Touch operation accurate recognition method and system of smart bracelet

Abstract

The application relates to the technical field of data processing, and discloses a touch operation accurate identification method and system of an intelligent bracelet, wherein the core of the method is to collect bracelet touch screen capacitance, motion sensing and pressure multidimensional data, establish an analysis time window, extract each dimension feature, complete timestamp alignment and normalization, and construct a multimodal feature vector; then a time sequence causal determination result is output through causal inspection, a joint determination space is constructed to obtain a spatial classification determination result, a dynamic threshold determination result is output by combining a motion intensity dynamic threshold, finally, three types of results are integrated based on a Bayesian decision theory to output a final identification result, feedback data is collected to optimize model parameters and update a sample library, and a closed loop identification mechanism is formed. Through multimodal data processing and multidimensional determination, the application effectively defines true and false touch operations, continuously improves identification accuracy through closed loop optimization, reduces misoperations, and greatly enhances the reliability of the intelligent bracelet touch control system.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a method and system for accurate recognition of touch operations on a smart bracelet. Background Technology

[0002] Smart bracelets, as portable wearable devices, are widely used in daily exercise and health monitoring scenarios, with their touchscreens being the core component for user interaction. However, during everyday physical activities such as running, walking, and arm swinging, the vibrations and acceleration changes generated by the movement are transmitted to the touchscreen, causing the touch sensor to receive interference signals from non-user-initiated touches. Current smart bracelet touch systems lack effective signal recognition and differentiation mechanisms, making it difficult to accurately distinguish between user-initiated touch operations and false touch signals caused by movement. This leads to frequent misoperations during exercise, severely impacting the user experience and device reliability. Summary of the Invention

[0003] The main purpose of this application is to provide a method and system for accurate recognition of touch operation on a smart bracelet, aiming to solve the technical problem that when a smart bracelet is in motion, interference signals are generated due to vibration and acceleration changes, and the touch system cannot distinguish between real and fake touches, which easily leads to misoperation and affects the user experience and reliability.

[0004] The first aspect of this application proposes a method for accurate recognition of touch operations on a smart bracelet, including: Collect capacitive signals, motion sensing data, and pressure data from the smart bracelet touchscreen. Establish an analysis time window based on the touch event trigger time. Extract frequency and time domain features from the capacitive signals. Filter the motion sensing data to extract motion features. Align and normalize all feature timestamps to form a multimodal feature vector. The multimodal feature vectors are aligned to a unified reference coordinate system along the time dimension. The temporal relationship between touch and motion signals is analyzed, and the causal direction between the two is determined by a causal test algorithm. The temporal causal determination result is then output. Based on the multimodal feature vector, touch trajectory, motion posture and signal energy features are extracted, a joint judgment space is constructed, a judgment boundary is constructed using historical samples, the current touch event features are mapped to the judgment space for classification, and the space classification judgment result is output. The motion intensity is determined based on the multimodal feature vector, the judgment threshold is dynamically adjusted, and the dynamic threshold judgment result is output by combining the analysis and judgment of the touch event timing pattern. By integrating the aforementioned three types of judgment results, calculating the posterior probability based on Bayesian decision theory and outputting the final recognition result, collecting feedback data to optimize model parameters and judgment thresholds, and updating the sample library to form a closed-loop recognition mechanism.

[0005] Furthermore, the steps of aligning the multimodal feature vectors to a unified reference coordinate system along the time dimension, analyzing the temporal relationship between touch and motion signals, determining the causal direction between the two using a causal testing algorithm, and outputting the temporal causality determination result include: Extract motion patterns, motion intensity, and duration of motion mutations, and establish mapping relationships between these three factors and initial delay thresholds, causality test thresholds, and time series analysis windows; According to the mapping relationship, set various parameters, locate the peak time of the touch signal and the sudden change time of the motion signal in the dynamic analysis window, calculate the time difference between the two to complete the initial timing determination, and simultaneously construct a multimodal weighted timing feature sequence. The multimodal weighted time series feature sequence is incorporated into the causal test vector autoregressive model. The model order is adjusted according to the initial time series judgment result, and the causal judgment result is obtained through the test algorithm. The initial time series determination result and the causal determination result are integrated. If the confidence level of the determination result does not reach the preset threshold, the threshold and the analysis window length are dynamically adjusted according to the motion intensity and iteratively optimized until the result is consistent, and the time series causal determination result is output.

[0006] Further, the steps of setting various parameters according to the mapping relationship, locating the peak moment of the touch signal and the moment of sudden change of the motion signal within the dynamic analysis window, calculating the time difference between the two to complete the initial timing determination, and simultaneously constructing a multimodal weighted timing feature sequence include: By integrating exercise patterns, exercise intensity, and historical data, a multi-dimensional mapping relationship is constructed, and a gradient optimization algorithm is used to iteratively optimize the dynamic analysis window and feature weight parameters; Based on the optimized parameters and combined with the pressure signal stability calibration feature weights, multimodal time series data are extracted to construct a correlation-enhanced weighted time series feature sequence. Within the dynamic analysis window, based on the effective contact time of pressure, the peak time of the touch signal and the abrupt change time of the motion signal are located, and the combined time difference between the two is calculated to complete the initial timing determination. The calibrated weighted time series feature sequence and the initial time series judgment result are input into the causality test process, and the feature attention is adjusted according to the motion intensity.

[0007] Further, the steps of extracting touch trajectory, motion posture, and signal energy features based on the multimodal feature vectors, constructing a joint judgment space, constructing a judgment boundary using historical samples, mapping the current touch event features to the judgment space for classification, and outputting the space classification judgment result include: The touch trajectory, motion posture, and signal energy feature weights are dynamically allocated according to the motion mode and motion intensity. Pressure features are integrated to construct a multi-dimensional feature set and establish a mapping relationship between motion intensity and kernel function. The multidimensional feature set is reduced to three dimensions using a feature dimensionality reduction algorithm, while retaining the core distinguishing information. It is then mapped to an enhanced three-dimensional joint decision space, and the support vector machine is initialized by selecting an adaptation kernel function based on the motion intensity. Dedicated decision boundaries are constructed based on touch intent, and the touch intent is predicted and matched with the corresponding boundary to complete the initial classification. The hyperplane parameters are optimized by cross-validation with other intent boundaries and combined with historical samples. Calculate the classification accuracy, pressure feature dimensionality reduction retention rate, and cross-validation conflict rate. If any indicator fails to meet the preset standard, readjust the feature weights, kernel function, and boundary parameters for iterative optimization until the indicator meets the standard. Map the current touch event features to the joint decision space to complete the classification and output the spatial classification decision result.

[0008] Furthermore, the steps of constructing dedicated judgment boundaries according to touch intent, predicting touch intents and matching corresponding boundaries to complete preliminary classification, and optimizing hyperplane parameters through cross-validation with other intent boundaries and combining historical samples include: The feature weights and kernel function parameters of touch intent are refined according to the intensity of motion, and the touch duration is added to assist in predicting touch intent; By combining real-time motion intensity with predicted touch intent, the specific boundary parameters of the corresponding intent are adjusted, and a support vector machine is used to complete the initial classification, while the classification confidence is calculated simultaneously. Cross-validate the preliminary classification results using other touch intent determination boundaries to verify the consistency of classification confidence and the matching degree between touch duration and touch intent; If the confidence consistency or duration intent matching degree does not meet the preset standard, the hyperplane penalty coefficient and kernel function parameters are optimized by combining historical misjudgment samples, the dedicated judgment boundary is retrained and iteratively optimized until the indicator meets the standard.

[0009] Furthermore, the steps of combining real-time motion intensity with the predicted touch intent, adjusting the corresponding boundary parameters of the intent, using a support vector machine to complete the initial classification, and simultaneously calculating the classification confidence include: Based on exercise intensity and exercise pattern, establish a mapping model between confidence threshold and misjudgment risk, and simultaneously build a correlation library between historical confidence and classification accuracy; By combining real-time motion intensity with the predicted touch intent, the specific boundary parameters of the corresponding intent and the appropriate kernel function type are adjusted with reference to the mapping model. Support vector machine is used to perform preliminary classification of the current touch event features, calculate the original confidence score of the classification, calibrate the original confidence score through the historical association database, and assess the risk of misjudgment based on the calibrated confidence score; If the classification confidence level meets the standard after calibration and the misjudgment risk does not meet the preset standard, proceed to other touch intent boundary cross-validation stages; if it does not meet the standard, mark it as a sample to be optimized and record the classification confidence fluctuation value, collect the sample to be optimized and fluctuation data, readjust the boundary parameters, confidence threshold and kernel function type, and iteratively train the model until the indicator meets the standard.

[0010] Further, the steps of determining motion intensity based on the multimodal feature vector, dynamically adjusting the judgment threshold, combining touch event timing pattern analysis, and outputting the dynamic threshold judgment result include: Based on multimodal feature vector recognition of motion patterns, specific motion intensity grading thresholds and threshold adjustment rules are set according to motion patterns, and an optimized sample library including historical threshold adjustments, judgment accuracy and boundary sample features is constructed simultaneously. Set the duration of the sliding window and continuously collect touch events. Calculate the statistical characteristics of motion intensity and the distribution of touch features within the window. Combine the optimized sample library to calibrate the initial judgment threshold. Dynamically adjust the threshold according to the motion mode and real-time motion intensity. Simultaneously label the boundary samples within the preset threshold range. Based on the adjusted judgment threshold, each touch event is initially judged. If the judgment results are consistent multiple times, the judgment conclusion is strengthened. If the judgment results are inconsistent or contain boundary samples, the periodicity of the motion pattern is analyzed and the judgment logic of the same historical scene in the sample library is matched. If a regular motion cycle exists, it is determined to be a systemic accidental touch due to motion interference; if there is no regularity, it is defined as a mixed touch scenario. All data, boundary samples, and judgment results are added to the optimization sample library, and the motion mode-specific threshold rules and calibration logic are iteratively optimized to output dynamic threshold judgment results.

[0011] Furthermore, the steps of setting the sliding window duration and continuously collecting touch events, calculating the statistical characteristics of motion intensity and the distribution of touch features within the window, calibrating the initial judgment threshold in conjunction with the optimized sample library, dynamically adjusting the threshold according to the motion mode and real-time motion intensity, and synchronously labeling boundary samples within the preset threshold range include: Subdivide the motion modes and construct an optimized sample library subset corresponding to each subdivision mode. Filter historical misjudged samples to establish a mapping relationship between misjudged scenarios and adaptation parameters. Set motion intensity fluctuation thresholds and window and threshold collaborative adaptation rules. Match real-time subdivided motion modes, combine the mapping relationship between the misjudged scenarios and the adaptation parameters, set the initial sliding window duration, continuously collect touch events, and calculate the statistical characteristics of motion intensity and the distribution of touch features within the window; If the fluctuation of exercise intensity exceeds the limit, the window duration and threshold adjustment range are adjusted, the initial judgment threshold is calibrated in combination with the sample library subset of the corresponding subdivision mode, the judgment threshold is dynamically adjusted according to the real-time exercise intensity, and the boundary samples in the preset threshold range are labeled simultaneously. The fluctuation adaptation effect, threshold deviation, boundary samples and misjudgment situations are fed back to the mapping relationship between the optimized sample library and the misjudgment scenario adaptation parameters, and the window, threshold rules and collaborative adaptation logic specific to the subdivision mode are iteratively optimized.

[0012] Furthermore, the steps of matching real-time subdivided motion patterns, combining the mapping relationship between the misjudged scenarios and adaptation parameters, setting the initial sliding window duration, continuously collecting touch events, and calculating the statistical features of motion intensity and the distribution of touch features within the window include: Historical misjudgment samples are clustered according to subdivided motion patterns, high-frequency misjudgment scenarios are given high weight, a weighted mapping relationship between misjudgment scenarios and window duration is constructed, historical statistical data of the same mode and intensity range are extracted, and touch feature pre-screening rules are set. Match real-time subdivided motion modes, combine the mapping relationship between the misjudged scene and the adaptation parameters and the weighted mapping between the misjudged scene and the window duration, set the initial sliding window duration, collect valid touch events according to the touch feature pre-screening rules, calculate the touch frequency synchronously, and verify and fine-tune the sliding window duration; Calculate the motion intensity statistical features and the filtered touch feature distribution within the fine-tuned sliding window, compare and calibrate them with the historical statistical data of the same intensity range in the corresponding sub-mode, and mark the feature data similar to the misjudged scenario. The similarity of features, motion intensity fluctuations, calibration deviations, and touch frequency data are fed back to the weighted mapping of misjudged scenarios and window duration and the sample library, and the window adaptation rules, touch feature pre-screening rules, and historical data migration logic are iteratively optimized.

[0013] A second aspect of this application also proposes a precise touch operation recognition system for a smart bracelet, characterized in that it includes: The feature extraction module is used to collect capacitive signals, motion sensing data and pressure data from the smart bracelet touch screen. It establishes an analysis time window based on the touch event trigger time, extracts frequency domain and time domain features from the capacitive signals, filters and extracts motion features from the motion sensing data, aligns and normalizes all feature timestamps, and forms a multimodal feature vector. The temporal causality determination module is used to align the multimodal feature vectors to a unified reference coordinate system according to the time dimension, analyze the temporal relationship between touch and motion signals, determine the causal direction of the two by combining the causality test algorithm, and output the temporal causality determination result. The spatial classification and determination module is used to extract touch trajectory, motion posture and signal energy features based on the multimodal feature vector, construct a joint determination space, construct a determination boundary using historical samples, map the current touch event features to the determination space for classification, and output the spatial classification and determination result. The dynamic threshold determination module is used to determine the motion intensity based on the multimodal feature vector, dynamically adjust the determination threshold, combine the touch event timing pattern analysis and determination, and output the dynamic threshold determination result. The fusion decision module is used to integrate the aforementioned three types of judgment results, calculate the posterior probability based on Bayesian decision theory and output the final recognition result, collect feedback data to optimize model parameters and judgment thresholds, and update the sample library to form a closed-loop recognition mechanism.

[0014] The first aspect of this plan brings the following benefits: This application collects multimodal data and extracts multidimensional features, and combines temporal causality tests to distinguish between touch and motion interference signals; it constructs a joint judgment space and dynamically adjusts thresholds to accurately define true and false touch; it integrates three types of judgment results and continuously improves recognition accuracy through closed-loop optimization, reduces misoperation, and significantly improves the reliability of smart bracelet touch systems and user experience. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for accurately recognizing touch operations on a smart bracelet according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a smart bracelet touch operation accurate recognition system according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of this application; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0018] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0019] Reference Figure 1 This application provides a method for accurate recognition of touch operations on a smart bracelet, including: S1: Collect capacitive signals, motion sensing data and pressure data from the smart bracelet touch screen. Establish an analysis time window based on the touch event trigger time. Extract frequency domain and time domain features from the capacitive signals. Filter the motion sensing data to extract motion features. Align and normalize all feature timestamps to form a multimodal feature vector. S2: Align the multimodal feature vectors to a unified reference coordinate system along the time dimension, analyze the temporal relationship between touch and motion signals, combine the causal test algorithm to determine the causal direction between the two, and output the temporal causal determination result; S3: Extract touch trajectory, motion posture and signal energy features based on the multimodal feature vector, construct a joint judgment space, construct a judgment boundary using historical samples, map the current touch event features to the judgment space for classification, and output the space classification judgment result; S4: Determine the motion intensity based on the multimodal feature vector, dynamically adjust the judgment threshold, combine the touch event timing pattern analysis and judgment, and output the dynamic threshold judgment result; S5: Integrate the aforementioned three types of judgment results, calculate the posterior probability based on Bayesian decision theory and output the final recognition result, collect feedback data to optimize model parameters and judgment thresholds, and update the sample library to form a closed-loop recognition mechanism.

