An oral cavity detection method based on smart braces

Abstract

This invention relates to the field of oral health monitoring technology and discloses an oral detection method based on smart braces. The method acquires raw oral pressure time series and frequency domain features using smart braces, determines optimized decomposition parameters, and performs multi-scale decomposition on the raw signal to generate multi-scale oral signal components. It extracts nonlinear dynamic features and obtains component complexity indices to screen out key signal components related to oral health. Personalized physiological response frequencies are generated by combining the user's oral baseline features, real-time occlusal status, and component complexity indices. Based on the key signal components and personalized physiological response frequencies, an oral physiological feature sequence is reconstructed, and low-frequency and high-frequency components are separated through modal decomposition. Oral abnormality categories are identified based on these two components, a mapping relationship is established with diagnostic results, and oral health detection instructions are generated. This method enables convenient and comprehensive oral detection, improves detection adaptability and accuracy, and provides support for oral health management.

Description

Technical Field

[0001] This invention relates to the field of oral health monitoring technology, specifically to an oral health monitoring method based on smart braces. Background Technology

[0002] With increasing awareness of oral health, early screening and dynamic monitoring of oral diseases have become important topics in modern preventive medicine. Traditional oral examinations rely on equipment from specialized medical institutions and the experience of physicians, resulting in problems such as long testing cycles, high costs, and reliance on subjective judgment, making it difficult to meet the public's demand for real-time, convenient, and personalized oral health management. Currently, oral examination technologies on the market are mainly divided into two categories: invasive and non-invasive. Invasive examinations, such as oral endoscopy and periodontal probes, can obtain high-precision data, but the operation process can easily cause discomfort to patients and cannot achieve long-term continuous monitoring. Non-invasive examinations, such as oral CT and ultrasound scanning, can provide three-dimensional structural information, but the equipment is bulky and expensive, only suitable for medical institutions, and difficult to popularize in home settings.

[0003] In recent years, the application of wearable devices in health monitoring has driven innovation in oral examination technology. Smart braces, as wearable devices that conform to the oral cavity structure, have the potential to continuously collect oral physiological signals, but significant technical bottlenecks remain in signal processing and health status recognition. Existing detection methods based on smart braces mostly rely on single-dimensional signal features (such as pressure peaks and occlusal frequency), ignoring the nonlinear and multi-scale characteristics of oral physiological signals—oral pressure signals contain both high-frequency components generated by dynamic movements such as chewing and swallowing, and low-frequency information about the steady-state changes of periodontal tissues. Single feature extraction methods are insufficient to comprehensively reflect oral health status.

[0004] Individual oral physiological characteristics vary significantly, such as bite force, tooth alignment, and periodontal tissue condition, all of which have individualized features. Generalized signal analysis models often result in insufficient detection accuracy, making it difficult to accurately identify early oral diseases (such as gingivitis, periodontitis, and malocclusion). Furthermore, existing technologies lack adaptive optimization mechanisms in signal decomposition and feature selection, making it difficult to effectively separate noise from valid signals. This leads to insufficient extraction of key physiological features, affecting the accuracy of subsequent health status assessments.

[0005] Developing an intelligent detection method that can adaptively process oral pressure signals, extract personalized physiological features, and accurately identify abnormal oral conditions is of great significance for improving the convenience, real-time performance, and accuracy of oral health monitoring, and is also an important research direction in the field of wearable medical devices. Summary of the Invention

[0006] The purpose of this invention is to provide an oral examination method based on smart braces to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an oral cavity detection method based on smart braces, the method comprising:

[0008] The original oral pressure time series and corresponding frequency domain features collected by the user wearing smart braces are obtained. The optimized decomposition parameters corresponding to the frequency domain features are determined. Based on the optimized decomposition parameters, the original oral pressure time series is decomposed into multi-scale signals to generate multi-scale oral signal components.

[0009] Extract the nonlinear dynamic features of the multi-scale oral signal components and obtain the component complexity index;

[0010] Based on the multi-scale oral signal components and component complexity index, key signal components related to oral health status are screened.

[0011] The system acquires the user's oral baseline features and real-time occlusal status, and combines them with component complexity indicators to generate personalized physiological response frequencies.

[0012] Based on the key signal components and personalized physiological response frequencies, the oral physiological feature sequence is reconstructed; the oral physiological feature sequence is then subjected to modal decomposition to separate the low-frequency components reflecting steady-state physiological characteristics and the high-frequency components reflecting dynamic changes.

[0013] Based on the low-frequency and high-frequency components, the user's oral abnormality category is identified, a mapping relationship between the abnormality category and the diagnostic result is established, and an oral health detection instruction is generated.

[0014] Preferably, the acquisition of the user's oral baseline features, real-time occlusal status, and component complexity index includes:

[0015] Collect multi-stage oral baseline data, including resting state, light occlusion state, and chewing task state;

[0016] Occlusal rhythm features are extracted from the multi-stage oral baseline data, and real-time occlusal state labels and component complexity indices are generated based on the occlusal rhythm features.

[0017] Based on the frequency band energy distribution, real-time occlusal state label and component complexity index of the oral baseline features, a matching score for a preset candidate frequency is calculated. The matching score includes physiological rhythm consistency score, state adaptability score and complexity correlation score.

[0018] An optimization algorithm is used to dynamically adjust the weight coefficients of the matching score to determine the personalized physiological response frequency.

[0019] Preferably, the step of performing multi-scale signal decomposition on the original oral pressure time series to generate multi-scale oral signal components includes:

[0020] The original oral pressure time series is processed using an adaptive filtering algorithm combined with the optimized decomposition parameters; the endpoint effect is suppressed at the sequence boundaries using periodic extension.

[0021] Extract the decomposed multi-scale coefficient set, and construct the multi-scale oral signal component based on the multi-scale coefficient set.

[0022] Preferably, the reconstructing of the oral physiological feature sequence based on the key signal components and personalized physiological response frequencies includes:

[0023] A time-frequency feature matrix is ​​constructed, the elements of which are 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.

[0024] Multi-scale time series analysis is performed on the time-frequency feature matrix, including short-scale analysis to capture transient features of occlusion, medium-scale analysis to track the evolution of physiological state, and long-scale analysis to monitor health trends.

