Deep learning personalized tutoring method based on real-time writing track

Abstract

The invention relates to the technical field of writing intelligent tutoring, and discloses a deep learning personalized tutoring method based on a real-time writing track. The method comprises the following steps: capturing writing motion data of a user in real time through a track sensing device, wherein the writing motion data comprises a pen point coordinate sequence, a timestamp sequence and a pressure intensity sequence; performing initial processing such as noise suppression and data alignment on the captured data; constructing a deep learning architecture based on the processed data to learn the writing feature representation; analyzing a real-time writing track sequence by applying the framework, and identifying writing deviation and a user habit mode; generating a customized tutoring instruction set according to the analysis result; and the deep learning architecture and the tutoring instruction set are dynamically updated by using the user interaction data and the continuously collected writing traces. According to the method, writing details can be comprehensively captured, writing problems and habits can be accurately identified through deep learning, personalized tutoring is provided, tutoring strategies can be dynamically adjusted, and the pertinence and effectiveness of writing tutoring are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent writing tutoring technology, specifically a deep learning-based personalized tutoring method based on real-time writing trajectories. Background Technology

[0002] In traditional writing instruction, whether in educational settings such as calligraphy practice and language writing learning, or in vocational training to cultivate specific writing standards, the methods often rely on manual observation and experience-based judgment. Teachers or tutors need to observe the learner's writing process with the naked eye and point out writing deviations based on personal experience, such as incorrect stroke order, improper pen pressure, and unbalanced character structure. This approach is not only limited by the tutor's subjective judgment ability, making it difficult to establish a unified and objective evaluation standard, but also fails to achieve detailed tracking of the writing process, resulting in many subtle writing habits and deviations going undetected in a timely manner.

[0003] With the development of digital technology, some writing aids have emerged, such as graphics tablets equipped with pressure sensors and writing software with trajectory recording functions. These tools can initially record writing trajectories and pressure information, but they have significant limitations in data processing and analysis. They typically only perform simple trajectory playback or pressure value statistics, unable to deeply explore the writing characteristics and habit patterns hidden within the data. Furthermore, the tutoring feedback provided by these tools is often general, lacking customized content for individual user differences, and failing to meet users' personalized learning needs.

[0004] In existing handwriting tutoring methods, the processing of handwriting data is mostly static, meaning that analysis and feedback are based on single or limited instances of handwriting data, failing to dynamically adjust tutoring strategies according to the user's continuous handwriting behavior. This static approach makes the tutoring process lack adaptability and consistency, making it difficult to effectively track the user's learning progress and improvement, thus affecting the improvement of tutoring effectiveness. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based personalized tutoring method based on real-time writing trajectories to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a deep learning-based personalized tutoring method based on real-time writing trajectories, the method comprising: The user's writing motion data is captured in real time by a trajectory sensing device. The writing motion data includes pen tip coordinate sequence, timestamp sequence and pressure intensity sequence. The captured writing motion data is initially processed, including noise suppression and data alignment operations; a deep learning architecture is built based on the processed writing motion data to learn writing feature representations; The application uses a deep learning architecture to analyze real-time writing trajectory sequences to identify writing deviations and user habit patterns; generates customized tutoring instruction sets based on the analysis results; and dynamically updates the deep learning architecture and tutoring instruction sets using user interaction data and continuously collected writing trajectories.

[0007] Preferably, the specific operations for initial processing of the captured writing motion data include: after the trajectory sensing device outputs the raw writing trajectory data, receiving the raw writing trajectory data and performing data cleaning to remove invalid points and device error points; smoothing the writing trajectory data using a sliding window mean algorithm to eliminate high-frequency noise; unifying the sampling interval of different data points using a time series resampling method; calculating the statistical distribution parameters of the writing trajectory data, defining a dynamic threshold range, and marking data points exceeding the threshold range as outliers; filling missing data points and replacing outliers using nearest neighbor interpolation technology; and finally normalizing the cleaned writing trajectory data to map it to a unified numerical range.

[0008] Preferably, the specific operations for constructing a deep learning architecture to learn handwriting feature representations include: receiving standardized handwriting trajectory data as input after the preprocessing module outputs it; designing a multi-layer recurrent neural network structure, which includes an input layer, a hidden layer, and an output layer; receiving the standardized handwriting trajectory data in the input layer and performing vector embedding transformation; applying a gating mechanism to the hidden layer to extract time-dependent features and capture long-term patterns in the handwriting sequence; generating handwriting feature vector representations in the output layer; dividing the user's historical handwriting trajectory dataset into a training subset and a validation subset; inputting the training subset into the multi-layer recurrent neural network for iterative training, optimizing the network weight parameters through a backpropagation algorithm; and using the validation subset to evaluate the network performance, stopping training when the error metric converges.

[0009] Preferably, the specific operations of the deep learning architecture used to analyze real-time writing trajectory sequences include: receiving the writing feature vector representation after the trained multi-layer recurrent neural network outputs it; loading the real-time writing trajectory sequence into the multi-layer recurrent neural network for forward propagation calculation; extracting the activation values ​​of the intermediate layers of the network as the writing pattern analysis result; performing a classification operation based on the writing pattern analysis result to identify the writing error type and habit intensity; outputting the error probability distribution and habit score; combining the error probability distribution and habit score to predict potential writing problem paths; and updating the classification weights of the multi-layer recurrent neural network according to the actual writing trajectory data.

[0010] Preferably, the specific operations for generating a customized tutoring instruction set include: receiving the error probability distribution and habit score after the writing diagnosis module outputs them; defining a tutoring rule base containing various tutoring action templates; matching the best tutoring action template based on the error probability distribution; adjusting the intensity parameters of the tutoring action template according to the habit score; generating an initial tutoring instruction sequence; optimizing the initial tutoring instruction sequence through a user preference database and adding personalized prompts; and outputting the final customized tutoring instruction set to the display device.

