoring indicators include the coefficient of variation of daily bite strength, high-frequency energy cumulative distribution, and other slowly varying parameters.

[0032] The dynamic weighted fusion algorithm integrates the results of multi-scale analysis, with weight allocation following the scale correlation principle. The short-time feature weight is positively correlated with the bite action frequency, the medium-time feature weight depends on the state transition probability, and the long-time feature weight is determined by the time consistency of health indicators. The fusion process introduces a conflict resolution mechanism, when different scale features appear contradictory, the scale conclusion with higher statistical significance is preferred. The final generated oral physiological feature sequence contains three types of data channels: the transient feature channel records micro-movement parameters, the state evolution channel encodes physiological phase markers, and the trend monitoring channel stores standardized values of health indicators. The time resolution of the sequence is unified to 1 second / frame, and missing data is filled by cubic spline interpolation.

[0033] The modal decomposition of oral physiological feature sequence adopts an improved extreme point detection algorithm. The extreme point positioning combines polynomial fitting and gradient analysis to avoid false extreme value labeling caused by noise. The mirror extension process symmetrically replicates the extreme point distribution at both ends of the sequence, and the extension length is 1-2 times the main period of the signal. The iterative screening process adopts an adaptive stopping criterion, which terminates the cycle when the mean square error change rate of the screening results of the last two times is less than 5%. The intrinsic modal component determination condition is relaxed to allow moderate harmonic aliasing to preserve the complex modulation characteristics of physiological signals. The envelope symmetry constraint is introduced in the screening process, which forces the upper and lower envelope lines to be symmetric about the zero mean, and suppresses non-physiological bias components.

[0034] The classification of multi-order modal components is based on two dimensions of average period and energy concentration. The average period is calculated by zero-crossing rate statistics, and the standard is set as 0.5 second period threshold. The energy concentration uses the frequency band energy ratio to measure, and calculates the energy proportion of the component in the 4-8Hz frequency band. The front-order short-period component usually presents dense oscillation characteristics, the average period is less than 0.5 seconds and the energy is concentrated in the high-frequency region, which is classified as high-frequency component. These components mainly reflect the rapid contraction of masticatory muscles, transient contact of teeth and other dynamic events. The back-order long-period component shows a slow fluctuation mode, the average period exceeds 0.5 seconds and the energy is distributed in the low-frequency region, which is classified as low-frequency component together with the residual term. This kind of component represents the salivary secretion rhythm, long-term occlusal adaptation and other steady-state physiological processes.

[0035] The analysis of high-frequency components focuses on the description of time-varying characteristics. The instantaneous change rate is calculated by the five-point differential formula, supplemented by median filtering to smooth the differential noise. The peak feature detection adopts a double-threshold strategy, the main threshold is used to identify significant peak groups, and the secondary threshold determines the peak boundary. The waveform distortion evaluation is based on the dynamic time warping algorithm, which calculates the morphological difference degree after aligning the real-time waveform with the standard template. The processing of low-frequency components focuses on periodic analysis. Time domain statistical features include moving average line, standard deviation band and other descriptive indicators. Energy distribution features are estimated by Welch periodogram, which highlights weak components using logarithmic coordinates. The correlation analysis of components calculates the phase synchronization index of signals at different sensor positions, and evaluates the coordination of oral movement.

[0036] The quality control of the reconstruction process sets multiple verification mechanisms. The completeness test of time-frequency feature matrix is realized by inverse reconstruction error, which requires the correlation coefficient of the original signal and the reconstructed signal to be greater than 0.85. The physiological reasonableness verification of modal components is based on the clinical knowledge base, and the abnormal components with average period exceeding the human physiological range (<0.1s or >10s) are removed. The stability test of multi-scale fusion results uses leave-one-out cross-validation to check the fluctuation amplitude of feature weights in different data subsets. The final output of oral physiological feature sequence needs to pass the integrity check to ensure that the transient, state and trend three types of channel data are time-aligned and have no logical contradictions.