[0025] By dynamically weighting and fusing the results of multi-scale time series analysis, an oral physiological feature sequence containing multi-modal physiological characteristics is generated.

[0026] Preferably, the modal decomposition of the oral physiological feature sequence to separate the low-frequency components reflecting steady-state physiological characteristics and the high-frequency components reflecting dynamic changes includes:

[0027] The extreme points of the oral physiological feature sequence are detected and mirror continuation processing is performed. Multi-mode components that satisfy the intrinsic conditions are extracted through iterative sieving.

[0028] Based on the average period and energy concentration of the multi-mode components, the first short-period components are classified as high-frequency components, and the second long-period components and residual terms are classified as low-frequency components.

[0029] Preferably, identifying the user's oral cavity abnormality category based on the low-frequency component and the high-frequency component includes:

[0030] Time-domain statistical features, energy distribution features, and component correlation features are extracted from the low-frequency components to construct a steady-state feature vector;

[0031] Instantaneous rate of change, peak features, and waveform distortion features are extracted from the high-frequency components to construct a transient feature vector;

[0032] A hierarchical classification strategy is used to classify the steady-state feature vectors into health states, and an adaptive detector is used to identify the abnormal events corresponding to the transient feature vectors.

[0033] By integrating health status classification results and abnormal events, and combining spatiotemporal continuity constraints for confidence verification, the oral cavity abnormality category is output.

[0034] Preferably, the acquisition of the original oral pressure time series collected by the user wearing smart braces includes:

[0035] Real-time capture of multi-point pressure data and auxiliary temperature data within the oral cavity;

[0036] The signals from the occlusal layer, lingual depressor layer, and buccal mucosa layer were classified and processed to analyze multi-source physiological parameters.

[0037] The processed data is integrated and standardized to generate a structured raw oral pressure time series.

[0038] Preferably, the step of dynamically adjusting the weight coefficients of the matching score using an optimization algorithm includes:

[0039] Initialize the position and velocity parameters of the particle swarm, and define the objective function as the weighted harmonic mean of the matching score;

[0040] Iteratively update particle positions and calculate the global optimal solution. When the objective function converges, output the optimized weight coefficients.

[0041] Preferably, the fusion of health status classification results and abnormal events includes:

[0042] A dynamic probabilistic graphical model is constructed, with the health status classification result as the prior probability of the node and abnormal events as observational evidence.

[0043] The node probability distribution is updated through Bayesian inference, and a joint probability output is generated by combining state transition constraints.

[0044] Preferably, after generating the oral health detection instruction, the process includes:

[0045] Real-time acquisition of instruction execution feedback data, including abnormal area pressure recovery rate, temperature fluctuation deviation value, and changes in health indicators;

[0046] The oral baseline features and component complexity index are dynamically updated based on the feedback data.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This oral detection method based on smart braces collects raw oral pressure time series and corresponding frequency domain features through smart braces, enabling real-time and dynamic monitoring of oral signals. It eliminates the reliance on traditional large-scale medical equipment, improves the convenience of detection, and facilitates users' oral health management in daily life.

[0049] In the signal processing stage, the optimization decomposition parameters are determined based on the frequency domain characteristics, and the original signal is decomposed into multi-scale components to generate multi-scale oral signal components. This multi-scale decomposition method can decompose complex oral signals into components of different scales, capture various feature information in oral signals in greater detail, and help with subsequent in-depth analysis of the signal.

[0050] Extracting the nonlinear dynamic features of multi-scale oral signal components and obtaining component complexity indices, and combining multi-scale oral signal components and component complexity indices to screen key signal components, can effectively focus on important information related to oral health status, reduce interference from irrelevant signals, and make subsequent analysis more targeted.

[0051] By acquiring the user's oral baseline features and real-time occlusal status, and combining them with component complexity indicators, personalized physiological response frequencies are generated. This fully considers the differences in individual physiological characteristics, making the detection more suitable for the actual situation of different users, and helping to improve the adaptability and accuracy of the detection.

[0052] Based on key signal components and personalized physiological response frequencies, the oral physiological feature sequence is reconstructed and modal decomposed to separate low-frequency components reflecting steady-state physiological characteristics and high-frequency components reflecting dynamic changes. This separation method can analyze the steady-state and dynamic characteristics of the oral cavity separately, and comprehensively grasp the health status of the oral cavity.

[0053] Based on low-frequency and high-frequency components, oral abnormality categories are identified and a mapping relationship is established with diagnostic results. Oral health testing instructions are generated, enabling the test results to be directly associated with diagnostic results. This provides a more intuitive and effective basis for oral health assessment, helps to detect potential oral problems in a timely manner, and assists in oral health management. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the working principle of the oral cavity detection method based on smart braces described in this invention.

[0055] Figure 2 This is a flowchart of multi-scale signal decomposition.

[0056] Figure 3 This is a flowchart of mode decomposition.

[0057] Figure 4 The flowchart for obtaining the original oral pressure time series. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1 This invention provides an oral cavity detection method based on smart braces, the method comprising:

[0060] The system collects raw oral pressure time-series data and frequency domain features from the user's mouth using smart braces. First, it determines the optimized decomposition parameters corresponding to the frequency domain features. Based on these parameters, it performs multi-scale signal decomposition on the raw oral pressure time-series, generating multi-scale oral signal components. Then, it extracts the nonlinear dynamic features of the multi-scale oral signal components and calculates the component complexity index. Based on the multi-scale oral signal components and the component complexity index, key signal components related to oral health status are selected. Combining the user's baseline oral features, real-time occlusal status, and the component complexity index, a personalized physiological response frequency is generated. Based on the key signal components and the personalized physiological response frequency, an oral physiological feature sequence is reconstructed, and modal decomposition is performed on this sequence to separate low-frequency and high-frequency components. Finally, based on the low-frequency and high-frequency components, the system identifies the user's oral abnormality categories, establishes a mapping relationship between abnormality categories and diagnostic results, and generates oral health detection instructions.