[0011] Preferably, the specific operations for dynamically updating the deep learning architecture and tutoring instruction set using user interaction data and continuously collected writing trajectories include: after the personalized tutoring generation module outputs a customized tutoring instruction set, it receives user feedback data on the tutoring instruction set; simultaneously, it receives newly collected writing trajectory data from the trajectory sensing device; it merges the feedback data and the newly collected writing trajectory data as an update dataset; it inputs the update dataset into a multi-layer recurrent neural network for incremental training and fine-tunes the network parameters; it recalculates the writing feature vector representation; it optimizes the action templates in the tutoring rule base based on the new writing feature vector representation; and iteratively executes the above operations until the tutoring instruction set is stable.

[0012] Preferably, the specific operations for optimizing action templates in the tutoring rule base include: receiving the updated writing feature vector representation after the dynamic optimization module outputs it; parsing the key feature dimensions in the writing feature vector representation; retrieving existing action templates in the tutoring rule base that are associated with the key feature dimensions; calculating the matching score between the existing action templates and the writing feature vector representation; when the matching score is lower than a preset standard, regrouping the action templates using a clustering algorithm; generating a new set of action templates and replacing the low-matching templates; and integrating the new set of action templates into the tutoring rule base.

[0013] Preferably, the specific operations for merging feedback data and newly acquired writing trajectory data as an updated dataset include: receiving feedback data after it is transmitted through the user interaction interface; simultaneously receiving newly acquired writing trajectory data from the trajectory sensing device; performing a structured transformation on the feedback data to generate a labeled dataset; performing the same cleaning and smoothing operations as the initial processing on the newly acquired writing trajectory data; fusing the labeled dataset and the cleaned newly acquired writing trajectory data; sorting the fused data in chronological order; dividing the fused data into batch update units; and outputting the batch update units to the incremental training process.

[0014] Preferably, the specific operations for regrouping action templates using a clustering algorithm include: receiving the matching score after the tutoring rule base optimization module outputs the matching score; loading the feature vectors of all action templates in the tutoring rule base; applying density clustering to group and analyze the feature vectors of all action templates; identifying low-density groups as elimination candidates; generating representative templates for high-density groups; constructing a new template set based on the representative templates; calculating the average similarity between the new template set and the written feature vector representation; and confirming the validity of the new template set when the average similarity reaches the target value.

[0015] Preferably, the specific operations for grouping and analyzing all action template feature vectors using density clustering include: receiving the action template feature vector after the template recombination module outputs the action template feature vector; defining the neighborhood radius parameter and the minimum number of points threshold; calculating the neighborhood density value of each action template feature vector; dividing core points and boundary points according to the neighborhood density value; connecting the core points to form an initial cluster group; merging adjacent cluster groups; outputting the final clustering result; and filtering high-density groups based on the final clustering result.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By capturing rich real-time writing motion data through trajectory sensing devices, including multi-dimensional information such as pen tip coordinates, timestamps, and pressure intensity, comprehensive and detailed raw materials are provided for subsequent analysis and processing. This data can accurately reflect every subtle movement of the user during the writing process, including the speed of pen strokes, changes in pressure, and the direction of strokes, allowing the interpretation of writing behavior to go beyond the surface level.

[0017] Noise suppression and data alignment of captured writing motion data can effectively improve data quality. Removing noise interference allows the data to more accurately reflect the user's actual writing process, while data alignment ensures the comparability of data from different writing stages, providing a reliable foundation for subsequent deep learning architecture learning. Processed data is more standardized and accurate, enabling deep learning models to better extract valuable writing features.

[0018] By building a deep learning architecture based on processed data to learn handwriting feature representations, it is possible to uncover potential patterns and features that are difficult for humans to detect. Deep learning models possess powerful learning and fitting capabilities, summarizing feature patterns corresponding to different handwriting styles and habits from large amounts of handwriting data, thus laying the foundation for accurately identifying handwriting deviations and user habit patterns. This feature representation, obtained through machine autonomous learning, is more comprehensive and accurate than manually defined features.

[0019] The application utilizes a deep learning architecture to analyze real-time writing trajectory sequences, enabling timely and accurate identification of deviations in the user's writing process, such as incorrect stroke order or structural proportion imbalances. It also clearly grasps the user's habitual patterns, such as pen stroke rhythm. Based on these analysis results, a customized set of tutoring instructions is generated, providing personalized guidance tailored to the user's specific problems and characteristics. This allows users to clearly understand their shortcomings and areas for improvement, avoiding the blind application of generalized tutoring.

[0020] By dynamically updating the deep learning architecture and tutoring instruction set using user interaction data and continuously collected writing trajectories, the entire tutoring system can continuously evolve along with the user's learning process. As user writing data accumulates and interactive feedback increases, the deep learning model continuously optimizes its understanding and recognition of writing features, and the tutoring instruction set is adjusted accordingly based on the user's progress and changes. This ensures that the tutoring content always matches the user's current state, maintaining the timeliness and relevance of the tutoring, and enabling the tutoring process to continuously adapt to the user's learning pace and changing needs. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the working principle of the deep learning-based personalized tutoring method based on real-time writing trajectory described in this invention. Figure 2 To write a flowchart of the initial processing of motion data; Figure 3 A flowchart for real-time trajectory sequence analysis; Figure 4 A flowchart generated for a customized tutoring instruction set; Figure 5 The flowchart for updating the dataset merging and preprocessing. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 This invention provides a deep learning-based personalized tutoring method based on real-time writing trajectories, the method comprising: The system captures real-time user writing motion data, including pen tip coordinate sequences, timestamp sequences, and pressure intensity sequences, using a trajectory sensing device. The captured writing motion data undergoes initial processing, including noise suppression and data alignment. A deep learning architecture is then constructed based on the processed data to learn writing feature representations. This architecture is applied to analyze real-time writing trajectory sequences, identifying writing deviations and user habit patterns. A customized set of tutoring instructions is generated based on the analysis results. The deep learning architecture and tutoring instruction set are dynamically updated using user interaction data and continuously collected writing trajectories.