[0037] The system implements a layered processing architecture. The bottom layer signal processing runs on the embedded processor of the smart mouthguard, completing the time-frequency matrix construction and short-time feature extraction. The mesoscale analysis is deployed on the mobile terminal, using the device GPU to accelerate neural network inference. The long-scale trend processing is executed on the cloud server, optimizing the model parameters in combination with multi-user data. Feature compression technology is used for data transmission, and the time-frequency matrix is compressed by more than 80% in volume through sparse coding. In terms of real-time performance, the end-to-end delay from signal acquisition to feature sequence output is controlled within 300 ms, meeting the real-time feedback requirements of clinical monitoring.

[0038] The dynamic updating mechanism continuously optimizes the feature extraction process. The frequency band division of the time-frequency matrix is re-evaluated monthly, and the sub-band boundaries are adjusted according to the spectral characteristics of the latest user data. The physiological modulation coefficients are updated quarterly, synchronizing with the latest clinical research conclusions. The adaptive learning of the modal decomposition parameters records the residual distribution of each decomposition, gradually optimizing the sensitivity of extreme point detection. The multi-scale weight strategy preserves historical adjustment records and establishes a weight-user behavior association rule base. These mechanisms enable the system to adapt to natural changes in user oral characteristics, maintaining the accuracy of long-term monitoring.

[0039] Embodiment 3: refer to Figure 4 The feature extraction process of the low-frequency component uses a multi-dimensional analysis method. The calculation of time-domain statistical features is based on a sliding window framework, and the window length is adjusted synchronously with the user's basic occlusion period. The mean parameter reflects the steady-state pressure level, and the long-term drift is eliminated through detrending processing. The variance analysis uses an adaptive threshold to dynamically distinguish between physiological fluctuations and abnormal variations. The skewness coefficient calculation introduces a robustness correction to reduce the interference of extreme values on distribution shape evaluation. The quantification of energy distribution features is achieved through improved wavelet packet decomposition, and the decomposition tree structure is dynamically optimized according to the spectral characteristics of the low-frequency component. The sub-band energy proportion calculation uses normalization processing to eliminate the differences in absolute energy values between individuals. The component correlation analysis constructs a three-dimensional feature space, including time delay cross-correlation, phase synchronization index, and nonlinear coupling strength.

[0040] The transient feature detection of the high-frequency component establishes a multi-stage processing pipeline. The calculation of instantaneous change rate uses a noise-resistant difference operator to suppress high-frequency noise amplification while maintaining slope accuracy. Peak feature recognition is preprocessed through morphological filtering to eliminate the influence of baseline fluctuations on peak detection. The peak parameter set includes time sequence indicators such as rise time, half-peak width, and peak interval. Waveform distortion evaluation uses dynamic template matching technology, and the template library contains twenty typical abnormal waveform patterns. The matching similarity is calculated as follows: where S represents the waveform similarity score, is the kth sample point value of the template signal, is the kth sample point value of real-time signal, N is the length of comparison window, is the position weight coefficient. The formula introduces amplitude normalization and spatial weighting mechanism to improve the physiological relevance of distortion detection.

[0041] The implementation of hierarchical classification system adopts hybrid architecture. The processing of steady-state feature vector uses ensemble learning framework, and the base classifiers include random forest, support vector machine and shallow neural network. The Gini importance and recursive feature elimination scores are calculated in the feature selection stage, and the feature subset with high cross-validation consistency is reserved. The fuzzy logic rule is used for classification decision fusion, and the probability transfer interval of different health states is defined. The adaptive detector is designed for transient feature vector, and the core is the variable threshold abnormal score mechanism. The threshold adjustment algorithm monitors the interquartile range of user's recent feature distribution, and dynamically sets the abnormal judgment boundary. The detector output includes event type, severity and spatiotemporal positioning three metadata.