[0061] Example 1: See Figure 2 The multi-stage baseline oral data acquisition process is achieved through a pressure sensor array built into the smart braces, covering three typical physiological scenarios: resting state, light occlusion state, and chewing task state. In the resting state, the user keeps their mouth naturally closed, and the sensors record the baseline pressure distribution under no occlusal load. The light occlusion state requires the user to close their teeth with a constant force, simulating unconscious everyday biting behavior. The chewing task state is activated through standardized chewing movements, and test materials of specific hardness and size are used to ensure data comparability. The sensor sampling frequency is set within the range of 200Hz to 1000Hz to balance high-frequency dynamic capture and data storage efficiency.

[0062] Occlusal rhythm features were extracted using a sliding window analysis method, with the window length adaptively adjusted based on the user's average occlusal cycle. Within each window, time-domain statistics of the pressure signal were calculated, including peak interval, amplitude variation coefficient, and rise slope. Resting state data were used to establish an individualized noise floor, light bite state data revealed the basic occlusal pattern, and chewing task data contained richer time-frequency features. By comparing the power spectral density differences among the three data sets, characteristic frequency bands related to functional movements were identified. Real-time occlusal state labels were generated using a finite state machine model, classifying continuous pressure signals into three discrete states: resting, light bite, or chewing. Component complexity indices were calculated based on a multi-scale entropy algorithm, reflecting the degree of irregularity of the signal in both the time and frequency domains.

[0063] The frequency energy distribution of oral baseline features was obtained through short-time Fourier transform, dividing it into four physiological frequency bands: delta, theta, alpha, and beta. The delta band corresponds to slow fluctuations of 0.5-4 Hz, the theta band covers the basal rhythm of 4-8 Hz, the alpha band (8-13 Hz) is associated with autonomic nervous regulation, and the beta band (13-30 Hz) reflects rapid muscle control. The energy proportion of each frequency band and the real-time occlusal state label constitute a three-dimensional feature space. Mahalanobis distance was used to measure the matching degree between candidate frequencies and the feature space. The physiological rhythm consistency score measures the degree of agreement between candidate frequencies and the user's historical dominant rhythm, the state adaptability score assesses the ability of frequencies to distinguish different occlusal states, and the complexity correlation score characterizes the correlation between frequency components and the complexity of oral motor functions.

[0064] The optimization algorithm employs an improved particle swarm optimization framework, where the particle dimension corresponds to the weight coefficients of the three matching scores. The initial population size is set to 50, particle positions are randomly initialized in the [0,1] interval, and velocity parameters are limited to the [-0.1, 0.1] range to prevent premature convergence. The objective function is defined as the weighted harmonic mean of the three scores, with a nonlinear decay factor introduced to balance the exploration and development phases. In each iteration, the particle updates its velocity vector based on its historical best solution and the global best solution. After position updates, boundary reflection processing is performed to avoid getting trapped in local optima. The convergence condition is set to the global best solution changing by less than 1e-5 over 20 consecutive generations, and the final output weight coefficient combination ensures the objective function reaches Pareto optimality.

[0065] Preprocessing of the raw oral pressure time series included baseline drift correction and motion artifact removal. The adaptive filter was designed using the LMS algorithm, with the step size dynamically adjusted based on the signal-to-noise ratio. The number of decomposition layers was determined by optimizing the decomposition parameters, and over-decomposition was automatically terminated through spectral flatness detection. Boundary processing employed a symmetric periodic extension strategy, mirroring the endpoint extrema to both ends of the sequence to form virtual extension segments. The multi-scale coefficient set was constructed using discrete wavelet transform, with Daubechies wavelet basis functions selected to ensure time-frequency localization. Detail coefficients and approximation coefficients were recombined after threshold denoising to form 6-8 components with clear physiological interpretations, corresponding to different sources such as the dominant frequency of occlusal force, harmonic components, and electromyographic interference.

[0066] The physical meaning of multi-scale oral signal components was verified through synchronized electromyography (EMG) signals. Surface EMG electrode arrays of the temporalis and masseter muscles provided independent references for coherence analysis with the pressure signal components. Components with a coherence coefficient exceeding 0.7 were labeled as valid physiological components, while the rest were considered environmental noise or device artifacts. Valid components were further classified through cluster analysis, and a component dictionary was 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.

[0067] The real-time bite state label update mechanism adopts an event-driven model, triggering reclassification when the pressure signal amplitude exceeds three times the resting state standard deviation. State duration statistical features are incorporated into the observed variables of the Hidden Markov Model, and the Viterbi algorithm decodes the most probable state sequence. The sliding window for the component complexity index is set to 5 seconds, updated every 1 second to balance real-time performance and stability. Index normalization employs a dynamic Z-score method, standardizing based on the moving mean and standard deviation of historical user data.

[0068] The final determination of personalized physiological response frequencies is achieved through multi-criteria decision-making. The candidate frequency list is sorted in descending order of total matching score, and the top 5% of frequencies are entered into an expert system for physiological rationality verification. Verification rules include whether the frequency falls within the typical range of human jawbone resonance (2-8Hz) and whether it has a known association with the user's past medical history. The verified frequencies are output as personalized physiological response frequencies to guide subsequent feature reconstruction. The entire process is implemented in an embedded system, with computational latency controlled within 50ms to meet real-time requirements.

[0069] Residual analysis of signal decomposition is used to monitor system health. When the residual energy exceeds 15%, a sensor calibration process is triggered, executing a self-test sequence via a built-in vibration motor. Adaptive adjustment of decomposition parameters employs a reinforcement learning framework, using the mutual information between the reconstructed and original signals as the reward signal. Optimization of the periodic extension length is achieved through cross-validation, selecting the extension ratio that minimizes the energy of the endpoint components. Storage of the multi-scale coefficient set utilizes sparse coding, retaining only the top 10% of coefficients in terms of amplitude to reduce transmission bandwidth requirements.

[0070] Dynamic maintenance of oral baseline features employs an incremental learning mechanism. One minute of standard occlusion data is automatically collected weekly, and distribution drift is detected using KL divergence. When the difference in distribution between new and old data exceeds a threshold, a baseline feature retraining process is triggered. Training data is weighted and fused from recent and historical samples, with exponential decay used to assign different weights to new and old data. Long-term trend analysis of component complexity indicators uses a seasonal decomposition method to separate the influence of slowly changing factors such as aging and wearing habits. This mechanism enables the system to continuously adapt to the natural evolution of the user's physiological state.