[0024] Example 1: See Figure 2 When a trajectory sensing device captures a user's writing motion, it generates a raw data stream containing multi-dimensional information. This data consists of the coordinate changes of the pen tip in three-dimensional space, millisecond-accurate timestamps, and contact force intensity collected by a pressure sensor. The primary task in the initial processing stage is to assess and clean this raw data. The data cleaning process employs a multi-level filtering mechanism. First, based on the physical characteristics of the device, outliers that significantly exceed the sensor's range are identified and removed. These outliers typically manifest as sudden jumps in coordinates to boundary values ​​or zero pressure intensity. Subsequently, a temporal continuity check identifies suspicious points where timestamps are discontinuous or coordinate abrupt changes exceed reasonable thresholds.

[0025] The sliding window mean algorithm employs a circular buffer structure, with the window size dynamically adjusted based on writing speed. When writing speed increases, the window size is automatically reduced to preserve detailed features; conversely, when writing speed decreases, the window size is expanded to enhance smoothing. During runtime, the algorithm calculates the statistical characteristics of the data within the window in real time, and points deviating from the center value by more than twice the standard deviation are weighted less. This adaptive smoothing strategy effectively suppresses high-frequency noise interference while preserving key features of the writing trajectory.

[0026] Time series resampling is based on a precise analysis of the original sampling frequency. The system first statistically analyzes the time interval distribution between consecutive data points to identify the main sampling period. For segments with intervals significantly longer than the average period, piecewise cubic Hermite interpolation is used to supplement intermediate points; for segments with excessively dense intervals, representative samples are extracted at equal intervals. The resampled data sequence has strictly uniform time intervals, establishing a unified time benchmark for subsequent feature analysis.

[0027] The dynamic threshold range is determined using a rolling calculation method. The system maintains a fixed-length historical data buffer and continuously updates statistical parameters. For coordinate data, both planar displacement and vertical motion characteristics are considered; for pressure data, the force distribution at different stages of writing—the beginning, movement, and end of the stroke—is differentiated. When a data point is detected to exceed the current threshold range, the system performs secondary verification based on the surrounding context to avoid misjudging reasonable changes in special writing strokes as abnormal.

[0028] The data imputation and correction process employs a context-based prediction method. For missing points, not only are the geometric relationships between adjacent points considered, but also the common trajectory patterns of that location in typical writing patterns are analyzed. Outlier replacement strategies vary depending on the outlier type: isolated outliers are corrected using linear interpolation; continuous outlier segments are reconstructed by referencing the typical writing patterns of the region. All correction operations are logged for reference by subsequent analysis modules.

[0029] The normalization process employs a combination of feature scaling and translation. Coordinate data normalization considers the actual range of the user's writing area, first calculating the minimum bounding rectangle of the writing plane, and then linearly mapping the coordinates to a standard interval. Pressure data processing is more refined; extreme values ​​at the start and end of the stroke are first excluded, and then a non-linear transformation is applied to the pressure distribution during the main stroke phases to ensure comparability of pressure characteristics across different users. Timestamps are converted to milliseconds offset relative to the start of writing, while retaining the original time information as metadata.

[0030] The entire initial processing flow adopts a parallel pipeline architecture, with data exchanged between processing modules via shared memory. The system maintains quality control indicators for each processing step and monitors the data processing effect in real time. When a decline in output quality is detected at any stage, parameter adjustments or switching to a backup algorithm are automatically triggered. Complete intermediate results and metadata are retained during processing, forming a traceable data processing chain.

[0031] The noise suppression algorithm combines time-domain and frequency-domain analysis methods. First, the original signal is decomposed into multi-resolution components to identify noise components in different frequency bands. Adaptive filters are used to suppress inherent high-frequency noise in electronic devices; low-frequency interference caused by uneven writing surfaces is eliminated through morphological processing. Data alignment ensures strict synchronization of data across spatial coordinates, time information, and pressure intensity, eliminating the effects of clock drift between sensors.

[0032] The data quality assessment module operates at each key node of the processing flow, continuously monitoring the integrity, consistency, and rationality of the data. Assessment metrics include trajectory continuity index, pressure distribution entropy value, and temporal consistency score. When data quality is found to be below predetermined standards, the system can trigger data re-acquisition or special processing procedures. For critical writing feature points, such as stroke transitions and starting / ending positions, enhanced processing algorithms are used for focused protection.

[0033] The processed data output uses a standardized encapsulation format, comprising three parts: raw data, processing records, and a quality assessment report. The system generates a unique tracking identifier for each written stroke, establishing a complete mapping from the original input to the final output. A data caching mechanism ensures recovery from the most recent checkpoint in the event of an interruption during processing, preventing data loss. The output interface supports multiple data format conversions to meet the input requirements of different analysis modules.

[0034] The entire initial processing system is designed with a configurable architecture, allowing core parameters to be dynamically adjusted via configuration files. The system provides a real-time monitoring interface, visually displaying the effects of each stage of data processing. A debug mode records detailed processing logs for algorithm optimization and problem diagnosis. The processing modules employ a plug-in design, supporting flexible expansion or replacement of specific functional components based on specific application scenarios. Performance optimization utilizes memory pre-allocation and computational parallelization strategies to ensure efficient operation even on resource-constrained embedded devices.

[0035] Example 2: See Figure 3 When standardized handwriting trajectory data is input into the deep learning architecture, the system first performs data format conversion and dimension alignment operations. The vector embedding transformation designed for the input layer uses a learnable projection matrix to map the preprocessed coordinate sequence to a high-dimensional space. This mapping preserves the geometric characteristics of the original trajectory while creating a richer representation space for subsequent feature extraction. The parameters of the embedding layer are initialized using a distribution that matches the statistical characteristics of the handwriting trajectory, accelerating the model convergence process. The input data undergoes batch standardization to eliminate scale differences between different handwriting samples.