[0042] The construction of dynamic probabilistic graph model adopts factor graph representation. The nodes are divided into two categories: observation nodes and hidden state nodes. The observation nodes correspond to real-time anomaly event detection results, and the hidden state nodes represent the potential health state. The prior probability distribution is initialized by historical diagnosis records, and the state transition matrix sets the basic parameters according to clinical guidelines. The observation likelihood function uses non-parametric estimation, and the bandwidth of kernel density function is adaptively adjusted according to the sparsity of user data. The Bayesian inference process uses approximate variational inference to balance the computational complexity and inference accuracy. The spatiotemporal continuity constraint is realized as a Markov random field, which defines the state consistency penalty term of adjacent time periods and adjacent sensor nodes.

[0043] The acquisition system of original oral pressure data uses a multi-modal sensing array. The occlusal layer monitoring uses an 8×8 matrix piezoresistive sensor with a spatial resolution of 2 square millimeters. The tongue pressure layer detection uses a flexible PVDF film sensor, and the sampling frequency is set to 500 Hz to capture rapid dynamic changes. The buccal mucosa layer monitoring combines impedance sensing and optical plethysmography technology to obtain auxiliary information of tissue microcirculation. The signal preprocessing includes four-stage cascade filtering: 50 Hz power frequency notch to eliminate environmental interference, 0.5 Hz high-pass filter to remove respiratory artifacts, 20 Hz low-pass filter to suppress electromyographic noise, and adaptive spectral subtraction algorithm to improve signal-to-noise ratio.

[0044] Multi-source data fusion adopts spatio-temporal registration technique. Temporal alignment is based on hardware synchronization signal, and the sampling clock deviation of each sensor channel is controlled within 100 microseconds. Spatial registration establishes an anatomical coordinate system to map the data of sensors at different positions to a standard oral model. Physiological parameter analysis algorithm includes three parallel processing flows: occlusal force center trajectory tracking, tongue movement pattern recognition, and mucosal blood perfusion index calculation. Data standardization adopts a hierarchical processing strategy, device-level calibration eliminates individual differences between sensors, session-level normalization compensates for the influence of environmental temperature, and user-level Z-score conversion achieves cross-individual comparability.

[0045] The generation of structured time series follows the ISO / TS 18234 standard. The data packet format includes header information, payload data, and integrity check. The header information records the acquisition timestamp, sensor ID, and signal quality index. The payload data uses differential encoding compression to reduce transmission bandwidth occupancy. The integrity check uses a combination of cyclic redundancy check and singular value decomposition mechanism to detect random errors and system bias simultaneously. Sequence storage uses a hierarchical database architecture, with the raw data layer retaining unprocessed signals, the feature layer storing extracted time series features, and the metadata layer recording processing parameters and quality control markers.

[0046] Post-processing of abnormal events includes credibility verification and clinical interpretation. Credibility verification checks the temporal persistence, spatial consistency, and multi-modal corroboration strength of abnormal events. The clinical interpretation module accesses the knowledge graph to map feature space anomalies to pathological mechanism chains. The interpretation output uses natural language generation technology to produce structured reports containing probability classification, differential diagnosis, and recommended measures. The system maintenance module continuously monitors feature extraction performance and automatically triggers sensor calibration procedures when the signal quality index falls below the threshold. The calibration sequence includes two standard protocols: step pressure test and frequency response test, which update sensor characteristic parameters by analyzing calibration data.

[0047] The user interaction interface realizes a closed-loop feedback mechanism. Real-time visualization displays pressure heat maps and abnormal warning markers with a refresh rate above 30Hz. Historical trend charts support multi-time scale zooming, observing pattern evolution from minute to month level. Interaction logs record user responses to warnings, and these behavior data are used to optimize abnormal detection sensitivity. The system configuration interface allows clinical professionals to adjust analysis parameters, and the modified parameters are protected by digital signature to ensure operation traceability. The data export module supports HL7 and FHIR medical information exchange standards, enabling seamless integration with electronic health record systems.