[0020] In step S1, through multi-source heterogeneous data synchronous acquisition technology and Gauss-Kalman two-level filtering algorithm, combined with mutual information entropy feature screening criteria, a dynamic analysis time window is constructed with the touch event trigger time as the time zero point. Multi-dimensional features of capacitance, motion, and pressure are extracted and timestamp alignment and normalization are completed, finally generating a multi-modal feature vector. Taking high-intensity outdoor running (exercise intensity ≥ 0.8g) as an example, the standard analysis window is 150ms before and after the touch trigger moment. After the smart bracelet collects the touch screen capacitance fluctuation signal generated by the user's arm swing, the high-frequency motion data of the three-axis accelerometer, and the contact data of the pressure sensor (including false touch signals caused by motion interference), the system first locks the precise timestamps of the three types of data through hardware triggering + software synchronization technology, synchronously identifies and marks the exercise mode as outdoor running, and generates the original data set—the set contains the voltage change curve of the capacitance signal, the three-axis acceleration value of the accelerometer, the angular velocity value of the gyroscope, and the contact pressure value of the pressure sensor. Subsequently, based on the exercise intensity adaptation feature extraction strategy, the high-frequency frequency domain features and peak amplitude are extracted from the capacitance signal using a short-time Fourier transform with a 5ms window. The motion sensing data is initially denoised using Gaussian filtering with σ-value adaptation, and then optimized by three iterations of Kalman filtering to extract motion features such as peak angular velocity. Simultaneously, contact peak features of the pressure signal are extracted. The feature dimensions are adapted based on the capacitance signal signal-to-noise ratio (set to 25dB, medium level). Redundant features with correlation <0.3 are eliminated using mutual information entropy, while effective features with correlation ≥0.7 in the pressure data are retained. Time offset is corrected using dynamic time warping based on the capacitance signal timestamp, and normalization is performed according to the feature weights for outdoor running (capacitance 45% + motion 45% + pressure 10%). Finally, the three types of features are concatenated by timestamp to form a high-discrimination multimodal feature vector adapted to the outdoor running scenario, providing a standardized feature data foundation for subsequent steps such as temporal causality determination in S2 and spatial classification determination in S3. This step enables the synchronous acquisition and precise preprocessing of multi-source data. By using dynamic windows and customized feature extraction to adapt to different motion scenarios, and removing redundant features to form a multimodal feature vector, it provides highly discriminative standardized data for subsequent judgment stages, thereby improving the accuracy of touch recognition from the source.

[0021] In step S2, the multimodal feature vectors generated in S1 are aligned to a unified reference coordinate system using time dimension alignment technology and gradient descent optimization algorithm, combined with Granger causality test algorithm. The temporal relationship between touch and motion signals is analyzed, and the causal direction is determined, ultimately outputting the temporal causality determination result. Taking the multimodal feature vectors of a high-intensity outdoor running scenario as an example, with an initial latency threshold of 5ms, an F threshold of 3.5, and a temporal analysis window of 35ms, the system first extracts the motion mode, intensity, and abrupt change duration. Parameters are set according to the mapping relationship, and a correlation-enhanced weighted temporal feature sequence calibrated for pressure stability is constructed. Using the effective pressure contact time as a benchmark, the peak value of touch capacitance and the abrupt change time of motion acceleration are located within the dynamic window, and the comprehensive time difference is calculated to complete the initial temporal determination. Subsequently, the weighted sequence is integrated into a Granger test vector autoregressive model. Based on the initial determination result, the model order is set to 3, and the causal relationship is determined through the F test. Based on the judgment result, and with a learning rate of 0.01 for the motion intensity, the window and weight parameters are iteratively optimized with the goal of minimizing the positioning error, verifying the convergence of parameters and stress adaptability. If the initial temporal and causal judgment confidence is lower than 0.8, the delay threshold and window length are dynamically adjusted according to the motion intensity for iterative optimization until the results are consistent. Finally, it is determined that in this outdoor running scenario, the touch peak occurs after the moment of motion change and the time difference is less than 5ms. Combined with the causal test results, the false touch judgment is strengthened, and the temporal causal judgment result is output as a false touch caused by motion, providing a temporal dimension judgment basis for subsequent S3, S4 judgments and S5 comprehensive decision-making. This step achieves accurate temporal positioning of touch and motion signals through dynamic parameter adaptation and weighted temporal analysis. Combined with the dual judgment of Granger causality test, the reliability of temporal causal judgment is greatly improved, providing accurate temporal basis for subsequent multi-dimensional judgments.

[0022] In step S3, the system extracts core features from the multimodal feature vectors generated in S1 using the KPCA feature dimensionality reduction algorithm and the support vector machine classification algorithm, combined with the confidence-misjudgment risk mapping criterion. This constructs a joint decision space, completes the touch event classification, and outputs the spatial classification result. Taking the multimodal feature vectors of a high-intensity outdoor running scenario as an example, the system first allocates weights according to the feature weights of the scenario (posture 40% + energy 35% + trajectory 25%), integrates pressure features to construct a four-dimensional feature set, matches the RBF kernel function, uses KPCA to reduce the four-dimensional features to three dimensions, maps them to the enhanced three-dimensional joint decision space, and initializes the support vector machine. Then, it predicts the touch intent as a point touch, adjusts the specific boundary parameters, performs preliminary classification of feature points, calculates the original confidence, and assesses the misjudgment risk after calibration using the historical confidence-accuracy association library. The confidence is then combined with pressure stability to determine the risk of misjudgment. The second calibration is incremented by 0.05. Subsequently, cross-validation is performed using the decision boundary of the swiping intent to verify the confidence consistency and duration-intent matching degree. Due to motion interference in this scenario, the confidence consistency is below 90%. The hyperplane penalty coefficient and kernel function σ value are optimized to 0.8 based on historical misjudgment samples. The boundary is retrained and reclassified. Finally, the feature points of this touch event are mapped to the joint decision space, determining that it falls within the mis-touch sample area. The output spatial classification result is a motion-induced false touch, echoing the S2 decision result and providing spatial dimension determination basis for the S5 comprehensive decision. This step, through feature dimensionality reduction and dynamic kernel function adaptation, constructs a highly discriminative joint decision space. Combined with hierarchical intent determination and cross-validation, it improves the accuracy of spatial classification and provides reliable spatial dimension determination results for comprehensive decision.

[0023] In step S4, by using sliding window statistical analysis technology and clustering algorithm, combined with loss-oriented feature pre-screening criteria, the motion intensity is determined based on the multimodal feature vector generated in S1 and the judgment threshold is dynamically adjusted. After analyzing the timing pattern of touch events, the dynamic threshold judgment result is output. Taking the multimodal feature vector of a high-intensity outdoor running scenario as an example, the system first identifies the exercise mode as outdoor running, matches the sample library subset of this sub-mode, sets the initial sliding window duration according to the weighted mapping of misjudged scenarios, collects continuous touch events and calculates the motion intensity fluctuation value and touch feature distribution within the window. Due to excessive motion intensity fluctuation, the window duration and threshold adjustment range are increased simultaneously, the initial pressure and energy ratio judgment thresholds are calibrated and appropriately increased, and the boundary samples in the ±5% range of the threshold are marked. Based on the adjusted threshold, the touch event is initially judged, and it is found that the judgment result is controversial and contains boundary samples. Further analysis of the periodicity of the exercise mode reveals that there is a regular arm swing motion cycle in outdoor running. Matching the judgment logic of similar scenarios in the sample library, the touch event is judged to be a systematic false touch caused by motion interference. Subsequently, the data, boundary samples and judgment results are added to the outdoor running sample library subset, and the window and threshold rules of this sub-mode are iteratively optimized. Finally, the dynamic threshold judgment result is output as a false touch caused by motion, which is consistent with the judgment results of S2 and S3, providing a threshold dimension judgment basis for scenario adaptation for S5 comprehensive decision-making. This step accurately identifies systematic accidental touches caused by motion interference by subdividing motion modes and dynamically adapting window thresholds. At the same time, it continuously iterates and optimizes the sample library rules, improving the adaptability of threshold judgment under different motion scenarios and providing scenario-based judgment results for comprehensive decision-making.

[0024] In step S5, the Bayesian decision fusion algorithm and clustering analysis algorithm are combined with the loss-weighted closed-loop optimization criterion to integrate the three types of judgment results: S2 temporal causality, S3 spatial classification, and S4 dynamic threshold. The posterior probability is calculated to output the final recognition result and complete the model closed-loop optimization. Taking the three types of judgment results in a high-intensity outdoor running scenario as an example, all three are judged as false touches caused by movement. The system first assigns dynamic weights according to the historical accuracy of the three types of judgments, quantifies the loss of false touches and missed judgments, and constructs a Bayesian decision model with a loss function, setting a posterior probability threshold that is dynamically adjusted with the loss. After inputting the three types of judgment results into the model, the posterior probability of the touch event being a real touch is calculated based on Bayesian decision theory. The result is much lower than the preset threshold, so the final recognition result is to refuse to respond to the false touch operation and record the event as a motion interference-type false touch. Then the loss is calculated. The weighted false positive and false negative rates are analyzed using a clustering algorithm to identify the characteristics of these difficult samples, edge samples are identified, and the pattern library is updated. New samples are added to the sample library in a stratified manner according to motion interference type. A differentiated incremental learning strategy is used to optimize the classification model. Finally, according to the parameter linkage rules, the time delay threshold of S2, the decision boundary of S3, and the adaptive threshold of S4 are simultaneously fine-tuned to optimize the loss function and dynamic weights of the Bayesian model. All optimized data are fed back to the feature extraction stage of S1, forming a closed-loop iterative optimization process. This provides better model parameters for touch recognition in similar outdoor running scenarios. This step achieves accurate fusion of multi-dimensional judgment results through Bayesian decision-making, outputting a final reliable touch recognition result. At the same time, relying on the closed-loop optimization mechanism, the model parameters are continuously iterated, allowing the touch recognition accuracy to continuously improve with use.

[0025] In one embodiment, the steps of acquiring capacitive signals, motion sensing data, and pressure data from the smart bracelet touchscreen, establishing an analysis time window based on the touch event trigger time, extracting frequency and time domain features from the capacitive signals, filtering and extracting motion features from the motion sensing data, and aligning and normalizing all feature timestamps to form a multimodal feature vector include: S10: Simultaneously collects touch screen capacitive signals, three-axis accelerometer data, gyroscope data and pressure data, locks precise timestamps, dynamically sets analysis time windows based on motion intensity, and identifies and marks motion patterns; S11: Customize the feature extraction strategy according to the motion mode, transform the capacitance signal to extract frequency domain and time domain features, and filter and denoise the motion sensing data to extract motion features. S12: Based on the signal-to-noise ratio adaptation feature dimension of the capacitor signal, use mutual information entropy to remove redundant features and combine the correlation of pressure data to retain effective features. S13: Based on the timestamp of the capacitor signal, the time offset is corrected by the time alignment algorithm, the normalization strategy is adapted according to the feature quality level, and the feature weights are concatenated according to the motion mode to form a multimodal feature vector.

[0026] In this embodiment, firstly, multi-source heterogeneous data from the smart bracelet is collected synchronously and the analysis window is dynamically set. Taking a high-intensity outdoor running scenario (exercise intensity ≥ 0.8g) as an example, the system uses hardware triggering + software synchronization technology to simultaneously collect the touch screen capacitor voltage change signal, the high-frequency acceleration data of the three-axis accelerometer, the angular velocity data of the gyroscope, and the contact pressure data of the pressure sensor. It accurately locks the unified timestamp of the four types of data, determines the exercise intensity ≥ 0.8g by calculating the root mean square value of the exercise data, and sets the analysis time window to 150ms before and after the touch event trigger time according to the rules. At the same time, the current exercise mode is identified as outdoor running by exercise feature matching and a special mark is made, which serves as the basis for defining the scene for subsequent customized feature extraction. Then, a feature extraction strategy was customized according to the outdoor running exercise mode. Addressing the need to extract high-frequency features in this scenario, a short-time Fourier transform with a 5ms window was used to perform frequency domain transformation on the capacitance signal, extracting high-frequency frequency domain features and time domain features such as peak amplitude and rising edge steepness. The motion sensing data was first denoised using Gaussian filtering with an adapted σ value, and then denoised a second time using three iterations of Kalman filtering to extract core motion features such as peak angular velocity and motion amplitude, achieving customized and accurate extraction of features across different dimensions. Next, feature dimensions were adapted and effective features were selected based on the capacitance signal-to-noise ratio (SNR). The capacitance signal SNR in the current scenario was detected to be 25dB, which is considered medium. The corresponding feature dimensions were adapted according to rules, and the correlation between each feature was calculated using the mutual information entropy algorithm. Redundant features with a correlation <0.3 were removed. Simultaneously, correlation verification was performed using pressure data, retaining only features with a correlation ≥0.7 with the pressure data, thus completing the accurate selection of effective features. Finally, using the timestamp of the capacitance signal as a benchmark, a dynamic time warping algorithm is employed to correct the time offset between the motion, pressure, and capacitance data, ensuring complete alignment of all feature time dimensions. Then, a normalization strategy is applied according to feature quality levels, and the selected effective features are weighted according to the feature weighting rules for outdoor running (capacitance 45% + motion 45% + pressure 10%). The three types of features are then sequentially concatenated by timestamp, ultimately forming a highly discriminative multimodal feature vector adapted to the high-intensity outdoor running scenario. This provides standardized feature data for subsequent time-series causal determination and spatial classification. This step achieves synchronous and accurate acquisition of multi-source data and scenario-specific customized processing. By eliminating redundant information through feature selection, and by using time alignment and weighted concatenation to form the feature vector, a highly accurate and discriminative standardized data foundation is provided for subsequent judgment stages.

[0027] In one embodiment, the steps of customizing the feature extraction strategy according to the motion mode, transforming the capacitance signal to extract frequency and time domain features, and filtering and denoising the motion sensing data to extract motion features include: S110: Based on motion pattern, extract feature frequencies and capacitance signal signal-to-noise ratio, and establish the mapping relationship between the two and capacitance signal transformation processing parameters and motion data filtering parameters; S111: Dynamically set transformation processing parameters according to the mapping relationship, extract the capacitance features of the customized frequency band, use a filtering algorithm to denoise the motion sensing data, and extract motion features; S112: The motion features are optimized using an iterative optimization algorithm. The correlation between the optimized motion features and the capacitance features is calculated. If the correlation does not meet the preset requirements, the transformation processing parameters and filtering parameters are adjusted to re-extract the features. S113: Combine the pressure features with the capacitance features and the optimized motion features for bidirectional cross-validation, eliminate low-correlation features, assign weights to various features according to motion modes, and form a customized high-discrimination feature set.

