A tactile-guided intelligent matching method and system for calligraphy templates

Abstract

This invention relates to the field of human-computer interaction and calligraphy teaching technology, and discloses a tactile-guided intelligent matching method and system for calligraphy templates. The method collects user writing actions and tactile response data through multimodal sensors, extracts tactile operation habit features, and constructs a personalized tactile profile. It uses a graph neural network to establish a user-template tactile association network, and combines meta-learning to construct a rapid adaptation model. Through transfer learning and domain adaptation, it calculates the similarity of tactile needs between templates, selectively transfers tactile habits from the source template to the target template, and designs a progressive transition strategy. During the user's copying process, progressive tactile feedback control is implemented, and the guidance strategy is continuously optimized in conjunction with online learning. Finally, based on the template switching effect analysis, adaptation ability is constructed, a template learning graph is generated, and a personalized recommendation strategy is produced. This invention enables smooth transfer of tactile habits when switching calligraphy templates, reduces user readaptation costs, and improves calligraphy learning efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human-computer interaction and calligraphy teaching, more specifically, it relates to a kind of intelligent matching method and system for touch guiding calligraphy template. BACKGROUND

[0002] As an important part of Chinese traditional culture, calligraphy learning process is highly dependent on accurate grasp of stroke strength, pen movement rhythm and space layout. With the development of digital teaching technology, the touch guiding system provides a new auxiliary means for calligraphy learning by providing real-time force feedback and vibration feedback to users. In actual use, users often need to switch between different calligraphy templates for practice, and different templates have obvious differences in stroke characteristics, movement rhythm and touch guiding requirements, which leads to the need for users to adapt to new touch guiding strategies after switching templates, affecting learning efficiency.

[0003] Existing calligraphy template matching technology is mainly based on static rules or simple similarity calculation, and usually uses a fixed font library for matching, without fully considering the user's personalized touch operation habits. Different users have different sensitivities to touch stimulus intensity, different degrees of dependence on touch guidance, and different preferences for touch feedback timing, and the same set of touch guiding strategies has obvious differences in effect for different users. Existing systems lack the ability to model user touch operation habits and cannot generate adaptive touch guiding schemes based on user's personalized characteristics.

[0004] When users switch calligraphy templates in existing touch guiding systems, they usually directly use the preset touch strategy of the target template, ignoring the touch operation habits established by the user on the source template, resulting in a mismatch between the touch strategy and the user's habits, and the user needs to adapt to the new touch mode from scratch. In addition, existing systems lack effective transfer learning mechanisms and cannot transfer the touch habit knowledge accumulated by the user on the familiar template to the new template, and also lack a gradual transition strategy to smooth the touch mode switching process, causing high user adaptation cost and low learning efficiency. SUMMARY

[0005] The present application provides a kind of touch guiding calligraphy template intelligent matching method and system, solve the technical problems that existing touch guiding system in related art lacks personalized touch habit modeling capability when user switches calligraphy template, cannot transfer source template touch habit to target template and lacks gradual transition strategy, resulting in high user adaptation cost.

[0006] The present application provides a kind of touch guiding calligraphy template intelligent matching method, including the following steps: S1 collects user writing action data and tactile interaction response data, and uses multimodal sensor synchronous acquisition and feature extraction methods to construct a personalized tactile operation profile including tactile operation habit characteristics, template usage history and cross-template habit change patterns; S2, obtain personalized tactile operation profiles, and use graph neural networks and meta-learning methods to construct a user-personalized tactile habit model that can predict users’ tactile preferences on specific templates and achieve rapid adaptation; S3, based on the user's personalized tactile habit model, uses transfer learning and domain adaptation methods to generate a tactile strategy intelligent matching scheme that includes template similarity calculation, selective transfer mechanism and progressive transition strategy; S4 receives the haptic strategy intelligent matching scheme and real-time writing data, adopts progressive haptic feedback control and online learning methods, outputs continuously optimized haptic guidance strategies and records template switching effect data; S5 acquires template switching effect data and user tactile operation profiles, and uses adaptive analysis and template learning graph construction methods to generate an optimization plan that includes personalized template recommendation strategies and learning suggestion reports.

[0007] In a preferred embodiment, S1 includes: A system for synchronously acquiring writing motion and tactile response is configured, based on a hardware platform of a smart writing pen and a haptic feedback glove, to collect writing motion data and tactile interaction response data; the smart writing pen integrates a six-axis inertial measurement unit to acquire the three-dimensional spatial trajectory, movement speed and posture angle of the pen tip, and a pressure sensor is installed on the pen tip to measure writing pressure, while a capacitive touch sensor array is embedded in the pen grip area to detect pen grip posture and grip force distribution; the haptic feedback glove has haptic actuators and response sensors installed in the fingers, palm and wrist respectively, with a pressure sensor and electromyography sensor configured next to each actuator; Based on the collected multimodal data, feature vectors of user tactile operation habits are extracted. The feature vectors include tactile sensitivity features, tactile dependence features, tactile timing preference features, and motor control ability features. A personalized tactile operation profile is created for each user, which includes habit characteristics, template usage history, and tactile preference configuration. The feature vector is dynamically updated using an exponentially weighted moving average method, and cross-template correlation analysis is performed on the differences in users' tactile operation habits on different calligraphy templates.

[0008] In a preferred embodiment, S2 includes: Extract template features and build a template feature library. Extract stroke geometric features, kinematic features and tactile guidance demand features for each template in the calligraphy template library, and organize them into a template feature vector. The tactile guidance demand features include directional guidance demand, force prompt demand, rhythm guidance demand and fluency support demand, which are quantified into zero to one values ​​to form a tactile demand vector. A user-template tactile habit association network is constructed. Based on multi-user, multi-template tactile operation data in personalized tactile operation profiles, a heterogeneous graph structure containing user nodes, template nodes, and tactile habit nodes is constructed using a graph neural network method. Imitation edges are established between user nodes and template nodes, preference edges are established between user nodes and tactile habit nodes, and demand edges are established between template nodes and tactile habit nodes. Graph convolutional networks are used for node representation learning. The training tasks include the main task of link prediction and the auxiliary task of tactile habit recommendation.

[0009] In a preferred embodiment, S2 further includes: To achieve rapid adaptation, a meta-learning model is trained based on a user-template tactile habit association network and a model-independent meta-learning algorithm is used to train the rapid adaptation model. Tasks are sampled from the user's template usage history. Each task contains a support set and a query set. The support set contains the user's previous writing data on the template, and the query set contains subsequent writing data. In the inner loop, gradient descent is performed on the support set to obtain task-specific parameters starting from the meta-parameters; in the outer loop, the performance of the adapted model is evaluated on the query set and the meta-parameters are updated; task similarity weighting is introduced in the meta-loss calculation, and the weight coefficient is inversely proportional to the cosine similarity of the feature vectors of the source template and the target template; the output of the meta-learning model is the adapted tactile parameter vector.

[0010] In a preferred embodiment, S3 includes: To calculate the tactile demand similarity between templates, the template feature vectors and tactile demand vectors of the source and target templates are retrieved from the template feature library. Geometric feature similarity, kinematic feature similarity, and tactile demand feature similarity are calculated separately. The three are then weighted and summed to obtain the overall similarity. The kinematic feature similarity is obtained by transforming the dynamic time warped distance using an exponential decay function. The tactile demand feature similarity is obtained by calculating the weighted average of each dimension using a Gaussian kernel function. The difference between the source template and the target template is calculated for the four dimensions of tactile needs. When the difference is less than the preset difference threshold, the corresponding dimension is determined to be transferable. When the difference is greater than or equal to the preset difference threshold, the corresponding dimension is determined to be adjusted, and a transfer marker vector is generated. A domain-adaptive-based haptic policy transfer scheme is generated. An adversarial domain adaptation method is used to construct a network containing a feature extractor, a haptic policy predictor, and a domain discriminator. A selective transfer mechanism is introduced, which directly applies the haptic policy of the source template for transferable dimensions, and adopts the policy predicted by the domain adaptation network for dimensions that need to be adjusted.

[0011] In a preferred embodiment, S3 further includes: A progressive tactile transition strategy is designed, dividing the adaptation process after template switching into an initial stage, a transition stage, and a stable stage. In the initial stage, the tactile strategy mainly uses the source template strategy, in the transition stage the proportion of the target template strategy is gradually increased, and in the stable stage the target template strategy is used entirely. The transition weight uses an S-shaped curve function that increases from zero to one with the number of writings. The speed parameter and midpoint parameter are adjusted according to the overall similarity. During the transition, the adaptability index is calculated. When the adaptability index is less than the first adaptability threshold, the transition speed is slowed down. When the adaptability index is greater than the first adaptability threshold but less than the second adaptability threshold, the current transition speed is maintained. When the adaptability index is greater than the second adaptability threshold, the transition speed is accelerated. Set an independent transition weight function for each dimension. Set a faster transition speed for dimensions marked as transferable in the migration marker vector, and a slower transition speed for dimensions marked as needing adjustment.