[0036] The main structure of the multilayer recurrent neural network adopts a bidirectional design. The forward network captures the temporal evolution of the writing process, while the backward network analyzes the inverse features of stroke formation. Each direction of the network contains three hidden layers, with residual connections between layers to alleviate the gradient vanishing problem in deep networks. The hidden units utilize an LSTM structure, and the memory cell states are customized according to the continuity of the writing trajectory. The forget gate bias in the gating mechanism is set to a positive number, tending to retain more historical information in the initial state. The activation functions of the input and output gates employ piecewise linear transformations to adapt to the feature change rates at different stages of the writing process.

[0037] The network training process employs a phased strategy, with the initial phase focusing on learning the basic spatiotemporal patterns of writing trajectories. The sample organization of the training subset considers the semantic relevance of the written content, placing multiple instances of the same character or stroke in the same batch. When calculating gradients using the backpropagation algorithm, differentiated weights are applied to different types of writing features, with greater weight given to the prediction errors of key stroke inflection points. The optimization process uses an adaptive learning rate algorithm, dynamically adjusting the learning step size of each parameter based on the parameter update history. Elastic constraints are introduced into the network weight updates to prevent large parameter changes from disrupting the learned stable features.

[0038] The evaluation of the validation subset not only focuses on the overall accuracy metric but also monitors the balance of detection rates for various types of handwriting errors. The validation process employs a sliding window technique, extracting multiple local segments from long handwriting sequences for independent evaluation. When validation metrics fluctuate, the system automatically initiates an early stopping mechanism, saving the current best model after several consecutive training epochs without improvement. Training configuration and data processing versions are recorded simultaneously during model saving to ensure consistency in subsequent loading. The validation phase also includes adversarial testing, applying controlled perturbations to the input data to evaluate the model's robustness.

[0039] When the trained network is applied to real-time writing trajectory analysis, the system employs a streaming processing mode. Input data undergoes the same preprocessing steps as in the training phase, then is fed into the network in fixed-length sliding windows. Forward propagation computation is optimized for real-time requirements, using a simplified computation mode with lower latency. The activation values ​​of the intermediate layers are reduced in dimensionality and normalized to form feature representations of the writing pattern. These feature representations capture the dynamic changes during the writing process, including multi-dimensional information such as stroke order, pen speed, and pressure variations.

[0040] The handwriting error classification module employs a hierarchical structure, first identifying broad error categories and then further subdividing each category into finer-grained subcategories. Classification decisions combine the original probabilities from the network output with context-based correction terms, which consider the coherence of the written content. The habit scoring system analyzes the stability of handwriting characteristics over time, and the score calculation integrates the performance of multiple handwriting examples to avoid interference from single fluctuations. Scores are calibrated before output to ensure comparability between different habit categories.

[0041] The potential handwriting problem prediction module establishes a Markov chain model to describe the transition probabilities between different types of errors. Model parameters are updated online based on the user's actual handwriting data, reflecting individual-specific error evolution patterns. Predicted results are represented as possible problem development paths, each accompanied by a confidence assessment. A classification weight adjustment mechanism monitors the consistency between the predicted results and actual observations; when the deviation exceeds a threshold, a weight rebalancing process is triggered. Weight updates are performed in small, incremental steps to maintain model stability.

[0042] The analysis of activation values ​​in the intermediate layers of the network employs an attention mechanism, automatically focusing on key regions in the writing trajectory. Visualizing the attention weights helps understand the model's decision-making process; this visualization is also used to verify whether the features learned by the network conform to writing patterns. For identified special writing patterns, the system establishes temporary memory units to enhance the detection sensitivity of related features in subsequent analyses.

[0043] During real-time analysis, the system continuously monitors computing resource usage and dynamically adjusts the network's computational depth and accuracy. When resources are strained, it automatically switches to a lightweight model or reduces the extraction of non-critical features. Analysis results are output in a standardized format, including structured data such as error type codes, habitual rating vectors, and confidence indices. The output interface supports multiple protocols, facilitating integration with different types of tutoring systems.

[0044] The entire analysis system's operational status is displayed in real-time via a dashboard, including core metrics such as data processing throughput, feature extraction quality, and classification performance. The system maintains detailed operation logs, recording the analysis process and results for each written sample, supporting subsequent auditing and optimization. A debug mode allows for in-depth tracking of feature propagation paths within the network, aiding in algorithm improvement and parameter tuning. The system design considers adaptability to different writing devices and scenarios, supporting multiple application modes through configuration switching.

[0045] Example 3: See Figure 4 When the error probability distribution and habit score output by the writing diagnosis module are input into the tutoring rule base system, the system first performs a multi-dimensional feature fusion operation. The tutoring rule base is stored using a graph structure, where nodes represent tutoring action templates and edges represent the transformation relationships between templates. Each template contains three dimensions: text prompts, visual demonstrations, and practice suggestions, which are dynamically combined using weight coefficients. The matching process between the error probability distribution and the rule base employs an improved nearest neighbor algorithm. This algorithm considers the hierarchical relationship of error types, setting different matching thresholds for parent and child errors. During the matching process, the system calculates the semantic distance between the input features and the template features.

[0046] in, This represents the i-th tutoring template. This represents the j-th input feature; Represents error type characteristics, Represents habitual characteristics, Represents contextual features; , , These are adjustable weighting coefficients, summing to 1. , , These represent the metric functions for the corresponding feature spaces. The distance calculation employs an approximate nearest neighbor search technique to accelerate the process, improving response speed while maintaining accuracy.

[0047] The habit rating adjustment mechanism employs a non-linear mapping function, converting ratings in the 0-1 range into intensity adjustment coefficients. For writing habits with ratings below 0.3, the system adopts a gentle, gradual tutoring strategy; for ratings between 0.3 and 0.7, it maintains standard tutoring intensity; and for habits above 0.7, it triggers a reinforced intervention mode. The intensity parameter affects the frequency, detail, and amount of practice in tutoring instructions, and a smooth transition algorithm is used during the adjustment process to avoid abrupt changes. After the initial tutoring instruction sequence is generated, the user preference database performs personalized adaptation operations, maintaining a preference feature vector formed from the user's historical interaction records.