[0028] In this embodiment, firstly, feature frequencies and the signal-to-noise ratio (SNR) of the capacitance signal are extracted based on the motion mode, and a parameter mapping relationship is established. Taking a high-intensity outdoor running scenario as an example, the system first extracts the high-frequency characteristic band of the capacitance (100-500Hz) that needs to be focused on in this mode. At the same time, the capacitance signal SNR is detected to be 25dB (medium level). Combining the parameter mapping rules in the technical disclosure, the correspondence between the two and the capacitance signal transformation processing parameters and motion data filtering parameters is established. The short-time Fourier transform window length is determined to be 5ms, the overlap rate is 70%, the Gaussian filter σ value is 0.8, and the Kelvin filter is 0.8. The Mann filter was iterated three times to define precise parameters for feature extraction. Then, according to this mapping relationship, various processing parameters were dynamically set. The capacitance signal was subjected to a short-time Fourier transform with a window length of 5ms for frequency domain transformation, accurately extracting frequency domain features of a customized frequency band of 100-500Hz. At the same time, time domain features such as peak amplitude and rising edge steepness were extracted. For the motion sensing data of the triaxial accelerometer and gyroscope, Gaussian filtering with a σ value of 0.8 was first used for initial noise reduction, and then motion features such as peak angular velocity, motion amplitude, and attitude angle change were extracted to achieve customized preliminary feature extraction. Secondly, the Kalman filter iterative optimization algorithm was used to perform three iterations to optimize the initially extracted motion features. The correlation between the optimized motion features and capacitance features was calculated using the Pearson correlation coefficient algorithm. The result was 0.75, which met the preset requirement of ≥0.7. No parameter adjustment and re-extraction were required. If the requirement was not met, the Fourier transform window length and Gaussian filter σ value were adjusted in reverse, and features were extracted again until the correlation met the requirement. Finally, pressure features such as contact peak value and pressure stability extracted from the pressure sensor were combined to perform bidirectional cross-validation on capacitance features and optimized motion features. The correlation between the two types of features and pressure features was calculated separately. Low correlation features with correlation <0.6 were eliminated. Then, according to the feature weight allocation rules for outdoor running scenarios, capacitance features were assigned a weight of 45%, motion features 45%, and pressure features 10%. The three types of highly correlated features were integrated to form a customized high-discrimination feature set adapted to high-intensity outdoor running scenarios. This laid a high-quality feature foundation for subsequent timestamp alignment and normalization processing, forming a complete logical closed loop with the previous multi-source data collection and subsequent feature selection steps. This step achieves customized feature extraction by establishing precise parameter mapping relationships. Iterative optimization and cross-validation enhance feature relevance, resulting in a highly discriminative feature set after weight allocation, providing high-quality feature materials for subsequent multimodal feature vector construction.

[0029] In one embodiment, the steps of aligning the multimodal feature vectors to a unified reference coordinate system along the time dimension, analyzing the temporal relationship between touch and motion signals, determining the causal direction between the two using a causal testing algorithm, and outputting the temporal causality determination result include: S20: Extract motion patterns, motion intensity, and duration of motion mutations, and establish the mapping relationship between these three factors and the initial time delay threshold, causality test threshold, and time series analysis window; S21: Set various parameters according to the mapping relationship, locate the peak time of the touch signal and the sudden change time of the motion signal in the dynamic analysis window, calculate the time difference between the two to complete the initial timing determination, and simultaneously construct a multimodal weighted timing feature sequence; S22: Integrate the multimodal weighted time series feature sequence into the causal test vector autoregressive model, adjust the model order according to the initial time series judgment result, and obtain the causal judgment result through the test algorithm; S23: Integrate the initial time series determination result and the causal determination result. If the confidence level of the determination result does not reach the preset threshold, dynamically adjust the threshold and the analysis window length according to the motion intensity and iteratively optimize until the result is consistent, and output the time series causal determination result.

[0030] In this embodiment, the exercise mode, exercise intensity, and duration of the sudden change in exercise intensity are first extracted and a parameter mapping relationship is established. Taking a high-intensity outdoor running exercise scenario as an example, the system extracts the exercise mode as outdoor running, the exercise intensity as ≥0.8g, and the duration of the sudden change in exercise intensity as 10ms from the constructed multimodal feature vector. Combining the parameter mapping rules in the technical disclosure, the system establishes a correspondence between these three parameters and the initial delay threshold, the causality test F threshold, and the time series analysis window, determining the initial delay threshold as 5ms, the F threshold as 3.5, and the time series analysis window as 35ms. This provides a precise parameter setting for subsequent time series determination and causality test. The system establishes a baseline; then, according to this mapping relationship, it sets various parameters. Within a 35ms dynamic analysis window, using the effective contact time of pressure as the calibration baseline, it accurately locates the peak time of the touch capacitance signal and the abrupt change time of the motion acceleration signal. The combined time difference between the two is calculated to be 3ms. Since this time difference is less than the initial delay threshold of 5ms, the initial timing is determined to be a suspected motion-induced false touch. At the same time, combined with the pressure signal stability calibration feature weight, the capacitance, motion, and pressure features are fused according to their weights to construct a correlation-enhanced multimodal weighted timing feature sequence, providing standardized timing data for causal testing. Secondly, the multimodal weighted temporal feature sequence was integrated into a vector autoregressive model based on Granger causality testing. Based on the initial temporal assessment of suspected false touches, the model order was adjusted to 3. The F-statistic was calculated to be 4.2, which is greater than the preset F-threshold of 3.5, indicating that motion signals can predict touch signals. This leads to a causal assessment result reinforcing the false touch assessment, consistent with the initial temporal assessment result. Finally, the initial temporal assessment result and the causal assessment result were integrated, and the combined confidence level was calculated to be 0.92, reaching the preset confidence level of 0.8. The confidence threshold does not require iterative optimization. If the confidence level is not met, a gradient descent learning rate of 0.01 is matched to the motion intensity, and the latency threshold and analysis window length are dynamically adjusted for re-judgment until the two judgment results are consistent and the overall confidence level is greater than or equal to the preset threshold. Finally, based on the integrated judgment result, the temporal causal judgment result is output as false touch caused by motion. This result forms a logical closed loop with the feature extraction and multimodal vector construction steps of the outdoor running scene described above, and provides accurate temporal causal dimension basis for subsequent spatial classification judgment and dynamic threshold judgment. This step achieves accurate parameter adaptation by establishing multi-dimensional parameter mapping relationship, and improves the reliability of the judgment result by combining temporal judgment and causal verification, providing accurate temporal causal basis for subsequent multi-dimensional joint judgment.

[0031] In one embodiment, the steps of setting various parameters according to the mapping relationship, locating the peak moment of the touch signal and the abrupt change moment of the motion signal within the dynamic analysis window, calculating the time difference between the two to complete the initial timing determination, and simultaneously constructing a multimodal weighted timing feature sequence include: S210: Integrates motion patterns, motion intensity, and historical data to construct a multi-dimensional mapping relationship, and uses a gradient optimization algorithm to iteratively optimize the dynamic analysis window and feature weight parameters; S211: Based on the optimized parameters and combined with the pressure signal stability calibration feature weights, extract multimodal time series data and construct a correlation-enhanced weighted time series feature sequence; S212: Within the dynamic analysis window, based on the effective contact time of pressure, locate the peak time of the touch signal and the sudden change time of the motion signal, and calculate the combined time difference between the two to complete the initial timing determination. S213: Input the calibrated weighted time series feature sequence and the initial time series judgment result into the causality test process, and adjust the feature attention according to the motion intensity.

[0032] In this embodiment, firstly, a multi-dimensional mapping relationship is constructed by integrating exercise mode, exercise intensity, and historical data, and the parameters are optimized. Taking a high-intensity outdoor running exercise scenario as an example, the system integrates the exercise mode in this scenario, exercise intensity ≥0.8g, and historical misjudgment sample data to construct a multi-dimensional mapping relationship between the three and the dynamic analysis window and feature weights. At the same time, the mutual information entropy of multimodal features is calculated to identify highly correlated feature pairs as optimization constraints. A gradient descent learning rate of 0.01 is matched according to high intensity, and the parameters are iteratively optimized with the goal of minimizing the positioning error. Finally, the dynamic analysis window is determined to be 35ms, the motion feature weight is 45%, and the capacitance feature weight is 45%. The pressure feature is set at 10%, and parameter convergence is ≤0.5% and pressure fit is ≥0.8, thus completing parameter optimization. Then, according to the optimized 35ms window and feature weight parameters, the stability of the pressure signal is detected in real time. If the stability is ≥0.8, the pressure weight is increased to 20%. Multimodal time-series data such as capacitance peak, motion angular velocity, and pressure contact value are extracted from the multimodal feature vector. The time-series data of each dimension are weighted and fused according to the optimized weights to construct a correlation-enhanced weighted time-series feature sequence, making the time-series data more consistent with the scene characteristics of outdoor running, and providing an accurate data foundation for subsequent peak positioning. Secondly, within a 35ms dynamic analysis window, using the effective pressure contact moment as the time reference, the peak moment of the touch capacitive signal and the abrupt change moment of the motion acceleration signal are accurately located through a feature peak detection algorithm. Combining the calculation rules of 60% peak difference and 40% difference of other dimension features, the comprehensive time difference between the two is calculated to be 3ms. This time difference is less than the preset 5ms delay threshold for outdoor running scenarios, and the initial timing judgment is completed accordingly. The result is suspected false touch caused by motion. Finally, the calibrated weighted timing feature sequence and the suspected false touch result of the initial timing judgment are input into Granger. The causal testing process adjusts the focus of features based on the high-intensity nature of outdoor running, strengthening the analytical weight of motion features in causal testing while weakening the weight of capacitance features. Specifically, it enhances the focus of motion features and weakens the focus of capacitance features according to motion intensity levels, making the analytical logic of causal testing more adaptable to the characteristics of motion interference scenarios. This provides scenario-based analytical guidance for subsequent model order adjustment and F-test calculations. All operations in this step form a logical closed loop with the parameter mapping and multimodal feature vector construction described earlier, laying a precise temporal and feature foundation for the subsequent causal determination results. This step achieves precise parameter adaptation through gradient optimization, constructs a time series by calibrating weights based on pressure stability, accurately locates peaks, and completes the initial determination, providing scenario-based accurate data and analytical guidance for causal testing.

[0033] In one embodiment, the step of constructing a multi-dimensional mapping relationship by fusing motion patterns, motion intensity, and historical data, and iteratively optimizing the dynamic analysis window and feature weight parameters using a gradient optimization algorithm includes: S2101: Integrate motion patterns, motion intensity, pressure stability and historical data to construct a multi-dimensional mapping relationship, calculate the mutual information entropy of multimodal features to identify highly correlated feature pairs, and use them as parameters to optimize constraints; S2102: Based on the learning rate of the motion intensity matching gradient optimization algorithm, the dynamic analysis window and feature weight parameters are iteratively optimized with the goal of minimizing the positioning error, to ensure that the parameters are adapted to the pressure constraints and highly correlated feature pairs. S2103: Verify the convergence and stress adaptability of the optimized parameters. Once the parameters meet the standards, determine the final parameters of the dynamic analysis window and feature weights. S2104: Construct a feature sequence using the final parameters. If the contribution of highly correlated features is insufficient or the decision conflict rate does not meet the preset standard in the subsequent causal test, the relevant data will be fed back to the parameter optimization stage for re-iteration.

[0034] In this embodiment, firstly, multi-dimensional information is integrated to construct a mapping relationship and determine parameter optimization constraints. Taking a high-intensity outdoor running scenario as an example, the system integrates the exercise mode, exercise intensity ≥0.8g, pressure signal stability 0.75, and historical misjudgment samples and judgment data in this scenario to construct a multi-dimensional mapping relationship between these four elements and the dynamic analysis window and feature weights. Simultaneously, the mutual information entropy between capacitance, motion, and pressure multi-modal features is calculated. Two highly correlated feature pairs are identified: high-frequency capacitance features and gyroscope angular velocity peak value, and pressure contact peak value and capacitance duration. These are used as the core constraints for parameter optimization to ensure that the subsequently optimized parameters closely match the high-intensity outdoor running scenario. The analysis of associated features defines clear boundaries for parameter iteration. Then, the learning rate of the gradient optimization algorithm is matched with the motion intensity and the parameter iteration is carried out. For the high intensity attribute of outdoor running ≥0.8g, a gradient descent learning rate of 0.01 is matched. The core optimization objective is to minimize the positioning error at the moment of touch peak and motion change. The duration of dynamic analysis window and the feature weight parameters of capacitance, motion and pressure are iteratively optimized. During the iteration process, the constraint conditions of highly associated feature pairs are always followed. At the same time, it is ensured that the optimized parameters are adapted to the requirements of pressure stability, so that the window and weight parameters meet the positioning accuracy requirements and fit the actual sensing data characteristics of the scene. Next, the convergence and stress adaptability of the optimized parameters were verified, and the final parameters were determined. After multiple iterations, optimized parameters were obtained with a dynamic analysis window of 35ms, a capacitance feature weight of 45%, a motion feature weight of 45%, and a stress feature weight of 10%. The convergence change value of the parameters was detected to be 0.3%, which meets the convergence requirement of ≤0.5%, and the stress adaptability was 0.82, which meets the preset standard of ≥0.8. Both parameter indicators met the standards, so this set of parameters was determined to be the final parameters for high-intensity outdoor running scenarios. If any indicator fails to meet the standard, iterative optimization will continue until the requirements are met. Finally, the final parameters were used to construct a feature sequence and establish a feedback iteration mechanism. The determined 35ms window and feature weight parameters were applied to the multimodal time series feature sequence. The construction of this system provides foundational data for subsequent causal testing. Simultaneously, monitoring rules are established. If, in subsequent Granger causality tests, the contribution of the aforementioned two highly correlated feature pairs is lower than a preset threshold, or the conflict rate between time-series and causal determinations exceeds 10%, then the feature data, determination results, parameter application effects, and other relevant data are fed back to the current parameter optimization stage. The learning rate and constraints are readjusted for iterative optimization, forming a closed loop of "parameter optimization - application - monitoring - feedback and further optimization." This step constructs a mapping relationship through multi-dimensional fusion, optimizes parameters using highly correlated features as constraints, ensures parameter effectiveness through dual-index verification, and establishes a feedback mechanism to ensure parameters continuously adapt to scenario requirements, providing accurate parameter support for subsequent determinations.

[0035] In one embodiment, the steps of extracting touch trajectory, motion posture, and signal energy features based on the multimodal feature vector, constructing a joint judgment space, constructing a judgment boundary using historical samples, mapping the current touch event features to the judgment space for classification, and outputting the space classification judgment result include: S30: Dynamically allocate the weights of touch trajectory, motion posture and signal energy features according to motion mode and motion intensity, integrate pressure features to construct a multi-dimensional feature set, and establish a mapping relationship between motion intensity and kernel function; S31: The multidimensional feature set is reduced to three dimensions using a feature dimensionality reduction algorithm, while retaining the core distinguishing information. It is then mapped to an enhanced three-dimensional joint decision space, and the support vector machine is initialized by selecting an adaptation kernel function based on the motion intensity. S32: Construct exclusive judgment boundaries according to touch intent, predict touch intent and match corresponding boundaries to complete preliminary classification, cross-validate with other intent boundaries, and optimize hyperplane parameters by combining historical samples; S33: Calculate the classification accuracy, pressure feature dimensionality reduction retention rate, and cross-validation conflict rate. If any indicator fails to meet the preset standard, readjust the feature weights, kernel function, and boundary parameters for iterative optimization until the indicator meets the standard. Map the current touch event features to the joint decision space to complete the classification and output the spatial classification decision result.