[0012] In a preferred embodiment, S4 includes: Implement progressive tactile feedback control, establish a tactile feedback control loop, read the current writing state in each control cycle, align the real-time writing trajectory with the standard trajectory of the target template using a dynamic time warping algorithm, and calculate the positional deviation, directional deviation, and force deviation; calculate the current tactile parameters according to the transition weight function of the progressive transition strategy, and obtain the current parameter value by linearly interpolating the source template parameter value and the target template parameter value according to the transition weight. A tactile feedback signal is generated based on the current tactile parameters and writing deviation. When the deviation is small, the tactile intensity is set to the currently calculated parameter value. When the deviation is large, the tactile intensity is appropriately increased. The control signal is sent to the tactile feedback device for execution through a low-latency communication interface, and the execution status of each tactile feedback is recorded.

[0013] In a preferred embodiment, S4 further includes: An online learning strategy is adopted to optimize haptic feedback. A comprehensive loss function is defined, which is a weighted sum of writing quality loss and user comfort loss, with the weight of writing quality loss being greater than that of user comfort loss. After each writing session, the gradient of the loss function with respect to the haptic parameters is calculated and updated online. An exponentially weighted moving average method is used to smooth the parameter updates. The parameter update frequency is dynamically adjusted according to the user's adaptation stage. In the initial stage, the parameters are updated after each writing session. In the transition stage, the update frequency is reduced, and in the stable stage, the update frequency is further reduced. When a continuous decline in writing quality is detected, the parameter update is triggered immediately. An evaluation mechanism for template switching effect was established, defining three types of evaluation indicators: adaptation time, writing quality, and user satisfaction. Paired t-tests were used to compare the differences between the experimental group and the control group.

[0014] In a preferred embodiment, S5 includes: To analyze users' cross-template adaptability, we extracted the adaptation curve for each template switch and obtained the adaptation speed slope and convergence value through linear regression fitting. We defined the adaptation speed index as the reciprocal of the number of writings required to achieve stable writing quality, and the adaptation stability index as the reciprocal of the standard deviation of writing quality during the adaptation period plus one. We then weighted the two to obtain the comprehensive adaptability index. A template learning graph is constructed, with all templates as nodes and the tactile similarity between templates as edge weights. A greedy algorithm is used to start from the user's current template and select the template with the highest similarity that has not been copied as the next recommended template. The recommendation strategy is dynamically adjusted based on the comprehensive adaptability index. When the comprehensive adaptability index is greater than the first ability threshold, the learning challenge is increased. When the comprehensive adaptability index is less than the second ability threshold, templates with similarity greater than the preset similarity threshold are selected. When the comprehensive adaptability index is between the second ability threshold and the first ability threshold, templates are recommended in order of standard similarity.

[0015] In a preferred embodiment, a tactile-guided calligraphy template intelligent matching system is used to perform the steps in the above-described tactile-guided calligraphy template intelligent matching method, including: The tactile operation profile construction module is used to collect user writing action data and tactile interaction response data. It adopts a multimodal sensor synchronous acquisition and feature extraction method to construct a personalized tactile operation profile that includes tactile operation habit characteristics, template usage history and cross-template habit change patterns. The tactile habit model building module is used to obtain personalized tactile operation profiles. It uses graph neural networks and meta-learning methods to build a user-personalized tactile habit model that can predict users’ tactile preferences on specific templates and achieve rapid adaptation. The haptic strategy matching module is used to generate an intelligent haptic strategy matching scheme based on the user's personalized haptic habit model, using transfer learning and domain adaptation methods. This scheme includes template similarity calculation, selective transfer mechanism and progressive transition strategy. The haptic feedback optimization module receives haptic strategy intelligent matching schemes and real-time writing data, adopts progressive haptic feedback control and online learning methods, outputs continuously optimized haptic guidance strategies and records template switching effect data; The template recommendation optimization module is used to acquire template switching effect data and user tactile operation profiles. It uses adaptive analysis and template learning graph construction methods to generate an optimization plan that includes personalized template recommendation strategies and learning suggestion reports.

[0016] The beneficial effects of this invention are as follows: By synchronously collecting user writing action data and tactile interaction response data using multimodal sensors, personalized features such as tactile sensitivity, tactile dependence, tactile timing preference, and motor control ability are extracted to construct a tactile operation profile that includes template usage history and cross-template habit change patterns. Based on this, a graph neural network is used to establish a user-template tactile habit association network, and a rapid adaptation model is trained using meta-learning methods. This enables the system to accurately predict the user's tactile preferences on specific templates and to quickly converge to the appropriate tactile strategy with only a small amount of writing data when the user switches to a new template, thus solving the problem of existing systems lacking the ability to model personalized tactile habits. Employing transfer learning and domain adaptation methods, this study identifies transferable dimensions by calculating the similarity of tactile needs between templates. A selective transfer mechanism is used to transfer tactile habit knowledge accumulated by users on source templates to target templates. Simultaneously, a progressive transition strategy based on an S-curve function is designed, dividing the adaptation process after template switching into three stages: initial, transition, and stable. The transition speed is dynamically adjusted based on real-time user adaptability indicators to achieve smooth switching of tactile modes. Combined with online learning methods, the tactile guidance strategy is continuously optimized, reducing the readaptation cost for users during template switching and improving calligraphy learning efficiency. Attached Figure Description

[0017] Fig. 1 This is a flowchart of a tactile-guided intelligent matching method for calligraphy templates according to the present invention; Fig. 2 This is a flowchart of a tactile-guided intelligent matching method for calligraphy templates according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a tactile-guided intelligent matching method for calligraphy templates, such as... Figs. 1-2 As shown, it includes the following steps: S1 collects user writing action data and tactile interaction response data, and uses multimodal sensor synchronous acquisition and feature extraction methods to construct a personalized tactile operation profile including tactile operation habit characteristics, template usage history and cross-template habit change patterns; In response to the individual differences in operating habits exhibited by different users when using the haptic guidance system for calligraphy practice, this step uses multimodal sensors to collect users' writing action data and haptic interaction response data to establish user haptic operation profiles. The individualized differences in users' tactile operating habits are mainly reflected in the following aspects: different users have different sensitivities to the intensity of tactile stimulation; some users need strong tactile cues to perceive them, while others react noticeably to slight tactile cues. Different users have different levels of reliance on tactile guidance; novice users rely heavily on tactile guidance, while experienced users prefer independent writing. Different users have different preferences for the timing of tactile feedback; some users prefer warning guidance before an action, while others prefer corrective feedback when the action deviates. Different users have different motor control abilities, affecting their response speed and adjustment accuracy to tactile guidance. These individualized differences lead to significant differences in the effectiveness of the same tactile guidance strategy for different users; specifically, the following steps are included: S11 is equipped with a system for synchronously acquiring writing action and tactile response data; based on a hardware platform of a smart writing pen and tactile feedback gloves, a system for synchronously acquiring writing action data and tactile interaction response data is constructed. In terms of writing motion acquisition, the smart writing pen integrates a six-axis inertial measurement unit, including a three-axis accelerometer and a three-axis gyroscope, with a sampling frequency of 200Hz, to acquire the three-dimensional spatial trajectory, movement speed, and attitude angle of the pen tip in real time; a pressure sensor is installed on the pen tip, with a sampling frequency of 100Hz, to measure the pressure during writing, with a pressure measurement range of 0.1 to 10 Newtons and a resolution of 0.05 Newtons; a capacitive touch sensor array is embedded in the pen grip area, with 8 sensors distributed around the pen barrel, with a sampling frequency of 50Hz, to detect the user's pen grip posture and grip force distribution, in order to reflect the user's writing motion characteristics; In terms of tactile response acquisition, the tactile feedback glove is equipped with tactile actuators and response sensors in the fingers, palm, and wrist. Three linear resonant actuators in the fingers are located on the fingertips of the thumb, index finger, and middle finger, respectively, to provide directional tactile guidance. Two eccentric rotor actuators in the palm are located in the center of the palm and the web between the thumb and index finger, respectively, to provide force cues. One pneumatic actuator in the wrist provides supportive tactile feedback. Each actuator is equipped with a pressure sensor and an electromyography (EMG) sensor. The pressure sensor measures the pressure of the tactile stimulus on the skin, and the EMG sensor measures the muscle response caused by the tactile stimulus. The sampling frequency of both is 100Hz. This is used to detect the intensity of the user's physiological response to the tactile stimulus. Regarding the collection of user subjective responses, at key moments after tactile feedback occurs, the user's subjective evaluation of tactile perception is collected through micro-buttons on the glove or voice commands. The evaluation dimensions include: whether the tactile intensity is appropriate, using a five-level rating: too weak, slightly weak, appropriate, slightly strong, and too strong; whether the tactile timing is appropriate, using a three-level rating: too early, appropriate, and too late; and whether the tactile guidance is effective, using a three-level rating: ineffective, partially effective, and effective. The subjective evaluation data is stored in association with the objective sensor data to build a user tactile perception model. A hardware clock synchronization mechanism is adopted to unify the time base of the writing pen sensor, glove sensor and haptic actuator to the system master clock, with the time synchronization accuracy controlled within 2 milliseconds. A data frame encapsulation protocol is established, and each data frame contains fields such as timestamp, writing pen position, posture, pressure, grip state, haptic actuator activation state, haptic response sensor data and user subjective evaluation. The data frames are transmitted to the cloud server via Bluetooth 5.0 with a transmission latency of less than 30 milliseconds.