[0048] The tutoring instruction optimization process considers multiple constraints, including cognitive load balance, time efficiency, and memory enhancement effect. The system employs an instruction combination optimization algorithm to select the optimal subset from the candidate instruction set. The optimization objective function is expressed as:

[0049] in, Indicates a subset of candidate instructions; Measure error coverage, Novelty assessment metrics Calculate personalized fit; , , The weight parameters are dynamically adjusted. This optimization problem is solved using a hybrid strategy combining a greedy algorithm and local search, obtaining an approximate optimal solution in a finite time.

[0050] The display device employs a responsive layout, automatically adjusting interface elements based on command complexity and terminal characteristics. Visual guidance includes dynamic stroke demonstrations, with the demonstration speed proportional to the user's writing speed. Text descriptions use a hierarchical expansion approach, prioritizing key points and displaying details through interactive actions. The position, size, and color of interface elements are optimized based on eye-tracking data to ensure a natural and smooth information acquisition experience.

[0051] The user feedback data collection system features a multi-channel input interface, supporting both explicit ratings and implicit behavioral feedback. Explicit feedback includes satisfaction evaluations and improvement suggestions for specific tutoring projects; implicit feedback analyzes user compliance and completion quality when implementing tutoring suggestions. After cleaning and labeling, the feedback data is synchronously integrated with continuously collected new writing trajectory data. The integration process employs a time alignment algorithm to ensure accurate matching between feedback and corresponding writing events.

[0052] The incremental training strategy employs an elastic weight consolidation method, assigning importance weights to core network parameters to prevent new data from overwriting already learned, stable knowledge. During training, the system dynamically adjusts the sampling ratio of new and old data, initially favoring the retention of existing knowledge and gradually increasing its influence as new data accumulates. Network parameter updates utilize a mini-batch incremental approach, evaluating the model's performance on the validation set after each training cycle to prevent overfitting. The incremental training trigger conditions comprehensively consider the amount of new data, the degree of model drift, and computational resource availability, achieving a balance between effectiveness and efficiency.

[0053] The optimization process for the tutoring rule base employs a two-stage strategy: first, local adjustments to individual template parameters, followed by a global restructuring of the template structure. Local adjustments are based on template usage statistics and performance feedback, modifying template content and applicable conditions. Global restructuring analyzes the association patterns between templates to optimize the rule base's topology. The optimized rule base is then version-controlled, retaining historical versions for rollback when necessary. The template matching index is reconstructed using the LSH approximate nearest neighbor algorithm, improving the retrieval efficiency of a large-scale rule base.

[0054] The system continuously monitors the effectiveness of tutoring instructions during operation, establishing an instruction-improvement correlation model. This model analyzes the correlation between specific types of tutoring and subsequent writing improvements, providing a basis for rule base optimization. Monitoring indicators include short-term compliance rate and long-term improvement trend, reflecting the immediate effect and lasting impact of instructions, respectively. An anomaly detection mechanism identifies tutoring projects with significantly declining effectiveness, triggering a special review process. The review process combines user feedback and changes in writing characteristics to determine the root cause of the problem and generate corrective solutions.

[0055] The dynamic update system employs a hierarchical caching structure, storing frequently updated content in a memory cache and persistently storing the complete dataset. Update operations utilize a transaction mechanism to ensure data consistency. The system resource management module balances the computational load, automatically reducing the computational precision of foreground services to maintain responsiveness when resource-intensive operations are performed in the background. Version compatibility checks ensure that updated components can work together, performing data format conversions and interface adaptations as necessary.

[0056] The entire tutoring and generation system adopts a microservice architecture, with core functional modules deployed independently and communicating asynchronously via message queues. A system status monitoring panel displays the operational metrics and interaction status of each module, triggering an alert mechanism for abnormal situations. The configuration management system records the current values ​​and modification history of all adjustable parameters, supporting the saving and reuse of parameter combinations. The testing and verification framework provides a simulated user environment, allowing for the evaluation of algorithm improvements without impacting the production system.

[0057] Example 4: See Figure 5, after the updated writing feature vector output by the dynamic optimization module enters the template parsing system, the system starts a multi-stage analysis process. Taking the user's writing of the Chinese character '永' as an example, the feature vector parsing process first performs a dimension importance ranking. The system records the changes in the feature vectors of the user's continuous 15 writings of the character '永' to form a simplified feature analysis table as shown in Table 1.

[0058] Table 1: Analysis of the importance of feature dimensions for the user's writing of the character '永'.

[0059]

[0060] Based on the data analysis of the table, the parsing system determines that F47 and F23 are the most distinctive feature dimensions when the user writes the character '永'. The system retrieves all action templates in the tutoring rule library that contain the keywords 'hook turning' and'starting point of the horizontal stroke', and finds a total of 8 relevant templates. The matching degree calculation process comprehensively considers three factors: the overlap degree of feature dimensions, the historical usage effect, and the user preference record. For example, for the hook turning training template numbered T205, the system finds that the overlap rate of its core feature dimensions with the user's current problem feature dimensions reaches 82%, but the user's completion rate for this template in the past 3 uses is only 60%, so the system automatically reduces its priority.

[0061] When the matching degrees of multiple consecutive templates are lower than the preset standard, the system starts the template recombination process. Taking the templates of the'starting point of the horizontal stroke' type as an example, there are 5 relevant templates in the existing library, and their feature distributions are discrete. Before clustering analysis, the system first standardizes the template feature vectors to eliminate the influence of different dimensions. During the density clustering process, the system identifies two core clustering groups: one group of templates focuses on the correction of the starting angle (Group A), and the other emphasizes the control of the starting force (Group B). Group A contains 3 templates, and Group B contains 2 templates. The feature distance between the two groups is significantly greater than the within-group distance.

[0062] The new template generation adopts the strategy of 'prototype + variation'. For the templates in Group A, the system calculates their centroid feature vectors as the basic prototype, and then makes a directional adjustment according to the specific feature deviation of the current user. For example, for the 15-degree starting angle deviation of this user, the newly generated T305 template retains the original angle training key points while adding a special visual prompt for small angle deviations. During the template variation process, the system refers to the successful correction cases of similar users and extracts their effective tutoring elements for combination.