[0036] In this embodiment, feature weights are dynamically allocated according to the exercise mode and intensity, and a multi-dimensional feature set is constructed. Taking the high-intensity outdoor running exercise scenario as an example, the system extracts touch trajectory, exercise posture, and signal energy features from the constructed multimodal feature vector. According to the weight rules of this scenario, the weights of posture features are allocated as 40%, energy features as 35%, and trajectory features as 25%. At the same time, pressure contact peak and stability features are fused to construct a four-dimensional feature set. Combining the mapping rules between exercise intensity and kernel function, the RBF kernel function adapted to the high-intensity scenario is determined, which provides the feature and algorithm benchmark for subsequent space construction and classification. Then, the KPCA feature dimensionality reduction algorithm is used to reduce the four-dimensional feature set to three dimensions. During the dimensionality reduction process, more than 95% of the core distinguishing information is retained. The processed three-dimensional features are mapped to the enhanced three-dimensional joint decision space. Then, the RBF kernel function is selected according to the high-intensity attribute of outdoor running to initialize the support vector machine, completing the construction of the decision space and classification model, so that the model can accurately adapt to the feature distinguishing requirements of the exercise scenario. Secondly, the system constructs dedicated judgment boundaries based on touch intent and performs preliminary classification and cross-validation. Touch intents are categorized into two types: taps and swipes. Dedicated judgment boundaries are constructed for each type. Based on the touch characteristics of an outdoor running scenario, the system predicts the current touch intent as a tap (i.e., constructing dedicated judgment boundaries based on tap and swipe intents, with added touch duration features to assist in predicting touch intent). Feature points are matched to the tap-specific boundary to complete the preliminary classification, indicating a suspected accidental touch. The result is then cross-validated using the swipe intent judgment boundary. Simultaneously, the hyperplane parameters of the support vector machine are optimized by combining historical accidental touch samples and normal touch samples, making the judgment boundaries more closely match the feature distribution of the actual scenario. Finally, three core indicators are calculated and iteratively optimized until the criteria are met, outputting the classification results. The system calculated that the classification accuracy was 93%, the pressure feature dimensionality reduction retention rate was 65%, and the cross-validation conflict rate was 8%, all of which met the preset standards of accuracy ≥92%, retention rate ≥60%, and conflict rate ≤10%, requiring no iterative optimization. If any indicator fails to meet the standard, the feature weights, kernel function parameters, and decision boundary parameters are readjusted and the system is trained again. After meeting the standard, the three-dimensional feature points of the current touch event are remapped to the joint decision space, and finally it is determined that it falls in the feature area of ​​the false touch sample. The output spatial classification result is a false touch caused by motion. This step builds a dedicated decision space through dynamic weight allocation and kernel function adaptation. After dimensionality reduction to retain core information, combined with intention hierarchical judgment and indicator verification, the accuracy of spatial classification is improved, providing a reliable basis for comprehensive decision-making.

[0037] In one embodiment, the steps of constructing dedicated judgment boundaries according to touch intent, predicting touch intents and matching corresponding boundaries to complete preliminary classification, and optimizing hyperplane parameters through cross-validation with other intent boundaries and combining historical samples include: S320: Refines the feature weights and kernel function parameters of touch intent according to motion intensity, and adds touch duration to assist in predicting touch intent; S321: Combining real-time motion intensity with predicted touch intent, adjust the corresponding boundary parameters of the intent, use support vector machine to complete the initial classification, and simultaneously calculate the classification confidence. S322: Use other touch intent determination boundaries to cross-validate the preliminary classification results, and verify the consistency of classification confidence with the matching degree between touch duration and touch intent; S323: If the confidence consistency or duration intent matching degree does not meet the preset standard, combine historical misjudgment samples to optimize the hyperplane penalty coefficient and kernel function parameters, retrain the dedicated judgment boundary and iteratively optimize until the indicator meets the standard.

[0038] In this embodiment, the feature weights and kernel function parameters of the touch intent are first refined according to the motion intensity to assist in the prediction of the intent. Taking a high-intensity outdoor running scenario as an example, the system refines the feature configurations for two types of touch intents, namely tap and swipe, for a motion intensity of ≥0.8g in this scenario. The tap intent is assigned feature weights of 35% curvature, 40% angular velocity, and 25% energy, and is adapted to an RBF kernel function with a σ value of 0.8. The trajectory length weight for the swipe intent is adjusted to 40%. At the same time, the touch duration feature of this touch event is extracted. The detected duration is only 0.1s, which is much lower than the basic duration of 0.3s for normal taps and 0.5s for swipes. This feature is used to assist in the prediction of the current touch event. The initial touch intent was a non-active suspected touch, which provided a precise direction for subsequent boundary matching and classification. Then, the boundary parameters were adjusted and a preliminary classification was completed by combining real-time motion intensity and predicted intent. Based on the high-intensity attributes of outdoor running and the predicted suspected touch intent, the system specifically increased the anti-interference parameters of the touch-specific judgment boundary and strengthened the distinguishing weight of motion posture features. The touch feature points in the three-dimensional joint judgment space were input into the support vector machine initialized by the RBF kernel function to complete the preliminary classification. The judgment result was a false touch caused by motion. At the same time, the original confidence of this classification was calculated to be 0.85 by feature matching degree and classification distance, which provides a numerical reference for subsequent cross-validation. Secondly, another intent boundary was used for cross-validation and verification of the two core indicators. The system called the judgment boundary of the swipe intent to cross-validate the preliminary classification results, mapping the same touch feature point to the judgment boundary of the swipe intent for secondary classification. The result was also determined to be non-active touch. Subsequently, the confidence consistency of the two classifications was calculated to be 0.92. At the same time, the matching degree between the touch duration and the predicted touch intent was verified. Because the actual duration of 0.1s deviated significantly from the duration feature of normal touch intent, the matching degree was 0.65, which did not meet the preset matching degree standard of 0.8. Finally, the boundary parameters were iteratively optimized for the unqualified indicators until they met the standards. Since the duration-intent matching degree did not meet the preset standard, the system retrieved the historical touch intents in the outdoor running scenario. For erroneous touch samples, the hyperplane penalty coefficient of the support vector machine is optimized by adjusting the coefficient from 2 to 3, while the RBF kernel function σ value is fine-tuned to 0.75. The touch-specific judgment boundary is retrained using the optimized parameters, and the touch feature points are input into the trained model for classification again. The duration-intent matching degree is recalculated to 0.82, and the confidence consistency remains at 0.92. Both indicators meet the standards, and the iteration stops. Finally, based on the optimized judgment results, it is confirmed that this touch is a false touch caused by motion. This step refines the intent feature configuration by motion intensity and assists in predicting intent by combining touch duration. The classification accuracy is improved through cross-validation and parameter iteration optimization, making the judgment boundary more suitable for touch features in motion scenarios.

[0039] In one embodiment, the steps of combining real-time motion intensity with the predicted touch intent, adjusting the corresponding boundary parameters for the intent, using a support vector machine to complete the initial classification, and simultaneously calculating the classification confidence include: S3210: Based on exercise intensity and exercise pattern, establish a mapping model between confidence threshold and misjudgment risk, and simultaneously build a correlation library between historical confidence and classification accuracy; S3211: Combining real-time motion intensity with the predicted touch intent, adjust the specific boundary parameters of the corresponding intent and the appropriate kernel function type with reference to the mapping model; S3212: Use support vector machine to perform preliminary classification of current touch event features, calculate the original confidence score of classification, calibrate the original confidence score through the historical association database, and assess the risk of misjudgment based on the calibrated confidence score; S3213: If the classification confidence level meets the standard after calibration and the misjudgment risk does not meet the preset standard, proceed to other touch intent boundary cross-validation steps; if it does not meet the standard, mark it as a sample to be optimized and record the classification confidence fluctuation value, collect the sample to be optimized and fluctuation data, readjust the boundary parameters, confidence threshold and kernel function type, and iteratively train the model until the indicator meets the standard.

[0040] In this embodiment, a confidence-related mapping model and association library are first established based on exercise intensity and mode. Taking a high-intensity outdoor running scenario as an example, the system establishes a mapping model between confidence threshold and misjudgment risk for the exercise mode and exercise intensity ≥0.8g in this scenario. A confidence threshold of 0.9 and a misjudgment risk threshold of 5% are set, meaning that when the confidence level is below 0.9, the misjudgment risk gradually increases as the value decreases. Simultaneously, all historical touch judgment data in this scenario are integrated to construct an association library between historical confidence and classification accuracy, recording the actual classification accuracy corresponding to different confidence intervals. The accuracy rate provides data support for subsequent confidence calibration and risk assessment. Then, the boundary and kernel function parameters are adjusted by combining real-time exercise intensity and predicted intent. Based on the high-intensity attributes of outdoor running and the previously predicted suspected touch intent, the system refers to the established mapping model, increases the weight of motion posture features in the touch-specific judgment boundary, strengthens the anti-interference parameter configuration, and matches the RBF kernel function with a σ value of 0.8 suitable for this scenario. The boundary parameters and kernel function type are adjusted in a targeted manner so that the model parameters are fully in line with the touch judgment requirements of the high-intensity outdoor running scenario. Secondly, a support vector machine (SVM) was used to complete the initial classification, calibrate the confidence level, and assess the risk of misjudgment. The 3D feature points of the current touch event were input into the adjusted SVM model to complete the initial classification. The result was a false touch caused by motion, and the original confidence level was calculated to be 0.85. The original confidence level was calibrated by using a correlation library of historical confidence levels and classification accuracy. Combined with the correlation data in this scenario, the confidence level was calibrated to 0.83. Then, based on the mapping model between the confidence level threshold and the risk of misjudgment, the risk of misjudgment in this classification was assessed to be 6%, slightly higher than the preset threshold of 5%. Finally, based on the confidence level and risk indicators, it was determined whether to proceed to the next stage or iterative optimization. After this calibration, the confidence level of 0.83 did not reach the preset threshold of 0.9, and the risk of misjudgment of 6% exceeded the standard of 5%. Therefore, the system marks the touch event as a sample to be optimized, and records the fluctuation value of the classification confidence score as 0.02. It collects the feature data, confidence score fluctuation data, and judgment results of the sample to be optimized, readjusts the hyperplane parameters of the touch-specific boundary, fine-tunes the confidence score threshold to 0.88, adjusts the kernel function σ value to 0.78, retrains the support vector machine model with the adjusted parameters, and classifies and calculates the confidence score of the sample again. After multiple iterations, the calibrated confidence score reaches 0.89, and the misjudgment risk is reduced to 4.5%. Both indicators meet the standards (i.e., the calibrated confidence score ≥ the preset threshold and the misjudgment risk ≤ 5%), and the iteration stops. This step achieves accurate calibration of confidence score and quantification of misjudgment risk by establishing a dedicated confidence score mapping model and association library. Through parameter iteration optimization, the model is adapted to scene features, improving the reliability of the initial classification.

[0041] In one embodiment, the step of using a support vector machine to perform preliminary classification of the current touch event features and calculating the original confidence score of the classification includes: S32120: Set the appropriate kernel function and confidence threshold according to the motion intensity, build a correlation library of motion intensity, touch intention, misjudgment features and confidence, and record historical calibration errors and parameters that exceed the standard fluctuation. S32121: Predict the current touch intent, combine real-time motion intensity and pressure characteristics, select low fluctuation parameters from the associated library, and adjust the judgment boundary hyperplane and anti-interference parameters; S32122: Using a support vector machine, the current touch event features are initially classified based on the adjusted parameters. The original confidence score of the classification is calculated simultaneously. The historical calibration error is corrected through the association library. The original confidence score is then calibrated a second time by combining the contribution of pressure features. S32123: Based on the calibrated classification confidence and the confidence threshold corresponding to the motion intensity, assess the risk of misjudgment, calculate the classification confidence fluctuation value of multiple consecutive touches, if the fluctuation exceeds the standard, the risk of misjudgment is added, if the risk reaches the preset standard, enter cross-validation, if the risk does not reach the preset standard, supplement the sample and fluctuation data to the association library, and readjust the kernel function, confidence threshold and boundary parameters for iterative optimization.

[0042] In this embodiment, a kernel function and confidence threshold are first set according to the exercise intensity, and a multi-dimensional association library is constructed. Taking a high-intensity outdoor running scenario (exercise intensity ≥ 0.8g) as an example, the system sets an appropriate RBF kernel function (σ value 0.8) and a confidence threshold of 0.9 according to this intensity. At the same time, historical touch data in this scenario is integrated to construct a multi-dimensional association library of exercise intensity, touch intent, misjudgment features, and confidence. The library records in detail the historical calibration error values ​​and parameters such as kernel functions and boundaries where confidence fluctuations exceed the standard, providing data for subsequent parameter selection and confidence calibration. Support; then predict the touch intention and combine multiple features to select low fluctuation parameters to adjust the model. The system combines the features extracted above, such as touch duration of 0.1s and low pressure contact peak, to predict that the current touch intention is a suspected point touch. Then, combined with the real-time high intensity of outdoor running and the feature of pressure signal stability of 0.75, it selects low fluctuation parameters for point touch judgment in this scenario from the association library. Based on this, it adjusts the hyperplane penalty coefficient of the support vector machine judgment boundary to 3, and strengthens the anti-interference parameter weight of motion posture features to make the model parameters more suitable for point touch judgment scenarios under motion interference. Secondly, a support vector machine (SVM) is used for initial classification and secondary calibration of the confidence level. The 3D feature points of the current touch event are input into the adjusted SVM model to complete the initial classification. The result is a false touch caused by motion, and the original confidence level is calculated to be 0.85. The historical calibration error of -0.02 for this scenario is retrieved from the associated library for correction. Then, the original confidence level is calibrated again by combining the pressure feature contribution (which is calculated as 10% according to the rules due to low pressure stability), resulting in a final calibrated confidence level of 0.83. Finally, the system assesses the risk of misjudgment based on the calibrated confidence level and determines the subsequent operation. Based on the 0.9 confidence level threshold corresponding to high-intensity outdoor running, the system assesses the risk of misjudgment of 0.83 after calibration as 6%. Simultaneously, the classification confidence fluctuation value of three consecutive touch events was calculated to be 0.03, which did not reach the fluctuation threshold of 0.05, so there was no need to add risk. However, since the misjudgment risk of 6% exceeded the preset standard of 5%, it did not meet the requirements for entering cross-validation. The system added the touch sample, confidence fluctuation data, and misjudgment risk value to the association library, readjusted the RBF kernel function σ value to 0.75, and fine-tuned the confidence threshold to 0.88. At the same time, the parameters of the decision boundary hyperplane were optimized, and the support vector machine model was retrained with the new parameters to form a closed loop of parameter optimization. This step achieves accurate selection of low-fluctuation parameters by building a multi-dimensional association library, and improves the confidence accuracy through secondary calibration. Combined with risk assessment and parameter iteration, the preliminary classification results are more in line with the actual judgment needs of motion scenarios.

[0043] In one embodiment, the step of determining motion intensity based on the multimodal feature vector, dynamically adjusting the judgment threshold, combining touch event timing pattern analysis, and outputting a dynamic threshold judgment result includes: S40: Based on multimodal feature vector recognition, motion patterns are identified, and exclusive motion intensity grading thresholds and threshold adjustment rules are set according to motion patterns. Simultaneously, an optimized sample library containing historical threshold adjustments, judgment accuracy, and boundary sample features is constructed. S41: Set the duration of the sliding window and continuously collect touch events, calculate the statistical characteristics of motion intensity and the distribution of touch features within the window, combine the optimized sample library to calibrate the initial judgment threshold, dynamically adjust the threshold according to the motion mode and real-time motion intensity, and simultaneously label the boundary samples within the preset threshold range. S42: Make a preliminary judgment on each touch event based on the adjusted judgment threshold. If the judgment results are consistent multiple times, the judgment conclusion is strengthened. If the judgment results are inconsistent or contain boundary samples, analyze the periodicity of the motion pattern and match the judgment logic of the same historical scene in the sample library. S43: If there is a regular motion cycle, it is judged as a systematic mis-touch due to motion interference; if there is no regularity, it is defined as a mixed touch scenario. All data, boundary samples and judgment results are added to the optimization sample library, the motion mode-specific threshold rules and calibration logic are iteratively optimized, and dynamic threshold judgment results are output.

