Orthodontic retainer deformation early warning method based on time series data anomaly detection

Abstract

This invention relates to an orthodontic retainer deformation early warning method based on time-series data anomaly detection, aiming to solve the problem of real-time monitoring and anomaly early warning of progressive deformation and stress concentration of retainers in complex oral environments. By integrating a multi-channel flexible pressure sensor array and a mandibular motion sensor, a mixed data stream of oral stress dynamics and physiological action timestamps is acquired. After spatiotemporal alignment and segmented semantic annotation, oral biomechanical features related to physiological behavior are extracted. A dynamic adaptive isolated forest model is constructed using principal component analysis and directional splitting hyperplanes to achieve accurate quantification and multi-level early warning of abnormal risks in non-stationary oral environments. This scheme can improve the robustness and individual adaptability of retainer status monitoring, providing intelligent decision support for material fatigue and follow-up visits.

Description

Technical Field

[0001] This invention relates to the field of oral biomechanical data anomaly detection and intelligent medical monitoring technology, and in particular to an orthodontic retainer deformation early warning method based on time-series data anomaly detection. Background Technology

[0002] Currently, in the field of orthodontic retainer status monitoring and deformation early warning, most algorithms are based on time-series data anomaly detection. Mainstream technical solutions include unsupervised anomaly detection methods such as statistical feature engineering and isolated forests, as well as online learning methods such as streaming concept drift adaptation, expert feedback iteration, or feature weighted fusion. Typical intelligent medical monitoring systems rely on multi-channel flexible pressure sensors and strain sensor arrays, combined with data segmentation, feature extraction, and static modeling, to achieve real-time identification of retainer stress state, deformation trends, or abnormal wearing behavior. With the improvement of oral biomechanical data acquisition accuracy and signal processing capabilities, data analysis models are becoming increasingly complex, moving towards high-dimensional feature fusion and multi-source heterogeneous signal mining. Some systems have introduced streaming anomaly detection and model adaptive update mechanisms to attempt to improve detection robustness in non-stationary physiological environments.

[0003] Currently, the Isolation Forest algorithm, as an efficient unsupervised anomaly detection tool, exhibits good performance in static or quasi-stationary environments. This algorithm randomly partitions the sample space using multiple isolated trees, demonstrating sensitivity to abnormal states. Applied to orthodontic retainer monitoring, it can initially identify deformation and stress anomalies. However, the inherent random feature selection and threshold setting structure of the Isolation Forest algorithm cannot effectively adapt to the periodic mechanical signal changes caused by complex oral physiological behaviors (such as chewing, swallowing, and speech). When faced with frequent switching of oral movement rhythms and mechanical states, the model splitting strategy ignores the locality of data distribution and dominant mechanical patterns, leading to high false positives, false negatives, and even failure to promptly capture progressive anomalies in retainers. Traditional techniques to compensate for this deficiency often employ external feature weighting, system entropy fusion, or adversarial sample injection, but these methods typically increase model complexity, rely on expert priors, or lack interpretability based on physiological semantics. Existing patented technologies often only support limited data streams, static environments, or low-complexity stress fields, failing to address the non-stationarity, rhythmic variations, and individual differences in real-world oral scenarios.

[0004] Existing technologies suffer from the following main shortcomings: First, time-series data on maintainer deformation under oral dynamic stress conditions exhibit significant non-stationary characteristics. The random splitting mechanism of isolated forests struggles to adaptively capture the data distribution corresponding to the dominant biomechanical patterns and rhythmic segments, and the initial abnormality scores lack reliable normalization standards. Second, a model structure strongly coupled with physiological movements is not formed, making it impossible to effectively segment and discriminate data using physiological behavioral labels such as occlusal loading, resting maintenance, and micro-motion adjustment. This manifests in the inability to locate false positive and false negative intervals. Third, model update strategies rely on full reconstruction, expert feedback, or external statistical features, lacking the ability for automatic iteration and individualized optimization driven by periodic rhythms, thus limiting the clinical applicability of the detection system.

[0005] Therefore, there is an urgent need to develop an anomaly detection algorithm system capable of actively adapting to the periodic changes of non-stationary stress signals in the complex dynamic physiological environment of the oral cavity. This system would enable the detection model to automatically identify the local structure and dominant mechanical patterns of data distribution, strengthen its linkage with rhythmic events such as chewing, swallowing, and resting, and improve the robustness of anomaly detection. Simultaneously, it needs to possess segmented and graded capabilities, rhythm consistency constraints, model thermal evolution, and low false alarm closed-loop feedback, overcoming the limitations of existing technologies in terms of detection stability and real-time error correction in dynamic oral stress fields. This would provide higher accuracy and clinically adaptable intelligent algorithm support for oral retainer deformation early warning systems. Summary of the Invention

[0006] This application provides a method for early warning of orthodontic retainer deformation based on anomaly detection of time-series data, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0007] The orthodontic retainer deformation early warning method based on time-series data anomaly detection provided in this application specifically includes: The system acquires raw time-series data from the multi-channel flexible pressure sensor array and micro-strain sensing unit of the orthodontic retainer, and simultaneously records the mandibular motion sensor signal as a rhythmic anchor point to generate a hybrid sensing data stream.

[0008] The hybrid sensor data stream is automatically segmented into bite loading segment time periods, unloading transition segment time periods, resting maintenance segment time periods, and micro-motion adjustment segment time periods with clear physiological semantics, in order to construct a segmented time-series dataset with physiological behavior labels.

[0009] For each sub-time period type in the segmented time series dataset, a sliding window is used to extract local data blocks and the high-dimensional sensing features are compressed using the principal component analysis algorithm. The principal component direction vectors corresponding to the dominant mechanical modes in each sub-time period are calculated to generate a directional feature embedding space.

[0010] By replacing the random feature selection strategy with the principal component direction vector in the directional feature embedding space, a directional splitting hyperplane aligned with the dominant mechanics mode of the current sub-period is constructed. Combined with the historical optimal subspace splitting parameter combination, an adaptive isolated forest model is established.

[0011] Acquire the data to be detected, identify the sub-time period type to which the data to be detected belongs, call the local path length calibration factor of the sub-time period type, and apply confidence weighting to the initial anomaly score in combination with the stability index of the three most recent detection results to generate a local anomaly risk quantification score.

[0012] Determine whether the local anomaly risk quantification score exceeds a preset safety threshold. If it is determined to exceed the threshold, generate a graded early warning instruction; otherwise, generate a normal state maintenance signal.

[0013] Based on the tiered early warning instructions, follow-up consultation prompts or replacement suggestions are pushed to mobile terminals. At the same time, abnormal response frequency and rhythm drift trend data of each sub-period of the day are aggregated to generate an individualized model optimization suggestion package containing information on the concentrated false alarm interval.

[0014] During the next initiation phase of orthodontic retainer wearing, the individualized model optimization suggestion package is input to update the subspace partitioning mechanism and splitting parameter benchmark of the adaptive isolated forest model.

[0015] The orthodontic retainer deformation early warning method based on time-series data anomaly detection provided in this application has the following beneficial effects: (1) By transforming the construction logic of isolated forests from static statistical modeling to dynamic structural adaptation based on oral physiological behavior rhythms, the sensitivity and specificity of anomaly detection in complex and variable oral biomechanical environments are significantly improved. Traditional methods usually rely on the assumption of global feature distribution. When faced with significantly different oral activity patterns such as chewing, swallowing, and resting, it is difficult to effectively distinguish between physiological fluctuations and pathological abnormalities, which easily leads to high false alarm rates or missed detections. This scheme uses short-term multi-channel sensor data to automatically identify mandibular movement rhythms and divides them into functional sub-segments with clear physiological semantics. This enables the model to establish localized and semantically aligned data representation systems for different biomechanical states. On this basis, a principal component orientation alignment directional splitting strategy is adopted to ensure that each node splitting is applied to the most sensitive feature dimension of the dominant biomechanical pattern of the current sub-segment. This greatly enhances the tree structure's ability to focus on key pathological signals, thereby achieving accurate capture of minor but clinically significant stress abnormalities (such as early signs of occlusal overload and nocturnal bruxism) without introducing external labels or complex feature engineering.

[0016] (2) The introduction of rhythm consistency constraints and lightweight retraining mechanism significantly improves the response sensitivity and computational efficiency of the algorithm in long-term wear scenarios, effectively overcoming the adaptation lag problem of existing streaming anomaly detection methods when facing individual rhythm drift and phased functional transitions. Unlike traditional isolated forests that require frequent reconstruction of the entire forest to cope with concept drift, this scheme reuses the historical optimal splitting parameter combination and directly activates the corresponding subspace configuration when a stable continuation of the rhythm type is detected, significantly reducing inference latency. When the rhythm mode changes, only local weight updates are triggered instead of full reconstruction, allowing the model to complete dynamic adaptation within hundreds of milliseconds, ensuring real-time performance while avoiding resource waste. Furthermore, in the anomaly scoring stage, a local c(n) calibration factor customized by sub-period is adopted, and confidence weighting is performed in combination with the recent detection stability index, so that the final output score not only reflects the degree of data outlier, but also integrates the dual judgment criteria of temporal continuity and physiological rationality, which is significantly better than the static scoring mechanism that relies solely on path length, especially showing stronger robustness in suppressing transient interference false alarms caused by swallowing actions.

[0017] (3) A closed-loop individualized model thermal evolution path is constructed, realizing the leap from single detection to long-term health management, with good scalability and clinical applicability. After each user completes daytime wear, the system automatically aggregates the abnormal frequency, false alarm concentration interval and rhythm evolution trend of each sub-time period, generates personalized subspace optimization suggestions, and seamlessly injects them into the model parameter system at the next startup, forming an adaptive loop of "perception-decision-feedback-evolution". This mechanism can continuously track the dynamic changes of individual oral function status (such as postoperative recovery period occlusal force reconstruction, chronic molar progression) without manual intervention or expert annotation, and adjust the detection sensitivity distribution accordingly to fully adapt to individual differences and disease progression. The overall architecture gets rid of the dependence on complex auxiliary modules such as feature weighting, entropy fusion, and adversarial training, and instead uses the human body's own movement rhythm as the internal driving force, making the algorithm structure highly isomorphic to the physiological process. This not only improves the interpretability of the detection results, but also provides a general technical paradigm for future expansion to other biomechanical monitoring scenarios (such as temporomandibular joint disorder assessment and orthodontic force monitoring).

[0018] The aforementioned technologies work together to construct a novel anomaly detection system driven by physiology, adapted to structure, and evolved by individuals. This system achieves a paradigm shift from "passive identification" to "active understanding." While ensuring high accuracy and low latency, it significantly enhances the stability and generalization ability of the model in real oral dynamic stress environments, providing solid support for the clinical application of intelligent oral health monitoring devices. Attached Figure Description

[0019] Figure 1 This is the main flowchart of an orthodontic retainer deformation early warning method based on time-series data anomaly detection.

[0020] Figure 2 This is a sub-flowchart of an orthodontic retainer deformation early warning method based on time-series data anomaly detection.

[0021] Figure 3 This is another sub-flowchart of the orthodontic retainer deformation early warning method based on time-series data anomaly detection. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] like Figure 1 As shown, this application provides a method for early warning of orthodontic retainer deformation based on anomaly detection of time-series data, specifically including: S1: Acquire the raw time-series data collected by the multi-channel flexible pressure sensor array and micro strain sensing unit of the orthodontic retainer, and simultaneously record the mandibular motion sensor signal as a rhythm anchor point to generate a hybrid sensing data stream.

[0025] S2: The hybrid sensor data stream is automatically segmented into bite loading segment periods, unloading transition segment periods, resting maintenance segment periods, and micro-motion adjustment segment periods with clear physiological semantics, in order to construct a segmented time-series dataset with physiological behavior labels.

[0026] S3: For each sub-time period type in the segmented time series dataset, a sliding window is used to extract local data blocks and the high-dimensional sensing features are compressed using the principal component analysis algorithm. The principal component direction vectors corresponding to the dominant mechanical modes in each sub-time period are calculated to generate a directional feature embedding space.

[0027] S4: By using the principal component direction vectors embedded in the directional feature embedding space to replace the random feature selection strategy, a directional splitting hyperplane aligned with the dominant mechanical mode of the current sub-period is constructed. Combined with the historical optimal subspace splitting parameter combination, an adaptive isolated forest model is established.

[0028] S5: Obtain the data to be detected, identify the sub-time period type to which the data to be detected belongs, call the local path length calibration factor of the sub-time period type, and apply confidence weighting to the initial anomaly score in combination with the stability index of the three most recent detection results to generate a local anomaly risk quantification score.

[0029] S6: Determine whether the local anomaly risk quantification score exceeds the preset safety threshold. If it is determined to exceed the threshold, generate a graded early warning instruction; otherwise, generate a normal state maintenance signal.

[0030] S7: Based on the graded early warning instructions, push follow-up consultation prompts or replacement suggestions to mobile terminals, and at the same time aggregate abnormal response frequency and rhythm drift trend data of each sub-period of the day to generate an individualized model optimization suggestion package containing information on the concentrated false alarm interval.