[0048] The long-term adaptation mechanism employs an incremental learning framework. Feature drift detection is automatically performed weekly, comparing the differences in KL divergence distribution between recent data and historical benchmarks. When a significant drift is detected, a model parameter fine-tuning process is triggered. This fine-tuning process utilizes elastic weight fixation technology, preserving learned important feature associations while adapting to new data. User-personalized profiles record the trajectory of physiological characteristic changes, providing objective quantitative evidence for clinical follow-up. The system maintenance log automatically records all parameter adjustments and algorithm updates, establishing a complete audit trail.

[0049] Example 4: The process of dynamically adjusting the matching score weight coefficients in the optimization algorithm adopts an improved particle swarm optimization framework. The initial particle swarm size is set to 60 particles, and the position vector of each particle contains three dimensions, corresponding to the weight coefficients of the physiological rhythm consistency score, state adaptability score, and complexity association score, respectively. The position initialization range is limited to the interval [0.1, 0.9] to avoid search space reduction caused by boundary values. The velocity vector is initialized using Gaussian distribution random sampling, and the standard deviation is set to 0.15 to maintain the diversity of the initial exploration. The particle memory mechanism not only records the individual's historical best position, but also maintains a sliding window of recent search trajectories to detect local convergence trends.

[0050] The objective function is constructed considering the synergistic effect of the three matching scores, using a weighted harmonic mean as the optimization criterion. This criterion strengthens the contribution of low-scoring items through nonlinear transformation, prompting the algorithm to improve all scores in a balanced manner rather than optimizing only a single indicator. A dynamic adjustment factor is introduced in the objective function calculation process, and its expression is: Where F represents the fitness value, For the first Each weighting coefficient It is an adjustment factor that decays over time. A small constant to prevent division by zero errors. The design follows an exponential decay law, initially allowing one weight to dominate the search direction, and later forcing the three weights to tend towards balance. The particle update rule introduces the concept of topological neighborhood, where each particle can only obtain the optimal information of neighboring particles within a specific radius. This radius decreases linearly with the number of iterations, achieving a gradual transition from global exploration to local development.

[0051] Convergence criterion adopts a compound condition strategy. The first re-determination is triggered when the global optimal solution improvement is less than 0.001 for 15 consecutive generations. The second re-determination checks the particle swarm diversity index. When the average Euclidean distance between particles is less than 5% of the search space diameter, convergence is confirmed. The post-processing smoothing is performed before the output of the weight coefficient. The cubic spline interpolation is used to fit the trajectory of the optimal weight of the previous generation. The extrapolated value of the tangent direction at the end of the curve is taken as the final result. This processing effectively suppresses the influence of random fluctuations on the stability of the output.

[0052] The construction of the dynamic probabilistic graph model adopts a combined structure of factor graph and Markov blanket. The node set contains three types of entities: the health state classification node represents the potential oral health state, the abnormal event observation node corresponds to the real-time detected transient feature, and the environmental context node records the occlusion state and device parameters and other auxiliary information. The edge set is divided into deterministic connection and probabilistic connection. The deterministic connection encodes the hard rules defined by clinical knowledge, and the probabilistic connection reflects the soft association obtained by statistical learning.

[0053] The prior probability distribution is established by fusing two data sources. The history diagnosis record provides explicit state labeling data, and the kernel density estimation method is used to construct the initial distribution. The user questionnaire data is transformed into the adjustment factor of the probability distribution by extracting the implicit theme through the latent Dirichlet allocation model. The design of the state transition matrix incorporates the time autocorrelation characteristics. The probability of the current state depends not only on the previous state but also on the pattern of the recent state sequence. This design can capture the inertia characteristics and periodicity of the oral health state.

[0054] The parameter estimation of the observation likelihood function uses the variational expectation-maximization algorithm. In the algorithm implementation, sparsity constraints are introduced to force the model to focus on strongly related features and ignore weak correlations. The inference process uses a parallel message passing mechanism to decompose the global probability graph into several sub-trees, each of which calculates independently and then merges the results through a consistency protocol. The spatiotemporal continuity constraint is implemented as an energy function, defining a penalty term for state jumps between adjacent time slices and a smoothing term for differences in spatially adjacent sensor readings.