[0044] In this embodiment, the system first identifies movement patterns based on multimodal feature vectors and builds a dedicated rule and sample library. Taking high-intensity outdoor running as an example, the system extracts movement features from the constructed multimodal feature vectors, accurately identifies the current movement pattern as outdoor running, and sets a dedicated exercise intensity grading threshold for this pattern (≥0.8g for high intensity, 0.4-0.8g for medium intensity, and <0.4g for low intensity). Simultaneously, threshold adjustment rules are established: for high intensity, the pressure and energy ratio thresholds are increased by 20%; for medium intensity, they are increased by 10%; and for low intensity, no adjustment is made. An optimized sample library for the outdoor running pattern is also constructed concurrently. The data in the library includes historical threshold adjustment records, the accuracy rate of judgment at different intensities, and various boundary sample characteristics. The system first identifies the characteristics of touch events to provide contextual support for subsequent threshold calibration and judgment. Then, it sets a sliding window to collect touch events and dynamically adjusts the threshold and marks boundary samples. The system sets the initial sliding window duration to 500ms for the outdoor running mode and continuously collects touch event data within the window. The statistical characteristics of the exercise intensity are calculated to be high intensity ≥0.8g with small fluctuations. The touch feature distribution shows irregular capacitance fluctuations. The initial judgment threshold is calibrated by combining historical data of similar high-intensity scenarios in the optimization sample library. According to the threshold adjustment rules, the pressure judgment threshold and the energy ratio judgment threshold are each increased by 20% to complete the dynamic adjustment of the threshold. At the same time, touch samples within ±5% of the adjusted threshold are marked as boundary samples and their feature data are recorded separately. Secondly, based on the adjusted threshold, the system initially judges and analyzes the periodicity of the movement. It compares the pressure and energy ratio characteristics of the current touch event with the adjusted threshold, initially identifying it as a suspected accidental touch due to motion interference. Since the judgment result includes labeled boundary samples, further analysis of the periodicity of the outdoor running movement pattern is conducted. By extracting the angular velocity data and acceleration variation patterns from the gyroscope, the system identifies a regular arm swinging motion cycle in this scenario, with a cycle duration of approximately 1.2 seconds. Subsequently, the system matches and optimizes the judgment logic for similar historical scenarios with high-intensity, regular motion cycles in outdoor running from the sample library, providing a reference for subsequent accurate judgments. Finally, the system determines the scene type, supplements data for iterative optimization, and outputs the results. Based on the identified... Based on the regular movement cycle, the touch event is determined to be a systematic false touch caused by motion interference, rather than a mixed touch scenario. Subsequently, the motion intensity data, touch event characteristics, boundary sample information, threshold adjustment records, and final judgment results are all added to the optimized sample library of the outdoor running mode. Based on the newly added data, the threshold adjustment rules and calibration logic specific to this mode are iteratively optimized to make the rules more in line with the touch characteristics of actual motion scenarios. Finally, the dynamic threshold judgment result is output as a false touch caused by motion. This step achieves dynamic adaptation of the judgment threshold by customizing the threshold rules and sample library for the motion mode. Combined with the periodic analysis of motion, it accurately identifies systematic false touches, and the iterative sample library ensures that the judgment rules continuously align with the actual scenario.

[0045] In one embodiment, the steps of setting the sliding window duration and continuously collecting touch events, calculating the statistical characteristics of motion intensity and the distribution of touch features within the window, calibrating the initial judgment threshold by combining the optimized sample library, dynamically adjusting the threshold according to the motion mode and real-time motion intensity, and synchronously labeling boundary samples within the preset threshold range include: S410: Subdivide the motion modes and construct an optimized sample library subset corresponding to each subdivision mode, filter historical misjudged samples to establish the mapping relationship between misjudged scenarios and adaptation parameters, and set motion intensity fluctuation thresholds and windows, threshold collaborative adaptation rules. S411: Match real-time subdivided motion modes, combine the mapping relationship between the misjudged scenarios and the adaptation parameters, set the initial sliding window duration, continuously collect touch events, and calculate the statistical features of motion intensity and the distribution of touch features within the window; S412: If the fluctuation of the exercise intensity exceeds the standard, adjust the window duration and threshold adjustment range, calibrate the initial judgment threshold in combination with the sample library subset of the corresponding subdivision mode, dynamically adjust the judgment threshold according to the real-time exercise intensity, and simultaneously label the boundary samples in the preset threshold range. S413: Feed back the current fluctuation adaptation effect, threshold deviation, boundary samples and misjudgment situation to the mapping relationship between the optimized sample library and the misjudgment scenario adaptation parameters, and iteratively optimize the window, threshold rules and collaborative adaptation logic specific to the subdivision mode.

[0046] In this embodiment, the exercise modes are first subdivided, and a dedicated sample library subset and mapping relationship are established. Taking the high-intensity outdoor running scenario as an example, the system subdivides the basic running mode into two sub-modes: outdoor running and indoor running. A dedicated optimized sample library subset is built for outdoor running. At the same time, historical misjudgment samples are screened from this subset, and typical misjudgment scenarios such as high-frequency arm swing and road bumps are identified. A mapping relationship is established between the misjudgment scenarios and the adaptation parameters of the sliding window duration and threshold adjustment range. The exercise intensity fluctuation threshold is set to 0.2g, and a collaborative adaptation rule for the window duration and threshold adjustment range is established: when the fluctuation exceeds the standard, the window duration is extended by 20%, and the threshold adjustment range is adjusted. The amplitude was increased by 10%, laying a scenario-based foundation for subsequent parameter adaptation. Then, the real-time subdivision mode was matched and relevant features were collected and calculated. The system identified the real-time subdivision motion mode as outdoor running from the multimodal feature vector. Combining the misjudgment scenario of high-frequency arm swing in this mode with the mapping relationship of adaptation parameters, the initial sliding window duration was set to 500ms. Touch event data was continuously collected in this window. By calculating the root mean square value of the motion data, the statistical feature of motion intensity within the window was found to be high intensity ≥0.8g, and the motion intensity fluctuation value was 0.25g. The touch feature distribution showed irregular instantaneous capacitance fluctuation characteristics, thus completing the accurate statistics of feature data. Next, it was determined whether the fluctuation of exercise intensity exceeded the standard, and parameters were dynamically adjusted and boundary samples were labeled. Upon comparison, the current exercise intensity fluctuation value of 0.25g exceeded the preset fluctuation threshold of 0.2g. According to the collaborative adaptation rule, the sliding window duration was extended by 20% to 600ms, and the threshold adjustment range was increased from the original 20% to 30%. Simultaneously, historical calibration data from the outdoor running optimized sample library subset was used to precisely calibrate the initial judgment thresholds for pressure and energy ratio. Then, based on the real-time high-intensity exercise attributes, the calibrated thresholds were increased by 30% to complete the dynamic adjustment. Simultaneously, touch samples within ±5% of the adjusted threshold were labeled as boundary samples, and their feature information was recorded. Finally, the feedback data was combined and iterated. The system optimizes the subdivided pattern rules, statistically analyzes the threshold judgment accuracy and threshold deviation after this fluctuation adaptation, organizes boundary sample features and cases without new misjudgments, and feeds this data back into the mapping relationship between the optimized sample library subset for outdoor running and the adaptation parameters for misjudged scenarios. For high-frequency arm swing scenarios, the corresponding sliding window duration and threshold adjustment range parameters are iteratively optimized, while the collaborative adaptation logic between the window and the threshold is improved, making the parameter adjustments more closely match the actual motion fluctuation characteristics of outdoor running. This step forms a logical closed loop with the previous motion pattern recognition and threshold rule setting. The adjusted parameters also provide accurate thresholds and window basis for the subsequent initial judgment of touch events. This step builds a dedicated sample library by subdividing motion patterns, realizes the collaborative dynamic adaptation of windows and thresholds, and accurately adjusts parameters by combining fluctuation monitoring. Data feedback iteration makes the rules more closely match the motion characteristics of subdivided scenarios.

[0047] In one embodiment, the steps of matching real-time subdivided motion patterns, combining the mapping relationship between the misjudged scenario and the adaptation parameters, setting the initial sliding window duration, continuously collecting touch events, and calculating the statistical features of motion intensity and the distribution of touch features within the window include: S4110: Cluster historical misjudged samples according to subdivided motion patterns, assign high weight to high-frequency misjudged scenarios, construct a weighted mapping relationship between misjudged scenarios and window duration, extract historical statistical data of the same mode and intensity range, and set touch feature pre-screening rules. S4111: Match the real-time subdivided motion mode, combine the mapping relationship between the misjudged scene and the adaptation parameter and the weighted mapping between the misjudged scene and the window duration, set the initial sliding window duration, collect valid touch events according to the touch feature pre-screening rules, calculate the touch frequency synchronously, and verify and fine-tune the sliding window duration. S4112: Calculate the motion intensity statistical features and the filtered touch feature distribution within the fine-tuned sliding window, compare and calibrate them with the historical statistical data of the same intensity range in the corresponding subdivision mode, and mark the feature data similar to the misjudged scenario. S4113: Feed back feature similarity, motion intensity fluctuation value, calibration deviation and touch frequency data to the weighted mapping of misjudged scenarios and window duration and sample library, and iteratively optimize window adaptation rules, touch feature pre-screening rules and historical data migration logic.

[0048] In this embodiment, historical misjudged samples are first clustered according to subdivided sports modes, and a weighted mapping is constructed. Pre-screening rules are then set. Taking the high-intensity outdoor running scenario as an example, the system performs cluster analysis on historical misjudged samples in the sample library for this subdivided mode, identifying three typical misjudged scenarios: high-frequency arm swing, road bumps, and sudden changes in arm swing amplitude. High-frequency arm swing accounts for 70% of these scenarios and is assigned a high weight of 0.7. Road bumps and sudden changes in arm swing amplitude are assigned weights of 0.2 and 0.1 respectively (i.e., high-frequency misjudged scenarios are assigned high weights based on misjudgment frequency and quantification of misjudgment loss). A weighted mapping relationship between misjudged scenarios and window duration is constructed, determining that high-frequency arm swing corresponds to an initial window of 500ms, road bumps to 550ms, and sudden changes in arm swing amplitude to 600ms. Simultaneously, historical statistical data for the high-intensity range of outdoor running (≥0.8g) is extracted, and touch feature pre-screening rules are set (excluding touch duration <0.05s and pressure peak <0.2N). (Invalid touch features): Invalid touch features with touch duration < 0.05s and pressure peak < 0.2N are removed to lay the foundation for subsequent accurate data collection; then, the real-time subdivision mode is matched and the initial sliding window duration is set and fine-tuned. The system accurately matches the real-time subdivision sports mode as outdoor running from the multimodal feature vector. Combining the mapping relationship between misjudged scenarios and adaptation parameters in this mode, and the above weighted mapping relationship, since the high-frequency swinging motion feature of the arm is detected, the initial sliding window duration is set to 500ms. Valid touch events in the window are collected according to the pre-screening rules. After removing 3 invalid touch data, the touch frequency in the window is statistically found to be 8 times / second, which belongs to high-frequency touch. The sliding window duration is fine-tuned to 450ms according to the rules to make the window duration adapt to the touch frequency feature. Next, the system calculates the feature data of the fine-tuned window and compares it with calibrated and labeled similar data. Within a 450ms fine-tuned sliding window, the system calculates the high-intensity feature with a mean of 0.9g and a fluctuation of 0.25g. Simultaneously, it extracts the filtered touch feature distribution: the capacitance signal exhibits irregular instantaneous fluctuations, and the pressure feature lacks a stable contact peak. This feature data is compared and calibrated with historical statistical data from outdoor running in the ≥0.8g high-intensity range, correcting the mean exercise intensity calibration deviation by 0.03g. Furthermore, through a feature similarity algorithm, the system marks the current touch feature data as having a 92% similarity to the features of a scenario where high-frequency arm swing is misjudged, thus completing the accurate labeling of similar data. Finally, the system provides feedback data and iteratively optimizes various rules and... Logically, the system feeds back all the data, including the calculated feature similarity of 92%, exercise intensity fluctuation of 0.25g, calibration deviation of 0.03g, and touch frequency of 8 times / second, to the weighted mapping relationship between misjudged scenarios and window duration in outdoor running mode and to the sample library. For high-frequency arm swing scenarios, its weight is increased to 0.75. At the same time, the window adaptation rules are iteratively optimized: the window duration is further shortened by 10% when there is high-frequency touch. The touch feature pre-screening rules and historical data migration logic are improved to make historical data more consistent with the feature analysis of real-time scenarios. This step achieves accurate window duration adaptation through weighted mapping. Combined with pre-screening and frequency fine-tuning, the effectiveness of data is improved. The data feedback iteration makes the rules and logic more consistent with the real-time characteristics of the subdivided scenarios.

[0049] In one embodiment, the steps of clustering historical misjudged samples by subdivided motion patterns, assigning high weights to high-frequency misjudged scenarios, constructing a weighted mapping relationship between misjudged scenarios and window duration, extracting historical statistical data of the same pattern and intensity range, and setting touch feature pre-screening rules include: S41101: Collect historical misjudgment samples from multiple devices and perform cross-device normalization processing. Cluster the historical misjudgment samples by subdividing motion modes and motion intensity ranges. Assign weights to each cluster based on the misjudgment frequency and the quantification results of misjudgment loss. Construct a loss weighted mapping relationship. Extract the historical statistical data of the same mode and intensity range after normalization. Set loss-oriented touch feature pre-screening rules. S41102: Analyze the misjudgment trend of each subdivided motion mode, adjust the weight of each cluster based on the misjudgment trend, further optimize the loss weighted mapping relationship, match the real-time subdivided motion mode with the motion intensity range, and verify the adaptability and trend avoidance effect of the optimized weighted mapping relationship and the touch feature pre-screening rule. S41103: Statistically analyze the high loss misjudgment avoidance rate and feature correlation retention rate after pre-screening, verify the accuracy of historical statistical data, and mark scenarios with adaptation bias and insufficient trend avoidance. S41104: Feed back the adaptation bias, avoidance rate and trend prediction results to the sample library, re-optimize the normalization algorithm, cluster weight and touch feature pre-screening rules, and iteratively improve the loss weighted mapping relationship and historical data extraction logic.

[0050] In this embodiment, historical misjudgment samples from multiple devices are first collected, normalized, and double-clustered to construct a weighted mapping and set pre-screening rules. Taking high-intensity outdoor running as an example, the system collects historical misjudgment samples related to outdoor running from smart bracelets of different brands and models. Cross-device normalization is completed through feature standardization to eliminate feature bias caused by differences in device hardware. Then, double clustering is performed according to the subdivided sports modes of outdoor / indoor running and the sports intensity range of ≥0.8g / 0.4-0.8g / <0.4g to obtain clusters of high-intensity range of outdoor running ≥0.8g. Combining the misjudgment frequency and misjudgment loss quantification results within the cluster, three types of misjudgment scenarios are assigned weights of 0.7, 0.2, and 0.1 respectively: high-frequency arm swing (misjudgment frequency 70%, loss ratio 65%), road bumps (misjudgment frequency 20%, loss ratio 25%), and sudden changes in arm swing amplitude (misjudgment frequency 10%, loss ratio 10%). A weighted mapping relationship between misjudgment scenarios and sliding window duration is constructed. Simultaneously, the system extracts normalized historical statistical data for outdoor running in the ≥0.8g range and sets loss-oriented touch feature pre-screening rules: removing invalid features such as touch duration <0.05s, pressure peak <0.2N, and capacitance fluctuation duration <0.1s to reduce the probability of high-loss misjudgment from the source. Then, it analyzes the misjudgment trend and adjusts the weights to optimize the mapping relationship, verifying the adaptability and trend avoidance effect. The system analyzes the misjudgment data of outdoor running sub-modes in the past 3 months and finds that the misjudgment trend of high-frequency arm swing is slightly increasing, while the misjudgment trend of sudden changes in arm swing amplitude is stable. Based on this, the cluster weight of high-frequency arm swing is adjusted from 0.7 to 0.75, while the weight of sudden changes in arm swing amplitude remains unchanged at 0.1. The loss weighted mapping relationship is further optimized, and then matched with the real-time outdoor running sub-mode and intensity range of ≥0.8g. The optimized mapping relationship and pre-screening rules are applied to real-time touch data processing, and the verification shows that the rule has a trend avoidance effect of 85% for high-loss misjudgment scenarios, with good overall adaptability. Secondly, the system statistically validates historical data for core indicators, marking deviations and deficiencies. Based on touch data from the optimized rules, the system found a high-loss misjudgment avoidance rate of 88% and a feature correlation retention rate of 92% after pre-screening, both meeting the preset requirements. Further comparison of real-time feature data with historical statistics revealed an overall accuracy of 90% for the historical data. However, the system also identified significant adaptation deviations in the bumpy road scenario under different road conditions, with a trend avoidance effect of only 70%, classifying it as a scenario with insufficient trend avoidance. Finally, the data is fed back to the sample database for iterative optimization of the algorithm and rule refinement. The system incorporates the statistically analyzed adaptation deviation data and high-loss misjudgment avoidance rate... The feature correlation retention rate and trend prediction results are all fed back to the outdoor running mode sample library. To address the adaptation bias in bumpy road scenarios, the cross-device normalization algorithm is re-optimized, feature data calibration factors for different road conditions are added, and the cluster weights and touch feature pre-screening rules for this scenario are adjusted. Road condition-related motion feature screening conditions are added, and the loss weighted mapping relationship and historical data extraction logic are iteratively improved to make historical data more in line with the real-time feature analysis needs of different scenarios. This step eliminates hardware differences through cross-device normalization, and loss weighted mapping makes parameter adaptation more in line with the actual impact of misjudgment. Combined with trend analysis and iterative optimization of rules, the ability to avoid high-loss misjudgment is effectively improved.