[0031] S8: Input the individualized model optimization suggestion package during the next orthodontic retainer wearing start-up phase to update the subspace partitioning mechanism and splitting parameter benchmark of the adaptive isolated forest model.

[0032] Step S1: Acquire the raw time-series data collected by the multi-channel flexible pressure sensor array and the micro-strain sensing unit of the orthodontic retainer, and simultaneously record the mandibular motion sensor signal as a rhythm anchor point to generate a hybrid sensing data stream. Specifically, this includes: S1.1: Initialize and configure the flexible pressure sensor array and micro strain sensing unit integrated in the key force-bearing parts of the orthodontic retainer to obtain the original multi-channel time-series voltage signal that reflects the dynamic characteristics of retainer deformation under the complex mechanical environment of the oral cavity.

[0033] A linked initialization sequence is executed on the flexible pressure sensor array and micro strain sensing unit integrated in the key stress parts of the orthodontic retainer. The bridge arm resistance, excitation voltage and gain parameters of each pressure sensing node and strain bridge circuit are loaded by relying on the microprocessor configuration interface to achieve precise setting of the basic hardware working state.

[0034] Zero-point calibration commands are applied to the output terminals of each sensor channel. The compensation coefficient is calculated using the output offset under static load conditions, and the compensation value is written into the sensor register space to eliminate the measurement error introduced by the initial bias.

[0035] A periodic stepped loading signal is applied to a flexible pressure sensor array. The sensitivity slope parameter is obtained by acquiring the response curve and using a linear fitting method. This parameter is then stored in the coefficient table of a specific channel to ensure that different channels have consistent calibration accuracy under relatively independent stress conditions.

[0036] For the miniature strain sensing unit, temperature compensation is adopted to read the output value of the temperature sensor in real time, dynamically correct the strain signal, and reduce the impact of oral cavity temperature fluctuations on the measurement.

[0037] A multi-channel sampling synchronous triggering mechanism is set up to ensure that the flexible pressure sensor array and the micro strain sensing unit complete parallel sampling under a unified clock beat, forming a raw time-series voltage signal matrix with time consistency.

[0038] Through the above initialization and calibration link, the sensor output, which is in a non-calibrated state, is converted into a multi-channel raw timing voltage signal with zero-point compensation, consistent sensitivity, and temperature stability, thus realizing the high-precision basic data source required for subsequent analog-to-digital conversion and data timestamp marking processing.

[0039] S1.2: Based on the original multi-channel timing voltage signal, perform analog-to-digital conversion and timestamp marking processing, and use a high-precision real-time clock module to assign an absolute time index to each sampling point to generate a digital sensing data sequence with precise timing labels.

[0040] S1.3: Perform posture calculation processing on the mandibular motion sensor synchronously mounted on the wearer, extract the acceleration peak and angular velocity zero-crossing features that represent natural physiological actions such as chewing, swallowing and speech, in order to generate a rhythmic anchor point event sequence that identifies the moment when the physiological behavior occurs.

[0041] It should be noted that rhythmic anchor points are time markers extracted from jaw movement sensor signals to identify the moments when physiological actions such as chewing and swallowing occur. For example, the time point corresponding to the zero-crossing of the angular velocity during a single swallowing action of the wearer serves as a rhythmic anchor point. These rhythmic anchor points constitute an event sequence, and the subsequent S2 step will segment the continuous sensor data stream based on this sequence, thereby assigning a clear physiological semantic to each data segment.

[0042] The raw triaxial acceleration and triaxial angular velocity signals collected by the synchronously mounted mandibular motion sensor are processed using coordinate system one, and an inertial measurement unit fusion algorithm is used to eliminate attitude deviations introduced by different installation angles, so as to obtain a drift-free attitude dataset in a fixed reference coordinate system.

[0043] The attitude dataset is input into the complementary filtering module, which suppresses the noise characteristics of a single sensor source and enhances the dynamic response accuracy by weighted fusion of high-frequency acceleration signals and low-frequency angular velocity signals.

[0044] A one-dimensional extremum search is performed on the fused acceleration signal, and local acceleration peaks are extracted based on the set mechanical action amplitude threshold. The peak positions are used as candidate points for mechanical impact events of the chewing action.

[0045] Zero-crossing detection is performed on the fused angular velocity signal, and the moment when the angular velocity changes from positive to negative or from negative to positive is marked by sign transformation, which serves as a candidate point for posture switching events of swallowing or speech actions.

[0046] Acceleration peak events and angular velocity zero-crossing events are merged in chronological order, and noise events with intervals smaller than the preset minimum physiological action period are filtered out using a time window to form a rhythmic anchor point event sequence.

[0047] Through the above posture calculation and feature extraction processing methods, the original posture signal of the previous step is transformed into a rhythmic anchor event sequence that can identify the time of occurrence of natural physiological actions such as chewing, swallowing and speech, so as to realize the physiological action time reference benchmark required for subsequent spatiotemporal alignment.

[0048] For example, in the monitoring of the wearer's mandibular movements, the sensor sampling frequency was configured to 200Hz, the acceleration threshold was set to 1.2g, and the angular velocity threshold was set to 45 degrees per second. Coordinate system processing employed Kalman filtering to fuse the three-axis data, with a complementary filter coefficient of 0.98 to ensure the stability of attitude estimation. In acceleration peak detection, the peak value was defined as the current sampling point value being greater than the values ​​of the previous and following five sampling points, and satisfying the amplitude threshold condition. In angular velocity zero-crossing detection, the zero-crossing point was defined as the current sampling point having the opposite sign to the previous sampling point and an amplitude exceeding the threshold. After filtering the fused event sequence through a time window (0.4 seconds), it retained daytime periodic chewing events, large-amplitude swallowing posture switching events, and rapid mandibular opening and closing events during specific phonation processes. The final output rhythmic anchor point sequence count corresponded to the actual physiological action count with no missed or false alarms, verifying the accuracy of this step in action recognition in a noisy oral environment.

[0049] S1.4: Based on the digital sensor data sequence with precise time-series labels and the rhythm anchor event sequence, perform spatiotemporal alignment operation of multi-source data streams. Use a linear interpolation algorithm to map the non-uniformly distributed rhythm anchors onto the sensor data axis with a fixed sampling rate to generate a synchronized hybrid data frame that includes the alignment relationship between oral biomechanical dynamics features and physiological action timestamps.

[0050] The input conditions are a digitized sensor data sequence with precise time-series labels and a rhythmic anchor event sequence. The digitized sensor data sequence is acquired by a multi-channel flexible pressure sensor array and a miniature strain sensor unit, and formed after analog-to-digital conversion and high-precision time indexing. The rhythmic anchor event sequence is obtained by attitude calculation from a mandibular motion sensor and includes time markers for physiological actions such as chewing, swallowing, and speech. Based on the above input, spatiotemporal alignment processing is performed on a data axis with a fixed sampling rate. First, the difference between the timestamp of each event in the rhythmic anchor event sequence and the sampling point time index of the digitized sensor data sequence is calculated to obtain the non-integer position index of the event occurrence time on the sampling axis. Second, a linear interpolation algorithm is used to construct a mapping relationship between adjacent sampling points. The interpolation formula can be expressed as:

[0051] in, The alignment value after interpolation. , These are the sensor values ​​of adjacent sampling points, For the rhythmic anchor point time, , Next, the interpolated event marker values ​​are used as additional fields to be inserted into the corresponding positions of the digitized sensor data sequence for association and binding, forming a preliminary draft of the hybrid data frame containing physiological action alignment information. Subsequently, the time index of each channel in the preliminary draft of the hybrid data frame is synchronously verified to ensure that all data channels share the same absolute time base and there are no cross-channel displacement differences. Finally, the field merging is performed on the verified hybrid data frame, combining pressure, strain, and action markers according to the consistency of sampling points to improve the accuracy of subsequent physiological semantic segmentation operations.

[0052] Through the above-mentioned spatiotemporal alignment and interpolation processing of multi-source data streams, the digital sensor data sequence and rhythmic anchor event sequence of the previous step are transformed into synchronized hybrid data frames that include the alignment relationship between oral biomechanical dynamic features and physiological action timestamps, thereby achieving the same-scale and synchronous fusion of physiological behavior and biomechanical changes.

[0053] For example, in one data collection cycle for a wearer, the sampling rate of the digital sensor data sequence is set to 200Hz. The rhythmic anchor event sequence includes 15 chewing actions and 5 swallowing actions. The maximum deviation between the timestamp of each event and the sensor data sampling point index is 0.045 seconds. During linear interpolation, for events within a 0.045-second deviation, adjacent sampling point indices t0=100 and t1=101 are selected, corresponding to pressure sensing values ​​y0=2.35MPa and y1=2.28MPa. Substituting these values ​​into the above formula yields the interpolation result y=2.33MPa. After inserting the interpolated event markers into the mixed data frame, time index synchronization verification ensures that all 15 chewing actions and 5 swallowing actions are aligned with the same absolute clock. During the structural merging stage, the pressure values, strain values, and action markers are combined into a mixed data frame with fixed fields. Validation results show that in subsequent physiological semantic segmentation, the accuracy of sub-segment recognition based on synchronized data frames is significantly improved, noise-triggered segmentation is reduced, and the input stability of the anomaly detection model is significantly enhanced.

[0054] S1.5: The synchronized mixed data frame is encapsulated and reassembled, and the pressure strain values ​​of each channel and the corresponding physiological action timestamp fields are packaged into a unified data structure to output a mixed sensing data stream containing oral mechanical dynamic features and physiological action timestamps that can be directly called by subsequent segmentation modules.

[0055] Step S2: The hybrid sensor data stream is automatically segmented into bite loading segments, unloading transition segments, resting maintenance segments, and micro-motion adjustment segments with clear physiological semantics to construct a segmented time-series dataset labeled with physiological behaviors. Specifically, this includes: S2.1: Perform peak detection and zero-crossing rate analysis on the mandibular motion sensor signal in the hybrid sensor data stream to generate a raw physiological action timestamp sequence containing the chewing cycle start point, swallowing trigger time, and resting maintenance interval.

[0056] It should be noted that the automatic segmentation is not arbitrary, but based on the mandibular motion sensor signals in the hybrid sensor data stream. Specifically, by detecting the peak value of mandibular acceleration and the zero-crossing point of angular velocity, the starting point of the chewing cycle, the swallowing trigger moment, and the resting maintenance interval are identified, thereby determining the segmentation boundary on the time axis.

[0057] Data parsing is performed on the jaw motion sensor signals in the hybrid sensing data stream to obtain continuous quantized sequences of triaxial acceleration and triaxial angular velocity.

[0058] Peak detection processing is performed on the acceleration sequence to filter out the set of peaks whose amplitude exceeds a preset threshold and whose interval between two peaks conforms to the characteristics of the chewing cycle, and these peaks are marked as the timestamp of the start point of the chewing cycle.

[0059] Zero-crossing rate analysis was performed on the angular velocity sequence, and the positive and negative transition points were marked as swallowing trigger times. Combined judgment was made with the vertical direction abrupt change characteristics of acceleration.

[0060] Continuous intervals where both acceleration and angular velocity are below the micromotion threshold are extracted and marked as resting maintenance intervals. This generates a raw physiological action timestamp sequence containing the start point of the chewing cycle, the swallowing trigger moment, and the resting maintenance interval.

[0061] By combining peak detection and zero-crossing rate analysis, the mixed data stream from the previous step is transformed into a sequence of original physiological action timestamps containing the start point of the chewing cycle, the swallowing trigger point, and the resting maintenance interval. This enables the structured marking of oral motor rhythm events for use in state machine logic.

[0062] For example, during the testing of an adult wearing an orthodontic retainer, the sampling frequency of the mandibular motion sensor was set to 100Hz, the acceleration amplitude threshold was 1.2g, the chewing cycle interval threshold was 0.8 to 1.2 seconds, and the angular velocity zero-crossing rate threshold was 3 to 5 times per second. During processing, peak values ​​were extracted from the triaxial acceleration signals, identifying 24 events with peak amplitudes between 1.25g and 1.35g and intervals meeting the cycle condition. Zero-crossing rate analysis was performed on the triaxial angular velocity signals, detecting an average of 4 zero-crossing events per second on the chewing-related axes, and matching these with the corresponding acceleration peak times, extracting 24 chewing start-point timestamps. For swallowing, a vertical acceleration abrupt change threshold of 2.0g was set, and the criterion was that the duration of the amplitude decreasing to within 0.3 rad / s after the angular velocity direction component crossed zero for more than 0.5 seconds. Three swallowing trigger timestamps were detected. The threshold for determining the resting period is an acceleration below 0.05g and an angular velocity below 0.02 rad / s. When there are no signal fluctuations exceeding the threshold within 30 consecutive seconds, the resting period is marked with two intervals. The processed original physiological action timestamp sequence completely covers the wearer's chewing, swallowing, and resting event distribution, ensuring that the subsequent state machine logic can perform sub-segment segmentation based on accurate time indexes.