[0055] The implementation of Bayesian update considers the real-time requirement. The complete probability update period is set to 200ms, and the observation evidence accumulated during this period is processed in chronological order. For high-frequency abnormal events, a sliding window aggregation strategy is used to combine consecutive events of the same type into a composite evidence. The update calculation uses logarithmic space operations to avoid numerical underflow caused by continuous multiplication. The integration of clinical knowledge rules is realized through soft constraints, which converts expert experience into a bias term of the probability distribution rather than a hard limit.

[0056] The decision stage of the joint probability output includes uncertainty quantification. In addition to calculating the maximum a posteriori state, the suboptimal hypothesis and its probability gap are also output simultaneously. When the confidence difference between the optimal and suboptimal states is less than a threshold, a fuzzy decision processing flow is triggered. This flow retrieves similar cases from the history database and adjusts the final judgment by referring to the majority voting result. The output interface generates a structured report containing three parts: probability distribution visualization, key evidence list, and decision basis explanation.

[0057] The system is implemented using a microservices architecture, with the optimization algorithm module deployed on edge computing nodes and the probability inference module running on cloud servers. The two modules communicate through an encrypted channel, transmitting reduced feature vectors rather than raw data, which protects user privacy and reduces communication overhead. The algorithm parameter management uses a version control system, retaining complete configuration snapshots for each update to support rapid rollback and difference comparison. The performance monitoring dashboard displays real-time running indicators such as computation delay, memory usage, and convergence curve, providing a basis for resource scheduling.

[0058] The adaptive learning mechanism continuously optimizes model performance. Weekly model diagnosis tests are automatically performed to check the performance of each matching score on the validation set. When performance degradation is detected, an incremental training process is started, using an elastic weight consolidation algorithm to balance new and old knowledge. The integration of user feedback data is designed as a dual-channel mechanism, with explicit feedback directly adjusting the probability graph parameters and implicit feedback indirectly optimizing the observation model through behavior log analysis. The long-term tracking database records the evolution history of weight coefficients, establishing an association analysis between parameter changes and changes in user physiological characteristics.

[0059] The exception handling mechanism ensures system robustness. The input data anomaly detection module identifies outliers that exceed the physiological range, triggering a sensor calibration process. The algorithm convergence anomaly monitoring module detects oscillation divergence and other conditions, automatically switching to a backup optimization strategy. The probability inference exception handling module activates a conservative decision-making mode when encountering contradictory evidence and requests manual review. The system health status is continuously monitored through a heartbeat mechanism, and any component anomalies will trigger a hierarchical alarm, escalating from automatic restart to manual intervention.

[0060] The visualization analysis tool supports deep parameter exploration. The three-dimensional space projection graph of weight coefficients displays the particle swarm optimization trajectory, with different colors marking the optimal solutions of each generation. The probability graph browser interactively displays the dependency strength between nodes, supporting hypothesis scenario simulation and counterfactual reasoning. The decision path tracing function reproduces the derivation chain of a specific conclusion, labeling the influence weight of key evidence. These tools are used not only for algorithm debugging and optimization but also as an auxiliary means for clinical teaching.

[0061] The knowledge update process is semi-automated. Newly published clinical research conclusions are extracted by natural language processing technology to extract key parameters and generate model adjustment recommendations. The expert review interface provides difference comparison and impact prediction functions to assist manual review and decision-making. Approved updates are verified in a shadow mode first, and the output differences between the new and old models are compared in a parallel environment to confirm safety and effectiveness before being deployed to the production system. The version change notification module automatically generates technical document update instructions and clinical use guidelines.