[0020] S12, Extract user tactile operation habit features; Based on the collected multimodal data, extract feature vectors reflecting user tactile operation habits; Tactile sensitivity feature extraction: Analysis of user response data to tactile stimuli of varying intensities; for each tactile feedback event, recording the activation intensity of the tactile actuator (normalized to the zero-to-one range), as well as the user's electromyographic (EMG) response intensity and subjective perception score; establishing a tactile sensitivity curve, with stimulus intensity on the horizontal axis and response intensity on the vertical axis. EMG response intensity is represented by root mean square (RMS) values ​​and normalized, and the subjective perception score uses a five-level scale. The response intensity R_response is calculated by weighting the EMG intensity R_emg and the subjective score R_subjective. Data points from multiple tactile events are fitted, and an S-curve function is used to model the sensitivity curve, with the function form being... ,in, Stimulus intensity The sensitivity coefficient, The sensitivity coefficient k represents the user's sensitivity to changes in tactile intensity, while the smaller the perception threshold T represents the user's lower tactile perception threshold. The sensitivity coefficient and perception threshold are extracted as tactile sensitivity features.

[0021] Haptic Dependence Feature Extraction: This section analyzes the correlation between user writing trajectories and haptic guidance. The writing process is divided into multiple time windows, each one second long. For each window, the haptic guidance activation rate is calculated (the proportion of haptic feedback activation time to the window duration); writing deviation is calculated (the average distance between the user's trajectory and the current standard trajectory of the calligraphy template, obtained from a calligraphy template library); and deviation improvement rate is calculated (the reduction ratio of deviation during haptic activation to deviation during inactivation), obtained by subtracting the average deviation during activation from the average deviation during inactivation, and then dividing by the average deviation during inactivation. A higher deviation improvement rate indicates more significant assistance from haptic guidance and a higher user dependence on haptics. The haptic dependence feature is obtained by averaging the deviation improvement rates across multiple windows. Simultaneously, writing fluency under haptic guidance is calculated. By analyzing speed changes before and after haptic activation, significant speed fluctuations indicate poor user adaptability to haptics, and a speed fluctuation coefficient is extracted as an adaptability feature.

[0022] Haptic Timing Preference Feature Extraction: This study analyzes user response to different haptic timings. Haptic guidance timings are categorized into three types: warning-type (activated after a preset time before the user's trajectory approaches a key control point); synchronization-type (activated when the user reaches the control point); and correction-type (activated after the user deviates from the standard trajectory by more than a threshold). Subjective evaluations and writing improvement effects under the three timings are statistically analyzed. An effectiveness score is calculated for each timing, weighted by the average subjective rating and the average deviation improvement rate under that timing. The effectiveness scores of the three timings are ranked to obtain the user's timing preference sequence, and the score of the optimal timing is recorded as the timing fit feature.

[0023] Motion control capability feature extraction: Analyzing the user's writing action characteristics. The average and standard deviation of writing speed are calculated; speed stability is represented by the coefficient of variation, which equals the speed standard deviation divided by the speed average. The curvature change of the writing trajectory is calculated, and the second derivative of curvature is extracted; curvature smoothness is represented by the root mean square of the second derivative. The stability of pen grip force is calculated by analyzing the temporal variation of force distribution using data from the grip sensor array. The stability of writing posture is calculated, represented by the standard deviation of posture angles. These features reflect the user's motion control precision; users with strong control capabilities have lower needs for haptic guidance, while users with weak control capabilities require more haptic assistance.

[0024] The extracted features are organized into a user tactile operation habit feature vector, which includes features such as sensitivity coefficient, perception threshold, tactile dependence, speed fluctuation coefficient, effectiveness score of three timings, speed variation coefficient, curvature smoothness, force stability and posture stability, comprehensively describing the user's tactile operation habits.

[0025] S13, Construct a database of user tactile operation profiles; Based on the tactile operation habit feature data accumulated from multiple writing training sessions, create a personalized tactile operation profile for each user; Archive Data Organization Structure: Each user is assigned a unique identifier, User_ID, and an independent data storage space is established. The archive includes a basic information section, recording background information such as the user's age, gender, calligraphy learning experience, and system usage time; a habit feature section, storing the tactile operation habit feature vector H_user extracted by S12, stored in a time-series format, recording the updated feature vector and its timestamp after each training session; a template usage history section, recording all calligraphy templates copied by the user, with each template record including template identifier, number of copies, cumulative duration, writing quality score, tactile guidance parameters, user satisfaction score, etc. The writing quality score is calculated by the deviation between the user's writing trajectory and the standard template trajectory, with the quality score equal to 100 minus the deviation value (mm) multiplied by 10, and the score range is 0 to 100; and a tactile preference configuration section, storing user-defined tactile parameter settings.

[0026] Dynamic file update mechanism: Each time a user completes a writing training session, the system automatically extracts the tactile operation habit features from that training session and compares them with historical features in the file. An exponentially weighted moving average method is used to update the feature vector. The new feature vector is calculated by weighting the features extracted during the current training session with historical features in the file. This update method reflects the latest changes in user habits while maintaining the stability of historical habits. When a significant change in user habits is detected, such as a change in the tactile sensitivity coefficient exceeding a preset threshold, the system records the change event and analyzes the cause. Possible causes include user learning progress, fatigue, and changes in device tightness. Corresponding measures are taken for different causes. For example, a decrease in dependence due to learning progress is normal, and the system automatically adjusts the tactile guidance intensity; a decrease in control ability due to fatigue is temporary, and the system prompts the user to rest.

[0027] Cross-template correlation analysis of user profiles: This analysis examines the differences in users' tactile operating habits across different calligraphy templates. Templates copied by users are clustered according to attributes such as calligraphy style, stroke complexity, and structural features. For each template category, the mean and variance of users' tactile habit characteristics are calculated. Smaller variance indicates stable user habits on that template category, while larger variance indicates fluctuating habits. By comparing the differences in habit characteristics between different template categories, it is possible to identify which template attributes lead to significant changes in users' tactile habits. For example, when switching from regular script to running script, users' writing speed increases, and their preference for tactile timing shifts from alert-based to synchronous. This cross-template habit change pattern is recorded in the user profile for subsequent template switching prediction.

[0028] This step produces a personalized tactile operation profile that includes the user's tactile operation habit characteristics, template usage history, and cross-template habit change patterns, providing a data foundation for subsequent tactile habit modeling and template matching.

[0029] S2, obtain personalized tactile operation profiles, and use graph neural networks and meta-learning methods to construct a user-personalized tactile habit model that can predict users’ tactile preferences on specific templates and achieve rapid adaptation; In response to the differences in tactile operating habits exhibited by users on different calligraphy templates, this step uses deep learning methods to construct a personalized tactile habit model for users. This model can predict users' tactile preferences on specific templates, providing a theoretical basis for the transfer of tactile strategies when switching templates.

[0030] The fundamental reason for the mismatch between tactile habits caused by template switching is that different calligraphy templates have significantly different stroke characteristics. Regular script strokes are regular and have clear turns, running script strokes are smooth and have strong continuity, and cursive script strokes are simplified and fast. These differences lead to different movement patterns when users write. Different templates also have different tactile guidance strategies. Regular script needs to provide clear directional guidance at turning points, running script needs to provide smooth support at connecting strokes, and cursive script needs to provide rhythmic prompts when writing quickly. When users switch from one template to another, the original tactile operation habits do not match the tactile guidance strategy of the new template, causing users to need to readjust and reducing learning efficiency.

[0031] S21, Extract template features and establish a template feature library; Based on the template data in the calligraphy template library, extract multi-dimensional feature vectors that reflect the writing features of the templates and establish a template feature library; The writing characteristics of the template determine the user's movement patterns and tactile needs when copying it. Extracting stroke geometric features: For each stroke in the template, calculate geometric parameters such as stroke length, mean curvature, variance of curvature, number of turning points, and distribution of turning angles. Stroke length reflects the spatial span of pen movement; longer strokes require a greater range of hand movement. The mean curvature reflects the degree of curvature of the stroke; straight strokes have small curvature, while curved strokes have large curvature. The variance of curvature reflects the complexity of the stroke shape; simple strokes have small curvature variations, while complex strokes have large curvature variations. The number of turning points and turning angles reflect the directional changes in the stroke; strokes with many turning points require higher coordination. For the entire character, calculate overall geometric features such as the number of strokes, the connection type between strokes (independent or connected), the character envelope area, and stroke density.

[0032] Kinematic Feature Extraction: Based on demonstration writing data from professional calligraphers, kinematic reference features of the templates are extracted. Demonstration writing data is obtained by inviting professional calligraphers to write templates using the S11-configured smart writing pen. Each template is written multiple times by several calligraphers, and the average value is used as the kinematic reference for that template. Writing speed curves are calculated, analyzing the mean, variance, acceleration, and jerk of the speed. The speed curve reflects the rhythm and fluency of the writing. Pressure variation curves are calculated, analyzing the mean, variance, frequency, and amplitude of pressure changes. The frequency of pressure changes is obtained by extracting the dominant frequency through Fourier transform. The pressure curve reflects the variations in writing pressure and the requirements for pressure control. Posture variation curves are calculated, analyzing the range and rate of change of the penholder's posture angle. Posture changes reflect the need for adjusting the pen grip and pen movement posture. For characters containing multiple strokes, the time interval between strokes is analyzed. The end and beginning of a stroke are determined by pressure sensor data. A stroke is considered to end when the pressure is below a first pressure threshold, and to begin when the pressure is above the first pressure threshold.