[0063] Cross-validation was employed during the template validation phase. The system mixed three newly generated templates with two existing original templates and organized 10 test users with similar issues for comparative trials. Multiple metrics were recorded during the trials: template comprehension time, action execution accuracy, immediate improvement effect, and subjective satisfaction. The validation data showed that the new template T305, while maintaining a similar comprehension time, improved execution accuracy by approximately 15 percentage points compared to the original template. Based on the validation results, the system officially added the T305 template to the database and marked the two unsatisfactory original templates as awaiting elimination.

[0064] Template replacement employs a gradual transition strategy. Instead of immediately deleting inefficient templates, the system first marks them as "under observation," while gradually guiding users to the new template. During the transition, the system continuously collects actual usage data of the old and new templates, generating a weekly performance comparison report. Only when the new template consistently outperforms the old template over three consecutive reporting periods will the system perform the final replacement. Metadata from the old template is retained during the replacement process as reference material for subsequent template optimization.

[0065] The rule base index reconstruction employs an incremental update technique. The system maintains an inverted index structure, recording the mapping relationship between each feature dimension and related templates. When a template is added or modified, the system only updates the index entries corresponding to the affected feature dimensions, rather than rebuilding the entire index. This design significantly reduces system overhead in high-frequency update scenarios. Index entries contain multi-level cache identifiers, and cache priority is automatically adjusted based on template usage frequency.

[0066] The template version control system uses a directed acyclic graph (DAG) structure to record the evolution history. Each template version saves a complete creation context, including: which features were analyzed, which user cases were referenced, and what verification processes were followed. Version rollback operations can accurately restore to any historical state while maintaining consistency with other templates. The system periodically performs version compression, merging similar intermediate versions to optimize storage efficiency.

[0067] The exception handling mechanism plays a crucial role in the template update process. When unexpected situations are detected during the use of a new template, such as a sudden increase in user confusion or a sharp drop in execution accuracy, the system automatically triggers a rollback process. The rollback decision considers multiple factors: the severity of the exception, the proportion of affected users, and whether there are alternative solutions. Each exception handling process generates an analysis report to guide subsequent template optimization.

[0068] The user feedback integration mechanism enables closed-loop optimization. A lightweight feedback collection component is embedded when each coaching template is presented. Users can express their opinions on the template through simple swipe ratings or tag selections. After semantic analysis, the feedback data is automatically categorized into the corresponding optimization dimensions of the template. For example, when multiple users mark "the prompts are not clear enough," the system will prioritize adjusting the visual prompts of that template. Feedback processing employs a tiered response strategy: frequently occurring issues are suggested for immediate optimization, while individual opinions are accumulated until a certain number are processed.

[0069] The template performance tracking system establishes a long-term observation mechanism. Each template is associated with a performance metric matrix, recording its performance across different usage scenarios, user groups, and time periods. The matrix data is automatically updated weekly, and trend analysis identifies signs of template performance decline. When a template's performance shows a continuous downward trend, the system adds it to the optimization queue, arranging for feature re-matching and content refresh.

[0070] The template parameter management system enables fine-grained control. Each template contains dozens of adjustable parameters, such as prompt intensity, presentation speed, and number of repetitions. The system organizes these parameters into a hierarchical structure, supporting batch adjustments and individual fine-tuning. Parameter modifications employ a canary release strategy, first testing new parameter combinations with a small user group to verify their effectiveness before gradually expanding their application. Parameter versions are kept synchronized with template versions to ensure the repeatability of historical analysis.

[0071] Inter-template network analysis reveals potential optimization opportunities. The system periodically calculates the content similarity and effect correlation between templates, constructing a template relationship graph. Graph analysis can identify two optimization scenarios: templates with redundant content can be considered for merging, and templates with complementary effects can be recommended for combined use. For example, the system found that using the horizontal stroke starting template and the vertical hook turning template alone is less effective than using them in a specific order, thus generating a new combined training scheme.

[0072] The template personalization engine enables dynamic adjustments. Based on real-time user feedback, the system fine-tunes various parameters of the currently presented template. For example, if it detects that a user is repeatedly replaying a segment while watching a demo video, the system automatically lowers the difficulty gradient of subsequent exercises and adds transition steps. Adaptation decisions consider the user's cognitive style characteristics; for instance, visual learners receive more visual cues, while kinesthetic learners receive more decomposed movement guidance. The adaptation process records detailed adjustment logs to optimize personalization strategies.

[0073] Example 5: When the tutoring rule base optimization module initiates the template reorganization process, it first performs a comprehensive feature analysis on the existing action templates. The system uses a stratified sampling method to extract representative template samples from the rule base, covering different error types and tutoring intensity levels. The feature representation of each template includes two parts: static attributes and dynamic indicators. Static attributes record the inherent features determined during template design, such as the type of writing error targeted and the applicable learning stage; dynamic indicators reflect the effectiveness data during the actual use of the template, including user completion rate and the persistence of improvement effects. This dual feature encoding method can comprehensively characterize the actual value of the template.

[0074] The density clustering algorithm employs an adaptive parameter setting method. The system analyzes the distribution of all template feature vectors, calculates the average distance and dispersion of sample points within the feature space, and derives appropriate neighborhood radius parameters accordingly. The minimum number of points threshold is determined using a logarithmic ratio based on the size of the rule base, ensuring reasonable clustering granularity for rule bases of different sizes. The parameter determination process utilizes a cross-validation strategy to avoid distorted clustering results due to inappropriate parameters.

[0075] The neighborhood density calculation employs a kernel density estimation method, smoothing the sample distribution around each template feature vector. This method effectively identifies dense regions in the feature space while reducing interference from outliers. Density calculation considers the importance weights of feature dimensions, allowing key features to have a greater impact on density evaluation. The system maintains a density distribution histogram to monitor the feature space's filling status in real time, providing a reference for subsequent clustering quality evaluation.

