Artificial intelligence-based method and system for detecting children's writing disorders

Abstract

The application provides an artificial intelligence-based children's writing disorder detection method and system, relating to the technical field of data processing, the method comprising: obtaining original handwriting data of a target child, and synchronously collecting environmental parameter data and emotional characteristic data at the corresponding time; segmenting the original handwriting data according to a time window, associating it with the environmental parameter data and emotional characteristic data of the same period; inputting into a recognition model, automatically analyzing the handwriting behavior of each segment and its corresponding environment and emotional characteristics, and discriminating the abnormalities appearing in the segmented handwriting behavior; performing abnormal behavior attribution processing, excluding the abnormalities caused by environmental or emotional influences, and marking the abnormalities not affected by the environment or emotions as suspected disorder behaviors; performing continuity and periodicity analysis to identify disorder behaviors that appear multiple times continuously; and generating a writing disorder detection report for the target child; the application improves the autonomy and accuracy of children's writing disorder detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a child writing disorder detection method and system based on artificial intelligence. BACKGROUND

[0002] At present, the detection of children's writing disorders mainly relies on rule-based evaluation methods, which collect and analyze the characteristics of handwriting trajectories and character structures in children's writing process through artificially formulated writing standard criteria. This kind of method usually uses a digital writing board or a handwriting input device to obtain original handwriting data, and then compares the stroke order, starting and ending positions, and overall character shape to determine whether there is a writing disorder. Some systems also combine speech or image information to realize auxiliary evaluation of children's writing ability through fixed expert experience models.

[0003] However, in actual school classroom scenarios, such rule-based detection methods are prone to misjudgment due to individual differences among children. For example, when students complete a large number of writing tasks in a short period of time, some children may speed up their writing due to psychological pressure or nervousness, resulting in shaky handwriting or unsmooth stroke connections. Existing technologies often misjudge such short-term abnormalities as writing disorders, failing to fully consider the impact of environmental changes and emotional factors on writing behavior, thereby affecting the accuracy of the detection results. SUMMARY

[0004] The purpose of the present application is to provide a child writing disorder detection method and system based on artificial intelligence, which aims to solve the problems mentioned in the background.

[0005] To solve the above technical problems, the technical solutions of the present application are as follows:

[0006] In a first aspect, a child writing disorder detection method based on artificial intelligence, the method comprising:

[0007] Obtaining original handwriting data of a target child when completing a specified writing task in a preset monitoring environment, and synchronously collecting environmental parameter data and emotional feature data at the corresponding time to obtain a synchronous multi-source original data set;

[0008] According to the synchronous multi-source original data set, data preprocessing is performed, the original handwriting data is segmented according to a time window, and the segmented handwriting data is associated with the environmental parameter data and emotional feature data of the same period to obtain a segmented associated data set;

[0009] Inputting the segmented associated data set into an identification model trained by historical big data, automatically analyzing the handwriting behavior of each segment and its corresponding environmental and emotional features, discriminating the abnormalities in the segmented handwriting behavior, and generating preliminary abnormality labeling data;

[0010] According to the preliminary anomaly labeling data, abnormal behavior attribution processing is performed, and the abnormal behaviors caused by environmental or emotional influences are excluded, and the abnormal behaviors not affected by the environment or emotions are marked as suspected disorder behaviors, to obtain a disorder behavior candidate data set;

[0011] According to the disorder behavior candidate data set, continuity and periodicity analysis is performed, and the disorder behaviors appearing multiple times are identified to form a disorder behavior set;

[0012] According to the disorder behavior set, a writing disorder detection report of the target child is generated, and the report content includes the specific type, frequency and associated context of the suspected disorder behavior.

[0013] Preferably, according to the synchronous multi-source original data set, data preprocessing is performed, and the original handwriting data is segmented according to a time window, and is associated with the environmental parameter data and emotional feature data in the same period to obtain a segmented associated data set, including:

[0014] According to the original handwriting data, the initial time window is pre-segmented to obtain an initial segmented handwriting data set;

[0015] For each initial segmented handwriting data, based on the behavior feature points, when any behavior feature point is detected, the position of the behavior feature point is taken as a new segmentation starting point, and the segmentation boundary is dynamically adjusted to form a dynamic segmented handwriting data set driven by behavior feature points, and the behavior feature points include handwriting speed change, pressure mutation and trajectory turning;

[0016] The dynamic segmented handwriting data set is associated with the environmental parameter data and emotional feature data in the corresponding time period to obtain a segmented associated data set.

[0017] Preferably, the segmented associated data set is input into an identification model trained by historical big data to automatically analyze the handwriting behavior of each segment and its corresponding environment and emotional features, and to identify the abnormality in the segmented handwriting behavior to generate preliminary anomaly labeling data, including:

[0018] For each segmented associated data, a multi-source input feature group containing handwriting behavior feature vector, environmental parameter feature vector and emotional feature vector is constructed, and the multi-source input feature group is taken as the input data of the identification model;

[0019] The identification model automatically adjusts the weight coefficients of the handwriting behavior features, environmental parameter features and emotional features according to the multi-source input feature group, and adopts an adaptive weight distribution method to make the influence factors of each feature category dynamically change when identifying abnormal behaviors in different scenarios;

[0020] The recognition model outputs an abnormal probability score and an abnormal type label for each segment of the associated data, the abnormal probability score quantifying the confidence of the segment handwriting behavior being abnormal, and the abnormal type label distinguishing between environment-influenced, emotion-influenced, or suspected disorder behavior;

[0021] According to the abnormal probability score and the abnormal type label, preliminary abnormal labeling data is generated.