[0033] Extracting Tactile Guidance Needs: Based on the geometric and kinematic features of the template, the tactile guidance needs of the template are derived. For templates with dense turning points, the need for directional guidance is high, requiring clear directional cues at turning points; for templates with frequent changes in pressure, the need for pressure cues is high, requiring pressure adjustment cues at pressure change points; for templates with fast writing speeds, the need for rhythm guidance is high, requiring tactile feedback synchronized with the writing rhythm; for templates with continuous strokes, the need for smoothness support is high, requiring continuous and stable tactile support. The needs for directional guidance, pressure cues, rhythm guidance, and smoothness support are quantified into values ​​from 0 to 1, forming the tactile demand vector of the template.

[0034] The extracted geometric, kinematic, and tactile demand features are organized into a template feature vector F_template, which contains three parts: geometric features, kinematic features, and tactile demand features. Feature vectors are extracted from all templates in the template library to build a template feature library. Each template record contains a template identifier Template_ID, template name, calligraphy style category (regular script, running script, cursive script, etc.), feature vector F_template, and tactile demand vector T_demand. The template feature library supports feature-based template retrieval and similarity calculation, uses a vector database for storage, and supports efficient nearest neighbor search.

[0035] S22, Construct a user-template tactile habit association network; Based on the multi-user, multi-template tactile operation data in the user tactile operation archive database constructed in S13, a graph neural network method is used to construct a user-template tactile habit association network, which can learn the complex relationship between user tactile habits and template features; Graph Neural Networks (GNNs) are deep learning models capable of processing graph-structured data, suitable for modeling multivariate relationships between users, templates, and tactile habits. A heterogeneous graph structure is constructed: the graph contains three types of nodes: user nodes, template nodes, and tactile habit nodes. The feature vector of a user node is the user tactile operation habit feature H_user extracted in S12; the feature vector of a template node is the template feature F_template extracted in S21; tactile habit nodes represent specific tactile operation patterns, such as high-intensity directional guidance, medium-intensity force cues, and low-frequency rhythmic guidance. The feature vector of each tactile habit node includes parameters such as tactile type, tactile intensity, tactile frequency, and tactile timing.

[0036] Establish edge connections in the graph: A copying edge is established between user nodes and template nodes, with the edge weight representing the number of times the user copies the template; a higher weight indicates greater familiarity with the template. A preference edge is established between user nodes and tactile habit nodes, with the edge weight representing the user's preference for that tactile habit, calculated using features such as tactile timing preference and tactile sensitivity extracted in S12. A demand edge is established between template nodes and tactile habit nodes, with the edge weight representing the template's demand for that tactile habit, calculated using tactile demand features extracted in S21. A ternary actual usage edge is established between user nodes, template nodes, and tactile habit nodes, representing the actual tactile habit the user adopts when copying a specific template; the edge attributes include the specific values ​​of tactile parameters and the user's satisfaction rating.

[0037] Graph convolutional networks (GCNNs) are used for node representation learning: GCNNs update node representations by aggregating information from neighboring nodes. For user nodes, information from template nodes they have imitated and nodes representing their preferred tactile habits is aggregated to learn a comprehensive tactile preference representation for the user. For template nodes, information from user nodes that have imitated the template and nodes representing tactile habits required by the template is aggregated to learn a comprehensive tactile requirement representation for the template. For tactile habit nodes, information from user nodes that prefer the habit and template nodes that require the habit is aggregated to learn a comprehensive applicability representation for the tactile habit. The update process of the GCNN is as follows: the representation of a node in the next layer is obtained by aggregating the representations of all its neighboring nodes in the current layer, followed by weight matrix transformation and activation function processing. The aggregation function uses weighted summation or attention mechanisms, and the activation function uses ReLU or LeakyReLU.

[0038] The training tasks for the graph neural network are as follows: The primary task is link prediction, which predicts whether a successful copying edge exists between a user node and a template node. Successful copying is defined as the user achieving a predefined standard of writing quality and high user satisfaction after copying the template. The link probability is calculated by the inner product of the user node representation and the template node representation, and then mapped to the zero-to-1 interval using a sigmoid activation function. The auxiliary task is tactile habit recommendation, which predicts the optimal tactile habit a user should adopt when copying a specific template. The recommendation score is calculated through a ternary interaction of the user node representation, the template node representation, and the tactile habit node representation. The interaction process is implemented through a learnable interaction weight matrix.

[0039] Contrastive learning is employed to enhance the model's generalization ability: Contrastive learning learns more robust representations by bringing similar samples closer together and pushing dissimilar samples further apart. Positive sample pairs are constructed: tactile habits of the same user on similar templates constitute positive sample pairs, and similar templates are determined by calculating cosine similarity using template features from S21. Negative sample pairs are constructed: tactile habits of the same user on significantly different templates constitute negative sample pairs, or tactile habits of different users on the same template constitute negative sample pairs. The loss function for contrastive learning is obtained by comparing the similarity of positive sample pairs with the similarity of all negative sample pairs. The similarity function uses cosine similarity, and a temperature parameter controls the smoothness of the distribution.

[0040] The model training employs a multi-task learning framework. The total loss function is obtained by weighted summation of link prediction loss, haptic habit recommendation loss, and contrastive learning loss, with weight coefficients determined through validation set tuning. Training data comes from a user haptic operation profile database built by S13, containing the haptic operation history of multiple users on multiple templates. The Adam optimizer is used, and hyperparameters such as learning rate, batch size, and number of training epochs are determined through validation set tuning. An early stopping mechanism is employed to prevent overfitting.

[0041] After training, the graph neural network learns the deep relationship between users' tactile habits and template features, and can predict users' tactile preferences on new templates, providing a basis for tactile strategy transfer when switching templates.

[0042] S23, train the meta-learning model to achieve rapid adaptation; based on the user-template tactile habit association network built in S22, the meta-learning method is used to train the rapid adaptation model, so that the model can quickly adjust the tactile strategy with only a small amount of data when the user switches to a new template; Meta-learning is a method of learning how to learn. It involves training on multiple related tasks to learn initialization parameters that allow the model to quickly adapt to new tasks. In this invention, each task corresponds to the user's tactile habits learned on a specific template. The goal of meta-learning is to learn an initial model state that enables the model to quickly converge to the optimal tactile strategy when faced with a new template.

[0043] The Model-Independent Meta-Learning (MAML) algorithm is employed: MAML uses second-order optimization to find a good model initialization parameter θ, such that starting from this parameter, it can adapt to new tasks with a small number of gradient updates. The algorithm flow is as follows: Task Sampling: A batch of tasks are sampled from the user's template usage history. Each task T_i corresponds to the user's tactile habit learning on template M_i. Each task contains a support set D_support and a query set D_query. The support set contains the user's previous writing data on the template for rapid adaptation; the query set contains subsequent writing data for evaluating the adaptation effect.

[0044] Inner loop adaptation: For each task, starting from the meta-parameters, several steps of gradient descent are performed on the support set to obtain task-specific parameters. The task-specific parameters are obtained by subtracting the learning rate multiplied by the gradient of the loss function with respect to the meta-parameters from the meta-parameters. The inner loop learning rate is an adjustable parameter, and the loss function calculates the mean square error between the tactile parameters predicted by the model and the tactile parameters actually preferred by the user.

[0045] Outer Loop Update: Evaluate the performance of the adapted model on the query set and calculate the meta-loss, which is the sum of the losses for all tasks on the query set. Update the meta-parameters using gradients; the new meta-parameters are obtained by subtracting the gradient of the outer loop learning rate multiplied by the meta-loss with respect to the meta-parameters from the current meta-parameters. This update process requires calculating the second derivative, meaning the gradient of the loss with respect to the meta-parameters includes the gradients of task-specific parameters with respect to the meta-parameters, resulting in high computational complexity. To reduce computational cost, a first-order approximate meta-learning algorithm is used, ignoring the second-order terms.

[0046] Meta-learning training data construction: The user's template switching history is extracted from the user's tactile operation profile database, and each template switch constitutes a meta-learning task. For example, when a user switches from template A to template B, the first three writing data on template B serve as the support set, and the last five writing data serve as the query set. To increase the diversity of the training data, data augmentation techniques are used to perform minor translation, rotation, and scaling transformations on the writing trajectory to generate new training samples.

[0047] Introducing task similarity weighting: The difficulty of switching tasks varies across different templates; switching from similar templates is easier to adapt to than switching from significantly different templates. The similarity between tasks is calculated based on the cosine similarity of the feature vectors of the source and target templates. Similarity weighting is introduced into the meta-loss calculation, where the loss for each task is multiplied by a corresponding weight coefficient. This weight coefficient is inversely proportional to the average similarity of the tasks; tasks with lower similarity have higher weights, encouraging the model to learn to handle template switching with significant differences.

[0048] Model Architecture Design: The meta-learning model employs a multilayer perceptron structure. The input is a concatenated vector of user tactile habit features and target template features. The output is a predicted tactile parameter vector, including parameters such as tactile intensity, tactile frequency, and tactile timing. The network contains multiple hidden layers, using ReLU activation, and the output layer employs sigmoid activation to ensure the output is within the range of zero to 1.