[0051] In one embodiment, the steps of integrating the aforementioned three types of judgment results, calculating the posterior probability based on Bayesian decision theory and outputting the final recognition result, collecting feedback data to optimize model parameters and judgment thresholds, and updating the sample library to form a closed-loop recognition mechanism include: S50: Assign dynamic weights based on historical accuracy to the time-series causal determination results, spatial classification determination results, and dynamic threshold determination results; quantify the loss from false triggers and missed detections; construct a Bayesian decision model containing a loss function; set a posterior probability threshold that is dynamically adjusted with the loss; and formulate parameter linkage adjustment rules. S51: Integrate the results of temporal causal judgment, spatial classification judgment, and dynamic threshold judgment into the Bayesian decision model, calculate the posterior probability based on Bayesian decision theory and make a decision, output the final touch recognition result, record false touch and missed judgment events, and statistically calculate the false judgment rate and missed judgment rate after loss weighting, and stratify difficult samples according to false touch type. S52: Clustering algorithm is used to analyze the characteristics of difficult samples in each stratum, identify marginal samples, trigger sample library update and add difficult and marginal samples to the sample library according to stratum, and use differential incremental learning strategy to optimize the classification model; S53: Combining the distribution of false touch features and the loss quantification results, various judgment thresholds and model parameters are adjusted synchronously according to the parameter linkage rules. The loss function, posterior probability threshold and dynamic weight of judgment results of the Bayesian decision model are optimized. All optimized data are fed back to the pre-initialization stage, and iterative optimization forms a closed-loop recognition mechanism.

[0052] In this embodiment, dynamic weights are first assigned to the three types of judgment results, and a Bayesian decision model is constructed. Taking a high-intensity outdoor running scenario as an example, the system retrieves the historical accuracy rates of the three types of judgments: temporal causality, spatial classification, and dynamic threshold. The historical accuracy rates for temporal causality are 94%, spatial classification is 93%, and dynamic threshold is 92%. Dynamic weights of 0.35, 0.33, and 0.32 are assigned to the three types of judgments according to their accuracy rates. At the same time, the losses from false triggers and false misses are quantified, with a false trigger loss coefficient of 1.2 and a false miss loss coefficient of 1.5. A Bayesian decision model incorporating this loss function is constructed, and the posterior probability threshold, which is dynamically adjusted with the loss, is set to 0.6. A parameter linkage adjustment rule is also formulated: when the accuracy of a single judgment module drops by more than 3%, the judgments of the other two modules are simultaneously fine-tuned. Determining thresholds and model parameters lays the foundation for multi-result fusion and decision-making. Then, the system integrates three types of judgment results into the model to complete the decision and collect relevant data. The system integrates the results of the three types of judgments in the outdoor running scenario mentioned above, which are all judged as false touches caused by movement, and inputs them into the Bayesian decision model. Based on Bayesian decision theory, the posterior probability of the touch event being a real touch is calculated to be 0.28, which is far lower than the threshold of 0.6. Based on this, the final touch recognition result is to refuse to respond to the false touch operation. At the same time, it records that there are no false touches or missed judgments in this judgment. The false judgment rate after loss weighting in this scenario is 2.1% and the missed judgment rate is 1.8%. The collected motion interference touch samples are stratified into subcategories such as arm swing interference and road bump interference according to the type of false touch, completing the stratification and sorting of difficult samples. Secondly, clustering algorithms are used to analyze the characteristics of difficult samples and optimize the classification model. The system uses K-means clustering to analyze the multimodal features of difficult samples in each stratum, accurately identifying edge samples whose features are at the decision boundary, triggering a sample library update mechanism. The stratified difficult samples and edge samples are added to the corresponding sample library subsets according to the high-intensity outdoor running scenario. Differentiated incremental learning strategies are adopted for samples of different accidental touch types. The training weight of the arm swing interference samples with the highest proportion is increased, and road condition features are added to the training of the road bump interference samples, thus completing the optimization of the classification model. Finally, parameters are adjusted and the Bayesian model is optimized according to the linkage rules to form a closed-loop recognition mechanism. The system combines the accidental touch feature distribution and loss quantification results of the outdoor running scenario, and synchronously fine-tunes the time delay threshold of S2 temporal causality judgment and S3 spatial classification according to the parameter linkage rules. The system determines the boundary hyperplane parameters and the intensity grading threshold of the S4 dynamic threshold. Simultaneously, it optimizes the loss function of the Bayesian decision model, fine-tunes the loss coefficients for false touches and missed detections, updates the posterior probability threshold to 0.58, and reassigns the dynamic weights of the three types of decisions to 0.34, 0.34, and 0.32. All parameter optimization data and model adjustment results are fed back to the pre-initialization stages, such as S1 multimodal feature extraction, ensuring that the feature extraction and parameter settings in the pre-initialization stages are adapted to the optimized decision model. This forms a closed-loop recognition mechanism encompassing "feature extraction - multi-dimensional decision - result fusion - parameter optimization - feedback iteration." This step achieves accurate decision-making across multiple decision results through dynamic weight allocation and loss function fusion. Combined with cluster analysis and incremental learning to optimize the model, the parameter feedback forms a closed loop, continuously improving the accuracy and scene adaptability of touch recognition.

[0053] In one embodiment, the steps of assigning dynamic weights based on historical accuracy to the temporal causality determination result, spatial classification determination result, and dynamic threshold determination result; quantifying false positives and false negatives; constructing a Bayesian decision model including a loss function; setting a posterior probability threshold that dynamically adjusts with the loss; and formulating parameter linkage adjustment rules include: S501: Statistically analyze the historical accuracy and hard sample identification rate of time-series causal determination, spatial classification determination, and dynamic threshold determination. Assign initial dynamic weights to the three types of determination results based on the double accuracy rate. Introduce a weight decay mechanism and clarify the weight decay triggering conditions. Quantify the loss of false triggers and false misses. Differentiate the loss difference between hard samples and ordinary samples. Construct a Bayesian loss function that includes a weight decay term and hard sample guidance to complete the preliminary construction of the Bayesian decision model. S502: Set a posterior probability threshold that is dynamically adjusted with loss, formulate hierarchical parameter linkage rules, distinguish between high-priority and low-priority linkage content, and clarify the execution logic of various linkage rules. S503: Verify the adaptability of dynamic weights, the rationality of loss quantification, and the effectiveness of hierarchical parameter linkage rules; statistically analyze the weight calibration error, loss deviation, and the loss reduction of high-priority rules; and mark evaluation indicators that do not meet the preset standards. S504: If any evaluation indicator fails to meet the preset standard, adjust the weight allocation, attenuation magnitude, loss quantification coefficient, and priority of the linkage rules for hierarchical parameters, optimize the Bayesian loss function and threshold adjustment logic, and feed all optimized data back to the previous steps to form a closed-loop iteration.

[0054] In this embodiment, the initial dynamic weights for the three-class decision are first calculated based on the double accuracy rate. A Bayesian loss function with weight decay and hard sample guidance is constructed. Taking a high-intensity outdoor running scenario as an example, the system first calculates the historical accuracy and hard sample recognition rate of temporal causality determination, spatial classification determination, and dynamic threshold determination. The historical accuracy rates of the three are 94%, 93%, and 92%, respectively, and the hard sample recognition rates are 92%, 90%, and 89%, respectively. The weights are calculated based on the double accuracy rate, assigning 0.35, 0.33, and 0.32 to the three-class decision. Initial dynamic weights were introduced, along with a weight decay mechanism. The trigger condition was set to reduce the weight by 10% when the recognition rate of difficult samples by a single judgment module was below 85% for 5 consecutive times. The losses from false hits and false misses were further quantified. The loss coefficients for false hits and false misses for ordinary samples were set to 1.2 and 1.5, while the loss coefficients for false hits and false misses for difficult samples were both increased to 2.0. After distinguishing the differences in sample losses, the weight decay term and the loss factor for difficult samples were integrated into the model to construct a Bayesian loss function that includes the weight decay term and the difficult sample-oriented loss function. This completed the initial model construction and provided a basic framework for subsequent decision-making. Then, the dynamic posterior probability threshold and hierarchical parameter linkage rules are set. The system sets the rule that the posterior probability threshold is dynamically adjusted with the loss: when the overall misjudgment loss exceeds the preset threshold, the posterior probability threshold is reduced from 0.6 to 0.55, and when the missed judgment loss exceeds the standard, it is increased to 0.65. At the same time, hierarchical parameter linkage rules are formulated, setting the time delay threshold of S2 temporal causality determination and the boundary hyperplane parameter of S3 spatial classification as high-priority linkage content, and the intensity hierarchical threshold of S4 dynamic threshold as low-priority linkage content. The logic of high-priority rules being executed immediately when the loss exceeds the standard and low-priority rules being executed iteratively in batches is clarified.

[0055] Next, the system verifies each indicator and marks the items that do not meet the standards. The system applies the initially constructed model and set rules to touch judgment in outdoor running scenarios to verify the adaptability of dynamic weights, the rationality of loss quantification, and the effectiveness of hierarchical parameter linkage rules. The statistics show that the weight calibration error is 1.2% and the loss deviation is 0.8%, both of which are lower than the preset standards of 3% and 1%, respectively. The loss reduction of high priority rules reaches 15%, which meets the preset requirement of ≥10%. All evaluation indicators meet the standards, and no indicators that do not meet the preset standards need to be marked. If any indicator does not meet the standards, the system will directly enter the parameter optimization stage. Finally, the model is iteratively optimized based on the verification results to form a closed loop. Since all evaluation indicators met the standards, the system only made minor adjustments to the Bayesian loss function, appropriately increasing the weight of the loss factor for difficult samples, and optimizing the threshold adjustment logic to make the adjustment of the posterior probability threshold more closely match real-time loss changes. If any indicator fails to meet the standards, the weight allocation ratio, weight decay magnitude, loss quantization coefficient, and priority of linkage rules are adjusted. After optimization, all optimized data are fed back to the preceding steps such as weight allocation and loss quantization, and all operations are re-executed, forming a closed loop of "parameter setting - model building - rule formulation - verification optimization - feedback iteration". This step, combined with the three types of judgment results mentioned above and the application of the Bayesian decision model, forms a logical closed loop. The optimized model and rules also provide a more accurate decision-making basis for the subsequent final touch recognition result output. This step constructs a Bayesian loss function that fits the scenario by allocating weights with dual accuracy and introducing a decay mechanism. Combined with hierarchical linkage rules and multi-indicator verification, it achieves accurate model optimization and provides reliable support for decision-making.

[0056] refer to Figure 2 A touch operation accuracy recognition device for a smart bracelet, comprising: The feature extraction module 100 is used to collect capacitive signals, motion sensing data and pressure data from the smart bracelet touch screen. It establishes an analysis time window based on the touch event trigger time, extracts frequency domain and time domain features from the capacitive signals, filters and extracts motion features from the motion sensing data, aligns and normalizes all feature timestamps, and forms a multimodal feature vector. The temporal causality determination module 200 is used to align the multimodal feature vectors to a unified reference coordinate system according to the time dimension, analyze the temporal relationship between touch and motion signals, determine the causal direction of the two by combining the causality test algorithm, and output the temporal causality determination result. The spatial classification and determination module 300 is used to extract touch trajectory, motion posture and signal energy features based on the multimodal feature vector, construct a joint determination space, construct a determination boundary using historical samples, map the current touch event features to the determination space for classification, and output the spatial classification and determination result. The dynamic threshold determination module 400 is used to determine the motion intensity based on the multimodal feature vector, dynamically adjust the determination threshold, combine the touch event timing pattern analysis and determination, and output the dynamic threshold determination result. The fusion decision module 500 is used to integrate the aforementioned three types of judgment results, calculate the posterior probability based on Bayesian decision theory and output the final recognition result, collect feedback data to optimize model parameters and judgment thresholds, and update the sample library to form a closed-loop recognition mechanism.

[0057] Furthermore, the aforementioned feature extraction module 100 includes: The multi-source data acquisition and window setting unit is used to simultaneously acquire touch screen capacitive signals, three-axis accelerometer data, gyroscope data and pressure data, lock precise timestamps, dynamically set the analysis time window based on motion intensity, and identify and mark motion patterns. A customized feature extraction unit is used to customize a feature extraction strategy according to the motion mode, perform transformation processing on the capacitance signal to extract frequency domain and time domain features, and perform filtering and noise reduction processing on the motion sensing data to extract motion features. The effective feature filtering unit is used to adapt the feature dimension based on the signal-to-noise ratio of the capacitance signal, use mutual information entropy to remove redundant features, and combine the correlation of pressure data to retain effective features. The multimodal feature vector construction unit is used to form a multimodal feature vector by using the timestamp of the capacitance signal as a reference, correcting the time offset through a time alignment algorithm, adapting the normalization strategy according to the feature quality level, and concatenating the feature vectors according to the motion mode and feature weights.

[0058] Furthermore, the aforementioned customized feature extraction unit includes: The parameter mapping relationship establishment unit is used to extract feature frequencies and capacitance signal signal-to-noise ratio based on motion patterns, and establish the mapping relationship between the two and capacitance signal transformation processing parameters and motion data filtering parameters. The preliminary feature extraction unit is used to dynamically set the transformation processing parameters according to the mapping relationship, extract the capacitance features of the customized frequency band, and use a filtering algorithm to denoise the motion sensing data and extract motion features. The motion feature iterative optimization unit is used to optimize the motion feature using an iterative optimization algorithm, calculate the correlation between the optimized motion feature and the capacitance feature, and adjust the transformation processing parameters and filtering parameters to re-extract the feature if the correlation does not meet the preset requirements. The feature cross-validation and weight allocation unit is used to perform bidirectional cross-validation of the capacitance feature and the optimized motion feature by combining the pressure feature, eliminating low-correlation features, and allocating the weights of various features according to the motion mode to form a customized high-discrimination feature set.