[0022] Preferably, according to the preliminary abnormal labeling data, abnormal behavior attribution processing is performed to exclude abnormal behaviors caused by environmental or emotional influences, and to mark abnormal behaviors not affected by environmental or emotional influences as suspected disorder behaviors, to obtain a disorder behavior candidate dataset, including:

[0023] According to the preliminary abnormal labeling data, correlation analysis is performed on the environmental parameter data and the emotional feature data corresponding to each segment of handwriting behavior, and a correlation score R between abnormal behavior and environmental or emotional changes is calculated, the correlation score R being used to measure the degree of influence of environmental or emotional factors on abnormal behavior;

[0024] According to the correlation score R, a plurality of attribution level thresholds are set to divide abnormal behaviors into strong correlation type, medium correlation type, and weak correlation type attribution labels, corresponding to R>T1, T2≤R≤T1, and R

[0025] Abnormal behaviors attributed to the strong correlation type and the medium correlation type are marked as environment-influenced or emotion-influenced abnormal behaviors, and abnormal behaviors attributed to the weak correlation type are marked as suspected disorder behaviors, to obtain a multi-level labeled disorder behavior candidate dataset;

[0026] For each abnormal behavior in the disorder behavior candidate dataset, the distribution of the attribution label in historical samples is counted to form attribution distribution data.

[0027] Preferably, according to the disorder behavior candidate dataset, persistence and periodicity analysis is performed to identify disorder behaviors that occur multiple times, to form a disorder behavior set, including:

[0028] For each suspected disorder behavior, based on a pre-set multi-scale time window, the occurrence frequency and duration of the behavior at different time scales are counted to obtain multi-scale behavior time series statistical data;

[0029] According to the multi-scale behavior time series statistical data, a behavior sequence clustering method is used to perform clustering analysis on the time series distribution of suspected disorder behaviors of the same category in the entire writing task process, to identify disorder behavior patterns with high frequency recurrence or periodicity characteristics;

[0030] According to the abnormal behavior mode, the adaptive trigger threshold of the abnormal behavior is automatically adjusted, when the suspected abnormal behavior appears continuously for a number of times or for a duration exceeding the adaptive trigger threshold within a preset multi-scale time window, the abnormal behavior is determined as a persistent or periodic abnormal behavior, and is included in the abnormal behavior set;

[0031] The abnormal behavior set containing the type of abnormal behavior, the time period of occurrence and the periodic statistical information is generated.

[0032] Preferably, for each initial segmented handwriting data, analysis is performed based on the behavior feature points, when any behavior feature point is detected, the position of the behavior feature point is taken as a new segmentation starting point, the segmentation boundary is dynamically adjusted, and a behavior feature point driven dynamic segmented handwriting data set is formed, including:

[0033] For each initial segmented handwriting data, the average speed, average pressure and trajectory angle change rate of the handwriting within the segment are calculated respectively;

[0034] Within the initial time window range, the speed change value, pressure change value and angle change value between adjacent data points are calculated, when any one of the speed change value, pressure change value or angle change value exceeds the corresponding preset dynamic threshold, the current data point is determined as a behavior feature point;

[0035] When a plurality of behavior feature points are continuously detected within a preset time threshold, a density clustering method is used to combine adjacent behavior feature points into one segmentation starting point;

[0036] The segmented data set segmented by the behavior feature points is subjected to noise filtering, and abnormal short segments with a length less than a set segmentation length threshold are removed, to obtain a dynamic segmented handwriting data set.

[0037] Preferably, the recognition model automatically adjusts the weight coefficients of the handwriting behavior features, environmental parameter features and emotional features according to the multi-source input feature group, and adopts an adaptive weight distribution method, so that the influence factors of each feature category dynamically change when abnormal behavior is discriminated in different scenarios, including:

[0038] For the handwriting behavior feature vector, the environmental parameter feature vector and the emotional feature vector, the corresponding weight coefficients are dynamically updated according to the correlation of the abnormal behavior discrimination results in the historical sample data 、 、 ;

[0039] The following weight dynamic updating formula is adopted:

[0040] ;

[0041] Wherein, is the weight coefficient of the i-th feature The mean score of the relevance of such features to the discrimination of abnormal behavior in historical samples, i∈{b,e,c}, represents handwriting behavior, environmental parameters and emotional features, respectively;

[0042] The recognition model adjusts the relevance score of each feature weight in real time based on the discrimination accuracy of the sample after each new sample analysis, forming a dynamic self-adaptive optimization mechanism.