[0049] Meta-learning training settings: Hyperparameters such as inner loop learning rate, inner loop update steps, outer loop learning rate, batch size, and number of training epochs are determined through validation set tuning. A validation set is used to monitor meta-learning performance; the validation set includes user switching tasks on the new template. Training stops when validation set performance no longer improves.

[0050] After training with meta-learning, the model obtains a well-defined initialization parameter θ_meta. When the user switches to a new template, the system starts from this initialization parameter and uses the user's previous writing data on the new template to quickly adapt. Only a small number of gradient updates are needed to converge to a haptic strategy suitable for the template, reducing the user's readaptation cost. The output of the meta-learning model is the adapted haptic parameter vector, which serves as an auxiliary input to the S3 transfer learning step. This vector is combined with the prediction results of the domain adaptation network to jointly determine the final haptic strategy.

[0051] This step yields a personalized tactile habit model, which includes a user-template tactile habit association network and a meta-learning fast adaptation model. This model predicts the user's tactile preferences on new templates and achieves rapid adaptation, providing a theoretical basis for subsequent template matching and tactile policy transfer. The model's output includes the user's predicted tactile preference vector on the target template and initialization parameters for rapid adaptation. These outputs are used in S3 to generate tactile policy transfer schemes.

[0052] S3, based on the user's personalized tactile habit model, uses transfer learning and domain adaptation methods to generate a tactile strategy intelligent matching scheme that includes template similarity calculation, selective transfer mechanism and progressive transition strategy; To address the issue of mismatched tactile operating habits when users switch calligraphy templates, this step employs a transfer learning approach. This method transfers the user's tactile habits from the source template to the target template, generating an adapted tactile guidance strategy and reducing the user's readjustment cost.

[0053] The core idea of ​​transfer learning is to use knowledge from the source domain to assist learning in the target domain. In this invention, the source domain is a calligraphy template that the user is already familiar with and its corresponding tactile strategy, while the target domain is the new template that the user is about to switch to. Transfer learning needs to solve two key problems: first, identifying the similarities and differences between the source and target templates to determine which tactile habits can be transferred and which need to be adjusted; second, designing transfer strategies that adapt to the feature requirements of the new template while maintaining the user's familiar tactile patterns.

[0054] S31, calculate the tactile requirement similarity between templates; based on the template features extracted in S21, calculate the tactile requirement similarity between the source template and the target template, and identify the transferable tactile habit dimensions and the dimensions that need to be adjusted; The similarity of tactile needs between templates reflects the degree of closeness in the tactile guidance needs of the two templates. High similarity means that users need similar types of tactile guidance on the two templates and tactile habits are easy to transfer; low similarity means that the tactile needs are very different and require a large strategy adjustment.

[0055] Obtain source and target template features: From the template feature library established in S21, retrieve the corresponding template feature vectors F_source and F_target, as well as haptic demand vectors T_source and T_target, based on the source template identifier and target template identifier. The haptic demand vector contains demand values ​​in four dimensions: directional guidance demand, force cues demand, rhythm guidance demand, and smoothness support demand, with each demand value ranging from zero to 1.

[0056] Calculate geometric feature similarity: Extract geometric feature sub-vectors from the source and target templates, and calculate the geometric similarity S_geo using cosine similarity. High geometric similarity indicates that the stroke shape features of the two templates are similar, and the user's motion patterns are similar.

[0057] Calculating kinematic feature similarity: Kinematic feature sub-vectors are extracted from the source and target templates. Kinematic features include writing speed distribution, force variation patterns, and posture variation range. Since kinematic features contain time-series data, Dynamic Time Warping (DTW) distance is used to calculate kinematic similarity. The DTW algorithm finds the optimal alignment path between two time series and calculates the cumulative distance after alignment, handling cases with different time series lengths and time distortion. The DTW distance D_DTW is normalized and converted into a similarity S_motion. The conversion uses an exponential decay function, with a scale parameter controlling the decay rate of the similarity. High kinematic similarity indicates that the writing rhythm and movement patterns of the two templates are similar.

[0058] Calculate the similarity of tactile requirement features: Extract the tactile requirement vectors Tsource and Ttarget from the source and target templates. Tactile requirements include four dimensions: directional guidance requirements, force cues, rhythm guidance requirements, and fluency support requirements. Calculate the similarity for each dimension i. , among them The kernel function bandwidth parameter is used to calculate a weighted average of the similarity across four dimensions. The weight of each dimension is determined based on its influence on the haptic strategy.

[0059] The overall tactile demand similarity is calculated as follows: The overall similarity S_total is obtained by weighted summation of geometric similarity S_geo, kinematic similarity S_motion, and tactile demand similarity S_tactile, with tactile demand features having the highest weight. The overall similarity S_total ranges from zero to 1, with a larger value indicating greater similarity between the two templates.

[0060] Identifying transferable and adjustable dimensions: For the four dimensions of tactile requirements, calculate the difference Δ_i between the source and target template requirements. The difference is the absolute value of the difference between the target template requirement value and the source template requirement value. Set a difference threshold δ. When the difference is less than the threshold, the dimension is considered transferable; when the difference is greater than or equal to the threshold, the dimension is considered to need adjustment. The threshold δ is set according to the user's adaptability; a larger threshold is set for users with strong adaptability, and a smaller threshold is set for users with weak adaptability. The determination results for the four dimensions are recorded as a transfer marker vector M. The i-th element M_i equal to 1 indicates that the dimension is transferable, and M_i equal to zero indicates that the dimension needs adjustment.

[0061] S32, Generate a haptic strategy transfer scheme based on domain adaptation; Based on the template similarity and transferable dimension recognition results calculated in S31, the domain adaptation method is used to generate a haptic strategy transfer scheme from the source template to the target template. Domain adaptation is a method of transfer learning that transfers knowledge from the source domain to the target domain by learning the mapping relationship between the source and target domains. In this invention, it is necessary to learn a mapping function from the source template haptic policy to the target template haptic policy.

[0062] Constructing a feature alignment network: The goal of feature alignment is to make the distributions of the source and target templates in the feature space similar, so that the haptic policy learned on the source template can be applied to the target template. An Adversarial Domain Adaptation method is employed, and the network consists of three components: a feature extractor G, a haptic policy predictor P, and a domain discriminator D.

[0063] The feature extractor maps template features to a shared feature space. The feature extractor employs a multilayer perceptron structure containing multiple hidden layers, and uses LeakyReLU as the activation function.

[0064] The tactile policy predictor predicts tactile parameters based on shared features, including tactile intensity, tactile frequency, and tactile timing. The predictor employs a multilayer perceptron structure, with the output layer using sigmoid activation to ensure the output is within the range of zero to 1.

[0065] The domain discriminator determines whether the shared features originate from the source domain or the target domain, and outputs a binary classification probability. The discriminator employs a multilayer perceptron structure, with the output layer using sigmoid activation.

[0066] Adversarial training process: The goal of feature extractor G is to extract domain-invariant features so that discriminator D cannot distinguish which domain the features come from; the goal of discriminator D is to accurately distinguish between source and target domain features. The two are trained adversarially, similar to the training process of a generative adversarial network.

[0067] The loss function is defined as follows: the total loss comprises three terms, a weighted sum of task loss, adversarial loss, and regularization loss. Task loss measures the accuracy of haptic policy prediction and is calculated on the source domain, since the source domain has labeled data (the user's actual haptic preferences, obtained from the user's haptic operation profile in S13, showing historical haptic parameters used by the user on the source template). Task loss is calculated using the mean squared error between the predicted and true values. Adversarial loss prompts the feature extractor to generate domain-invariant features, measured by the degree to which the discriminator misclassifies the target domain features as originating from the source domain. Regularization loss prevents overfitting and is calculated using the sum of squared parameters of the feature extractor and predictor. The weights for adversarial loss and regularization loss are determined through validation set tuning.

[0068] Training strategy: Alternating optimization is employed, with each iteration consisting of two steps. The first step fixes the discriminator, updates the feature extractor and predictor, and minimizes the total loss. The second step fixes the feature extractor and predictor, updates the discriminator, and maximizes the discrimination accuracy. The Adam optimizer is used, and hyperparameters such as learning rate, batch size, and number of training epochs are determined through validation set tuning.

[0069] A selective transfer mechanism is introduced: For dimensions identified as transferable by S31, the tactile strategy of the source template is directly applied; for dimensions requiring adjustment, a strategy predicted by the domain adaptation network is adopted. Specifically, a transfer weight vector M is defined, containing four elements corresponding to the four tactile demand dimensions. M is the transfer label vector calculated by S31. When a dimension is transferable, M_i equals 1; when a dimension needs adjustment, M_i equals zero. The source template tactile strategy vector P_source is obtained. This vector is extracted from the user's tactile operation file in S13 and contains the tactile parameters historically used by the user on the source template, including tactile intensity, tactile frequency, and tactile timing parameters for the four dimensions. The target template tactile strategy vector P_adapt predicted by the domain adaptation network is obtained. This vector is calculated by the feature extractor G and the tactile strategy predictor P based on the target template features. The final tactile strategy vector P_target of the target template is calculated through selective transfer. For each dimension i, P_target_i equals M_i multiplied by P_source_i plus 1 minus M_i multiplied by P_adapt_i. This selective transfer mechanism preserves familiar tactile patterns while adjusting for differences, reducing the user's cognitive load. The results of selective transfer serve as the target policy input for the S33 progressive transition strategy.