[0063] S2.2: Based on the original physiological action timestamp sequence, perform state machine logic judgment processing to generate an event tag set labeled with bite loading state, unloading transition state, resting maintenance state and micro-motion adjustment state.

[0064] For each record in the original physiological action timestamp sequence, a finite state machine-based logical discrimination operation is performed. The continuous acceleration peak events and angular velocity zero-crossing events are pattern matched according to a preset state transition diagram to identify the corresponding oral biomechanical behavior category.

[0065] The time between the start point of the chewing cycle and the next peak event in the matching results is compared with the preset occlusal loading state time threshold. If the continuous loading characteristic condition is met, the interval is marked as "occlusal loading state".

[0066] The sequence from the end of the chewing cycle to the swallowing trigger point is compared with the preset unloading transition state feature template. If the correlation coefficient is greater than the threshold, it is determined to be an "unloading transition state" and a corresponding event marker is generated.

[0067] The long-term, peakless event interval that meets the characteristics of the resting maintenance state after the swallowing trigger point is determined as the "resting maintenance state" by comparing the interval length with the resting threshold, and this state is recorded in the event set.

[0068] Short intervals with frequent low-amplitude peaks in the resting maintenance phase are identified as "micro-adjustment states" by comparing the frequency statistics with the preset micro-adjustment state discrimination factor, and this state label is added to the event label set.

[0069] By using the state machine logic discrimination processing method described above, the original physiological action timestamp sequence is transformed into an event tag set containing occlusal loading state, unloading transition state, resting maintenance state and micro-motion adjustment state, thereby achieving a clear classification of oral dynamic stress patterns.

[0070] S2.3: Use event tag sets to perform boundary cutting and fragment reassembly processing on the original multi-channel time series data to generate initial segmented time series data blocks with clear physiological semantic definitions.

[0071] The event tag set generated based on state machine logic is traversed and parsed to extract the start and end timestamp fields corresponding to each type of physiological behavior event, which are used as time boundary parameters for cutting the original multi-channel time series data.

[0072] The extracted time boundary parameters are mapped to the digital time-series data sequence acquired by the multi-channel flexible pressure sensor array and the micro strain sensing unit to locate the index range of various physiological behavioral events in the data stream, and record the synchronous sampling values ​​of all channels within the range.

[0073] Within the located index range, boundary cutting operations are performed on the data of each channel to extract time-series segments covering the start and end times of the event, forming a consistent data block corresponding to the physiological action.

[0074] Each segment is recombined according to the order of events. Channel data that spans multiple events are split and recombined according to the boundaries to ensure that the content of the segment data contains only mechanical features of the same type of physiological semantics.

[0075] The recombined segments were grouped into four categories: occlusal loading segment period, unloading transition segment period, resting maintenance segment period, and micro-motion adjustment segment period, to construct an initial segmented temporal data block set with clear physiological semantics.

[0076] Through the above-mentioned boundary cutting and fragment recombination processing methods, the event tag set of the previous step is transformed into a multi-channel initial segmented time series data block that can be directly used for segmented modeling, realizing a one-to-one correspondence between data and physiological semantics and providing structured input for subsequent standardization processing and feature extraction.

[0077] For example, in a complete chewing cycle monitoring of the wearer, the state machine logic identifies the start time of the chewing loading state as 12.53 seconds and the end time as 13.27 seconds, and the start time of the unloading transition state as 13.27 seconds and the end time as 13.89 seconds. Based on the correspondence between the absolute timestamps of the start and end times and the pressure data sampling rate, the index range of the sampling points corresponding to the chewing loading state data is calculated to be 2506 to 2654, and the index range of the unloading transition state is 2654 to 2778. In the boundary segmentation stage, the data corresponding to the index ranges of channels 1 to 6 are extracted into independent segments and labeled with event type. In the segment reassembly stage, the chewing loading state data of each channel are spliced ​​into a synchronous data block with a length of 148 sampling points, and the unloading transition state data are spliced ​​into a synchronous data block with a length of 124 sampling points. The initial segmented time-series data block set output contains two types: "biting loading segment" and "unloading transition segment". All channel signals in each type are fully aligned in the time dimension and contain only the mechanical pattern features corresponding to the physiological semantics. Subsequent length unification and filtering will be carried out on this basis, which is expected to significantly improve the pattern purity and robustness of the segmented data in the feature compression stage.

[0078] S2.4: Perform length normalization and noise suppression filtering on the initial segmented time-series data blocks to generate standardized candidate segmented time-series data units.

[0079] S2.5: Perform metadata binding and encapsulation processing on the standardized candidate segmented time series data units and their corresponding event tags to generate segmented time series datasets with physiological behavior labels.

[0080] like Figure 2 As shown, step S3: For each sub-time period type in the segmented time series dataset, a sliding window is used to extract local data blocks, and principal component analysis is used to compress high-dimensional sensing features. The principal component direction vectors corresponding to the dominant mechanical modes in each sub-time period are calculated to generate a directional feature embedding space. Specifically, this includes: S3.1: Perform sliding window truncation processing on the original signals of specific sub-time periods in the segmented time series dataset with physiological behavior labels to generate a sequence of local data blocks containing continuous time segments, ensuring that each local data block covers the complete oral mechanical fluctuation cycle and retains time correlation characteristics.

[0081] S3.2: Based on the local data block sequence, perform high-dimensional feature extraction operations on the multi-channel flexible pressure sensor array and the micro strain sensing unit to generate a high-dimensional sensing feature matrix containing orthogonal pressure components, shear stress components and deformation gradients, thereby realizing the mapping transformation from the original time domain signal to the multi-dimensional mechanical state space.

[0082] The original time-domain signals from the multi-channel flexible pressure sensor array and the micro strain sensing unit in the local data block sequence are normalized to eliminate the influence of the range difference between different sensing channels on the subsequent mechanical characteristic calculation.

[0083] The normalized multi-channel pressure signal is mapped to a three-dimensional orthogonal coordinate system according to the sensor layout. The orthogonal pressure components corresponding to the three-dimensional mechanical axes of the oral cavity are calculated respectively, and a three-dimensional pressure component vector group is formed.

[0084] For the orthogonal component signals, the gradient of the lateral and longitudinal force distribution is calculated by the difference between adjacent nodes of the sensor array, and the local shear stress component matrix is ​​derived to capture the shear deformation mode of the retainer under different behaviors.

[0085] The output of the micro strain sensing unit is subjected to joint spatial and temporal differential processing to calculate the deformation gradient field and form a high-dimensional strain feature vector set describing the local strain change rate and direction of the material.

[0086] The orthogonal pressure components, shear stress components, and deformation gradient features are spliced ​​together according to the channel index and time slice index to construct a high-dimensional sensing feature matrix containing various mechanical state parameters, thereby realizing the complete mapping and transformation of the original multi-channel time domain signal to a multi-dimensional mechanical state space.

[0087] Through the above processing method, the local data block sequence of the previous step is transformed into a high-dimensional sensing feature matrix that can be called by principal component analysis, realizing multi-dimensional encoding of mechanical state information, and providing an input basis for noise suppression and physical semantic enhancement for subsequent extraction of dominant mechanical patterns.

[0088] For example, in data processing for the biting loading segment, the sampling frequency of each channel of the flexible pressure sensor array is set to 200Hz, with a range of 0-500kPa. The strain measurement range of the micro-strain sensing unit is ±2%, and the sampling frequency is 100Hz. The two sampling frequencies differ because the pressure signal changes rapidly (chewing impact is on the order of milliseconds), requiring a higher sampling rate to capture instantaneous peak values; while the deformation signal changes relatively slowly, and reducing the sampling rate can reduce the amount of data and power consumption. After spatiotemporal alignment (linear interpolation) as described in S1.4, the data of different frequencies are unified to the same number of sampling points on the time axis (e.g., all equivalent to 200Hz), thereby forming a synchronized high-dimensional feature matrix. The normalized pressure signal is mapped using three-dimensional coordinates to obtain three orthogonal pressure components: X, Y, and Z. The X-axis is aligned with the tongue-lip direction, the Y-axis with the mesial-distal direction, and the Z-axis with the direction perpendicular to the bite. The shear stress component is calculated using the Z-direction pressure gradient and the lateral displacement gradient, employing a central difference formula.

[0089] in Indicates the pressure component value. The distance between adjacent nodes. Indicates the position coordinates as The pressure component value measured at the sensor node, Indicates the position coordinates as The pressure component values ​​are measured at the sensor nodes. The deformation gradient is obtained by spatial difference of the strain field in the X, Y, and Z directions, and a 9-dimensional gradient feature is synthesized. Finally, the three-dimensional pressure component, the three-dimensional shear stress component, and the 9-dimensional deformation gradient feature are concatenated to form a 15-dimensional high-dimensional mechanical feature vector, which is accumulated over each sliding window time slice to form a feature matrix.

[0090] S3.3: Principal component analysis algorithm is used to decompose the high-dimensional sensing feature matrix into covariance matrix and sort the eigenvalues ​​to generate a set of principal component eigenvectors arranged in descending order of variance contribution rate, thereby completing the filtering of redundant noise information and the initial locking of key mechanical variation directions.

[0091] The high-dimensional sensing feature matrix is ​​centered, and the mean of each feature dimension is used as a reference for vector translation to eliminate the DC bias of the output values ​​of different sensing channels and ensure the consistency of subsequent covariance calculation.

[0092] The centered feature matrix is ​​standardized to scale the variance of each feature to a unit scale, ensuring the comparability of the contributions of each mechanical component to the eigenvalues ​​during the covariance matrix decomposition process.

[0093] A covariance matrix is ​​constructed based on the feature matrix that has been centered and standardized (i.e., subtracting the mean and dividing the standard deviation so that the mean of each feature is 0 and the variance is 1). This eliminates the differences in the dimensions and numerical ranges of different features, thereby accurately characterizing the joint variation relationship between each pressure component, shear component and deformation gradient, and realizing an unbiased quantitative mapping of feature correlation.

[0094] Eigenvalue decomposition is performed on the covariance matrix, decomposing it into eigenvectors and their corresponding eigenvalue sets. The variance contribution of each principal component is then obtained through mathematical operations, as shown in the following formula:

[0095] in, Let covariance matrix be the variance matrix. The eigenvector matrix, To diagonalize the eigenvalue matrix, It is the transpose of the eigenvector matrix.

[0096] Sort the eigenvalues ​​in descending order of magnitude, and divide each eigenvalue by the sum of all eigenvalues ​​to calculate the variance contribution rate, as shown in the following formula:

[0097] in, For the first The variance contribution rate of each principal component For the first The eigenvalues ​​are 1, and the denominator is the sum of all eigenvalues.

[0098] By sorting the variance contribution rates in descending order, principal components with contribution rates below a preset noise threshold are filtered out, and a set of key feature vectors that can significantly characterize the direction of change in the current mechanical mode are identified.

[0099] Through the above covariance matrix decomposition and eigenvalue sorting process, the high-dimensional mechanical state features extracted in the previous step are transformed into a set of principal component feature vectors arranged in descending order of variance contribution rate, thereby achieving the filtering of redundant noise information and the initial locking of key mechanical variation directions.

[0100] For example, during the orthodontic retainer occlusion loading segment, a 5-dimensional feature matrix is ​​obtained, consisting of the three orthogonal components X, Y, and Z from the pressure sensor, the shear component from the strain sensor, and the deformation gradient index. The sampling number is 1000 time points. After centering, standardization is performed to make the variance of each feature equal to 1, forming a standardized feature matrix. The covariance matrix is ​​constructed, and eigenvalue decomposition is performed to obtain eigenvalues ​​λ of 3.2, 1.5, 0.8, 0.3, and 0.2. After sorting, the variance contribution rates r are calculated to be 0.53, 0.25, 0.13, 0.05, and 0.03, respectively. The cumulative contribution rate of the first two principal components exceeds 0.75, while principal components with a contribution rate below 0.05 are discarded. Finally, the feature vectors of the first two principal components are locked for the current mechanical pattern recognition. The output feature vector set significantly improves the anomaly detection stability of the isolated forest in the subsequent directional embedding space construction.

[0101] S3.4: Based on the eigenvector corresponding to the largest eigenvalue whose cumulative variance contribution rate in the principal component eigenvector set exceeds a preset threshold, perform dominant mechanical pattern recognition to generate principal component direction vectors that represent the most sensitive change direction of oral stress state in the current sub-period, and establish the core deformation axis under this physiological behavior.

[0102] For the principal component eigenvector set obtained by principal component analysis and sorted by variance contribution rate, each eigenvector and its corresponding eigenvalue are read sequentially. The variance contribution rate of each eigenvalue is calculated (eigenvalue divided by the sum of all eigenvalues), and cumulative summation is performed. When the cumulative variance contribution rate reaches or exceeds a preset threshold, accumulation is stopped, and the corresponding eigenvector is locked. This eigenvector is used as the principal component direction vector representing the dominant mechanical mode of the current sub-period, and it is normalized to a magnitude of 1. The normalized vector is the direction of the core deformation axis under this physiological behavior, which is used to construct the directional feature embedding space in subsequent steps.