[0059] Furthermore, the aforementioned time-series causality determination module 200 includes: The parameter mapping relationship construction unit is used to extract motion patterns, motion intensity, and duration of motion mutations, and establish the mapping relationship between these three factors and the initial time delay threshold, causality test threshold, and time series analysis window. The initial timing determination and feature sequence construction unit is used to set various parameters according to the mapping relationship, locate the peak time of the touch signal and the sudden change time of the motion signal within the dynamic analysis window, calculate the time difference between the two to complete the initial timing determination, and simultaneously construct a multimodal weighted timing feature sequence. The causal determination result acquisition unit is used to integrate the multimodal weighted time series feature sequence into the causal test vector autoregressive model, adjust the model order according to the initial time series determination result, and obtain the causal determination result through the test algorithm; The judgment result integration and optimization unit is used to integrate the initial time series judgment result and the causal judgment result. If the confidence level of the judgment result does not reach the preset threshold, the threshold and the analysis window length are dynamically adjusted according to the motion intensity and iteratively optimized until the result is consistent, and the time series causal judgment result is output.

[0060] Furthermore, the aforementioned initial time series determination and feature sequence construction unit includes: The multi-dimensional mapping construction and parameter optimization unit is used to integrate motion patterns, motion intensity and historical data to construct multi-dimensional mapping relationships, and uses gradient optimization algorithm to iteratively optimize dynamic analysis window and feature weight parameters; The weighted time series feature sequence construction unit is used to extract multimodal time series data and construct a correlation-enhanced weighted time series feature sequence according to the optimized parameters and the feature weights calibrated by the stability of the pressure signal. The initial timing determination unit is used to locate the peak moment of the touch signal and the moment of sudden change of the motion signal within the dynamic analysis window, based on the effective contact moment of pressure, and calculate the combined time difference between the two to complete the initial timing determination. The data input and feature attention adjustment unit is used to input the calibrated weighted time series feature sequence and the initial time series judgment result into the causality test process and adjust the feature attention according to the motion intensity.

[0061] Furthermore, the aforementioned multi-dimensional mapping construction and parameter optimization unit includes: The multi-dimensional mapping construction and constraint determination unit is used to integrate motion patterns, motion intensity, pressure stability and historical data to construct multi-dimensional mapping relationships, calculate the mutual information entropy of multi-modal features to identify highly correlated feature pairs, and use them as parameters to optimize constraint conditions. The parameter iterative optimization unit is used to iteratively optimize the dynamic analysis window and feature weight parameters according to the learning rate of the gradient optimization algorithm based on the motion intensity, with the goal of minimizing the positioning error, to ensure that the parameters are adapted to the pressure constraints and highly correlated feature pairs. The optimization parameter verification and determination unit is used to verify the convergence and stress adaptability of the optimized parameters. After the parameters meet the standards, the final parameters of the dynamic analysis window and feature weights are determined. The parameter application and feedback iteration unit is used to construct the feature sequence using the final parameters. If the contribution of highly correlated features is insufficient or the judgment conflict rate does not meet the preset standard in the subsequent causal test, the relevant data will be fed back to the parameter optimization stage for re-iteration.

[0062] Furthermore, the aforementioned spatial classification determination module 300 includes: The feature weight allocation and multi-dimensional feature set construction unit is used to dynamically allocate the feature weights of touch trajectory, motion posture and signal energy according to motion mode and motion intensity, integrate pressure features to construct a multi-dimensional feature set, and establish the mapping relationship between motion intensity and kernel function; The joint decision space construction and model initialization unit is used to reduce the multidimensional feature set to three dimensions using a feature dimensionality reduction algorithm, retain the core distinguishing information, map it to the enhanced three-dimensional joint decision space, and select the appropriate kernel function to initialize the support vector machine according to the motion intensity. The decision boundary construction and preliminary classification unit is used to construct exclusive decision boundaries in layers according to touch intentions, predict touch intentions and match the corresponding boundaries to complete the preliminary classification, and optimize the hyperplane parameters by cross-validation with other intention boundaries and combining historical samples. The classification index verification and result output unit is used to calculate the classification accuracy, pressure feature dimensionality reduction retention rate and cross-validation conflict rate. If any index fails to meet the preset standard, the feature weights, kernel functions and boundary parameters are readjusted and iteratively optimized until the index meets the standard. The current touch event features are mapped to the joint judgment space to complete the classification and output the spatial classification judgment result.

[0063] Furthermore, the aforementioned determination boundary construction and preliminary classification unit includes: The touch intent prediction auxiliary unit is used to refine the feature weights and kernel function parameters of touch intent according to the motion intensity, and adds touch duration to assist in predicting touch intent; The preliminary classification and confidence calculation unit is used to combine real-time motion intensity and predicted touch intent, adjust the dedicated boundary parameters of the corresponding intent, use support vector machine to complete the preliminary classification, and simultaneously calculate the classification confidence. The classification result cross-validation unit is used to cross-validate the preliminary classification results using the decision boundaries of other touch intentions, and to verify the consistency of classification confidence and the matching degree between touch duration and touch intention; The boundary parameter iterative optimization unit is used to optimize the hyperplane penalty coefficient and kernel function parameters by combining historical misjudgment samples if the confidence consistency or duration intent matching degree does not meet the preset standard, retrain the dedicated judgment boundary and iteratively optimize it until the indicator meets the standard.

[0064] Furthermore, the aforementioned preliminary classification and confidence calculation unit includes: The confidence-related model and association library construction unit is used to establish a mapping model between confidence threshold and misjudgment risk based on motion intensity and motion pattern, and simultaneously build an association library between historical confidence and classification accuracy. The boundary and kernel function parameter adjustment unit is used to combine real-time motion intensity and predicted touch intent, and adjust the exclusive boundary parameters and the appropriate kernel function type of the corresponding intent with reference to the mapping model. The preliminary classification and confidence calibration unit is used to perform preliminary classification of the current touch event features using a support vector machine, calculate the original confidence of the classification, calibrate the original confidence through the historical association database, and assess the risk of misjudgment based on the calibrated confidence. The classification index verification and model iteration unit is used to proceed to other touch intent boundary cross-validation stages if the classification confidence meets the standard after calibration and the misjudgment risk does not meet the preset standard; if it does not meet the standard, it is marked as a sample to be optimized and the classification confidence fluctuation value is recorded. The sample to be optimized and the fluctuation data are collected, the boundary parameters, confidence threshold and kernel function type are readjusted, and the model is iteratively trained until the index meets the standard.

[0065] Furthermore, the aforementioned boundary and kernel function parameter adjustment unit includes: The kernel function and confidence threshold setting and association library construction unit is used to set the appropriate kernel function and confidence threshold according to the motion intensity, build an association library of motion intensity, touch intention, misjudgment features and confidence, and record historical calibration errors and parameters exceeding the standard. The low-fluctuation parameter selection and boundary parameter adjustment unit is used to predict the current touch intention, combine real-time motion intensity and pressure characteristics, select low-fluctuation parameters from the associated library, and adjust the judgment boundary hyperplane and anti-interference parameters. The preliminary classification and confidence level secondary calibration unit is used to perform preliminary classification of the current touch event features based on the adjusted parameters using a support vector machine, simultaneously calculate the original confidence level of the classification, correct historical calibration errors through the association library, and perform secondary calibration of the original confidence level in combination with the pressure feature contribution. The misjudgment risk assessment and parameter iteration optimization unit is used to assess the misjudgment risk based on the calibrated classification confidence and the confidence threshold corresponding to the motion intensity, calculate the classification confidence fluctuation value of multiple consecutive touches, and add the misjudgment risk if the fluctuation exceeds the standard. If the risk reaches the preset standard, it will enter cross-validation. If it does not reach the preset standard, the sample and fluctuation data will be added to the association library, and the kernel function, confidence threshold and boundary parameters will be readjusted for iterative optimization.

[0066] Furthermore, the aforementioned dynamic threshold determination module 400 includes: The motion pattern recognition and threshold rule construction unit is used to recognize motion patterns based on multimodal feature vectors, set exclusive motion intensity grading thresholds and threshold adjustment rules according to motion patterns, and simultaneously build an optimized sample library that includes historical threshold adjustments, judgment accuracy and boundary sample features. The sliding window setting and threshold dynamic adjustment unit is used to set the duration of the sliding window and continuously collect touch events, calculate the statistical characteristics of motion intensity and the distribution of touch features within the window, calibrate the initial judgment threshold in combination with the optimized sample library, dynamically adjust the threshold according to the motion mode and real-time motion intensity, and simultaneously label the boundary samples within the preset threshold range. The touch event preliminary judgment and periodic analysis unit is used to make preliminary judgments on each touch event based on the adjusted judgment threshold. If the judgment results are consistent multiple times, the judgment conclusion is strengthened. If the judgment results are inconsistent or contain boundary samples, the periodicity of the motion pattern is analyzed and the judgment logic of the same historical scene in the sample library is matched. The accidental touch type definition and sample library iteration unit is used to determine systematic accidental touches with regular motion cycles as motion interference, and those without regularity as mixed touch scenarios. It supplements all data, boundary samples and judgment results into the optimization sample library, iteratively optimizes the motion mode-specific threshold rules and calibration logic, and outputs dynamic threshold judgment results.

[0067] Furthermore, the aforementioned sliding window setting and threshold dynamic adjustment unit includes: The subdivided mode sample library construction and rule setting unit is used to subdivide motion modes and build an optimized sample library subset corresponding to each subdivided mode, filter historical misjudged samples to establish the mapping relationship between misjudged scenarios and adaptation parameters, and set motion intensity fluctuation thresholds and window and threshold collaborative adaptation rules. The initial window setting and feature statistics unit is used to match real-time subdivided motion modes. Combining the mapping relationship between the misjudged scenario and the adaptation parameters, the initial sliding window duration is set, touch events are continuously collected, and the motion intensity statistics and touch feature distribution within the window are calculated. The window and threshold dynamic adjustment and boundary sample labeling unit is used to adjust the window duration and threshold adjustment range if the motion intensity fluctuation exceeds the standard, calibrate the initial judgment threshold in combination with the sample library subset of the corresponding subdivision mode, dynamically adjust the judgment threshold according to the real-time motion intensity, and simultaneously label the boundary samples in the preset threshold range. The data feedback and rule iteration optimization unit is used to feed back the current fluctuation adaptation effect, threshold deviation, boundary samples and misjudgment situation to the mapping relationship between the optimization sample library and the misjudgment scenario adaptation parameters, and iteratively optimize the window, threshold rules and collaborative adaptation logic specific to the subdivided mode.

[0068] Furthermore, the aforementioned initial window setting and feature statistics unit includes: The misjudged sample clustering and weighted mapping construction unit is used to cluster historical misjudged samples according to subdivided motion patterns, assign high weights to high-frequency misjudged scenarios, construct a weighted mapping relationship between misjudged scenarios and window duration, extract historical statistical data of the same pattern and intensity range, and set touch feature pre-screening rules. The initial sliding window setting and fine-tuning unit is used to match the real-time subdivided motion mode. Combining the mapping relationship between the misjudged scene and the adaptation parameters and the weighted mapping between the misjudged scene and the window duration, the initial sliding window duration is set, valid touch events are collected according to the touch feature pre-screening rules, the touch frequency is calculated synchronously, and the sliding window duration is verified and fine-tuned. The feature statistics and calibration unit within the window is used to calculate the motion intensity statistical features and the filtered touch feature distribution within the fine-tuned sliding window, compare and calibrate them with the historical statistical data of the same intensity range in the corresponding subdivision mode, and mark feature data similar to misjudged scenarios. The data feedback and rule iteration optimization unit is used to feed back feature similarity, motion intensity fluctuation value, calibration deviation and touch frequency data to the weighted mapping and sample library of misjudged scenarios and window duration, and iteratively optimize window adaptation rules, touch feature pre-screening rules and historical data migration logic.

[0069] Furthermore, the aforementioned misjudged sample clustering and weighted mapping construction unit includes: The multi-device misjudgment sample processing and weighted mapping construction unit is used to collect historical misjudgment samples from multiple devices and perform cross-device normalization processing. It clusters historical misjudgment samples by subdividing motion modes and motion intensity ranges, assigns weights to each cluster by combining misjudgment frequency and misjudgment loss quantification results, constructs loss weighted mapping relationship, extracts historical statistical data of the same mode and intensity range after normalization, and sets loss-oriented touch feature pre-screening rules. The misjudgment trend analysis and weighted mapping optimization unit is used to analyze the misjudgment trend of each subdivided motion mode, adjust the weight of each cluster based on the misjudgment trend, further optimize the loss weighted mapping relationship, match the real-time subdivided motion mode and motion intensity range, and verify the adaptability and trend avoidance effect of the optimized weighted mapping relationship and touch feature pre-screening rules. The core indicator statistics and scenario labeling unit is used to statistically analyze the high loss misjudgment avoidance rate and feature correlation retention rate after pre-screening, verify the accuracy of historical statistical data, and label scenarios with adaptation bias and insufficient trend avoidance. The data feedback and full-process iterative optimization unit is used to feed back the adaptation bias, avoidance rate and trend prediction results to the sample library, re-optimize the normalization algorithm, cluster weight and touch feature pre-screening rules, and iteratively improve the loss weighted mapping relationship and historical data extraction logic.

[0070] Furthermore, the aforementioned fusion decision module 500 includes: The dynamic weight allocation and Bayesian model building unit is used to assign dynamic weights based on historical accuracy to the time-series causal judgment results, spatial classification judgment results, and dynamic threshold judgment results, quantify the loss of false hits and missed judgments, build a Bayesian decision model containing a loss function, set the posterior probability threshold that is dynamically adjusted with the loss, and formulate parameter linkage adjustment rules. The multi-judgment result integration and decision-making unit is used to integrate the temporal causal judgment result, spatial classification judgment result and dynamic threshold judgment result into the Bayesian decision model, calculate the posterior probability based on Bayesian decision theory and make a decision, output the final touch recognition result, record false touch and missed judgment events, and statistically calculate the false judgment rate and missed judgment rate after loss weighting, and stratify difficult samples according to false touch type. The difficult sample analysis and classification model optimization unit is used to analyze the characteristics of difficult samples in each stratum using clustering algorithms, identify marginal samples, trigger sample library updates and add difficult and marginal samples to the sample library according to strata, and optimize the classification model using a differentiated incremental learning strategy. The parameter linkage adjustment and closed-loop mechanism construction unit is used to combine the distribution of false touch features and loss quantification results, and synchronously adjust various judgment thresholds and model parameters according to the parameter linkage rules. It optimizes the loss function, posterior probability threshold and dynamic weight of judgment results of the Bayesian decision model, and feeds all optimization data back to the pre-initialization stage to iteratively optimize and form a closed-loop recognition mechanism.

[0071] Furthermore, the aforementioned dynamic weight allocation and Bayesian model construction unit includes: The initial weight allocation and preliminary Bayesian model construction unit is used to statistically analyze the historical accuracy and hard sample identification rate of time-series causal determination, spatial classification determination, and dynamic threshold determination. Initial dynamic weights are allocated to the three types of determination results based on the double accuracy. A weight decay mechanism is introduced and the triggering conditions for weight decay are clarified. The loss of false triggers and false misses is quantified. The loss difference between hard samples and ordinary samples is distinguished. A Bayesian loss function containing a weight decay term and hard sample guidance is constructed to complete the preliminary construction of the Bayesian decision model. The posterior probability threshold setting and hierarchical linkage rule formulation unit is used to set the posterior probability threshold that is dynamically adjusted with loss, formulate hierarchical parameter linkage rules, distinguish between high-priority and low-priority linkage content, and clarify the execution logic of various linkage rules. The multi-index verification and non-compliance item marking unit is used to verify the adaptability of dynamic weights, the rationality of loss quantification, and the effectiveness of hierarchical parameter linkage rules. It also calculates the weight calibration error, loss deviation, and the loss reduction of high-priority rules, and marks the evaluation indicators that do not meet the preset standards. The indicator optimization and closed-loop iteration unit is used to adjust the weight allocation, attenuation magnitude, loss quantification coefficient and priority of hierarchical parameter linkage rules if any evaluation indicator fails to meet the preset standard, optimize the Bayesian loss function and threshold adjustment logic, and feed all optimization data back to the previous steps to form a closed-loop iteration.