[0070] S33, Design a progressive haptic transition strategy; Based on the haptic strategy migration scheme generated in S32, design a progressive haptic transition strategy to enable users to smoothly transition from the haptic mode of the source template to the haptic mode of the target template. A sudden switch in haptic strategy can increase the cognitive burden and cause operational confusion for users. A gradual transition strategy adjusts haptic parameters in stages, allowing users to gradually adapt to the haptic guidance of the new template.

[0071] Define the transition phases: The adaptation process after template switching is divided into three phases. Initial Phase: The tactile strategy primarily uses the source template strategy, supplemented by the target template strategy, allowing users to begin imitating the new template within a familiar tactile environment, reducing initial difficulty. Transition Phase: Gradually increase the proportion of the target template strategy and decrease the proportion of the source template strategy, guiding users to adapt to the tactile requirements of the new template. Stable Phase: Fully adopt the target template strategy; users have largely adapted to the tactile guidance of the new template.

[0072] Design the time evolution function of tactile parameters: For each tactile parameter (such as tactile intensity, tactile frequency, etc.), define its evolution law in the transition process. The current tactile parameter value is obtained by linear interpolation of the source template policy parameter value and the target template policy parameter value according to the transition weight. The transition weight changes with the number of writings to control the transition speed from the source policy to the target policy.

[0073] Design the transition weight function: An S-shaped curve function is adopted, where the transition weight Wtransition increases from zero to 1 according to the number of writing iterations t, following an S-shaped curve. The function form is as follows: Where 'a' is the velocity parameter and 'b' is the midpoint parameter. The transition parameters are adjusted based on the template similarity S_total. The velocity parameter 'a' is directly proportional to the overall similarity S_total, while the midpoint parameter 'b' is inversely proportional to the overall similarity S_total. A higher similarity results in a faster transition, while a lower similarity results in a slower transition.

[0074] Introducing user adaptive feedback adjustment: During the transition process, the user's writing quality is monitored in real time. The user's writing deviation D_current on the target template is calculated, and the deviation D_first from the first writing is recorded as a baseline. An adaptive index A_adapt is defined as 1 minus D_current divided by D_first; a larger adaptive index indicates better adaptation. When A_adapt is less than the first adaptive threshold, adaptation is considered difficult, and the transition speed is slowed down; when A_adapt is greater than the first adaptive threshold but less than the second adaptive threshold, adaptation is considered normal, and the current transition speed is maintained; when A_adapt is greater than the second adaptive threshold, adaptation is considered good, and the transition speed is accelerated. Parameter adjustments are performed after each writing session.

[0075] Design a multi-dimensional transition strategy: For dimensions identified as transferable by S31, the transition speed can be faster; for dimensions requiring adjustment, the transition speed should be slower. Set an independent transition weight function w_transition_i for each dimension i. Obtain the transition marker vector M from S31. When M_i equals 1, the velocity parameter of that dimension is set larger, and the midpoint parameter is set smaller; when M_i equals zero, the velocity parameter of that dimension is set smaller, and the midpoint parameter is set larger. The current tactile parameter P_i of each dimension is calculated by interpolation using the transition weight w_transition_i for that dimension. The calculation formula is: P_i = 1 - w_transition_i multiplied by the source template parameter P_source_i + w_transition_i multiplied by the target template parameter P_target_i.

[0076] A transition effect evaluation mechanism was established: After the transition phase, the user's adaptation to the target template's tactile strategy was assessed. The user's writing quality, tactile satisfaction, and learning curve slope on the target template were calculated and compared with a control group that switched directly (without a transition strategy). When the gradual transition strategy significantly outperformed the direct switch, the effectiveness of the method was verified; when the effect was not significant, the reasons were analyzed, which may have been due to improper transition parameter settings or significant individual user differences, leading to targeted adjustments to the transition strategy.

[0077] After this step, a tactile strategy intelligent matching scheme based on transfer learning is obtained. This scheme includes template similarity calculation, domain adaptation network, selective transfer mechanism and progressive transition strategy. When users switch templates, the tactile habits of the source template can be smoothly transferred to the target template, which significantly reduces the user's readjustment cost and improves the efficiency of calligraphy learning.

[0078] S4 receives the haptic strategy intelligent matching scheme and real-time writing data, adopts progressive haptic feedback control and online learning methods, outputs continuously optimized haptic guidance strategies and records template switching effect data; Based on the haptic strategy transfer scheme generated by S3 and the real-time writing data continuously collected by S1, progressive haptic feedback is executed during the user's copying of the target template, and the haptic strategy is continuously optimized according to the user's actual performance. S41 implements progressive haptic feedback control; based on the progressive transition strategy designed in S33, the haptic feedback parameters are dynamically adjusted during the user's imitation of the target template to achieve a smooth transition from the haptic mode of the source template to the haptic mode of the target template. Establish a haptic feedback control loop: The control loop executes at a preset frequency. In each control cycle, the system performs the following steps: Read the current writing status: Obtain real-time data such as the user's current pen tip position, writing speed, force applied, and pen grip posture from the S1 data acquisition system.

[0079] Calculate writing deviation: Compare the real-time writing trajectory with the standard trajectory of the target template. The standard trajectory is obtained from the template feature library established in S21. The time series is aligned using a dynamic time warping algorithm to calculate positional deviation, directional deviation, and force deviation. The deviation calculation method is the same as that of traditional haptic guidance systems, but the interpretation and response strategy of the deviation need to consider that the user is in the adaptation period of template switching.

[0080] Determine the current transition phase: Based on the number of times the user writes on the target template (t), determine whether the current stage is the initial stage, the transition stage, or the stable stage. The haptic feedback strategies differ for each stage.

[0081] Calculate the current tactile parameters: Based on the transition weight function in S33, calculate the tactile parameters at the current moment. Obtain the transition weight w_transition corresponding to the current writing count t from S33. The transition weight is calculated using an S-curve function. For the case of a uniform transition strategy, the current parameter value P_current is obtained by linear interpolation of the source template parameter value P_source and the target template parameter value P_target according to the transition weight. The calculation formula is P_current = (1 - w_transition) x P_source + w_transition x P_target. For the case of a multi-dimensional transition strategy, the current parameter value P_current_i of each dimension i is calculated by interpolation using the transition weight w_transition_i of that dimension. For each parameter, such as tactile intensity, tactile frequency, and tactile timing, interpolation is performed separately. The tactile intensity parameter controls the activation amplitude of the tactile actuator, with a value ranging from zero to 1; the tactile frequency parameter controls the activation frequency of tactile feedback; and the tactile timing parameter controls the time offset of tactile feedback relative to the writing action.

[0082] Generating tactile feedback signals: Based on the calculated tactile parameters and the current writing deviation, a control signal is generated and sent to the tactile actuator. The generation of the control signal follows these rules: when the deviation is small, the tactile intensity is set to the currently calculated parameter value; when the deviation is large, the tactile intensity is appropriately increased, and the increase coefficient is determined according to the magnitude of the deviation.

[0083] Executing haptic feedback: Control signals are sent to the haptic feedback device via a low-latency communication interface, and the haptic actuator generates corresponding haptic stimuli based on the signal parameters. The execution of each haptic feedback event is recorded, including the trigger time, haptic parameters, writing deviation, and user response. This data is used for subsequent strategy optimization.

[0084] In the initial stage, since users are just beginning to imitate the new template and are not yet familiar with its stroke characteristics and writing rhythm, the tactile feedback is mainly based on the familiar patterns of the source template, providing users with stable tactile support and reducing cognitive burden. At the same time, the system will provide tactile cues unique to the target template at key locations (such as stroke starting points and turning points) to allow users to gradually perceive the characteristics of the new template.

[0085] During the transition phase, tactile feedback gradually shifts from the source template mode to the target template mode. Users begin to adapt to the tactile guidance of the new template, and the system dynamically adjusts the transition speed based on the user's adaptive feedback. When users exhibit difficulties adapting (increased writing deviation, decreased tactile satisfaction), the system slows down the transition speed and prolongs the retention time of the familiar tactile mode; when users adapt well, the system accelerates the transition speed, switching to the target template mode more quickly.

[0086] During the stabilization phase, haptic feedback was entirely based on the target template strategy, and users had largely adapted to the haptic guidance of the new template. The system continued to monitor user writing quality and satisfaction to ensure the effectiveness of the haptic strategy.

[0087] S42 employs online learning to optimize tactile strategies; based on the user's real-time writing performance on the target template, online learning methods are used to continuously optimize the tactile strategies, making tactile guidance more aligned with the user's actual needs; The difference between online and offline learning lies in the fact that online learning updates the model instantly as the data stream arrives, without waiting for the complete dataset. In template switching scenarios, the system immediately adjusts the haptic strategy based on the effect of each writing action completed by the user, enabling rapid iterative optimization.

[0088] Define the comprehensive loss function: The loss function is obtained by weighted summation of writing quality loss L_quality and user comfort loss L_comfort. Writing quality loss is calculated using the root mean square error of writing deviation; user comfort loss is calculated by combining subjective ratings and physiological signals (electromyographic intensity, skin conductance change rate). The total loss function L_total is obtained by weighted summation of the two, with writing quality loss having a greater weight than user comfort loss to ensure priority is given to writing quality.