[0103] By using the above-mentioned cumulative variance judgment and feature vector extraction processing method based on preset thresholds, the principal component analysis results of the previous step are transformed into directional parameters that quantitatively characterize the sensitive changes in oral stress state, thereby achieving accurate identification of the dominant mechanical mode and providing a unique directional reference for the next stage of projection transformation.

[0104] For example, when performing principal component analysis on the high-dimensional sensor feature matrix of the resting maintenance period, a cumulative variance contribution rate threshold of 0.85 is set. Five eigenvalues ​​are obtained through covariance matrix decomposition: 3.2, 1.1, 0.8, 0.4, and 0.3. Their calculated variance contribution rates are 0.53, 0.18, 0.13, 0.07, and 0.05, respectively, and cumulative summation is performed. When the cumulative contribution rate reaches 0.84, a fourth eigenvalue is added, resulting in a cumulative contribution rate of 0.91. This fourth eigenvalue is then selected as the largest eigenvalue, and its corresponding eigenvector is extracted. Normalization is applied to adjust the magnitude of this eigenvector to 1. For example, if the original vector components are (0.45, 0.77, 0.44), normalization yields (0.43, 0.74, 0.42). This vector corresponds to the direction of change of the vertical compression dominant mode in the mechanical state space. This direction vector can be used to form a low-dimensional embedding space in subsequent steps, enabling precise localization of the sensitive deformation axis in the resting maintenance segment, and effectively improving the robustness of the adaptive isolated forest model in anomaly detection under this mode.

[0105] The cumulative variance contribution rate is calculated using the following formula:

[0106] in, For eigenvalues, Indicates the first 1 eigenvalue, Indicates the first 1 eigenvalue, This represents the total number of feature dimensions (i.e., the total number of eigenvalues) of the high-dimensional sensing feature matrix. For the front The number of principal components. The molecule is the first... The sum of all eigenvalues, with the denominator being all The sum of all eigenvalues. This formula is used to determine whether the cumulative variance contribution rate meets a preset threshold in order to identify the dominant mode eigenvector.

[0107] S3.5: Based on the principal component direction vector, the original high-dimensional sensing feature matrix is ​​projected and compressed to generate a directional feature embedding space that retains only the dominant mechanical mode information, forming low-dimensional structured input data for the subsequent isolated forest model to construct a directional splitting hyperplane.

[0108] Based on the identified principal component direction vectors, the high-dimensional sensing feature matrix extracted in the previous step is centered and normalized. The mean of each feature component is adjusted to zero and scaled proportionally according to the standard deviation to eliminate the difference in dimensions between different channels, so that the projection transformation calculation is performed on a uniform scale.

[0109] The transformation matrix (i.e. projection matrix) is formed by the direction vectors of the principal components. The normalized feature matrix is ​​multiplied by the transformation matrix to achieve the orthogonal projection operation from the high-dimensional mechanical state space to the low-dimensional feature space, ensuring that only the feature components that contribute the most to the mechanical mode of the current sub-period and have the least noise sensitivity are retained.

[0110] During the projection process, only the first few principal component directions whose cumulative variance contribution rate reaches a preset threshold are selected for projection, while low contribution directions are discarded, thereby limiting the effective dimension of the embedding space. The filtered projection results are encapsulated into a directional feature embedding spatial data structure, where each dimension corresponds to a dominant mechanical change direction, and original temporal indexes and physiological behavior labels are attached to form low-dimensional structured input data that can be directly used for subsequent directional splitting hyperplane construction.

[0111] By using matrix projection and dimensionality compression, the high-dimensional sensing feature matrix from the previous step is transformed into a directional feature embedding space. This process preserves the dominant mechanical mode information while eliminating redundant noise, resulting in low-dimensional structured data that meets the sensitivity requirements of the dynamic subspace adaptive isolated forest for input dimension and direction.

[0112] For example, a high-dimensional sensing feature matrix for the biting loading segment includes 16 channels of pressure components, 8 channels of shear stress components, and 4 channels of deformation gradient components, with a matrix size of 28×200. Centering and normalization are performed to ensure that the mean of all feature components is 0 and the standard deviation is 1. The high-dimensional sensing feature matrix is ​​multiplied by the identified principal component direction vector (length 28) to obtain a projection component of length 1, corresponding to a cumulative contribution rate of 0.88, exceeding the preset threshold of 0.8. The cumulative contribution rate is calculated using the following formula: ,in The eigenvalue is represented by the sum of the first eigenvalues, and the denominator is the sum of all eigenvalues. After filtering, only the dominant component is retained as a one-dimensional feature embedding space, and the time index and "biting load" label for this sub-period are added to generate low-dimensional structured input data. When this input is used in the Isolation Forest model, the splitting plane is aligned with this direction, significantly improving the path length stability and reducing the false alarm rate when detecting anomalies in this sub-period. In clinical validation, the detection performance shows a significant improvement in anomaly identification accuracy.

[0113] like Figure 3 As shown, step S4: Using the principal component direction vectors in the directional feature embedding space to replace the random feature selection strategy, a directional splitting hyperplane aligned with the dominant mechanics mode of the current sub-period is constructed. Combined with the historically optimal subspace splitting parameter combination, an adaptive isolated forest model is established. Specifically, this includes: S4.1: Based on the principal component direction vector of the spatial output of the directional feature embedding, the candidate split dimension set of the isolated tree node is constrained and mapped. The original random feature selection range is reduced to only include the principal component axis corresponding to the dominant mechanical mode of the current sub-period, so as to generate a constrained feature candidate set and ensure that subsequent splitting operations focus on the physical dimension most sensitive to the oral stress state.

[0114] Based on the directional feature embedding space and its principal component direction vector output by step S3.5, the low-dimensional structured input data corresponding to the current sub-time period type is selected as the initial object for filtering the split dimension of isolated tree nodes.

[0115] Perform a feature dimension scan on the low-dimensional structured input data, calculate the variance contribution of each dimension in the current sub-period, and generate a dimension contribution sequence in descending order.

[0116] Features whose variance contribution rate in the dimensional contribution sequence is lower than a preset sensitivity threshold are removed. The threshold is determined based on the threshold setting of the cumulative variance contribution rate in the principal component analysis stage.

[0117] The remaining high-contribution dimensions are mapped and scored using the principal component direction vectors. The scoring formula is as follows:

[0118] in, For the mapping score, Principal component direction vector, The feature dimension is a unit vector. A higher mapping score indicates a stronger physical alignment between that dimension and the direction of the principal component.

[0119] Based on the mapping scores, several dimensions that are most strongly coupled with the current mechanical mode are selected in descending order, and the indices of these dimensions are used to form a restricted feature candidate set.

[0120] By binding the dimensional index with the original feature table, a structured index table for the restricted feature candidate set is established for use in step S4.2.

[0121] Through the above constraint mapping process, the directional feature embedding space of the previous step is transformed into a restricted feature candidate set containing only the axis of the dominant mechanical mode, so as to realize the focus of the node splitting operation on the physical dimension most sensitive to the oral cavity stress state.

[0122] For example, for the resting maintenance segment, the principal component direction vector in the directional feature embedding space is [0.87, 0.49, 0.02]. The low-dimensional structured input data contains three feature dimensions: vertical compression component, horizontal shear component, and orthogonal pressure component, with variance contribution values ​​of 0.68, 0.52, and 0.04, respectively. Under the condition that the cumulative variance contribution rate threshold is set to 0.9, the orthogonal pressure component is removed. The mapping score is calculated between the principal component direction vector and the remaining two-dimensional unit vectors. The score of the vertical compression component is |0.87·1.0|=0.87, and the score of the horizontal shear component is |0.49·1.0|=0.49. After sorting, the vertical compression component is selected as the only restricted feature candidate dimension. This dimension is used as the normal reference for the directional splitting hyperplane in the subsequent S4.2 node splitting. Validation results show that this strategy significantly improves the stability of anomaly detection and reduces the disturbance of non-stationary noise to path length calculation within the resting maintenance segment.

[0123] S4.2: Using the principal component direction vector in the restricted feature candidate set as the normal reference, the geometric orientation of the node split hyperplane is oriented and constructed to replace the traditional random cutting method perpendicular to the coordinate axis. This generates an oriented split hyperplane that is strongly coupled with the current occlusal loading or resting maintenance and other physiological behaviors, thereby achieving isomorphic alignment between the data space partitioning logic and the intrinsic structure of oral biomechanics.

[0124] Normal reference determination processing is performed on the principal component direction vectors contained in the restricted feature candidate set, and the direction vector is used as the sole normal reference of the splitting plane to eliminate the direction drift caused by traditional random cutting.

[0125] During the construction of node geometric orientation, the dominant mechanical mode type corresponding to the current sub-time period is retrieved based on its identifier. A spatial planar relationship is then established between the direction vector mapped to this type and the center point of the node data distribution. This spatial planar relationship means using the principal component direction vector as the normal direction of the splitting hyperplane, and ensuring that this hyperplane passes through the center point of the node data distribution.

[0126] The splitting hyperplane is constructed using the vector-point method, and the binding of the direction vector to the plane is achieved through the following formula:

[0127] in, Principal component direction vector, For any point in the feature space, These are the coordinates of the data distribution center point.

[0128] Based on this construction formula, the coefficients of the plane equation are calculated, and the plane geometric orientation is locked in a mechanical direction that is strongly coupled with physiological behaviors such as current occlusal loading or resting maintenance.

[0129] Perform cross-validation between the directional plane and the dataset. By calculating the distribution balance of samples on both sides of the plane, adjust the plane threshold position to balance the volume ratio of the split subspaces.

[0130] The modified directional splitting hyperplane is partitioned into node spaces, and the restricted feature candidate set from the previous step is transformed into a segmentation unit isomorphic to the intrinsic structure of oral biomechanics, thus achieving consistency between the isolated forest structure and physical semantics.

[0131] This directional construction process transforms the restricted feature candidate set from the previous step into a node splitting geometry with directional constraints, thereby achieving mechanical isomorphism of the spatial partitioning of the isolated forest model in a non-stationary oral environment.

[0132] S4.3: Based on the construction results of the directional splitting hyperplane, retrieve and call the historical best subspace splitting parameter combination bound to the sub-time period type. If a rhythmic pattern shift is detected, trigger the lightweight retraining process to update the isolated tree weight distribution, so as to generate a dynamic splitting parameter benchmark with temporal evolution capability, avoid the computational redundancy brought by full model reconstruction and maintain the ability to track non-stationary data.

[0133] The historical optimal subspace splitting parameter combination refers to a set of parameters obtained through lightweight retraining or offline validation over several previous detection periods (e.g., the past 7 days) for the same sub-time type (such as bite loading segment), resulting in the highest anomaly detection accuracy or lowest false positive rate for the isolated tree on the validation set. This combination includes at least the segmentation threshold of the directional splitting hyperplane, node depth limits, and a baseline path length. For example, the historical optimal parameters for a wearer's bite loading segment are: a segmentation threshold of 12.5 MPa, a maximum depth of 12 layers, and a baseline path length of 7.4. These parameters are stored in a dynamic parameter library for direct use during subsequent detections to maintain model stability. The geometric construction results of the directional splitting hyperplane are mapped using parameter indexes. A unique identifier for the current sub-period type is extracted as a retrieval key to access the historical parameter library of the adaptive isolated forest model. Hash matching and index jump operations are performed on the retrieval key to locate the historically optimal subspace splitting parameter combination bound to that sub-period type. This combination includes a multi-dimensional parameter set such as splitting threshold, node depth constraint, path length baseline, and feature distribution weights. Consistency checks are performed on the historically optimal subspace splitting parameter combination. The rhythm pattern discrimination module calculates the similarity between the principal component direction vector collected in the current sub-period and the historical principal component direction vectors, using the cosine similarity formula.

[0134] in, This represents the direction vector of the principal components in the current time period. The historical direction vector is used; if the similarity is lower than a preset threshold, it is determined that the rhythm pattern has shifted. For sub-time periods determined to have shifted rhythm patterns, a lightweight retraining process is invoked. The normal vector of the original directional splitting hyperplane remains unchanged; only the splitting threshold and path length baseline on isolated tree nodes are updated. By adjusting the splitting threshold and regulating the parameter θ, the standard deviation of the path length of all samples within the current node is minimized, thereby achieving adaptive correction of the weight distribution of the isolated tree. A new dynamic splitting parameter benchmark is generated by combining historical optimal parameters (either through lightweight retraining or direct invocation), and injected into the current subspace partitioning mechanism through structured encapsulation to achieve continuous tracking capability for non-stationary data.

[0135] Through the above-mentioned retrieval, verification, offset determination and lightweight retraining processing, the directional splitting hyperplane construction result of the previous step is transformed into a dynamic splitting parameter benchmark with temporal evolution capability, thereby realizing the stability and computational redundancy optimization of the model in a non-stationary oral environment.