[0062] The feedback data collection system after the generation of oral health detection instructions uses a multi-channel asynchronous processing architecture. The monitoring of the pressure recovery rate of the abnormal area is achieved through a high dynamic range pressure sensing array, which continuously tracks the changes in the mechanical properties of the abnormal marker area at a sampling frequency of 100 Hz. The recovery rate calculation uses a piecewise linear regression method to identify the different stage rates of pressure value regression to the baseline. The temperature fluctuation deviation value is obtained in combination with a distributed infrared temperature measurement module, establishing a 3x3 temperature measurement grid around the abnormal area, and the baseline temperature value is taken from the moving percentile of the user's historical data. The quantification of the health index change amount is based on time series difference analysis, comparing the offset amplitude of key coordinates in the feature space before and after the execution of the instruction.

[0063] The preprocessing of feedback data includes two key steps: outlier filtering and time series alignment. Outlier filtering uses a density-based clustering algorithm to identify and eliminate outliers that deviate significantly from the main distribution. Time series alignment considers sensor response delays and determines the optimal time offset for each data stream through cross-correlation analysis. The data quality assessment module calculates the integrity index and signal-to-noise ratio of each feedback parameter, and channel data below the threshold will be marked and excluded from analysis. The preprocessed feedback data stream is divided into three parallel processing branches, which are used to update the oral baseline features, adjust the component complexity index, and optimize the detection instruction parameters.

[0064] The dynamic update mechanism of the oral baseline features uses an incremental learning strategy. The update speed of the resting state baseline is slower, and the exponential weighted moving average algorithm is used, with the new data weight not exceeding 10%. The update speed of the functional state baseline (such as chewing, light biting) is faster, and the new data weight can reach 30%, to capture changes in user behavior patterns more quickly. The baseline feature library maintains multiple time dimension reference values, including weekly average, monthly trend line, and quarterly benchmark. The update decision is realized through hypothesis testing, and when the KL divergence of the new data distribution and the historical baseline exceeds the preset threshold, the baseline feature recalibration process is triggered. The recalibration process includes a manual review link, requiring the user to perform a standardized action sequence to verify the reliability of the automatic update results.

[0065] The adjustment of the component complexity indicator adopts a parameter adaptive framework. The short-term fluctuations are smoothed by Kalman filtering, preserving the trend changes while suppressing random noise. The tracking of long-term evolution uses time series decomposition techniques, decomposing the complexity changes into seasonal and trend components. The indicator normalization reference system establishes a dynamic expected range based on user demographics such as age, gender, etc. Additional clinical confirmation is required for adjustments beyond the expected range. The coupling relationship between the complexity indicator and the physiological response frequency is periodically evaluated by cross-spectral analysis, and the re-matching process is started when the coherence coefficient is below 0.5.

[0066] The user interaction interface provides multi-level feedback visualization. The real-time monitoring view displays the stress recovery process in the form of a heat map, with color gradients representing the recovery progress in different regions. The historical comparison view superimposes the current and past health indicator change trajectories, highlighting statistically significant differences. The early warning management interface allows users to confirm or question system detection results, and these interaction data are used as important feedback signals for model optimization. The mobile application synchronously pushes a concise summary, describing the clinical implications of the feedback data in standardized terminology.

[0067] Data security and privacy protection mechanisms are integrated throughout the feedback processing flow. The collected raw feedback data are anonymized at the sensor end, removing direct personal identifiers. End-to-end encryption is used during transmission, and strict access control policies are implemented for stored data. The use of feedback data is limited to algorithm optimization and personalized service improvement, and without explicit authorization, it cannot be used for any secondary purposes. Data retention strategies implement hierarchical management, with raw signals retained for only 7 days, feature-level data retained for 1 year, and aggregated analysis results saved for a long time.