[0076] The core point identification process introduces a fuzzy decision mechanism. Traditional core point determination uses a binary decision, while this system employs continuous membership representation, reflecting the probability of a sample point becoming a core point. This design makes the cluster boundaries smoother, avoiding the forced classification of templates in a critical state. The boundary point allocation strategy has also been improved, considering not only the distance to the nearest core point but also analyzing its bridging role in the feature space, retaining valuable transitional templates.

[0077] The initial clusters are formed incrementally. The system starts with the highest-density core points and gradually expands the cluster range. Each time a new point is added, the compactness and separation of the current cluster are reassessed. When a decline in cluster quality is detected, expansion of the current cluster is stopped, and a new cluster is built starting from the next high-density core point. This dynamic cutoff strategy maximizes the representativeness of each category while maintaining cluster purity.

[0078] The cluster merging phase employs a multi-scale analysis approach. The system examines the relationships between clusters at different granularity levels, first merging small clusters with significant overlap, and then processing large clusters with ambiguous boundaries. Merging decisions are based not only on geometric distance but also on semantic similarity, avoiding the forced merging of templates with similar features but vastly different guidance objectives. Each merging operation generates detailed log records, including the reasons for merging and impact assessments, for subsequent analysis.

[0079] The selection criteria for high-density groups integrate multiple indicators. In addition to the traditional number of sample points, quality dimensions such as intra-group feature consistency and historical performance stability are also considered. The system calculates a comprehensive score for each candidate group, and the scoring function balances scale effects and quality requirements. The selection process employs a multi-round refinement strategy, first retaining the larger groups with the highest scores, and then selecting the most representative core subset from among them.

[0080] The representative template generation process emphasizes maintaining diversity. For each high-density group, the system not only selects the geometric center point as the primary representative template but also chooses several edge points with complementary features as auxiliary representatives. This multi-representation strategy can better cover the range of feature variations within the group and avoid pattern rigidity caused by over-centralization. The metadata of the representative template records the statistical characteristics of its source group in detail, providing contextual reference for subsequent template use.

[0081] The validation of the new template set employs an adversarial testing approach. The system constructs a set of test samples containing typical writing problems and edge cases. These samples are then processed using both the old and new templates, and the rationality and completeness of the tutoring suggestions are compared. The testing focuses on the templates' adaptability in different contexts, particularly the quality of their response to atypical writing patterns. Deficiencies discovered during validation trigger fine-tuning of template parameters or content supplementation, forming an iterative optimization loop.

[0082] Low-density groups are handled using a tiered retention strategy. Completely isolated sparse templates are moved to the observation area, temporarily disabled but with their metadata retained; groups with a certain size but insufficient density are marked as needing enhancement, and the system actively seeks writing examples with similar characteristics to attempt to supplement relevant templates. This differentiated approach not only cleanses the rule base but also avoids losing potentially valuable coaching patterns.

[0083] The integration process following template reorganization focuses on knowledge transfer. The system analyzes the correspondence between the old and new template systems, establishing a version mapping table to ensure that users' historical tutoring records are correctly linked to the new templates. For templates that have undergone significant changes, the system generates transitional instructions to help users understand the adjusted tutoring points. The integration process pays special attention to the synergistic relationships between templates, reconstructing the tutoring sequence recommendation network to ensure that new templates can be organically integrated into the existing tutoring system.

[0084] Clustering quality monitoring continues throughout the entire reorganization process. The system defines multiple quality indicators, such as intra-cluster similarity, inter-cluster discriminability, and boundary clarity, and tracks the trends of these indicators in real time. When a quality anomaly is detected, the current operation is automatically paused and switched to diagnostic mode. The quality report is presented in a visual format, highlighting key decision points that require manual review, and supports interactive parameter adjustment and result optimization.

[0085] The template relationship graph is reconstructed after recombination. The system calculates the feature similarity and tutoring complementarity between all pairs of templates in the new template set, constructing a weighted directed graph structure. The graph analysis algorithm identifies potential combination patterns between templates, generating multiple sets of candidate tutoring sequences. After these sequences are evaluated by the effect prediction model, the optimal solution is included in the standard recommendation list of the rule base, enriching the selection space for personalized tutoring.

[0086] A controlled experimental design was used to track the long-term effectiveness of the reorganization. Some users continued to use the old rule base, while the remaining users switched to the new version. The value of the reorganization was evaluated by comparing the changes in coaching effectiveness between the two groups. The experimental design strictly controlled confounding variables to ensure the comparability of the results. The tracking data was not only used to verify the effectiveness of this reorganization but also to accumulate experience for subsequent optimization and gradually improve the methodology of template reorganization.

[0087] An anomaly handling mechanism ensures the safety of the reorganization process. When unexpected behavior is detected after reorganization, such as a decrease in tutoring coverage for certain writing problems or fluctuations in user satisfaction exceeding expectations, the system automatically initiates a rollback analysis. Rollback decisions consider multiple factors, including the scope, severity, and possible causes of the anomaly, implementing only the minimum necessary adjustments while ensuring tutoring continuity. Each anomaly handling process generates a case report, serving as a basis for improving the reorganization strategy.

[0088] User feedback channels remained open throughout the redesign. A dedicated feedback interface was designed to allow users to easily express their experience with the new templates. Feedback was analyzed using natural language processing technology to extract key opinions and sentiments, automatically categorizing them into corresponding template improvement dimensions. The system regularly generates feedback summary reports, highlighting frequently asked questions and key suggestions to guide subsequent template optimization.

[0089] Once the reorganized rule base enters a stable operating phase, the system automatically initiates a preventative maintenance plan. This plan periodically checks the distribution changes of template usage data and detects potential feature drift. Maintenance operations include template parameter calibration, feature weight adjustment, and small-scale local reorganization to maintain the timeliness of the rule base in a gradual manner and avoid the performance overhead of triggering large-scale reorganization again. The maintenance process is executed silently in the background, without affecting normal front-end tutoring services.