[0136] For example, during the detection of bite loading segments, the historical optimal subspace splitting parameter combination includes a splitting threshold of 0.85, a node depth limit of 12 layers, a path length baseline of 7.4, and a feature distribution weight ratio of 1.5:1.0:0.8. The cosine similarity between the currently collected principal component direction vector and the historical direction vector is 0.72, which is lower than the preset threshold of 0.8, indicating a shift in the rhythm pattern. A lightweight retraining process is invoked, keeping the normal vector direction unchanged. By optimizing the θ parameter in the formula, the path length standard deviation is adjusted from 1.2 to 0.9, thus resetting the node splitting threshold to 0.83 and updating the path length baseline to 7.1. The updated dynamic splitting parameter baseline is injected into the isolated forest construction process. In the next detection cycle, when identifying anomalies in similar data blocks, the path length deviation is significantly reduced, and the anomaly scoring stability index is greatly improved, verifying that this method reduces the computational overhead of full model reconstruction while maintaining high detection accuracy.

[0137] S4.4: Based on the dynamic splitting parameter benchmark and the directional splitting hyperplane, a recursive binary splitting operation is performed to construct multiple isolated trees with physiological semantic guidance. In each splitting layer, the normal direction of the splitting surface is forced to be consistent with the direction vector of the principal component to generate an adaptive isolated forest model, so that the statistical characteristics of the abnormal path length can truly reflect the degree of deformation deviation of the retainer under a specific oral movement rhythm.

[0138] Based on the dynamic splitting parameter benchmark and the input conditions of the directional splitting hyperplane, a recursive binary splitting operation is performed on the construction process of the isolated forest. The input includes a constrained set of candidate features after being constrained by the principal component direction vectors and their corresponding geometric orientation information. A directional splitting condition is introduced for the root node of each isolated tree. According to the splitting threshold set in the dynamic splitting parameter benchmark, the directional hyperplane is used as the splitting interface of the root node, dividing the first-level subspace region within the low-dimensional directional feature embedding space. For each newly generated subspace node, the corresponding directional splitting hyperplane normal vector is called as a parallel orientation constraint during the second-level split. The dataset inside the node is subjected to a binary splitting operation according to the dynamic threshold, ensuring that each level of split strictly maintains the perpendicular relationship between the splitting plane and the direction vector of the preceding principal component.

[0139] The path length within each node of the recursive split is calculated to ensure that the statistical characteristics of the path length completely correspond to the spatial distribution of the oral biomechanical rhythm. The node results from all split levels are assembled into an isolated tree structure according to the topological hierarchy, and each isolated tree follows a directional splitting rule (i.e., the principal component direction vector is the normal direction) during splitting, thus forming an adaptive isolated forest model. This model ensures that the distribution of abnormal path lengths accurately reflects the biomechanical deviation of the retainer under specific oral movement rhythms. Through recursive binary splitting and parallel orientation retention processing, the principal component direction vector results from the previous step are transformed into a forest topology structure with physiological semantic guidance, achieving structural adaptation and improved robustness of the anomaly detection model in non-stationary oral environments.

[0140] S4.5: Perform integrity verification and convergence testing on the initial topology of the adaptive isolated forest model to confirm that the directional splitting mechanism for each sub-period type has been correctly injected and the weight update logic is correct. Output the adaptive isolated forest model to provide a robustly optimized core discrimination engine for the calculation of local anomaly risk quantification scores for subsequent new inflow data.

[0141] Based on the initial topology of the constructed adaptive isolated forest model, the configuration parameters of the directional splitting mechanism bound to the corresponding sub-time period type and the weight distribution updated by lightweight retraining are invoked to form a complete forest instance containing the geometric orientation of the splitting planes of each isolated tree, the node depth distribution, and the statistical features of path length. During the verification phase, the node splitting rules of each isolated tree in the initial topology are checked one-to-one with the input principal component direction vectors. The vector inner product consistency judgment method is used to evaluate the parallelism between the normal direction of the splitting plane and the dominant mechanics mode of the sub-time period, and splitting nodes that do not meet expectations are marked by setting a parallelism threshold. For the weight update logic, the weight values ​​of each isolated tree node are read, and the Euclidean distance between the current node weight vector and the historical baseline weight vector is calculated as a weight deviation metric. The following formula is used to complete the weight consistency verification:

[0142] in, These are the current model node weight values. The historical baseline weight value, As a measure of weight deviation, This represents the total number of isolated trees participating in the weight consistency check in the isolated forest model (i.e., the number of trees in the isolated forest). To ensure the integrity of the topology, the proportion of leaf nodes, maximum depth, and sample coverage of each isolated tree in the forest are checked sequentially. The convergence stability of the model output on a fixed input sample set is verified through iterative path length analysis based on the convergence criterion. The convergence criterion is expressed as follows:

[0143] in, This represents the average path length in the current iteration. This represents the average path length calculated by the model for all training samples in the previous iteration. To determine the convergence rate of change, the model is compared with a preset convergence threshold to determine whether it has reached a stable state. Through integrity verification and convergence testing, the initial topology of the adaptive isolated forest model from the previous step is transformed into a confirmed model instance, realizing a core engine for anomaly detection with high robustness and stability.

[0144] For example, in the validation process of an adaptive isolated forest model for an 18-year-old orthodontic patient, the input conditions were initial topological structure data constructed from four physiological sub-time periods, containing 32 isolated trees, each with a maximum depth of 15 and a leaf node ratio of no less than 0.25. The parallelism threshold was set to 0.95. Cosine similarity was calculated between the splitting plane direction of each node and the principal component direction vector. Three nodes were found to have thresholds below the set value, automatically triggering split adjustments. In the weight deviation metric calculation, the Euclidean distance D between the historical baseline weight vector and the current weight vector was 0.12, below the stability threshold of 0.2, indicating that the weight change was acceptable. In the convergence criterion calculation, the mean path length changed from 12.4 in the 9th iteration to 12.35 in the 10th iteration, with a change rate C of 0.004, below the set convergence threshold of 0.01, confirming model convergence stability. In subsequent detection, this model instance reduced the fluctuation range of abnormal scoring output for non-stationary oral stress data to ±0.05, effectively improving the stability and clinical adaptability of abnormality identification.

[0145] Step S5: Obtain the data to be detected, identify the sub-time period type to which the data to be detected belongs, call the local path length calibration factor of the sub-time period type, and apply a confidence weight to the initial anomaly score in combination with the stability index of the three most recent detection results to generate a local anomaly risk quantification score. Specifically, this includes: S5.1: Based on the segmented time-series dataset with physiological behavior labels, perform timestamp matching processing on the newly incoming mixed sensor data stream to determine the type of the current data, such as the biting loading segment, unloading transition segment, resting maintenance segment, or micro-motion adjustment segment, and generate the data block to be detected.

[0146] A pre-retrieval was performed on the segmented time-series dataset with physiological behavior labels to confirm that the input hybrid sensor data stream had complete timestamp fields and corresponding multi-channel pressure and strain values. The difference between the absolute time index of each sampling point in the hybrid sensor data stream and the timestamp set of each sub-time period boundary in the segmented time-series dataset was calculated, and a bidirectional nearest neighbor matching algorithm was used to map the current sampling point to the nearest sub-time period boundary. For the sampling point set that had been matched to the sub-time period boundary, interval assignment was determined. Based on the preset time range templates for the occlusal loading segment, unloading transition segment, resting maintenance segment, and micro-motion adjustment segment, the physiological semantic assignment type of the current data was determined. Consistency verification of the assignment type determination results was performed. When cross-segment switching occurred due to continuous sampling point assignment, a rhythmic pattern priority rule was introduced to ensure that cross-segment switching only took effect when the peak or zero-crossing feature of mandibular movement coincided with the boundary. Segment extraction was performed on the consistent-verified sampling point set, and its corresponding multi-channel pressure and strain values ​​and timestamp fields were packaged into a unified data block format and bound with physiological semantic assignment labels to form the data block to be detected. By using the above processing methods based on timestamp matching, interval attribution determination and consistency verification, the results of the previous step are transformed into data blocks to be detected that have clear physiological semantic attribution and can be directly entered into the anomaly detection scoring calculation stage, thereby achieving accurate sub-time period type positioning of newly inflowing data in a non-stationary oral environment.

[0147] For example, in the testing scenario of an orthodontic retainer wearer, the sampling rate of the hybrid sensor data stream is 200Hz. Each sampling point records three channels of flexible pressure values ​​and two channels of micro-strain values, and the absolute time index is generated by a high-precision real-time clock. The resting maintenance segment boundary timestamps in the segmented time-series dataset are 12:15:20 and 12:15:40, and the occlusal loading segment boundary timestamps are 12:15:40 and 12:15:50. The time index of newly incoming data sampling points is matched with the above boundary timestamp set. A bidirectional nearest neighbor matching algorithm is used to obtain the nearest boundary difference within 0.05 seconds. When the difference is less than the threshold, the current segment is assigned to the occlusal loading segment type. During the consistency verification process, the peak mandibular motion acceleration of this segment was detected at 12:15:41, which meets the rhythm pattern priority rule requirement; therefore, the cross-segment switching determination is valid. All sampling points within the interval from 12:15:40 to 12:15:50 were extracted. The pressure three-channel values ​​ranged from 0.8 to 1.3 MPa, and the strain two-channel values ​​ranged from 1.2 to 2.0 μm. These values ​​were uniformly encapsulated into a data block structure and bound with a biting loading segment label. This data block was then used as input to the isolated forest anomaly detection scoring calculation stage, which significantly improved the accuracy of sub-time period type localization and detection stability in this scenario.

[0148] S5.2: Utilize the physiological semantic attribution type of the data block to be detected, retrieve and call the local path length calibration factor that is strictly bound to the type from the pre-stored adaptive isolated forest model parameter library, so as to construct a normalized benchmark reference quantity for the current oral stress state distribution characteristics.

[0149] The input conditions include the data block to be detected, whose physiological semantics have been determined by S5.1, and the index table of multiple local path length calibration factors stored in the adaptive isolated forest model parameter library.

[0150] For the physiological semantic attribution type of the data block to be detected, parameter library retrieval and positioning processing is performed. Specifically, the target parameter record is accurately located by matching the unique label identifier of the type with the key-value pair in the index table.

[0151] The retrieved target parameter records are subjected to integrity verification and version consistency verification to ensure that the local path length calibration factor is consistent with the current model topology and subspace partitioning mechanism.

[0152] The verified local path length calibration factor is invoked and loaded into the normalization calculation module of the current detection process as a reference value to form a standard path length scale for the mechanical state distribution of this sub-period.

[0153] The normalized baseline quantity is associated with and stored with the data block to be detected for use in the subsequent S5.3 path length standardization calculation, thereby achieving comparability and consistency of path length statistics under different physiological action modes.

[0154] By using a parameter retrieval and calibration factor call processing method based on physiological semantics, the data block to be detected in the previous step is transformed into a structured input with a normalized benchmark of the current oral stress state distribution characteristics, thereby realizing a unified evaluation scale for path length.

[0155] For example, during the testing of an orthodontic retainer wearer, the data block to be tested was classified as the occlusal loading segment time period, and this type was labeled "CLD". The local path length calibration factor corresponding to "CLD" in the parameter library was set to 1 / ln(n), where... Let ln be the isolated tree sample size, and ln represent the natural logarithm. The retrieval and localization process matches the "CLD" key-value pair to obtain the calibration factor value. Assuming the current isolated tree sample size is 512, the calibration factor value is calculated using the formula 1 / ln(512), and the result is loaded into the normalization calculation module. During subsequent path length standardization, the traversal path length is divided by this calibration factor value to achieve comparability across different sub-time periods. Verification shows that this method significantly improves the stability of abnormal scoring under complex oral stress environments, reduces the bias in cross-sub-time period comparisons, and significantly enhances the clinical reference value of risk identification results.

[0156] S5.3: Based on the normalized baseline reference, the traversal path length of the data block to be detected on the directional splitting hyperplane is standardized to obtain an initial anomaly score that reflects the degree of deviation of the mechanical state at the current moment.

[0157] After invoking the local path length calibration factor bound to the current physiological semantic attribution type, the traversal path length of the data block to be detected in the directional splitting hyperplane is used as the execution object. The input conditions include the normalized reference value, the feature vector of the data block, and the adaptive isolated forest model structure information. The traversal path length sequence is accumulated node by node to form the original total path length, and the normal vector direction of the splitting hyperplane is recorded during the accumulation process to ensure orientation consistency. The original total path length is scaled using the normalized benchmark reference value. The scaling factor is determined by the calibration factor c(n) corresponding to the sub-time period type to ensure that the data dimensionality differences under different physiological action modes are eliminated. When performing the standardization calculation, the original path length is set to L. raw The calibration factor is c(n), and the normalized path length is... Calculate using the following formula:

[0158] in, The original path length of the sample to be tested on the isolated tree. This is a local path length calibration factor used to eliminate the impact of sample size differences across different sub-time periods on path length. Based on the standardized path length, the Isolation Forest scoring function is called to convert the path length into an initial outlier score. The formula is:

[0159] in Indicates the standardized path length. This is a function for the average path length. The distribution characteristics corresponding to physiological movement patterns are compared to confirm the representativeness of the score under the current oral stress state. Through the above standardization and scoring processing, the normalized reference value and path statistics results from the previous step are transformed into an initial anomaly score reflecting the degree of deviation from the current mechanical state, thus achieving a cross-modal consistent detection benchmark.