[0068] The system maintenance module establishes a complete feedback loop monitoring system. Data pipeline health monitoring includes throughput, delay, and error rate, etc. The algorithm performance dashboard tracks the evolution trend of key indicators over time, with automatic warning lines set. Hardware state monitoring records sensor precision degradation, battery consumption, and other device parameters, providing predictive maintenance requirements. All monitoring data are included in the system health score model, and when the comprehensive score is below the threshold, a hierarchical response mechanism is triggered, from automatic resource allocation to manual intervention inspection.

[0069] The clinical integration module realizes seamless integration of feedback data and medical processes. Cases with continuously deteriorating abnormalities automatically generate referral recommendations, accompanied by complete trend analysis charts. Cases with significant improvements produce rehabilitation progress reports, marking key turning points and possible influencing factors. All clinical outputs use a standardized medical terminology system and are compatible with general electronic health record systems. The doctor review interface highlights the uncertainty areas identified by the system, requesting focused attention on these boundary conditions.

[0070] The long-term adaptation framework ensures the system continuously adapts to user changes. The annual comprehensive evaluation process recalibrates all underlying parameters, referencing the latest demographic information and health status of the user. The quarterly feature drift detection analyzes the progressive changes in oral patterns, identifying algorithm modules that need adjustment. The monthly performance audit verifies the effectiveness of the feedback mechanism, ensuring the optimization direction aligns with clinical goals. These adaptation mechanisms at different time scales form a three-dimensional self-updating system, enabling the system to maintain accuracy and utility over years of use.

[0071] The exception handling protocol covers special scenarios throughout the feedback data lifecycle. Data completion in sensor failure cases employs generative adversarial networks, synthesizing reasonable values based on adjacent node data and historical patterns. Data recovery during transmission interruptions uses differential synchronization protocols, retransmitting only missing or corrupted data segments. The algorithm anomaly detection module identifies feedback patterns that do not conform to physiological rules, triggering data reacquisition processes. All exception events are recorded in audit logs, including occurrence time, handling measures, and result verification information.

[0072] The user education component helps understand the value and limitations of the feedback mechanism. The introductory tutorial demonstrates the correct response to typical feedback scenarios. Regular knowledge pushes explain the impact of the latest optimization's detection parameters on user experience. The FAQ library covers various levels from technical operations to clinical implications. These educational resources are presented in multimedia formats, dynamically adjusting content depth and presentation methods based on user learning progress and preferences.

[0073] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action without necessarily requiring or implying that any such entity or action are in fact mutually exclusive or directional. Moreover, the terms "include", "have", or any variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0074] While the embodiments of the application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made therein without departing from the spirit and scope of the application, which is defined by the following claims and their equivalents.

Claims

1. A smart mouthguard-based oral detection method, characterized in that, The method comprises the following steps: acquiring original oral cavity pressure time series and corresponding frequency domain features collected by a user wearing a smart dental brace, determining optimization decomposition parameters corresponding to the frequency domain features, performing multi-scale signal decomposition on the original oral cavity pressure time series based on the optimization decomposition parameters, and generating multi-scale oral cavity signal components; extracting nonlinear dynamic features of the multi-scale oral cavity signal components, and acquiring component complexity indicators; screening key signal components related to the oral cavity health status according to the multi-scale oral cavity signal components and the component complexity indicators; acquiring oral cavity baseline features of the user, real-time occlusion states, and combining the component complexity indicators to generate individual physiological response frequencies; reconstructing oral cavity physiological feature sequences based on the key signal components and the individual physiological response frequencies; performing modal decomposition on the oral cavity physiological feature sequences to separate low-frequency components reflecting steady-state physiological characteristics and high-frequency components reflecting dynamic changes; identifying oral cavity abnormal categories of the user according to the low-frequency components and the high-frequency components, establishing a mapping relationship between the abnormal categories and diagnostic results, and generating oral cavity health detection instructions.