[0089] Online updates are employed using gradient descent: after each writing operation, the gradient of the loss function with respect to the haptic parameters is calculated. Gradient calculation uses numerical differentiation. The update formula is: new parameter values ​​equal old parameter values ​​minus the learning rate multiplied by the gradient. To prevent system instability, an upper limit is set on parameter variation.

[0090] A forgetting mechanism is introduced: an exponentially weighted moving average method is used to smooth parameter updates. The smoothed parameter value θ_smooth is obtained by multiplying the current updated parameter value by the forgetting factor, adding 1 minus the forgetting factor multiplied by the parameter value of the previous smoothing, and ensuring that the forgetting factor is less than 0.5 so that the model pays more attention to recent data.

[0091] Set an adaptive update frequency: Dynamically adjust the parameter update frequency based on the user's adaptation stage and the trend of handwriting quality changes. In the initial stage, when the user is unfamiliar with the new template, parameters are updated after each writing session to quickly adjust the tactile strategy. In the transition stage, as the user gradually adapts, the update frequency is reduced to avoid frequent adjustments affecting user adaptation. In the stable stage, when the user has basically adapted, the update frequency is further reduced, with only minor adjustments to maintain strategy stability. Simultaneously, when a continuous decline in handwriting quality is detected, regardless of the stage, a parameter update is immediately triggered to promptly correct tactile strategy deviations.

[0092] Establish a parameter update recording mechanism: The system records detailed information for each parameter update, including the update time, parameter value before update, parameter value after update, loss function value, gradient value, etc. These records are used to analyze the parameter optimization trajectory, identify parameter convergence trends, and provide reference data for subsequent template switching by users.

[0093] S43, Establish a template switching effect evaluation mechanism; Based on the user's writing data before and after template switching, evaluate the effect of haptic strategy transfer and gradual transition, and verify the effectiveness of the method; The evaluation index system is defined as follows: Three categories of evaluation indicators are established. The first category is the adaptation time indicator, which refers to the number of writing attempts required for a user to reach a stable writing quality level after switching templates. A stable level is defined as consistently achieving a quality score above a predetermined percentage of the target template's average quality across multiple writing attempts. The second category is the writing quality indicator, where the overall deviation is obtained by taking the square root of the weighted sum of the squares of three components: positional deviation, directional deviation, and pressure deviation. Positional deviation is pre-weighted with the highest weight. The third category is the user satisfaction indicator, measured through a questionnaire with multiple questions, each using a five-point rating scale.

[0094] A controlled experiment was set up: Users were randomly divided into an experimental group and a control group. The experimental group used the tactile strategy transfer and progressive transition method of this invention, while the control group used the direct switching method, i.e., no transfer learning, directly using the standard tactile strategy of the target template, which was derived from the tactile demand vector in the template feature library. The adaptation time, writing quality, and user satisfaction of the two groups were compared. The experimental design used a paired design, with each user participating in both the experimental and control groups, alternating between the two methods in different template switching tasks to eliminate the influence of individual differences. The experimental sample size was set at thirty users, with each user completing ten template switching tasks, five of which used the experimental group method and five of which used the control group method, for a total of three hundred template switching samples collected.

[0095] Statistical analysis methods: Paired t-tests were used to compare the differences between the experimental and control groups. For each evaluation indicator, the average score of each user in the experimental and control groups was calculated, and then the paired difference was calculated. A t-test was performed on the paired difference to test the null hypothesis that there was no difference between the two groups, and the alternative hypothesis that the experimental group was better than the control group. When the p-value was less than the preset threshold, the difference was considered statistically significant, the null hypothesis was rejected, and the alternative hypothesis was accepted. Simultaneously, the effect size was calculated to measure the practical significance of the difference; the effect size was equal to the difference between the two group means divided by the pooled standard deviation.

[0096] Analysis of experimental results: The experimental results show that the user adaptation time of the method of the present invention is shorter than that of the control group, the writing quality during the adaptation period is better than that of the control group, and the user satisfaction is higher than that of the control group.

[0097] This step completes the execution and continuous optimization of real-time haptic feedback. Through gradual transition, online learning, and effect evaluation, it enables users to smoothly switch from the source template to the target template, significantly reducing the readjustment cost.

[0098] S5: Obtain template switching effect data and user tactile operation profiles, and use adaptive analysis and template learning graph construction methods to generate an optimization plan that includes personalized template recommendation strategies and learning suggestion reports; Based on the template switching data accumulated from S1 to S4, we evaluate the user's adaptation effect, analyze the key factors affecting adaptation, generate personalized optimization suggestions, and provide improvement directions for subsequent template switching.

[0099] S51, Analyze users' cross-template adaptability; Based on users' performance during multiple template switching, analyze users' cross-template adaptability and identify users' adaptation patterns; Extracting the adaptation curve: For each template switch, plot the curve of how the user's handwriting quality changes with the number of writing attempts. Handwriting quality is represented by the quality score defined in S43, with the horizontal axis representing the number of writing attempts and the vertical axis representing the quality score. The slope of the curve reflects the adaptation speed, obtained by fitting the quality data of the first five writing attempts using linear regression; a larger slope indicates faster adaptation. The convergence value of the curve reflects the final mastery level, taken as the average of the last three writing attempts; a higher convergence value indicates better final mastery. By comparing the adaptation curves for different template switches, we can identify the user's adaptation pattern. Some users adapt quickly but with large fluctuations, characterized by a large slope but a large quality standard deviation; others adapt slowly but steadily, characterized by a small slope but a small quality standard deviation.

[0100] Calculate the adaptability index: Define the adaptation speed index V_adapt as 1 divided by the number of writing operations N_stable required to achieve stable writing quality; define the adaptation stability index S_adapt as 1 divided by the standard deviation of writing quality during the adaptation period σ_quality plus 1; the comprehensive adaptability index C_adapt is calculated by weighting the adaptation speed index and the adaptation stability index, with the adaptation speed weight being greater than the adaptation stability weight. The higher the comprehensive adaptability index, the stronger the user's cross-template adaptability.

[0101] Identifying influencing factors: A multiple linear regression analysis was used to establish an adaptation time prediction model. The independent variables included template similarity S_total, user proficiency (quantified by cumulative usage time), and haptic strategy differences (quantified by the Euclidean distance between the source strategy and the target strategy). The dependent variable was adaptation time N_stable.

[0102] S52, Generate personalized template recommendation strategies; Based on user adaptability analysis, generate personalized template recommendation strategies and optimize template learning paths; Constructing a template learning graph: A template learning graph G_learning is constructed, with all templates as nodes and the similarity between templates as edge weights. The node set V in the graph contains all templates in the template library, with N_template nodes; the edge set E contains connections between all template pairs, and the weight w_ij of edge (i, j) is the tactile similarity S_total between template i and template j, calculated using method S31. A graph traversal algorithm is used to plan the optimal template learning path for the user, prioritizing templates with high similarity to the current template to reduce switching difficulty. Specifically, a greedy algorithm is used, starting from the template the user is currently copying, and each time selecting the template with the highest similarity to the current template that the user has not yet copied as the next recommended template, repeating this process until all templates are recommended. The recommended path is recorded as a template sequence R equal to template1, template2, ..., templateN_template, where template1 is the user's current template, and templatei + 1 is the uncopyable template with the highest similarity to templatei.

[0103] Dynamically adjust the recommendation strategy: Obtain the user's comprehensive adaptability index C_adapt from S51. When C_adapt is greater than the first ability threshold, the user is judged to have strong adaptability, and the template with the highest similarity is selected to increase the learning challenge. When C_adapt is less than the second ability threshold, the user is judged to have difficulty adapting, and the template with the highest similarity is selected, and the similarity must be greater than the preset similarity threshold. When C_adapt is greater than the second ability threshold but less than the first ability threshold, the user is judged to have moderate adaptability, and templates are recommended according to the standard similarity order.

[0104] Generate learning suggestion reports: Provide users with personalized learning suggestions, including suggested template switching frequency, suggested practice counts, and key tactile dimensions to focus on. For users who adapt quickly, it's recommended to switch to each template after practicing it a few times; for users who adapt slowly, it's recommended to switch to each template after practicing it more times. Suggestions are presented in a visual report format, including a learning path map, an adaptability radar chart, and bar charts for performance across various dimensions.

[0105] This step completes the evaluation of the template switching adaptation effect and generates optimization suggestions, providing users with personalized learning path planning and improving the overall efficiency of calligraphy learning.

[0106] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A tactile-guided intelligent matching method for calligraphy templates, characterized in that, Includes the following steps: S1 collects user writing action data and tactile interaction response data, and uses multimodal sensor synchronous acquisition and feature extraction methods to construct a personalized tactile operation profile including tactile operation habit characteristics, template usage history and cross-template habit change patterns; S2, obtain personalized tactile operation profiles, and use graph neural networks and meta-learning methods to construct a user-personalized tactile habit model that can predict users’ tactile preferences on specific templates and achieve rapid adaptation; S3, based on the user's personalized tactile habit model, uses transfer learning and domain adaptation methods to generate a tactile strategy intelligent matching scheme that includes template similarity calculation, selective transfer mechanism and progressive transition strategy; S4 receives the haptic strategy intelligent matching scheme and real-time writing data, adopts progressive haptic feedback control and online learning methods, outputs continuously optimized haptic guidance strategies and records template switching effect data; S5 acquires template switching effect data and user tactile operation profiles, and uses adaptive analysis and template learning graph construction methods to generate an optimization plan that includes personalized template recommendation strategies and learning suggestion reports.

2. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S1 includes: A system for synchronously acquiring writing motion and tactile response is configured, based on a hardware platform of a smart writing pen and a haptic feedback glove, to collect writing motion data and tactile interaction response data; the smart writing pen integrates a six-axis inertial measurement unit to acquire the three-dimensional spatial trajectory, movement speed and posture angle of the pen tip, and a pressure sensor is installed on the pen tip to measure writing pressure, while a capacitive touch sensor array is embedded in the pen grip area to detect pen grip posture and grip force distribution; the haptic feedback glove has haptic actuators and response sensors installed in the fingers, palm and wrist respectively, with a pressure sensor and electromyography sensor configured next to each actuator; Based on the collected multimodal data, feature vectors of user tactile operation habits are extracted. The feature vectors include tactile sensitivity features, tactile dependence features, tactile timing preference features, and motor control ability features. A personalized tactile operation profile is created for each user, which includes habit characteristics, template usage history, and tactile preference configuration. The feature vector is dynamically updated using an exponentially weighted moving average method, and cross-template correlation analysis is performed on the differences in users' tactile operation habits on different calligraphy templates.

3. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S2 includes: Extract template features and build a template feature library. Extract stroke geometric features, kinematic features and tactile guidance demand features for each template in the calligraphy template library, and organize them into a template feature vector. The tactile guidance demand features include directional guidance demand, force prompt demand, rhythm guidance demand and fluency support demand, which are quantified into zero to one values ​​to form a tactile demand vector. A user-template tactile habit association network is constructed. Based on multi-user, multi-template tactile operation data in personalized tactile operation profiles, a heterogeneous graph structure containing user nodes, template nodes, and tactile habit nodes is constructed using a graph neural network method. Imitation edges are established between user nodes and template nodes, preference edges are established between user nodes and tactile habit nodes, and demand edges are established between template nodes and tactile habit nodes. Graph convolutional networks are used for node representation learning. The training tasks include the main task of link prediction and the auxiliary task of tactile habit recommendation.

4. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S2 further includes: To achieve rapid adaptation, a meta-learning model is trained based on a user-template tactile habit association network and a model-independent meta-learning algorithm is used to train the rapid adaptation model. Tasks are sampled from the user's template usage history. Each task contains a support set and a query set. The support set contains the user's previous writing data on the template, and the query set contains subsequent writing data. In the inner loop, gradient descent is performed on the support set to obtain task-specific parameters starting from the meta-parameters; in the outer loop, the performance of the adapted model is evaluated on the query set and the meta-parameters are updated; task similarity weighting is introduced in the meta-loss calculation, and the weight coefficient is inversely proportional to the cosine similarity of the feature vectors of the source template and the target template; the output of the meta-learning model is the adapted tactile parameter vector.

5. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S3 includes: To calculate the tactile demand similarity between templates, the template feature vectors and tactile demand vectors of the source and target templates are retrieved from the template feature library. Geometric feature similarity, kinematic feature similarity, and tactile demand feature similarity are calculated separately. The three are then weighted and summed to obtain the overall similarity. The kinematic feature similarity is obtained by transforming the dynamic time warped distance using an exponential decay function. The tactile demand feature similarity is obtained by calculating the weighted average of each dimension using a Gaussian kernel function. The difference between the source template and the target template is calculated for the four dimensions of tactile needs. When the difference is less than the preset difference threshold, the corresponding dimension is determined to be transferable. When the difference is greater than or equal to the preset difference threshold, the corresponding dimension is determined to be adjusted, and a transfer marker vector is generated. A domain-adaptive-based haptic policy transfer scheme is generated. An adversarial domain adaptation method is used to construct a network containing a feature extractor, a haptic policy predictor, and a domain discriminator. A selective transfer mechanism is introduced, which directly applies the haptic policy of the source template for transferable dimensions, and adopts the policy predicted by the domain adaptation network for dimensions that need to be adjusted.

6. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S3 further includes: A progressive tactile transition strategy is designed, dividing the adaptation process after template switching into an initial stage, a transition stage, and a stable stage. In the initial stage, the tactile strategy mainly uses the source template strategy, in the transition stage the proportion of the target template strategy is gradually increased, and in the stable stage the target template strategy is used entirely. The transition weight uses an S-shaped curve function that increases from zero to one with the number of writings. The speed parameter and midpoint parameter are adjusted according to the overall similarity. During the transition, the adaptability index is calculated. When the adaptability index is less than the first adaptability threshold, the transition speed is slowed down. When the adaptability index is greater than the first adaptability threshold but less than the second adaptability threshold, the current transition speed is maintained. When the adaptability index is greater than the second adaptability threshold, the transition speed is accelerated. Set an independent transition weight function for each dimension. Set a faster transition speed for dimensions marked as transferable in the migration marker vector, and a slower transition speed for dimensions marked as needing adjustment.

7. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S4 includes: Implement progressive tactile feedback control, establish a tactile feedback control loop, read the current writing state in each control cycle, align the real-time writing trajectory with the standard trajectory of the target template using a dynamic time warping algorithm, and calculate the positional deviation, directional deviation, and force deviation; calculate the current tactile parameters according to the transition weight function of the progressive transition strategy, and obtain the current parameter value by linearly interpolating the source template parameter value and the target template parameter value according to the transition weight. A tactile feedback signal is generated based on the current tactile parameters and writing deviation. When the deviation is small, the tactile intensity is set to the currently calculated parameter value. When the deviation is large, the tactile intensity is appropriately increased. The control signal is sent to the tactile feedback device for execution through a low-latency communication interface, and the execution status of each tactile feedback is recorded.

8. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S4 further includes: An online learning strategy is adopted to optimize haptic feedback. A comprehensive loss function is defined, which is a weighted sum of writing quality loss and user comfort loss, with the weight of writing quality loss being greater than that of user comfort loss. After each writing session, the gradient of the loss function with respect to the haptic parameters is calculated and updated online. An exponentially weighted moving average method is used to smooth the parameter updates. The parameter update frequency is dynamically adjusted according to the user's adaptation stage. In the initial stage, the parameters are updated after each writing session. In the transition stage, the update frequency is reduced, and in the stable stage, the update frequency is further reduced. When a continuous decline in writing quality is detected, the parameter update is triggered immediately. An evaluation mechanism for template switching effect was established, defining three types of evaluation indicators: adaptation time, writing quality, and user satisfaction. Paired t-tests were used to compare the differences between the experimental group and the control group.

9. The tactile-guided calligraphy template intelligent matching method according to claim 1, characterized in that, S5 includes: To analyze users' cross-template adaptability, we extracted the adaptation curve for each template switch and obtained the adaptation speed slope and convergence value through linear regression fitting. We defined the adaptation speed index as the reciprocal of the number of writings required to achieve stable writing quality, and the adaptation stability index as the reciprocal of the standard deviation of writing quality during the adaptation period plus one. We then weighted the two to obtain the comprehensive adaptability index. A template learning graph is constructed, with all templates as nodes and the tactile similarity between templates as edge weights. A greedy algorithm is used to start from the user's current template and select the template with the highest similarity that has not been copied as the next recommended template. The recommendation strategy is dynamically adjusted based on the comprehensive adaptability index. When the comprehensive adaptability index is greater than the first ability threshold, the learning challenge is increased. When the comprehensive adaptability index is less than the second ability threshold, templates with similarity greater than the preset similarity threshold are selected. When the comprehensive adaptability index is between the second ability threshold and the first ability threshold, templates are recommended in order of standard similarity.

10. A tactile-guided calligraphy template intelligent matching system, used to perform the steps in the tactile-guided calligraphy template intelligent matching method as described in any one of claims 1-9, characterized in that, include: The tactile operation profile construction module is used to collect user writing action data and tactile interaction response data. It adopts a multimodal sensor synchronous acquisition and feature extraction method to construct a personalized tactile operation profile that includes tactile operation habit characteristics, template usage history and cross-template habit change patterns. The tactile habit model building module is used to obtain personalized tactile operation profiles. It uses graph neural networks and meta-learning methods to build a user-personalized tactile habit model that can predict users’ tactile preferences on specific templates and achieve rapid adaptation. The haptic strategy matching module is used to generate an intelligent haptic strategy matching scheme based on the user's personalized haptic habit model, using transfer learning and domain adaptation methods. This scheme includes template similarity calculation, selective transfer mechanism and progressive transition strategy. The haptic feedback optimization module receives haptic strategy intelligent matching schemes and real-time writing data, adopts progressive haptic feedback control and online learning methods, outputs continuously optimized haptic guidance strategies and records template switching effect data; The template recommendation optimization module is used to acquire template switching effect data and user tactile operation profiles. It uses adaptive analysis and template learning graph construction methods to generate an optimization plan that includes personalized template recommendation strategies and learning suggestion reports.

Images