[0160] For example, during the detection process of a wearer's resting maintenance period, the original traversal path length L of the data block to be detected is... raw The measured value was 12.8, and the corresponding local path length calibration factor C for this sub-period was... cal The recorded value is 0.180, and the standardized path length L is calculated according to the formula. std=71.1. Using the average path length from the Isolation Forest theory as the normalization benchmark and substituting it into the scoring function, the initial anomaly score S is calculated. init The score is approximately 0.012, which is low, indicating a significant deviation from the normal mechanical state and placing it in a high-deformation-risk zone. In another interlocking loading segment test, L... raw =15.6, C cal If it is 0.144, then L std ≈108.3, S init The score of ≈0.003 also indicates a deviation in the mechanical state of the current bite loading segment.

[0161] Through the above standardization and scoring process, the system maintains a consistent detection benchmark under different physiological action modes and significantly improves the comparability and robustness of abnormal scores.

[0162] S5.4: Extract the initial anomaly score sequence generated within the three most recent consecutive detection cycles, use the moving variance algorithm to calculate its fluctuation dispersion to characterize the stability index of the detection process, and use the stability index as a confidence weight coefficient to apply to the current initial anomaly score to generate a weighted anomaly score that has been corrected by time series smoothing.

[0163] Based on the initial anomaly score sequence output from the preceding steps, the score values ​​from the three most recent consecutive detection periods are selected to form an analysis window. The arithmetic mean of the three score values ​​within this window is calculated and used as the central reference value. The difference between each score value and the central reference value is calculated, and the square of each difference is obtained, resulting in three squared values. The three squared values ​​are summed, divided by (the sample size within the window minus one, i.e., 2), and then the square root is taken to obtain the sample standard deviation of the score values ​​within the window, denoted as the stability index. This stability index reflects the degree of fluctuation and dispersion of the results from three consecutive detections for the same sub-time period type. This stability index is used as a confidence weight coefficient and multiplied by the initial anomaly score of the current detection period to generate a weighted anomaly score after time-series smoothing correction.

[0164] By using the above processing method, the initial abnormality score of the previous step is transformed into a weighted risk score with stability correction capability, thereby achieving adaptive control of the confidence of abnormal detection results in non-stationary oral environment.

[0165] For example, during a real-time monitoring process of an orthodontic retainer wearing, the initial abnormality scores for the three most recent consecutive monitoring cycles were 2.4, 2.8, and 3.1, respectively. With a window size of 3, the calculated window mean μ was (2.4 + 2.8 + 3.1) / 3 = 2.766. Squaring the difference between each score and the mean yielded (2.4 - 2.766). 2 =0.134、(2.8-2.766) 2=0.00115、(3.1-2.766) 2 =0.111. Summing the squared differences yields 0.24615. With a window sample size of n = 3 and a normalized denominator of n-1 = 2, the stability index is calculated using the above formula, resulting in a stability index of 0.35015. This stability index is used as a confidence weight to product and correct the initial abnormality score of 3.1 for the current cycle, resulting in a weighted abnormality score of 1.08547. This weighted abnormality score, combined with rhythm consistency constraints in subsequent step S5.5, can significantly improve the stability of the score and the accuracy of clinical risk assessment. Especially in environments with large fluctuations in the wearer's rhythm, this method can effectively suppress the interference of occasional spikes in abnormality determination.

[0166] S5.5: The integrated weighted abnormality scoring and rhythm consistency constraint mechanism performs logical verification and final fusion processing on the scoring results with drastic fluctuations, so as to output a quantitative score of local abnormality risk that can accurately reflect the actual deformation risk of the retainer in a non-stable oral environment.

[0167] Step S6: Determine whether the local anomaly risk quantification score exceeds a preset safety threshold. If it exceeds the threshold, generate a tiered early warning instruction; otherwise, generate a normal state maintenance signal. Specifically, this includes: S6.1: Obtain the local anomaly risk quantification score after confidence weighting and read the hierarchical safety threshold set stored in the individualized dynamic model. The hierarchical safety threshold set includes the mild deformation warning threshold, the moderate stress concentration warning threshold, and the severe material fatigue warning threshold generated based on the statistical analysis of historical false alarm intervals, in order to construct a benchmark reference system for condition judgment.

[0168] It should be noted that the preset safety threshold can be a single threshold or a set of multiple graded thresholds. In this embodiment, a graded early warning strategy is adopted, with a pre-set early warning threshold of 2.5 for mild deformation, 3.5 for moderate stress concentration, and 4.5 for severe material fatigue (these thresholds can be dynamically updated based on historical false alarm statistics). When the quantification score of local anomaly risk exceeds any of the thresholds, the corresponding level of early warning instruction is triggered; if the score is lower than the minimum threshold, a normal state maintenance signal is output.

[0169] The system obtains the local anomaly risk quantification score after confidence weighting as the input object, parses the numerical structure of the score and locks the physiological behavior sub-period type to which it belongs, and calls the storage interface of the individualized dynamic model at the data access layer to read the set of hierarchical safety thresholds associated with the sub-period type.

[0170] An index retrieval is performed on the graded safety threshold set, and the numerical entries of the mild deformation warning threshold, moderate stress concentration warning threshold, and severe material fatigue warning threshold are transferred to the calculation cache area and arranged in order of risk level to form a threshold sequence.

[0171] For each threshold entry, source metadata is attached to indicate that the threshold is generated by the statistical results of the historical false alarm set interval and the rhythm drift correction factor, and the threshold is fine-tuned during the calibration phase using the stability index within the historical detection period.

[0172] A numerical comparison engine is used to construct a state discrimination benchmark reference system. The threshold sequence is mapped to the comparison matrix of the score to be detected in a fixed format, and the difference calculation module is initialized for each matrix unit for accurate discrimination during subsequent interval mapping.

[0173] By using data binding, the quantitative score of local anomaly risk after confidence weighting is bidirectionally associated with the benchmark reference system, so that when the state discrimination logic is called, the difference structure between the score value and the corresponding threshold can be read at the same time, thus achieving seamless connection of the discrimination process.

[0174] Through the above processing method, the local anomaly risk quantification score of the previous step is transformed into a benchmark reference system that can be directly compared with the set of graded safety thresholds, thereby achieving more accurate risk level judgment and dynamic adaptation.

[0175] For example, in the testing scenario of a long-term orthodontic retainer wearer, the confidence-weighted local anomaly risk quantification score is 3.85, belonging to the occlusal loading segment. The set of graded safety thresholds associated with this type in the individualized dynamic model includes a mild deformation warning threshold of 2.5, a moderate stress concentration warning threshold of 3.5, and a severe material fatigue warning threshold of 4.5. The three thresholds are arranged in order of risk level to generate a sequence [2.5, 3.5, 4.5], with additional source metadata recording that the thresholds are derived from the statistical analysis of concentrated false alarm intervals of test samples over the past 90 days and a rhythm drift correction factor of 0.05. During the calibration phase, a stability index of 0.92 is applied to the values ​​of the threshold sequence, resulting in a mild deformation warning threshold of 2.5 + 0.05 = 2.55, a moderate stress concentration warning threshold of 3.5 + 0.05 = 3.55, and a severe material fatigue warning threshold of 4.5 + 0.05 = 4.55. After initializing the benchmark reference matrix, the risk score of 3.85 is compared with the aforementioned thresholds to calculate the difference: 1.30 with the slight deformation threshold, 0.30 with the moderate stress concentration threshold, and -0.70 with the severe material fatigue threshold. This difference structure is bidirectionally bound to the graded threshold set, achieving precise alignment of the state discrimination benchmark. After interval mapping, it is determined that the current risk level belongs to the moderate stress concentration interval. Subsequently, S6.2 can execute the risk level mapping and subsequent early warning generation process.

[0176] S6.2: Based on the hierarchical safety threshold set, perform multi-level interval mapping processing on the local anomaly risk quantification score. By comparing the numerical relationship between the local anomaly risk quantification score and the threshold values ​​at each level, determine the risk level interval of the current oral biomechanical environment, so as to generate a state discrimination label containing the risk level identifier.

[0177] S6.3: Execute branch logic judgment based on the status discrimination label. If the status discrimination label indicates that the risk level exceeds the preset threshold, call the graded early warning generation module. Combine the current sub-time period type and the stability index of the three most recent test results to construct a graded early warning instruction that includes suggestions on the urgency of follow-up visits, so as to output an active intervention signal for the risk of retainer failure.

[0178] Based on the state discrimination label generated through multi-level interval mapping, a conditional matching operation is performed on the risk level value to determine whether it exceeds a preset threshold. The discrimination logic module is called to compare the input risk level label with the threshold index table. If the label level is greater than or equal to the threshold number corresponding to the threshold, the active intervention signal generation branch is initiated. The physiological behavior sub-period type information bound to the current risk level label is read, and the sub-period type is used as the index key to retrieve matching records in the early warning parameter library, extracting the response strategy parameter set used to construct the graded early warning. The stability index value sequence within the three most recent detection cycles is obtained in parallel, and mean filtering is performed on it to remove short-term fluctuations. The mean result is used as a confidence weight parameter and bound to the response strategy parameter set of the current sub-period type. Using the above-bound response strategy parameter set, the graded early warning generation module is called to construct the output instruction, including a risk level description field, a follow-up visit time suggestion field, a maintainer replacement suggestion field, and a confidence weight field. The constructed early warning instruction is formatted and encapsulated to generate a binary data packet and marked as an active intervention signal. This signal triggers the early warning push process of the subsequent communication module. By using branch logic judgment and early warning content construction, a graded early warning instruction is formed based on risk level labels, physiological behavior sub-period types, and stability index, so as to achieve timely and proactive intervention in the risk of retainer failure.

[0179] For example, in a single testing cycle, the risk level label determined by S6.2 mapping is "severe material fatigue risk." The corresponding number for this level in the critical value index table is 3. The threshold matching operation result is greater than or equal to the condition, thus entering the graded early warning branch. The current data block belongs to the "biting loading segment" sub-period type. The response strategy parameter set for this type is retrieved from the early warning parameter library and configured as follows: follow-up consultation recommendation within 2 days, replacement recommendation type is high-strength retainer material, and the initial confidence weight coefficient is 0.85. The stability index sequences for the three most recent tests are 0.82, 0.84, and 0.79. After mean filtering, the mean value is 0.8166. This value is used as the corrected confidence weight and bound to the response strategy parameter set. Based on this, the graded early warning generation module constructs the instruction content field: the risk level is described as "severe material fatigue," the follow-up consultation recommendation is "follow-up consultation as soon as possible within 2 days," the retainer replacement recommendation is "replace the current retainer with high-strength material," and the confidence weight field is 0.8166. After binary encapsulation, an active intervention signal is obtained and pushed to a mobile terminal via a Bluetooth Low Energy communication module. Detailed warning information is displayed on the user's end, guiding subsequent clinical interventions.

[0180] S6.4: If the state discrimination label indicates that the risk level is within the preset critical value range, a normal state maintenance signal is generated, and the current local abnormal risk quantification score is used as a new historical data sample to update the sliding window buffer to optimize the calculation benchmark of the stability index in the next detection, so as to ensure the continuity and smoothness of the real-time identification results.

[0181] The local anomaly risk quantification score, corrected by rhythm consistency constraints, is used as input. The state discrimination label logic module is called to confirm that the risk level is within the preset critical value range, and then the normal state signal construction processing is performed.

[0182] The current local anomaly risk quantification score is formatted and converted to ensure compatibility with the data structure of the sliding window buffer, so as to ensure that the numerical type, time index and physiological behavior label conform to the historical sample entry specifications.

[0183] The timestamp binding key is generated based on the formatted score data, and an insertion operation is appended to the end of the buffer to extend the sequence without overwriting the original historical samples.

[0184] A window overflow detection mechanism is used to determine whether the number of samples after insertion exceeds the upper limit of the buffer capacity. If it does, the earliest sample is deleted according to the first-in-first-out strategy to keep the window length stable.

[0185] During the sample entry process, a pre-calculated benchmark for the stability index of the next testing cycle is simultaneously calculated, including the mean, variance, and dispersion indices, to provide an immediate reference for subsequent confidence weighting.

[0186] By using window sample updates and stability index pre-calculation processing, the local anomaly risk quantification score from the previous step is transformed into a historical data benchmark that can be used for stability evaluation in the next cycle, thus achieving continuity and smoothness of real-time identification results.