2. The method of claim 1, wherein, The method comprises the following steps: acquiring oral cavity baseline features of the user, real-time occlusion states, and component complexity indicators, comprising: collecting multi-stage oral cavity baseline data, wherein the multi-stage includes a resting state, a light occlusion state, and a chewing task state; extracting occlusion rhythm features from the multi-stage oral cavity baseline data, generating real-time occlusion state labels and component complexity indicators based on the occlusion rhythm features; calculating matching degree scores of preset candidate frequencies based on the frequency band energy distribution of the oral cavity baseline features, the real-time occlusion state labels, and the component complexity indicators, wherein the matching degree scores include physiological rhythm consistency scores, state adaptability scores, and complexity correlation scores; 3. The method of claim 1, wherein, dynamically adjusting weight coefficients of the matching degree scores by using an optimization algorithm to determine individual physiological response frequencies. The method comprises the following steps: processing the original oral cavity pressure time series by using an adaptive filtering algorithm combined with the optimization decomposition parameters; and suppressing end effect by using a period extension method at a sequence boundary; 4. The method of claim 1, wherein, extracting a multi-scale coefficient set after decomposition, and constructing the multi-scale oral cavity signal components based on the multi-scale coefficient set. The method comprises the following steps: constructing a time-frequency feature matrix, wherein elements of the time-frequency feature matrix are generated by nonlinear mapping of amplitude values of the key signal components, frequency band weights corresponding to the individual physiological response frequencies, and preset physiological modulation coefficients; performing multi-scale time series analysis on the time-frequency feature matrix, including short-time scale analysis to capture occlusion transient features, medium-time scale analysis to track physiological state evolution, and long-time scale analysis to monitor health trends; 5. The method of claim 1, wherein, generating oral cavity physiological feature sequences containing multi-mode physiological features by dynamically weighting and fusing multi-scale time series analysis results. The method comprises the following steps: Detecting extreme points of the oral physiological feature sequence and performing mirror extension processing, extracting multi-order modal components satisfying intrinsic conditions through iterative screening; According to the average period and energy concentration degree of the multi-order modal components, the short-period components of the first order are classified as high-frequency components, and the long-period components of the last order and the residual terms are classified as low-frequency components.

6. The method of claim 1, wherein, According to the low-frequency components and high-frequency components, the user's oral cavity abnormality category is identified, including: Extracting time-domain statistical features, energy distribution features, and component correlation features from the low-frequency components to construct a steady-state feature vector; Extracting instantaneous change rate, peak feature, and waveform distortion feature from the high-frequency components to construct a transient feature vector; Using a hierarchical classification strategy to classify the steady-state feature vector into a health state, and identifying the abnormal event corresponding to the transient feature vector through an adaptive detector; Fusing the health state classification result and the abnormal event, combining the spatiotemporal continuity constraint for confidence verification, and outputting the oral cavity abnormality category.

7. The method of claim 1, wherein, The obtained original oral pressure time sequence collected by the user wearing a smart dental cover includes: Real-time capture of multi-point pressure data and temperature auxiliary data in the oral cavity; Classifying and processing occlusal layer, tongue pressure layer, and buccal mucosa layer signals to analyze multi-source physiological parameters; Integrating and standardizing the processed data to generate a structured original oral pressure time sequence.

8. The method of claim 2, wherein, The weight coefficient of the matching degree score is dynamically adjusted by using an optimization algorithm, including: Initializing the position and speed parameters of the particle swarm, and defining the target function as the weighted harmonic mean of the matching degree score; Iteratively updating the particle position and calculating the global optimal solution, and outputting the weight coefficient optimization result when the target function converges.

9. The method of claim 6, wherein, The fusion of health state classification results and abnormal events includes: Constructing a dynamic probability graph model, taking the health state classification result as the node prior probability, and taking the abnormal event as the observation evidence; Update the node probability distribution through Bayesian inference, and generate joint probability output combined with state transition constraints.

10. The method of claim 1, wherein, After generating the oral health detection instruction, including: Real-time acquisition of instruction execution feedback data, including abnormal area pressure recovery rate, temperature fluctuation deviation value, and health index change amount; According to the feedback data, dynamically update the oral baseline features and component complexity indicators.

Images