[0071] Example 2: See Figure 3 The construction of the time-frequency feature matrix is ​​based on the amplitude characteristics of key signal components, the frequency band weights of personalized physiological response frequencies, and preset physiological modulation coefficients. The amplitude of the key signal components is extracted using Hilbert transform to obtain the instantaneous envelope, eliminating the interference of high-frequency oscillations on amplitude estimation. The frequency band weights of personalized physiological response frequencies are obtained from analysis of user historical data, reflecting the ability of different frequency bands to represent individual oral conditions. The physiological modulation coefficients are derived from large-scale clinical research data, describing the statistical correlation between typical oral abnormalities and energy changes in specific frequency bands. These three elements generate matrix elements through a nonlinear mapping function, which employs a combination of piecewise linear approximation and Sigmoid activation, enhancing numerical stability while preserving the physical meaning of the features.

[0072] The matrix's row dimensions correspond to discrete sampling points in the time series, while the column dimensions cover the physiological frequency band from 0.5Hz to 30Hz. The generation process for each matrix element includes a dynamic normalization step, transforming the original eigenvalues ​​to the [0,1] interval to eliminate dimensional differences. A non-uniform sampling strategy is employed on the time axis, increasing sampling density in regions of rapid signal change and reducing the sampling rate in stable regions to optimize storage efficiency. Frequency band division uses an overlapping sub-band design, with adjacent bands exhibiting 15%-20% frequency overlap, avoiding the feature truncation effect caused by strict band division. The application of physiological modulation coefficients introduces adaptive gain control, dynamically adjusting the contribution weights of different frequency bands based on the user's real-time biting force.

[0073] Multi-scale time series analysis is implemented at three levels: short-term, medium-term, and long-term. Short-term analysis focuses on a 50-200ms time window, capturing transient features of occlusal movements using differential operators, including microscopic parameters such as pressure rise rate and peak hold time. A sliding window strategy is employed, with a window step size of 10ms to ensure feature continuity. Transient feature extraction is combined with morphological filtering techniques to eliminate spurious transients caused by sensor noise. Medium-term analysis covers a 5-30 second time range, using recurrent neural networks to model the evolution trajectory of physiological states. Hidden layer units store temporal dependencies, and gating mechanisms control information transmission paths, distinguishing between normal physiological fluctuations and abnormal state transitions. Long-term analysis processes data blocks longer than 5 minutes, using trend decomposition algorithms to separate periodic patterns from baseline drift. Monitoring indicators for health trends include slowly varying parameters such as the daily average occlusal strength coefficient of variation and high-frequency energy accumulation distribution.

[0074] A dynamic weighted fusion algorithm integrates multi-scale analysis results, with weight allocation following the principle of scale correlation. Short-term feature weights are positively correlated with the frequency of biting movements, medium-term feature weights depend on the state transition probability, and long-term feature weights are determined by the temporal consistency of health indicators. A conflict resolution mechanism is introduced into the fusion process; when features at different scales contradict each other, the scale conclusion with higher statistical significance is prioritized. The final generated oral physiological feature sequence contains three data channels: a transient feature channel records micro-motion parameters, a state evolution channel encodes physiological stage markers, and a trend monitoring channel stores standardized values ​​of health indicators. The temporal resolution of the sequence is uniformly set to 1 second / frame, and missing data is imputed using cubic spline interpolation.

[0075] Modal decomposition of oral physiological feature sequences employs an improved extremum detection algorithm. Extremum point localization combines polynomial fitting and gradient analysis to avoid false extremum markers caused by noise. Mirror continuation symmetrically replicates the extremum distribution at both ends of the sequence, with the continuation length being 1-2 times the signal's main period. The iterative screening process uses an adaptive stopping criterion, terminating the loop when the mean square error change rate of two consecutive screening results is less than 5%. The criteria for intrinsic modal components are relaxed to allow for moderate harmonic aliasing to preserve the complex modulation characteristics of physiological signals. Envelope symmetry constraints are introduced during screening, forcing the upper and lower envelopes to be symmetrical about the zero mean, suppressing non-physiological bias components.

[0076] The classification of multimodal components is based on two dimensions: mean period and energy concentration. The mean period is calculated statistically using the zero-crossing rate, with a 0.5-second period threshold as the distinguishing criterion. Energy concentration is measured using the frequency band energy ratio, calculating the energy proportion of the component in the 4-8Hz frequency band. Early-order short-cycle components typically exhibit dense oscillation characteristics, with a mean period less than 0.5 seconds and energy concentrated in the high-frequency region; these are classified as high-frequency components. These components primarily reflect dynamic events such as rapid contraction of masticatory muscles and transient tooth contact. Later-order long-cycle components display a slow fluctuation pattern, with a mean period exceeding 0.5 seconds and energy distributed in the low-frequency region; these, along with the residual term, are classified as low-frequency components. These components characterize homeostatic physiological processes such as salivary secretion rhythm and long-term occlusal adaptation.

[0077] The analysis of high-frequency components focuses on characterizing time-varying properties. Instantaneous rate of change is calculated using a five-point differential formula, supplemented by median filtering to smooth differential noise. Peak feature detection employs a dual-threshold strategy: the primary threshold identifies significant peak clusters, and the secondary threshold determines peak boundaries. Waveform distortion assessment is based on a dynamic time warping algorithm, aligning the real-time waveform with a standard template to calculate morphological differences. Low-frequency component processing focuses on periodic analysis. Time-domain statistical features include descriptive indicators such as moving averages and standard deviation bands. Energy distribution characteristics are estimated using Welch periodograms, with logarithmic coordinates highlighting weak components. Component correlation analysis calculates the phase synchronization index of signals from different sensor positions to assess the coordination of oral cavity movements.

[0078] The reconstruction process employs a multi-layered verification mechanism for quality control. The completeness of the time-frequency feature matrix is ​​verified through reverse reconstruction error, requiring a correlation coefficient exceeding 0.85 between the original and reconstructed signals. The physiological rationality of the modal components is verified using a clinical knowledge base, eliminating anomalous components with average periods exceeding the human physiological range (<0.1s or >10s). Stability testing of the multi-scale fusion results utilizes leave-one-out cross-validation to check the fluctuation range of feature weights across different data subsets. The final output oral physiological feature sequence undergoes integrity verification, ensuring temporal alignment and the absence of logical contradictions in the transient, state, and trend channels.