[0187] S6.5: Integrate the generated graded early warning instructions or normal state maintenance signals, encapsulate them into a standardized real-time identification result data packet. The data packet contains timestamps, risk level identifiers, trigger threshold types, and confidence weight information to complete the final output for the progressive deformation or stress concentration state of the retainer and provide it for subsequent communication modules to call.

[0188] Step S7: Based on the tiered early warning instructions, push follow-up consultation prompts or replacement suggestions to the mobile terminal. Simultaneously, aggregate abnormal response frequency and rhythm drift trend data for each sub-period of the day to generate an individualized model optimization suggestion package containing information on the false alarm concentration interval. Specifically, this includes: S7.1: Generate graded warning instructions based on the comparison results of the local anomaly risk quantification score and the preset safety threshold. Use Bluetooth Low Energy communication protocol to encapsulate the graded warning instructions to generate a mobile terminal push data packet containing follow-up visit prompts or replacement suggestions.

[0189] S7.2: For each physiological semantic sub-period completed on the day, extract the corresponding abnormal response frequency data and rhythm drift trend data, and use time series aggregation algorithm to perform multi-source fusion processing on the abnormal response frequency data and rhythm drift trend data to generate a comprehensive statistical feature vector characterizing the stability of daytime mechanical behavior.

[0190] For each physiological semantic sub-segment completed on the day, the local anomaly risk quantification scores generated in the previous steps are grouped by sub-segment type. The number of abnormal responses within each sub-segment type is extracted to form an anomaly response frequency data matrix. The timestamp sequences of each sub-segment are cross-correlated with the local anomaly risk score sequences to calculate the rhythm drift trend vector of each sub-segment within the entire day's detection cycle. The trend vector is jointly defined based on the rate of change of the anomaly response center time and the rate of change of amplitude over the continuous detection cycle. The anomaly response frequency data and the rhythm drift trend vector are used as parallel inputs, aligned on a unified time axis by a time series aggregation algorithm, and the two types of data are simultaneously fused using a weighted accumulation strategy. The weight coefficients are preset according to the physiological action importance index of the sub-segment type. Feature extraction is performed on the fused multi-source time series to generate a comprehensive statistical feature vector containing a whole-day mechanical stability index, a rhythm consistency index, and an anomaly response concentration index. The mechanical stability index is calculated by the following formula:

[0191] in The frequency of abnormal responses in each detection cycle. This represents the average frequency of abnormal responses throughout the day. This represents the total number of detection cycles. Through time series aggregation and comprehensive feature vector construction, the local anomaly response results from the previous step are transformed into multidimensional statistical data that can characterize the stability of daytime mechanical behavior, enabling accurate input for subsequent false alarm concentration interval location.

[0192] S7.3: Based on the distribution of high response intervals in the comprehensive statistical feature vector, cluster analysis algorithm is applied to locate the false alarm concentration intervals. The spatial mapping mechanism is used to associate and match the false alarm concentration intervals with specific sub-time period type labels to generate a false alarm distribution map that identifies the weak links of the model.

[0193] Based on the high-response interval distribution input in the comprehensive statistical feature vector, density peak localization processing is performed in the two-dimensional feature space, and the bivariate data consisting of abnormal response frequency and rhythm drift amplitude is mapped to the standardized coordinate system for unified analysis.

[0194] In a standardized coordinate system, a hierarchical clustering algorithm based on Euclidean distance weights is applied to iteratively merge neighboring high-response sample points according to the criterion of minimizing intra-cluster variance, forming a set of several candidate false alarm clusters with close feature associations.

[0195] For each candidate false alarm cluster set, calculate its centroid position, number of members, and mean square anomaly magnitude within the cluster, and select clusters whose centroid position is within the high response interval threshold and whose cluster density exceeds the preset lower limit as the final false alarm set interval candidates.

[0196] By using a spatial mapping mechanism, the centroid positions of candidate intervals in the false alarm set are projected onto the physiological semantic sub-period label space. By matching the projected coordinates with the sub-period label index, a one-to-one correspondence is established, thereby realizing the association and matching between the false alarm set intervals and specific sub-period types.

[0197] Based on the matching results, a false alarm distribution map is constructed. In the map, the label index is used as the horizontal axis and the centroid position of the high response interval is used as the vertical axis. The cluster density and mean square anomaly amplitude parameters are labeled to form structured and visualized data that can intuitively reflect the weak links of the model under different sub-time periods.

[0198] Through the above clustering analysis and spatial mapping processing, the comprehensive statistical feature vector results of the previous step are transformed into a false alarm distribution map that identifies the weak links of the model, thereby achieving accurate localization and visualization of the areas prone to false alarms in the isolated forest detection algorithm under non-stationary oral environment.

[0199] S7.4: Combining mobile terminal push data packets, comprehensive statistical feature vectors, and false alarm distribution maps, structured data packaging technology is used to integrate and encode multi-source information to generate individualized model optimization suggestion packages containing information on the concentrated false alarm intervals.

[0200] Step S8: During the startup phase of the next orthodontic retainer wearing, input the individualized model optimization suggestion package to update the subspace partitioning mechanism and splitting parameter baseline of the adaptive isolated forest model. Specifically, this includes: S8.1: Parse the individualized model optimization suggestion package stored in the mobile terminal cache to extract the sub-period type label set containing false alarm set interval information, rhythm drift trend vector and historical best subspace split parameter combination, and generate the model configuration data stream to be updated.

[0201] S8.2: Based on the set of sub-period type labels in the model configuration data stream to be updated, the initial subspace partitioning boundary of the adaptive isolated forest model is remapped to correct the sub-period segmentation deviation caused by individual rhythm drift and generate a dynamic subspace partitioning mechanism that adapts to the current physiological behavior characteristics.

[0202] Based on the set of sub-segment type labels contained in the model configuration data stream to be updated, each label is loaded into the decoding module to form a sub-segment type index table, which serves as the reference source for mapping the segmentation boundaries of the current model state. The boundary coordinates of each sub-segment within the initial subspace partitioning mechanism are precisely read and matched with the target sub-segment labels recorded in the index table to generate a boundary difference matrix. This matrix reflects the offset distribution between the current model segmentation boundary and the ideal boundary after individual rhythm drift. A bidirectional offset calculation is performed on the boundary difference matrix, representing the offset as positive and negative displacement values ​​on the time axis. The displacement values ​​are dynamically corrected in conjunction with the rhythm drift trend vector to ensure that the correction coefficients conform to the individual's current chewing rhythm frequency and resting segment duration parameters. A linear mapping algorithm is used to convert the corrected displacement values ​​into corresponding feature space segmentation point coordinates. A coordinate replacement operation is then used to remap the boundaries of each segment of the initial subspace partitioning mechanism, aligning the spatial distribution of the mechanism with actual physiological behavior patterns. After remapping, each subspace undergoes continuity and boundary interval checks to ensure that the boundary difference between any two adjacent subspaces is greater than the minimum sampling period and satisfies the logical order constraints of the sub-segment, thereby generating a dynamic subspace partitioning mechanism that adapts to the current physiological behavioral characteristics. Through the above remapping process, the sub-segment type label set obtained in the previous step is transformed into a dynamic subspace segmentation structure with individualized rhythm compensation capabilities, enabling the model to adapt to individual differences in oral biomechanics.

[0203] For example, for an orthodontic retainer wearer, the individualized model optimization suggestion package includes a set of labels for the occlusal loading segment, the resting maintenance segment, and the micro-adjustment segment. The rhythm drift trend vector is that the average lengthening of the resting segment during the day is 0.15 seconds, and the average shortening of the occlusal loading segment is 0.1 seconds. During this step, the initial subspace partitioning mechanism uses 1250 sample points for the resting maintenance segment boundary and 750 sample points for the occlusal loading segment boundary. In the boundary difference matrix generated after loading the label set, the offset of the resting maintenance segment is +15 sample points, and the offset of the occlusal loading segment is -10 sample points. Combined with the rhythm drift trend vector, the offsets are dynamically corrected, resulting in a corrected value of +18 sample points for the resting maintenance segment and -12 sample points for the occlusal loading segment. A linear mapping algorithm is used to convert the corrected values ​​into feature space segmentation coordinates, adjusting the resting maintenance segment boundary to 1268 sample points and the occlusal loading segment boundary to 738 sample points. During the verification phase, the adjusted boundary differences were all greater than the minimum sampling period of 5 sampling points, and the segmentation logic was correct. The output dynamic subspace partitioning mechanism can significantly improve the model's segmentation accuracy for the wearer in a non-stationary oral environment in subsequent detection, reduce the segmentation deviation caused by rhythm drift, and improve the stability and real-time performance of anomaly detection.

[0204] S8.3: Using the historical optimal subspace splitting parameter combination in the model configuration data stream to be updated, the directional splitting hyperplane normal vectors corresponding to each sub-period within the dynamic subspace partitioning mechanism are weighted and iteratively updated to replace the random initialization parameters of the previous period, generating a set of directional splitting hyperplanes with individualized prior knowledge.

[0205] The historical optimal subspace splitting parameter combination in the model configuration data stream to be updated is used for parameter calculation. The principal component direction vectors of each sub-time period included in the combination are indexed and paired with their corresponding splitting thresholds according to type labels, forming a parameter matching table for the current iteration. Based on this parameter matching table, a vector weighting operation is performed on the directional splitting hyperplane normal vectors corresponding to each sub-time period within the dynamic subspace partitioning mechanism. The weighting coefficients are derived from the normalized results of the historical detection stability index and the current rhythm drift deviation value. The corrected normal vector value is calculated using the following weighted update formula: in, This is the normal vector of the directional splitting hyperplane from the previous cycle. The normal vector in the historical optimal parameters. The weighting coefficients are used. The updated normal vectors are normalized to maintain unit length characteristics and ensure the stability of geometric orientation. After normalization, the normal vectors are re-injected into the corresponding node structures of the dynamic subspace partitioning mechanism according to the sub-time period type, replacing the randomly initialized parameters of the previous cycle. Consistency constraint checks are performed on the replaced node structures to verify whether the hyperplane orientation is parallel to the principal component direction vector of the sub-time period and whether there is any angular deviation exceeding the limit. Through this iterative update chain, the partitioning boundary of the previous step is fused with the historical optimal splitting parameter combination to generate a set of oriented splitting hyperplanes with individualized prior knowledge, achieving structural adaptation optimization of the model in a dynamic oral environment.

[0206] For example, in the historical optimal parameter combination of a certain individual during the resting maintenance phase, the principal component direction vector is (0.612, 0.487, 0.623), the splitting threshold is 2.15, the detection stability index is 0.82, the rhythm drift bias is 0.12, and the normalized weight coefficient ω takes the value of 0.87 in the current iteration. The normal vector update result is calculated according to the formula. The previous cycle normal vector (0.598, 0.502, 0.621) is weighted and synthesized with the historical optimal normal vector to obtain the corrected normal vector (0.610, 0.492, 0.622). This vector is then normalized, maintaining a length of 1.000. After replacing the normal vector of the resting maintenance phase node in the dynamic subspace partitioning mechanism, the direction deviation is confirmed to be less than 0.5° by a consistency check, and the model update is complete. This operation significantly improves the stability of anomaly detection in actual testing, significantly reduces the number of false alarms during the resting maintenance phase, and significantly enhances the ability of the retainer to identify deformation risks during long-term wear.

[0207] S8.4: Based on the rhythmic drift trend vector in the model configuration data stream to be updated, perform dynamic offset compensation calculation on the local path length calibration factor benchmark value associated with the directional split hyperplane set to eliminate systematic measurement errors caused by long-term wear and generate a local path length calibration factor library that has been thermally corrected.

[0208] Based on the rhythmic drift trend vector in the model configuration data stream to be updated, trend analysis calculations are performed on the local path length calibration factor baseline value associated with each principal component direction in the directional splitting hyperplane set. The rhythmic drift vector is projected into the physiological action sub-segment space to obtain the quantified drift amplitude value for each sub-segment. An offset compensation coefficient matrix is ​​constructed using the quantified drift amplitude value, and the mapping relationship model between the compensation coefficient and the drift amplitude is solved using linear regression or polynomial fitting methods. This mapping relationship model is used to dynamically adjust the local path length calibration factor baseline value. Specifically, a compensation coefficient corresponding to the current drift amplitude is applied to the calibration factor for each sub-segment type, achieving numerical correction of systematic measurement errors introduced by long-term wear. For the compensation coefficient calculation, the root mean square error minimization criterion is used, and the mean offset caused by drift is calculated using the following formula:

[0209] in, This is the path length observation for the current period. As the baseline value, The sample size for this sub-period is given. The mean offset is then used as a compensation coefficient to superimpose the calibration factor baseline value, resulting in the corrected local path length calibration factor. After performing the above offset compensation process on all sub-periods, a thermally evolved corrected local path length calibration factor library is formed. Through dynamic offset compensation calculation and full-sub-period calibration factor updates, the results of the previous step are transformed into factor library data that eliminates measurement errors, thereby improving the long-term stability of the model in non-stationary oral environments.