[0072] Reference Figure 3 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, this computer device includes a processor, memory, network interface, and database connected via a bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores operations, computer programs, and the database. The internal memory provides an environment for the operation and execution of the computer programs stored in the non-volatile storage media. The database stores data such as the accurate touch operation recognition method for the smart bracelet. The network interface is used for communication with external terminals via a network connection. When executed by a processor, this computer program implements a method for accurately recognizing touch operations on a smart bracelet. The method includes: acquiring capacitive signals, motion sensing data, and pressure data from the smart bracelet's touchscreen; establishing an analysis time window based on the touch event trigger moment; extracting frequency and time domain features from the capacitive signals; filtering the motion sensing data to extract motion features; aligning and normalizing all feature timestamps to form a multimodal feature vector; aligning the multimodal feature vector to a unified reference coordinate system along the time dimension; analyzing the temporal relationship between the touch and motion signals; using a causality test algorithm to determine the causal direction between the two; and outputting a temporal causality determination result. The process involves: extracting touch trajectory, motion posture, and signal energy features based on the multimodal feature vectors; constructing a joint judgment space; using historical samples to construct judgment boundaries; mapping the current touch event features to the judgment space for classification; and outputting the spatial classification judgment result. The motion intensity is determined based on the multimodal feature vectors, and the judgment threshold is dynamically adjusted. Combined with the temporal pattern analysis of the touch event, a dynamic threshold judgment result is output. The three judgment results are integrated, and the posterior probability is calculated based on Bayesian decision theory, resulting in the final recognition result. Feedback data is collected to optimize model parameters and judgment thresholds, and the sample library is updated to form a closed-loop recognition mechanism.

[0073] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for accurate recognition of touch operations on a smart bracelet, including: acquiring capacitive signals, motion sensing data, and pressure data from the smart bracelet's touchscreen; establishing an analysis time window based on the touch event trigger moment; extracting frequency and time domain features from the capacitive signals; filtering and extracting motion features from the motion sensing data; aligning and normalizing all feature timestamps to form a multimodal feature vector; aligning the multimodal feature vector to a unified reference coordinate system along the time dimension; analyzing the temporal relationship between the touch and motion signals; and combining a causal test algorithm to determine... The causal direction of the two factors is determined, and the temporal causal determination result is output. Based on the multimodal feature vector, touch trajectory, motion posture, and signal energy features are extracted to construct a joint determination space. Historical samples are used to construct the determination boundary, and the current touch event features are mapped to the determination space for classification, outputting the spatial classification determination result. The motion intensity is determined according to the multimodal feature vector, the determination threshold is dynamically adjusted, and the determination is combined with the temporal pattern analysis of the touch event to output the dynamic threshold determination result. The above three types of determination results are integrated, the posterior probability is calculated based on Bayesian decision theory, and the final recognition result is output. Feedback data is collected to optimize model parameters and determination thresholds, and the sample library is updated to form a closed-loop recognition mechanism.

[0074] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0075] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for accurate recognition of touch operation on a smart bracelet, characterized in that, include: Collect capacitive signals, motion sensing data, and pressure data from the smart bracelet touchscreen. Establish an analysis time window based on the touch event trigger time. Extract frequency and time domain features from the capacitive signals. Filter the motion sensing data to extract motion features. Align and normalize all feature timestamps to form a multimodal feature vector. The multimodal feature vectors are aligned to a unified reference coordinate system along the time dimension. The temporal relationship between touch and motion signals is analyzed, and the causal direction between the two is determined by a causal test algorithm. The temporal causal determination result is then output. Based on the multimodal feature vector, touch trajectory, motion posture and signal energy features are extracted, a joint judgment space is constructed, a judgment boundary is constructed using historical samples, the current touch event features are mapped to the judgment space for classification, and the space classification judgment result is output. The motion intensity is determined based on the multimodal feature vector, the judgment threshold is dynamically adjusted, and the dynamic threshold judgment result is output by combining the analysis and judgment of the touch event timing pattern. The system integrates three types of judgment results, calculates the posterior probability based on Bayesian decision theory and outputs the final recognition result, collects feedback data to optimize model parameters and judgment thresholds, and updates the sample library to form a closed-loop recognition mechanism. The steps of extracting touch trajectory, motion posture, and signal energy features based on the multimodal feature vectors, constructing a joint judgment space, using historical samples to construct judgment boundaries, mapping the current touch event features to the judgment space for classification, and outputting the spatial classification judgment result include: The touch trajectory, motion posture, and signal energy feature weights are dynamically allocated according to the motion mode and motion intensity. Pressure features are integrated to construct a multi-dimensional feature set and establish a mapping relationship between motion intensity and kernel function. The multidimensional feature set is reduced to three dimensions using a feature dimensionality reduction algorithm, while retaining the core distinguishing information. It is then mapped to an enhanced three-dimensional joint decision space, and the support vector machine is initialized by selecting an adaptation kernel function based on the motion intensity. We construct exclusive decision boundaries based on touch intent, classify touch intent into two categories: tap and swipe, introduce touch duration features to assist in the prediction of touch intent, match the predicted touch intent with the corresponding boundary to complete the preliminary classification, call the decision boundaries corresponding to other touch intents to carry out cross-validation, and combine historical mis-touch samples and normal touch samples to jointly optimize the hyperplane parameters of support vector machine. Calculate the classification accuracy, pressure feature dimensionality reduction retention rate, and cross-validation conflict rate. If any indicator fails to meet the preset standard, readjust the feature weights, kernel function, and boundary parameters for iterative optimization until the indicator meets the standard. Map the current touch event features to the joint decision space to complete the classification and output the spatial classification decision result.

2. The method for accurate recognition of touch operation on a smart bracelet according to claim 1, characterized in that, The steps of aligning the multimodal feature vectors to a unified reference coordinate system along the time dimension, analyzing the temporal relationship between touch and motion signals, determining the causal direction between the two using a causal testing algorithm, and outputting the temporal causal determination result include: Extract motion patterns, motion intensity, and duration of motion mutations, and establish mapping relationships between these three factors and initial delay thresholds, causality test thresholds, and time series analysis windows; According to the mapping relationship, set various parameters, locate the peak time of the touch signal and the sudden change time of the motion signal in the dynamic analysis window, calculate the time difference between the two to complete the initial timing determination, and simultaneously construct a multimodal weighted timing feature sequence. The multimodal weighted time series feature sequence is incorporated into the causal test vector autoregressive model. The model order is adjusted according to the initial time series judgment result, and the causal judgment result is obtained through the test algorithm. The initial time series determination result and the causal determination result are integrated. If the confidence level of the determination result does not reach the preset threshold, the threshold and the analysis window length are dynamically adjusted according to the motion intensity and iteratively optimized until the results are consistent, and the time series causal determination result is output.

3. The method for accurate recognition of touch operation on a smart bracelet according to claim 2, characterized in that, The steps of setting various parameters according to the mapping relationship, locating the peak moment of the touch signal and the moment of sudden change of the motion signal within the dynamic analysis window, calculating the time difference between the two to complete the initial timing determination, and simultaneously constructing a multimodal weighted timing feature sequence include: By integrating exercise patterns, exercise intensity, and historical data, a multi-dimensional mapping relationship is constructed, and a gradient optimization algorithm is used to iteratively optimize the dynamic analysis window and feature weight parameters; Based on the optimized parameters and combined with the pressure signal stability calibration feature weights, multimodal time series data are extracted to construct a correlation-enhanced weighted time series feature sequence. Within the dynamic analysis window, based on the effective contact time of pressure, the peak time of the touch signal and the abrupt change time of the motion signal are located, and the combined time difference between the two is calculated to complete the initial timing determination. The calibrated weighted time series feature sequence and the initial time series judgment result are input into the causality test process, and the feature attention is adjusted according to the motion intensity.

4. The method for accurate recognition of touch operation on a smart bracelet according to claim 1, characterized in that, The steps of constructing dedicated decision boundaries according to touch intent, classifying touch intents into two categories: tap and swipe, introducing touch duration features to assist in the prediction of touch intents, matching the predicted touch intents with corresponding boundaries to complete preliminary classification, calling decision boundaries corresponding to other touch intents for cross-validation, and simultaneously optimizing the hyperplane parameters of the support vector machine by combining historical mis-touch samples and normal touch samples, include: The feature weights and kernel function parameters of touch intent are refined according to the intensity of motion, and the touch duration is added to assist in predicting touch intent; By combining real-time motion intensity with predicted touch intent, the specific boundary parameters of the corresponding intent are adjusted, and a support vector machine is used to complete the initial classification, while the classification confidence is calculated simultaneously. Cross-validate the preliminary classification results using other touch intent determination boundaries to verify the consistency of classification confidence and the matching degree between touch duration and touch intent; If the confidence consistency or duration intent matching degree does not meet the preset standard, the hyperplane penalty coefficient and kernel function parameters are optimized by combining historical misjudgment samples, the dedicated judgment boundary is retrained and iteratively optimized until the indicator meets the standard.

5. The method for accurate recognition of touch operation of a smart bracelet according to claim 4, characterized in that, The steps of combining real-time motion intensity with predicted touch intent, adjusting the corresponding boundary parameters for the intent, using a support vector machine to complete preliminary classification, and simultaneously calculating classification confidence include: Based on exercise intensity and exercise pattern, establish a mapping model between confidence threshold and misjudgment risk, and simultaneously build a correlation library between historical confidence and classification accuracy; By combining real-time motion intensity with the predicted touch intent, the specific boundary parameters of the corresponding intent and the appropriate kernel function type are adjusted with reference to the mapping model. Support vector machine is used to perform preliminary classification of current touch event features, calculate the original confidence score of the classification, calibrate the original confidence score through a historical association database, and assess the risk of misjudgment based on the calibrated confidence score; If the classification confidence level meets the standard after calibration and the misjudgment risk does not meet the preset standard, proceed to other touch intent boundary cross-validation stages; if it does not meet the standard, mark it as a sample to be optimized and record the classification confidence fluctuation value, collect the sample to be optimized and fluctuation data, readjust the boundary parameters, confidence threshold and kernel function type, and iteratively train the model until the indicator meets the standard.

6. The method for accurate recognition of touch operation of a smart bracelet according to claim 1, characterized in that, The steps of determining motion intensity based on the multimodal feature vector, dynamically adjusting the judgment threshold, combining touch event timing pattern analysis, and outputting the dynamic threshold judgment result include: Based on multimodal feature vector recognition of motion patterns, specific motion intensity grading thresholds and threshold adjustment rules are set according to motion patterns, and an optimized sample library including historical threshold adjustments, judgment accuracy and boundary sample features is constructed simultaneously. Set the duration of the sliding window and continuously collect touch events. Calculate the statistical characteristics of motion intensity and the distribution of touch features within the window. Combine the optimized sample library to calibrate the initial judgment threshold. Dynamically adjust the threshold according to the motion mode and real-time motion intensity. Simultaneously label the boundary samples within the preset threshold range. Based on the adjusted judgment threshold, each touch event is initially judged. If the judgment results are consistent multiple times, the judgment conclusion is strengthened. If the judgment results are inconsistent or contain boundary samples, the periodicity of the motion pattern is analyzed and the judgment logic of the same historical scene in the sample library is matched. If a regular motion cycle exists, it is determined to be a systemic accidental touch due to motion interference; if there is no regularity, it is defined as a mixed touch scenario. All data, boundary samples, and judgment results are added to the optimization sample library, and the motion mode-specific threshold rules and calibration logic are iteratively optimized to output dynamic threshold judgment results.

7. The method for accurate recognition of touch operation of a smart bracelet according to claim 6, characterized in that, The steps of setting the sliding window duration and continuously collecting touch events, calculating the statistical characteristics of motion intensity and the distribution of touch features within the window, calibrating the initial judgment threshold in conjunction with an optimized sample library, dynamically adjusting the threshold according to the motion mode and real-time motion intensity, and synchronously labeling boundary samples within the preset threshold range include: Subdivide the motion modes and construct an optimized sample library subset corresponding to each subdivision mode. Filter historical misjudged samples to establish a mapping relationship between misjudged scenarios and adaptation parameters. Set motion intensity fluctuation thresholds and window and threshold collaborative adaptation rules. Match real-time subdivided motion modes, combine the mapping relationship between the misjudged scenarios and the adaptation parameters, set the initial sliding window duration, continuously collect touch events, and calculate the statistical characteristics of motion intensity and the distribution of touch features within the window; If the fluctuation of exercise intensity exceeds the limit, the window duration and threshold adjustment range are adjusted, the initial judgment threshold is calibrated in combination with the sample library subset of the corresponding subdivision mode, the judgment threshold is dynamically adjusted according to the real-time exercise intensity, and the boundary samples in the preset threshold range are labeled simultaneously. The fluctuation adaptation effect, threshold deviation, boundary samples and misjudgment situations are fed back to the mapping relationship between the optimized sample library and the misjudgment scenario adaptation parameters, and the window, threshold rules and collaborative adaptation logic specific to the subdivision mode are iteratively optimized.

8. The method for accurate recognition of touch operation of a smart bracelet according to claim 7, characterized in that, The steps of matching real-time subdivided motion patterns, combining the mapping relationship between the misjudged scenarios and adaptation parameters, setting the initial sliding window duration, continuously collecting touch events, and calculating the statistical features of motion intensity and the distribution of touch features within the window include: Historical misjudgment samples are clustered according to subdivided motion patterns, high-frequency misjudgment scenarios are given high weight, a weighted mapping relationship between misjudgment scenarios and window duration is constructed, historical statistical data of the same mode and intensity range are extracted, and touch feature pre-screening rules are set. Match real-time subdivided motion modes, combine the mapping relationship between the misjudged scene and the adaptation parameters and the weighted mapping between the misjudged scene and the window duration, set the initial sliding window duration, collect valid touch events according to the touch feature pre-screening rules, calculate the touch frequency synchronously, and verify and fine-tune the sliding window duration; Calculate the motion intensity statistical features and the filtered touch feature distribution within the fine-tuned sliding window, compare and calibrate them with the historical statistical data of the same intensity range in the corresponding sub-mode, and mark the feature data similar to the misjudged scenario. The similarity of features, motion intensity fluctuations, calibration deviations, and touch frequency data are fed back to the weighted mapping of misjudged scenarios and window duration and the sample library, and the window adaptation rules, touch feature pre-screening rules, and historical data migration logic are iteratively optimized.

9. A precise touch operation recognition system for a smart bracelet, used to implement the method described in any one of claims 1-8, characterized in that, include: The feature extraction module is used to collect capacitive signals, motion sensing data and pressure data from the smart bracelet touch screen. It establishes an analysis time window based on the touch event trigger time, extracts frequency domain and time domain features from the capacitive signals, filters and extracts motion features from the motion sensing data, aligns and normalizes all feature timestamps, and forms a multimodal feature vector. The temporal causality determination module is used to align the multimodal feature vectors to a unified reference coordinate system according to the time dimension, analyze the temporal relationship between touch and motion signals, determine the causal direction of the two by combining the causality test algorithm, and output the temporal causality determination result. The spatial classification and determination module is used to extract touch trajectory, motion posture and signal energy features based on the multimodal feature vector, construct a joint determination space, construct a determination boundary using historical samples, map the current touch event features to the determination space for classification, and output the spatial classification and determination result. The dynamic threshold determination module is used to determine the motion intensity based on the multimodal feature vector, dynamically adjust the determination threshold, combine the touch event timing pattern analysis and determination, and output the dynamic threshold determination result. The fusion decision module integrates the three types of judgment results, calculates the posterior probability based on Bayesian decision theory and outputs the final recognition result, collects feedback data to optimize model parameters and judgment thresholds, and updates the sample library to form a closed-loop recognition mechanism.

Images