[0079] The system implementation employs a layered processing architecture. The bottom-level signal processing runs on the embedded processor of the smart braces, completing time-frequency matrix construction and short-term feature extraction. Mesoscale analysis is deployed on the mobile terminal, utilizing the device's GPU to accelerate neural network inference. Long-scale trend processing is executed on a cloud server, optimizing model parameters using multi-user data. Data transmission employs feature compression technology, compressing the time-frequency matrix by over 80% through sparse coding. Regarding real-time performance, the end-to-end latency from signal acquisition to feature sequence output is controlled within 300ms, meeting the immediate feedback requirements of clinical monitoring.

[0080] A dynamic update mechanism continuously optimizes the feature extraction process. The frequency band division of the time-frequency matrix is ​​reassessed monthly, adjusting sub-band boundaries based on the spectral characteristics of the latest user data. Physiological modulation coefficients are updated quarterly, incorporating the latest clinical research findings. Adaptive learning of modality decomposition parameters records the residual distribution of each decomposition, progressively optimizing the sensitivity for extreme point detection. A multi-scale weighting strategy retains historical adjustment records, establishing a rule base for the association between weights and user behavior. These mechanisms enable the system to adapt to natural changes in the user's oral cavity characteristics, maintaining long-term monitoring accuracy.

[0081] Example 3: See Figure 4 The feature extraction process for low-frequency components employs a multi-dimensional analysis method. Time-domain statistical features are calculated using a sliding window framework, with the window length adjusted synchronously with the user's baseline engagement cycle. The mean parameter reflects the steady-state pressure level, and detrending processing eliminates the impact of long-term drift. Analysis of variance uses an adaptive threshold to dynamically distinguish between physiological fluctuations and abnormal variations. The skewness coefficient calculation incorporates robustness correction to reduce the interference of extreme values ​​on distribution pattern assessment. The quantification of energy distribution features is achieved through improved wavelet packet decomposition, with the decomposition tree structure dynamically optimized based on the spectral characteristics of low-frequency components. Subband energy proportion calculation employs normalization processing to eliminate differences in absolute energy values ​​between individuals. Component correlation analysis constructs a three-dimensional feature space, including three orthogonal dimensions: time-delay cross-correlation, phase synchronization index, and nonlinear coupling strength.

[0082] A multi-stage processing pipeline is established for transient feature detection of high-frequency components. The calculation of the instantaneous rate of change employs a noise-resistant differential operator to suppress high-frequency noise amplification while maintaining slope accuracy. Peak feature identification utilizes morphological filtering preprocessing to eliminate the influence of baseline fluctuations on peak detection. The peak parameter set includes temporal indices such as rise time, half-width at half-maximum (HWHM), and inter-peak interval. Waveform distortion assessment employs dynamic template matching technology, with a template library containing twenty typical anomalous waveform patterns. The matching similarity is calculated as follows:

[0083]

[0084] Where S represents the waveform similarity score, The value of the kth sampling point of the template signal. Let N be the value of the k-th sampling point of the real-time signal, and N be the length of the comparison window. This represents the positional weighting coefficient. This formula introduces amplitude normalization and spatial weighting mechanisms to improve the physiological relevance of distortion detection.

[0085] The hierarchical classification system employs a hybrid architecture. Steady-state feature vector processing utilizes an ensemble learning framework, with the base classifiers including random forests, support vector machines, and shallow neural networks. In the feature selection phase, Gini importance and recursive feature elimination scores are calculated, retaining a subset of features with high cross-validation consistency. Classification decision fusion uses fuzzy logic rules to define probability transition intervals for different health states. An adaptive detector is designed for transient feature vectors, with a core variable threshold anomaly scoring mechanism. The threshold adjustment algorithm monitors the interquartile range of the user's recent feature distribution and dynamically sets the anomaly judgment boundary. The detector output includes three metadata items: event type, severity, and spatiotemporal location.

[0086] The dynamic probabilistic graphical model is constructed using factor graphs. Nodes are divided into two categories: observation nodes and hidden state nodes. Observation nodes correspond to real-time abnormal event detection results, while hidden state nodes represent potential health states. The prior probability distribution is initialized using historical diagnostic records, and the state transition matrix is ​​set with basic parameters based on clinical guidelines. The observation likelihood function is estimated nonparametrically, and the bandwidth of the kernel density function is adaptively adjusted according to the sparsity of user data. The Bayesian inference process employs approximate variational inference to balance computational complexity and inference accuracy. The spatiotemporal continuity constraint is implemented as a Markov random field, defining state consistency penalties for adjacent time periods and adjacent sensor nodes.

[0087] The raw oral pressure data acquisition system employs a multimodal sensor array. Occlusal layer monitoring utilizes an 8×8 matrix piezoresistive sensor with a spatial resolution of 2 square millimeters. Tongue pressure layer detection employs a flexible PVDF thin-film sensor with a sampling frequency set to 500Hz to capture rapid dynamic changes. Buccal mucosa layer monitoring combines impedance sensing and optical volumetric imaging techniques to obtain auxiliary information on tissue microcirculation. Signal preprocessing includes a four-stage cascaded filtering process: a 50Hz power frequency notch filter to eliminate environmental interference, a 0.5Hz high-pass filter to remove respiratory artifacts, a 20Hz low-pass filter to suppress electromyographic noise, and an adaptive spectral subtraction algorithm to improve the signal-to-noise ratio.

[0088] Multi-source data fusion employs spatiotemporal registration technology. Time alignment is based on hardware synchronization signals, with sampling clock deviations for each sensor channel controlled within 100 microseconds. Spatial registration establishes an anatomical coordinate system, mapping data from sensors at different locations to a standard oral cavity model. The physiological parameter analysis algorithm comprises three parallel processing streams: occlusal force center trajectory tracking, tongue movement pattern recognition, and mucosal blood perfusion index calculation. Data standardization employs a hierarchical processing strategy: device-level calibration eliminates individual sensor differences, session-level normalization compensates for the influence of ambient temperature, and user-level Z-score conversion achieves cross-individual comparability.