[0210] S8.5: The dynamic subspace partitioning mechanism, the set of directional splitting hyperplanes, and the local path length calibration factor library are atomically encapsulated and loaded into the memory runtime environment to complete the state reset of the adaptive isolated forest model and generate an updated adaptive isolated forest model with real-time individual adaptation capabilities.

[0211] Preferably, after the user removes the orthodontic retainer and leaves it for a period of time before wearing it again, the initial stress, occlusal stress, and resting stress data after re-wearing can be quickly collected through a multi-channel flexible pressure sensor array and a micro strain sensing unit. Combined with the mandibular movement rhythm anchor point, physiological behavior segmentation and directional feature extraction are completed. By using an adaptive isolated forest model to compare with the historical normal baseline, abnormal stress distribution, local stress concentration, and stress mutation caused by minute deformation are accurately identified. An abnormal risk score is generated through confidence weighting. When the score exceeds the safety threshold, a deformation warning is immediately triggered, realizing real-time detection and timely reminder of minute deformations that are not visible to the naked eye after removal and re-wearing.

[0212] For example, after a user removes their orthodontic retainer and leaves it unsupported at room temperature for two consecutive days, minute, invisible deformations may occur due to factors such as tooth deformation, stress release from the retainer material, and environmental temperature fluctuations. Upon re-wearing, a multi-channel flexible pressure sensor array and a micro-strain sensing unit are immediately activated to simultaneously collect the initial stress, contact pressure distribution, and deformation gradient time-series data during the resting maintenance period after re-wearing. The posture stabilization signal output from the mandibular motion sensor is used as a rhythmic anchor point to complete spatiotemporal alignment, generating a hybrid sensor data stream.

[0213] The system automatically categorizes this data segment into the resting maintenance sub-period, extracts the current dominant mechanical features through sliding window truncation and principal component analysis, and generates a directional feature embedding space. Subsequently, it calls upon a directional splitting hyperplane and adaptive isolated forest model matched to the resting maintenance segment to compare the real-time data with the user's historical normal resting stress baseline (60–100 kPa). The detection shows that the current local stress peak reaches 195 kPa, and the average stress offset is 88 kPa, significantly exceeding the resting segment's safety threshold of 140 kPa, and the pressure distribution asymmetry deviates significantly from the normal range. After local path length calibration and weighting with the stability index of the three most recent tests, the system outputs a local abnormality risk quantification score of 3.7, exceeding the moderate deformation warning threshold of 3.0. It then determines that the retainer has undergone minor deformation and triggers a graded warning, pushing a notification to the mobile terminal indicating abnormal retainer fit and local stress concentration. The user is advised to promptly check the fit or have a follow-up appointment to avoid tooth displacement or accelerated retainer damage due to continuous abnormal force.

[0214] Preferably, during the long-term repeated wearing and removal of orthodontic retainers and daily use, time-series data such as occlusal stress amplitude, resting stress stability, and strain response consistency are continuously collected. Fatigue-sensitive features such as stress drift trend, stress cycle decay, and strain consistency deviation are extracted through physiological segmentation and principal component analysis. The long-term mechanical signals are trend analyzed and anomalies are quantified based on an adaptive isolated forest model. The material fatigue risk score is generated by combining anomaly frequency, rhythmic drift, and historical stability index. The graded warning of material performance degradation, fatigue deterioration, and even failure risk is output according to the score level. The material fatigue state is identified in advance before the retainer shows obvious damage, and replacement or re-examination suggestions are given.

[0215] The material fatigue risk quantification score is generated based on a time-series data anomaly detection method. The system first accumulates stress and strain sensing time-series data for each sub-period during long-term wear, including bite loading, resting maintenance, unloading transition, and micro-adjustment, establishing an individual normal mechanical baseline. Then, principal component analysis is used to extract fatigue-sensitive features such as stress attenuation rate, baseline drift, distribution consistency deviation, and cyclic fluctuation dispersion, generating a directional feature embedding space. Subsequently, an adaptive isolated forest model and a directional splitting hyperplane are used to calculate the anomaly path length, which is standardized using a material fatigue-specific local path length calibration factor to obtain an initial anomaly score. This score is then combined with the stability index of the most recent 7 days of testing results for confidence weighting, and integrates the daily anomaly response frequency and rhythmic drift trend to ultimately form a comprehensive material fatigue risk quantification score. This score accurately reflects the fatigue state of the retainer due to long-term use, including decreased elasticity, stress relaxation, and deterioration of mechanical properties. The system judges based on preset grading thresholds; when the score reaches the severe fatigue threshold, a replacement warning is immediately triggered.

[0216] For example, after more than 10 months of repeated daily wearing and removal, cyclical occlusal impact, oral temperature and humidity cycles, and material aging, orthodontic retainers gradually exhibit fatigue performance degradation, such as decreased elastic modulus, stress relaxation, and decreased mechanical response. During long-term wear, multi-channel time-series data were continuously collected across four sub-periods: occlusal loading, resting maintenance, unloading transition, and micro-adjustment. Key indicators included occlusal stress amplitude, resting stress stability, strain response consistency, and mechanical cycle decay rate. The system performed feature extraction segmented according to physiological behavior, and used principal component analysis to identify fatigue-sensitive features such as stress drift trends, cyclic load decay, and long-term baseline shift. An adaptive isolated forest model was used to compare trends and quantify anomalies in historical time-series data. The test revealed that the peak stress in the occlusal segment of the retainer decreased by 32% compared to the initial wearing stage, the resting segment stress baseline continuously drifted by 41 kPa, the stress cycle consistency deviation exceeded the normal range, and the frequency of abnormal responses increased significantly in the past 7 days. The system comprehensively calculates the material fatigue risk score by weighting the daily anomaly frequency, rhythm drift trend, and historical stability index. The score is 4.3, which exceeds the severe material fatigue warning threshold of 4.0. The system determines that the retainer has experienced significant performance degradation and is at risk of failure. The system then triggers a graded material fatigue warning and pushes a reminder to the mobile terminal that the retainer is fatigued and aging and should be replaced as soon as possible. At the same time, the fatigue characteristics and false alarm interval data are incorporated into the individualized model optimization package. The model parameters are then updated hot during the next wear and startup, continuously improving the long-term monitoring accuracy and warning reliability.

[0217] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0218] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0219] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for early warning of orthodontic retainer deformation based on anomaly detection of time-series data, characterized in that, include: The raw time-series data collected by the multi-channel flexible pressure sensor array and micro strain sensing unit of the orthodontic retainer are acquired, and the mandibular motion sensor signal is recorded simultaneously as a rhythm anchor point to generate a hybrid sensing data stream. The hybrid sensing data stream is automatically segmented into bite loading segment periods, unloading transition segment periods, resting maintenance segment periods, and micro-motion adjustment segment periods with clear physiological semantics, in order to construct a segmented time-series dataset with physiological behavior labels; For each sub-time period type in the segmented time series dataset, a sliding window is used to extract local data blocks and the high-dimensional sensing features are compressed using the principal component analysis algorithm. The principal component direction vectors corresponding to the dominant mechanical modes in each sub-time period are calculated to generate a directional feature embedding space. By replacing the random feature selection strategy with the principal component direction vector in the directional feature embedding space, a directional splitting hyperplane aligned with the dominant mechanical mode of the current sub-period is constructed. Combined with the historical optimal subspace splitting parameter combination, an adaptive isolated forest model is established. Acquire the data to be detected, identify the sub-time period type to which the data to be detected belongs, call the local path length calibration factor of the sub-time period type, and apply confidence weighting to the initial anomaly score in combination with the stability index of the three most recent detection results to generate a local anomaly risk quantification score. Determine whether the local anomaly risk quantification score exceeds a preset safety threshold. If it is determined to exceed the threshold, generate a graded early warning instruction; otherwise, generate a normal state maintenance signal.

2. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 1, characterized in that, After generating the tiered early warning instruction, the method further includes: Based on the aforementioned tiered early warning instructions, follow-up consultation prompts or replacement suggestions are pushed to mobile terminals. At the same time, abnormal response frequency and rhythm drift trend data of each sub-period of the day are aggregated to generate an individualized model optimization suggestion package containing information on the concentrated false alarm interval. During the next initiation phase of orthodontic retainer wearing, the individualized model optimization suggestion package is input to update the subspace partitioning mechanism and splitting parameter benchmark of the adaptive isolated forest model.

3. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 1, characterized in that, The establishment of the adaptive isolated forest model includes: Based on the principal component direction vector of the directional feature embedding space output, the candidate split dimension set of isolated tree nodes is subjected to constraint mapping processing, which reduces the originally random feature selection range to only include the principal component axis corresponding to the dominant mechanical mode of the current sub-period, so as to generate a restricted feature candidate set. Using the principal component direction vector in the restricted feature candidate set as the normal reference, the geometric orientation of the node split hyperplane is oriented and constructed to replace the traditional random cutting method perpendicular to the coordinate axis, so as to generate an oriented split hyperplane that is strongly coupled with the current occlusal loading or resting maintenance and other physiological behaviors. Based on the construction results of the directional splitting hyperplane, the historical optimal subspace splitting parameter combination bound to the sub-time period type is retrieved and called. If a shift in rhythm pattern is detected, a lightweight retraining process is triggered to update the isolated tree weight distribution in order to generate a dynamic splitting parameter benchmark with temporal evolution capability. Based on the dynamic splitting parameter benchmark and the directional splitting hyperplane, a recursive binary splitting operation is performed to construct multiple isolated trees with physiological semantic guidance. In each splitting layer, the normal direction of the splitting surface is forced to be consistent with the direction vector of the principal component to generate an adaptive isolated forest model. The initial topology of the adaptive isolated forest model is checked for integrity and convergence. The directional splitting mechanism for each sub-time period is confirmed to be correctly injected and the weight update logic is correct. The adaptive isolated forest model is then output.

4. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 1, characterized in that, The generated local anomaly risk quantification score includes: Based on the segmented time-series dataset with physiological behavior labels, timestamp matching processing is performed on the newly incoming hybrid sensor data stream to determine the type of the current data, which is the biting loading segment, unloading transition segment, resting maintenance segment, or micro-motion adjustment segment, and to generate a data block to be detected. Using the physiological semantic attribution type of the data block to be detected, a local path length calibration factor that is strictly bound to the type is retrieved and called from the pre-stored adaptive isolated forest model parameter library to construct a normalized benchmark reference quantity for the current oral stress state distribution characteristics. Based on the normalized reference value, the traversal path length of the data block to be detected on the directional splitting hyperplane is standardized and calculated to obtain an initial anomaly score reflecting the degree of deviation of the mechanical state at the current moment. The initial anomaly score sequence generated in the three most recent consecutive detection cycles is extracted, and its fluctuation dispersion is calculated using the sliding variance algorithm to characterize the stability index of the detection process. The stability index is then used as a confidence weight coefficient to apply to the current initial anomaly score to generate a weighted anomaly score that has been corrected by time-series smoothing. By combining the weighted anomaly scoring and rhythm consistency constraint mechanism, the scoring results with drastic fluctuations are logically verified and finally fused to output the local anomaly risk quantification score that can accurately reflect the actual deformation risk of the retainer in a non-stationary oral environment.

5. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 1, characterized in that, The process of generating a hybrid sensing data stream includes: initializing, converting analog data to digital data and timestamping the multi-channel flexible pressure sensor array and the micro strain sensing unit to generate a digital sensing data sequence; performing attitude calculation on the mandibular motion sensor to obtain a rhythmic anchor event sequence; and then encapsulating the digital sensing data sequence and the rhythmic anchor event sequence after spatiotemporal alignment.

6. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 1, characterized in that, The construction of the segmented time-series dataset with physiological behavior labels includes: performing peak and zero-crossing detection and state machine judgment on the mandibular motion signal in the hybrid sensor data stream, generating an event label set, and using the event label set to segment, reassemble, filter and label the hybrid sensor data stream to obtain the segmented time-series dataset.

7. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 1, characterized in that, The process of generating the directional feature embedding space includes: performing sliding window truncation on each sub-time period in the segmented time series dataset, extracting a high-dimensional feature matrix containing orthogonal pressure, shear stress, and deformation gradient, obtaining principal component direction vectors through principal component analysis, and generating the directional feature embedding space through projection compression.

8. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 7, characterized in that, The high-dimensional feature matrix includes orthogonal pressure components, shear stress components, and deformation gradient.

9. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 1, characterized in that, Both the tiered early warning command and the normal state maintenance signal include timestamps, physiological behavior labels, trigger threshold parameters, and confidence weight information.

10. The orthodontic retainer deformation early warning method based on time-series data anomaly detection according to claim 2, characterized in that, The individualized model optimization suggestion package includes sub-period type labels, rhythm drift trend vectors, and historical best split parameter sets for adjusting subspace partition boundaries. These are used to seamlessly inject the model runtime environment when the orthodontic retainer is worn for the next start-up, completing the closed-loop parameter thermal evolution.

Images