[0090] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based personalized tutoring method based on real-time writing trajectory, characterized in that, include: The user's writing motion data is captured in real time by a trajectory sensing device. The writing motion data includes pen tip coordinate sequence, timestamp sequence and pressure intensity sequence. The captured writing motion data is initially processed, including noise suppression and data alignment operations; a deep learning architecture is built based on the processed writing motion data to learn writing feature representations; The application uses a deep learning architecture to analyze real-time writing trajectory sequences to identify writing deviations and user habit patterns; generates customized tutoring instruction sets based on the analysis results; and dynamically updates the deep learning architecture and tutoring instruction sets using user interaction data and continuously collected writing trajectories.

2. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 1, characterized in that, The specific operations for initial processing of the captured writing motion data include: receiving the raw writing trajectory data after the trajectory sensing device outputs it, performing data cleaning to remove invalid points and device error points; smoothing the writing trajectory data using a sliding window mean algorithm to eliminate high-frequency noise; unifying the sampling interval of different data points using a time series resampling method; calculating the statistical distribution parameters of the writing trajectory data, defining a dynamic threshold range, and marking data points exceeding the threshold range as outliers; filling missing data points and replacing outliers using nearest neighbor interpolation; and finally normalizing the cleaned writing trajectory data to map it to a unified numerical range.

3. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 2, characterized in that, The specific operations for constructing a deep learning architecture to learn handwriting feature representations include: receiving standardized handwriting trajectory data as input after the preprocessing module outputs it; designing a multi-layer recurrent neural network structure, which includes an input layer, hidden layers, and an output layer; receiving the standardized handwriting trajectory data and performing vector embedding transformation at the input layer; applying a gating mechanism to the hidden layer to extract time-dependent features and capture long-term patterns in the handwriting sequence; generating handwriting feature vector representations at the output layer; dividing the user's historical handwriting trajectory dataset into a training subset and a validation subset; inputting the training subset into the multi-layer recurrent neural network for iterative training, optimizing the network weight parameters through a backpropagation algorithm; and using the validation subset to evaluate the network performance, stopping training when the error metric converges.

4. The deep learning personalized tutoring method based on real-time writing trajectory according to claim 3, characterized in that, The specific operations of the deep learning architecture used in the application to analyze real-time handwriting trajectory sequences include: receiving the handwriting feature vector representation after the trained multi-layer recurrent neural network outputs it; loading the real-time handwriting trajectory sequence into the multi-layer recurrent neural network for forward propagation calculation; extracting the activation values ​​of the intermediate layers of the network as the handwriting pattern analysis results; performing classification operations based on the handwriting pattern analysis results to identify handwriting error types and habit strengths; outputting error probability distribution and habit scores; combining the error probability distribution and habit scores to predict potential handwriting problem paths; and updating the classification weights of the multi-layer recurrent neural network based on the actual handwriting trajectory data.

5. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 4, characterized in that, The specific operations for generating a customized tutoring instruction set include: receiving the error probability distribution and habit score after the writing diagnosis module outputs them; defining a tutoring rule base containing various tutoring action templates; matching the best tutoring action template based on the error probability distribution; adjusting the intensity parameters of the tutoring action template according to the habit score; generating an initial tutoring instruction sequence; optimizing the initial tutoring instruction sequence through a user preference database and adding personalized prompts; and outputting the final customized tutoring instruction set to the display device.

6. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 5, characterized in that, The specific operations for dynamically updating the deep learning architecture and tutoring instruction set using user interaction data and continuously collected writing trajectories include: after the personalized tutoring generation module outputs a customized tutoring instruction set, it receives user feedback data on the tutoring instruction set; simultaneously, it receives newly collected writing trajectory data from the trajectory sensing device; it merges the feedback data and the newly collected writing trajectory data as an update dataset; it inputs the update dataset into a multi-layer recurrent neural network for incremental training and fine-tunes the network parameters; it recalculates the writing feature vector representation; it optimizes the action templates in the tutoring rule base based on the new writing feature vector representation; and iteratively executes the above operations until the tutoring instruction set is stable.

7. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 6, characterized in that, The specific operations for optimizing action templates in the tutoring rule base include: receiving the updated writing feature vector representation after the dynamic optimization module outputs it; parsing the key feature dimensions in the writing feature vector representation; retrieving existing action templates in the tutoring rule base that are associated with the key feature dimensions; calculating the matching score between the existing action templates and the writing feature vector representation; when the matching score is lower than the preset standard, using a clustering algorithm to regroup the action templates; generating a new set of action templates and replacing the low-matching templates; and integrating the new set of action templates into the tutoring rule base.

8. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 7, characterized in that, The specific operations for merging feedback data and newly acquired writing trajectory data as an updated dataset include: receiving feedback data after it is transmitted through the user interface; simultaneously receiving newly acquired writing trajectory data from the trajectory sensing device; performing structured transformation on the feedback data to generate a labeled dataset; performing the same cleaning and smoothing operations as the initial processing on the newly acquired writing trajectory data; fusing the labeled dataset and the cleaned newly acquired writing trajectory data; sorting the fused data in chronological order; dividing the fused data into batch update units; and outputting the batch update units to the incremental training process.

9. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 8, characterized in that, The specific steps for regrouping action templates using clustering algorithms include: receiving the matching score output by the tutoring rule base optimization module; loading the feature vectors of all action templates in the tutoring rule base; applying density clustering to group and analyze the feature vectors of all action templates; identifying low-density groups as elimination candidates; generating representative templates for high-density groups; constructing a new template set based on the representative templates; calculating the average similarity between the new template set and the written feature vector representations; and confirming the validity of the new template set when the average similarity reaches the target value.

10. The deep learning-based personalized tutoring method based on real-time writing trajectory according to claim 9, characterized in that, The specific operations for grouping and analyzing all action template feature vectors using density clustering include: receiving the action template feature vector after the template recombination module outputs it; defining the neighborhood radius parameter and the minimum number of points threshold; calculating the neighborhood density value of each action template feature vector; dividing the core points and boundary points according to the neighborhood density value; connecting the core points to form the initial cluster group; merging adjacent cluster groups; outputting the final clustering result; and filtering high-density groups based on the final clustering result.

Images