[0089] The generation of structured time series follows the ISO / TS18234 standard. The data packet format includes three parts: header information, payload data, and integrity verification. The header information records the acquisition timestamp, sensor ID, and signal quality indicators. The payload data uses differential coding compression to reduce transmission bandwidth consumption. Integrity verification employs a combined mechanism of cyclic redundancy check and singular value decomposition, which can simultaneously detect random errors and systematic biases. The sequence storage adopts a hierarchical database architecture: the raw data layer retains the unprocessed signal, the feature layer stores the extracted time series features, and the metadata layer records processing parameters and quality control markers.

[0090] Post-processing of anomalous events comprises two stages: credibility verification and clinical interpretation. Credibility verification checks the temporal persistence, spatial consistency, and multimodal corroboration strength of the anomalous events. The clinical interpretation module accesses a knowledge graph, mapping feature space anomalies to pathological mechanism chains. The interpretation output employs natural language generation technology to produce a structured report containing probability grading, differential diagnosis, and recommended interventions. The system maintenance module continuously monitors feature extraction performance, automatically triggering sensor calibration procedures when the signal quality index falls below a threshold. The calibration sequence includes two standard protocols: stepped stress testing and frequency response testing, updating sensor characteristic parameters through calibration data analysis.

[0091] The user interface implements a closed-loop feedback mechanism. Real-time visualization displays pressure heatmaps and abnormal warning markers, with a refresh rate maintained above 30Hz. Historical trend graphs support multi-timescale scaling, evolving from minute-level to month-level observation modes. Interaction logs record user responses to warnings; this behavioral data is used to optimize anomaly detection sensitivity. The system configuration interface allows clinical professionals to adjust analysis parameters; modified parameters are protected by digital signatures to ensure traceability. The data export module supports both HL7 and FHIR medical information exchange standards, enabling seamless integration with electronic health record systems.

[0092] 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.

[0093] 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.

[0094] 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:

[0095]

[0096] 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.

[0097] The convergence determination employs a composite conditional strategy. The first criterion is triggered when the improvement in the global optimum is less than 0.001 for 15 consecutive generations. The second criterion checks the particle swarm diversity index; convergence is confirmed when the average Euclidean distance between particles is less than 5% of the search space diameter. Post-processing smoothing is performed before the final weight coefficients are output. Cubic spline interpolation is used to fit the trajectory of the optimal weights from each generation, and the extrapolated value of the tangent direction at the end of the curve is taken as the final result. This processing effectively suppresses the impact of random fluctuations on output stability.

[0098] The dynamic probabilistic graphical model is constructed using a combination of factor graphs and Markov blankets. The node set contains three types of entities: health status classification nodes representing potential oral health states, abnormal event observation nodes corresponding to transient features detected in real time, and environmental context nodes recording auxiliary information such as occlusal status and device parameters. The edge set is divided into deterministic connections and probabilistic connections. Deterministic connections encode hard rules defined by clinical knowledge, while probabilistic connections reflect soft associations obtained through statistical learning.

[0099] The prior probability distribution is established by fusing two data sources. Historical diagnostic records provide explicit state-labeled data, and an initial distribution is constructed using kernel density estimation. User questionnaire data extracts latent topics through a latent Dirichlet allocation model, which are then transformed into adjustment factors for the probability distribution. The state transition matrix incorporates time autocorrelation characteristics; the probability of the current state depends not only on the previous state but also on the patterns of recent state sequences. This design captures the inertial characteristics and periodic patterns of oral health status.

[0100] The parameter estimation of the observation likelihood function employs a variational expectation-maximization algorithm. The algorithm implementation introduces sparsity constraints, forcing the model to focus on strongly correlated features while ignoring weak correlations. The inference process utilizes a parallel message-passing mechanism, decomposing the global probability graph into several subtrees. Each subtree is computed independently, and the results are fused through a consensus protocol. The spatiotemporal continuity constraint is implemented as an energy function, defining a penalty term for state transitions between adjacent time slices and a smoothing term for differences in readings between spatially adjacent sensors.

[0101] The implementation of Bayesian updates takes real-time requirements into account. The full probability update period is set to 200ms, during which accumulated observational evidence is processed in a timestamp-ordered manner. For high-frequency anomalous events, a sliding window aggregation strategy is adopted to merge consecutively occurring similar events into composite evidence. Update calculations use logarithmic space operations to avoid numerical underflow caused by continuous multiplication. The integration of clinical knowledge rules is achieved through soft constraints, transforming expert experience into bias terms of the probability distribution rather than hard constraints.

[0102] The decision-making phase of the joint probability output includes uncertainty quantification. In addition to calculating the maximum a posteriori probability state, it also outputs the suboptimal hypothesis and its probability difference. 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 historical cases and adjusts the final judgment based on the majority voting results. The output interface generates a structured report containing three parts: probability distribution visualization, a list of key evidence, and an explanation of the decision-making basis.

[0103] The system employs a microservice architecture, with the optimization algorithm module deployed on edge computing nodes and the probabilistic inference module running on cloud servers. The two modules communicate via an encrypted channel, transmitting simplified feature vectors instead of raw data, protecting user privacy and reducing communication overhead. Algorithm parameter management utilizes a version control system, retaining a complete configuration snapshot with each update, supporting rapid rollback and difference comparison. A performance monitoring dashboard displays real-time operational metrics such as computation latency, memory usage, and convergence curves, providing a basis for resource scheduling.

[0104] An adaptive learning mechanism continuously optimizes model performance. Weekly automated model diagnostic tests check the performance of various matching scores on the validation set. When performance degradation is detected, an incremental training process is initiated, employing an elastic weight merging algorithm to balance new and old knowledge. User feedback data integration is designed with a dual-channel mechanism: explicit feedback directly adjusts probability map parameters, while implicit feedback indirectly optimizes the observation model through behavioral log analysis. The evolution history of weight coefficients is tracked in a database over a long period, establishing a correlation analysis between parameter changes and alterations in user physiological characteristics.

[0105] An anomaly handling mechanism ensures system robustness. The input data anomaly detection module identifies outliers exceeding physiological ranges, triggering sensor calibration. The algorithm convergence anomaly monitoring module detects oscillations and divergences, automatically switching to backup optimization strategies. The probabilistic reasoning anomaly handling module, when encountering contradictory evidence, initiates a conservative decision-making mode and requests manual review. System health is continuously monitored via a heartbeat mechanism; any component malfunction triggers tiered alarms, escalating from automatic restart to manual intervention.

[0106] The visualization analysis tools support in-depth parameter exploration. A 3D spatial projection plot of weighted coefficients displays the particle swarm optimization trajectory, with different colors marking the optimal solutions for each generation. The probabilistic graph browser interactively displays the strength of dependencies between nodes, supporting hypothetical scenario simulation and counterfactual reasoning. The decision path tracing function recreates the derivation chain of specific conclusions, annotating the influence weights of key evidence. These tools are used for algorithm debugging and optimization, and can also serve as auxiliary tools for clinical teaching.

[0107] The knowledge update process is semi-automated. Newly published clinical research findings are used to extract key parameters using natural language processing to 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 first validated in a shadow mode, running in a parallel environment to compare the output differences between the old and new models. Only after confirming safety and effectiveness are they deployed to the production system. The version change notification module automatically generates technical document update instructions and clinical usage guidelines.

[0108] Example 5: The feedback data acquisition system after the oral health testing command is generated adopts a multi-channel asynchronous processing architecture. Monitoring of the pressure recovery rate in abnormal areas is achieved through a high dynamic range pressure sensor array, which continuously tracks the changes in the mechanical properties of the abnormally marked areas at a sampling frequency of 100Hz. The recovery rate is calculated using a piecewise linear regression method to identify different stages of the pressure value's return to the baseline. Temperature fluctuation deviation values ​​are obtained by combining a distributed infrared thermometry module, establishing a 3×3 temperature measurement grid around the abnormal area, with the reference temperature value taken from the moving percentile of the user's historical data. The quantification of changes in health indicators is based on temporal differential analysis, comparing the offset amplitude of key coordinates in the feature space before and after command execution.

[0109] The preprocessing of feedback data includes two key steps: outlier filtering and temporal alignment. Outlier filtering employs a density-based clustering algorithm to identify and remove outliers that significantly deviate from the main distribution. Temporal 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 for each feedback parameter; channel data below the threshold are flagged and excluded from analysis. The preprocessed feedback data stream is divided into three parallel processing branches, used to update oral baseline features, adjust component complexity indices, and optimize detection command parameters, respectively.

[0110] The dynamic update mechanism for oral baseline features employs an incremental learning strategy. The resting state baseline updates slowly, using an exponentially weighted moving average algorithm, with new data weighting not exceeding 10%. Functional state baselines (such as chewing and light biting) update faster, with new data weighting reaching up to 30%, to more quickly capture changes in user behavior patterns. The baseline feature library maintains reference values ​​across multiple time dimensions, including weekly averages, monthly trend lines, and quarterly benchmarks. Update decisions are made through hypothesis testing; when the KL divergence between the new data distribution and the historical baseline exceeds a preset threshold, a baseline feature recalibration process is triggered. The recalibration process includes a manual review step, requiring users to perform a standardized sequence of actions to verify the reliability of the automatic update results.

[0111] The adjustment of component complexity indicators employs a parametric adaptive framework. Short-term fluctuations are smoothed using Kalman filtering to preserve trend changes while suppressing random noise. Long-term evolution tracking utilizes time-series decomposition techniques to break down complexity changes into seasonal and trend components. A dynamic expected range for indicator normalization is established based on demographic characteristics such as user age and gender; adjustments exceeding this range require additional clinical confirmation. The coupling relationship between complexity indicators and physiological response frequencies is periodically assessed using cross-spectral analysis; a re-matching process is initiated when the coherence coefficient falls below 0.5.

[0112] The user interface provides multi-level feedback visualization. The real-time monitoring view displays the stress recovery process as a heatmap, using color gradients to indicate the recovery progress of different areas. The historical comparison view overlays the changes in current and past health indicators, highlighting statistically significant differences. The alert management interface allows users to confirm or question system test results; this interactive data serves as important feedback signals in model optimization. The mobile application simultaneously pushes concise summaries, describing the clinical meaning of the feedback data using standardized terminology.

[0113] Data security and privacy protection mechanisms are implemented throughout the entire feedback processing flow. Raw feedback data is anonymized at the sensor end, removing direct personal identifiers. End-to-end encryption is used during transmission, and strict access control policies are enforced for stored data. The use of feedback data is limited to algorithm optimization and personalized service improvement; it is not used for any secondary purposes without explicit authorization. A tiered data retention strategy is implemented: raw signals are retained for only 7 days, feature-level data for 1 year, and aggregated analysis results are stored long-term.

[0114] The system maintenance module establishes a complete feedback loop monitoring system. Data pipeline health monitoring includes operational metrics such as throughput, latency, and error rate. The algorithm performance dashboard tracks the evolution trends of key indicators over time and sets automatic early warning thresholds. Hardware status monitoring records equipment parameters such as sensor accuracy degradation and battery wear, predictively indicating maintenance needs. All monitoring data is incorporated into the system health scoring model; when the overall score falls below a threshold, a tiered response mechanism is triggered, ranging from automatic resource allocation to manual intervention.

[0115] The clinical integration module seamlessly integrates feedback data with medical processes. Cases exhibiting abnormal and persistent deterioration automatically generate referral recommendations, accompanied by comprehensive trend analysis charts. Cases showing significant improvement produce recovery progress reports, highlighting key turning points and potential influencing factors. All clinical outputs utilize standardized medical terminology and are compatible with common electronic health record systems. The physician review interface highlights areas of uncertainty in the system's judgment, requesting focused attention on these boundary cases.

[0116] A long-term adaptation framework ensures the system continuously adapts to changes in user behavior. An annual comprehensive evaluation process recalibrates all fundamental parameters, referencing the user's latest demographic information and health status. Quarterly feature drift detection analyzes the gradual changes in oral patterns, identifying algorithm modules requiring adjustment. Monthly performance audits validate the effectiveness of the feedback mechanism, ensuring optimization directions remain aligned with clinical goals. These adaptation mechanisms across different time scales form a comprehensive self-updating system, enabling the system to maintain accuracy and usability over many years of use.

[0117] The anomaly handling protocol covers special scenarios throughout the entire lifecycle of feedback data. Data completion in case of sensor failure employs a generative adversarial network, synthesizing reasonable values ​​based on neighboring node data and historical patterns. Data recovery during transmission interruptions uses a differential synchronization protocol, retransmitting only missing or corrupted data segments. The algorithm's anomaly detection module identifies feedback patterns that do not conform to physiological patterns, triggering a data re-acquisition process. All anomaly events are recorded in the audit log, including the time of occurrence, handling measures, and result verification information.

[0118] The user education component helps users understand the value and limitations of feedback mechanisms. Introductory tutorials demonstrate the correct responses to typical feedback scenarios. Regular knowledge updates explain the impact of the latest optimized testing parameters on user experience. A FAQ library covers everything from technical operation to clinical implications. These educational resources are presented in multimedia formats, dynamically adjusting content depth and presentation based on user learning progress and preferences.

[0119] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An oral cavity detection method based on smart braces, characterized in that, include: The original oral pressure time series and corresponding frequency domain features collected by the user wearing smart braces are obtained. The optimized decomposition parameters corresponding to the frequency domain features are determined. Based on the optimized decomposition parameters, the original oral pressure time series is decomposed into multi-scale signals to generate multi-scale oral signal components. Extract the nonlinear dynamic features of the multi-scale oral signal components and obtain the component complexity index; Based on the multi-scale oral signal components and component complexity index, key signal components related to oral health status are screened. Acquire the user's oral baseline features, real-time occlusal status, and component complexity index to generate personalized physiological response frequencies; Based on the key signal components and personalized physiological response frequencies, the oral physiological feature sequence is reconstructed. Modal decomposition is performed on the oral physiological feature sequence to separate the low-frequency component reflecting steady-state physiological characteristics and the high-frequency component reflecting dynamic changes. Based on the low-frequency and high-frequency components, the user's oral abnormality category is identified, a mapping relationship between the abnormality category and the diagnostic result is established, and an oral health detection instruction is generated. The step of identifying the user's oral cavity abnormality category based on the low-frequency and high-frequency components includes: Temporal statistical features, energy distribution features, and component correlation features are extracted from the low-frequency components to construct a steady-state feature vector; instantaneous rate of change, peak features, and waveform distortion features are extracted from the high-frequency components to construct a transient feature vector; a hierarchical classification strategy is used to classify the steady-state feature vector for health status, and an adaptive detector is used to identify the abnormal events corresponding to the transient feature vector; the health status classification results and abnormal events are fused, and confidence is verified by combining spatiotemporal continuity constraints to output the oral cavity abnormality category; The acquisition of the user's oral baseline features, real-time occlusal status, and component complexity index includes: Multi-stage oral baseline data is collected, including resting state, light occlusion state, and chewing task state. Occlusal rhythm features are extracted from the multi-stage oral baseline data, and real-time occlusal state labels and component complexity indices are generated based on the occlusal rhythm features. According to the frequency band energy distribution of the oral baseline features, the real-time occlusal state labels, and the component complexity indices, a matching score for preset candidate frequencies is calculated. The matching score includes physiological rhythm consistency score, state adaptability score, and complexity correlation score. An optimization algorithm is used to dynamically adjust the weight coefficients of the matching score to determine personalized physiological response frequencies.

2. The method according to claim 1, characterized in that, The step of performing multi-scale signal decomposition on the original oral pressure time series to generate multi-scale oral signal components includes: The original oral pressure time series is processed using an adaptive filtering algorithm combined with the optimized decomposition parameters; the endpoint effect is suppressed at the sequence boundaries using periodic extension. Extract the decomposed multi-scale coefficient set, and construct the multi-scale oral signal component based on the multi-scale coefficient set.

3. The method according to claim 1, characterized in that, The process of reconstructing the oral physiological feature sequence based on the key signal components and personalized physiological response frequencies includes: A time-frequency feature matrix is ​​constructed, the elements of which are 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. Multi-scale time series analysis is performed on the time-frequency feature matrix, including short-scale analysis to capture transient features of occlusion, medium-scale analysis to track the evolution of physiological state, and long-scale analysis to monitor health trends. By dynamically weighting and fusing the results of multi-scale time series analysis, an oral physiological feature sequence containing multi-modal physiological characteristics is generated.

4. The method according to claim 1, characterized in that, The modal decomposition of the oral physiological feature sequence to separate the low-frequency components reflecting steady-state physiological characteristics and the high-frequency components reflecting dynamic changes includes: The extreme points of the oral physiological feature sequence are detected and mirror continuation processing is performed. Multi-mode components that satisfy the intrinsic conditions are extracted through iterative sieving. Based on the average period and energy concentration of the multi-mode components, the first short-period components are classified as high-frequency components, and the second long-period components and residual terms are classified as low-frequency components.

5. The method according to claim 1, characterized in that, The acquisition of the raw oral pressure time series collected by the user while wearing smart braces includes: Real-time capture of multi-point pressure data and auxiliary temperature data within the oral cavity; The signals from the occlusal layer, lingual depressor layer, and buccal mucosa layer were classified and processed to analyze multi-source physiological parameters. The processed data is integrated and standardized to generate a structured raw oral pressure time series.

6. The method according to claim 1, characterized in that, The step of dynamically adjusting the weight coefficients of the matching score using an optimization algorithm includes: Initialize the position and velocity parameters of the particle swarm, and define the objective function as the weighted harmonic mean of the matching score; Iteratively update particle positions and calculate the global optimal solution. When the objective function converges, output the optimization results of the weight coefficients.

7. The method according to claim 1, characterized in that, The fusion of health status classification results and abnormal events includes: A dynamic probabilistic graphical model is constructed, with the health status classification result as the prior probability of the node and abnormal events as observational evidence. The node probability distribution is updated through Bayesian inference, and a joint probability output is generated by combining state transition constraints.

8. The method according to claim 1, characterized in that, After generating the oral health testing instruction, the following is included: Real-time acquisition of instruction execution feedback data, including abnormal area pressure recovery rate, temperature fluctuation deviation value, and changes in health indicators; The oral baseline features and component complexity index are dynamically updated based on the feedback data.

Images