[0043] Preferably, the attribution distribution data is used for self-adaptive adjustment of the recognition model and individualized intervention, specifically including:

[0044] According to the attribution label distribution of each abnormal behavior in the candidate data set of the target child, the frequency of each attribution type in the historical data of the target child is calculated to generate an individualized attribution frequency vector;

[0045] According to the individualized attribution frequency vector, the attribution type with the highest frequency in the historical data of the target child is identified, and the weight coefficient of the corresponding feature category is set to a first preset weight, and the weight coefficients of other feature categories are set to a second preset weight and a third preset weight, respectively. The first preset weight is greater than the second preset weight, and the second preset weight is greater than the third preset weight;

[0046] When the recognition model discriminates a new sample, the first, second and third preset weights are used to weight each feature category, the first preset weight is used to weight the high-sensitive feature category, and the second and third preset weights are used to weight the medium-sensitive and low-sensitive feature categories, respectively.

[0047] Preferably, according to the multi-scale behavior time series statistical data, a behavior sequence clustering method is used to cluster analyze the time series distribution of suspected abnormal behaviors of the same category in the entire writing task process, and identify abnormal behavior patterns with high frequency recurrence or periodicity, including:

[0048] According to the multi-scale behavior time series statistical data, for each suspected abnormal behavior, the time series similarity of the behavior in different time periods is calculated using the dynamic time warping algorithm, and the behavior sequences with a time series similarity higher than a preset time series similarity threshold are divided into the same behavior cluster to identify the repeated abnormal behavior pattern;

[0049] For each abnormal behavior sequence included in each behavior cluster, a sliding autocorrelation analysis method is used to calculate the period length and period intensity of each abnormal behavior sequence, wherein the period length is the mean value of the behavior interval, and the period intensity is the maximum autocorrelation value obtained by the sliding autocorrelation analysis method.

[0050] When the frequency or cycle intensity of a behavior in a behavior cluster continuously exceeds the corresponding preset mode threshold, the behavior cluster is determined as a disorder behavior mode with periodic characteristics, and the corresponding cycle length and cycle intensity are output as the statistical results of the periodic characteristics.

[0051] In a second aspect, the system comprises:

[0052] A data acquisition module is configured to acquire original handwriting data of a target child when completing a designated writing task in a preset monitoring environment, and synchronously acquire environmental parameter data and emotional feature data at the corresponding time, to obtain a synchronous multi-source original data set.

[0053] A data preprocessing module is configured to perform data preprocessing according to the synchronous multi-source original data set, segment the original handwriting data according to a time window, associate the original handwriting data with the environmental parameter data and emotional feature data in the same period, and obtain a segmented associated data set.

[0054] An abnormality discrimination module is configured to input the segmented associated data set into an identification model trained by historical big data, automatically analyze handwriting behavior of each segment and corresponding environmental and emotional features, discriminate abnormalities in the segmented handwriting behavior, and generate preliminary abnormality labeling data.

[0055] An abnormality attribution module is configured to perform abnormal behavior attribution processing according to the preliminary abnormality labeling data, exclude abnormalities caused by environmental or emotional influences, mark abnormalities not affected by environmental or emotional influences as suspected disorder behaviors, and obtain a disorder behavior candidate data set.

[0056] A behavior analysis module is configured to perform continuity and periodicity analysis according to the disorder behavior candidate data set, identify disorder behaviors appearing multiple times, and form a disorder behavior set.

[0057] A report generation module is configured to generate a writing disorder detection report of the target child according to the disorder behavior set, and the report content includes specific types, occurrence frequencies and associated contexts of suspected disorder behaviors.

[0058] The above-mentioned scheme of the present application at least has the following beneficial effects:

[0059] The present application realizes multi-source data fusion analysis by acquiring original handwriting data of a target child in a preset monitoring environment and synchronously acquiring environmental parameter data and emotional feature data, can comprehensively reflect writing behavior characteristics of children in real situations and dynamic changes of the writing behavior characteristics affected by external influences.

[0060] After data segmentation and correlation analysis, the system can not only make fine-grained judgments on each writing behavior, but also identify short-term abnormalities caused by environmental changes or emotional fluctuations, thereby effectively reducing false judgments caused by non-disability fluctuations. For example, when a child temporarily appears handwriting shaking in a noisy environment, the system can automatically attribute it to environmental noise changes, avoiding mistakenly judging such behavior as a writing disorder.

[0061] Intelligent analysis and attribution processing of the segmented and correlated data set by the recognition model trained by big data can accurately separate the real disorder behavior from the incidental abnormality, greatly improving the accuracy and reliability of disorder detection. At the same time, through the continuity and periodicity analysis of the disorder behavior candidate data, the system can track and identify the repeated disorder behaviors, forming a systematic disorder behavior set, providing an objective basis for individualized intervention and dynamic assessment.

[0062] Finally, the detection report generated by the present application not only contains the type and frequency of suspected disorder behavior, but also reflects the specific environment and emotional context when the behavior occurs, providing more targeted guidance and auxiliary decision-making for educators and parents. These beneficial effects enable the present application to significantly improve the scientificity, accuracy and individualization level of children's writing disorder detection in actual teaching and rehabilitation training scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 is a flowchart of the writing disorder detection method for children based on artificial intelligence provided by an embodiment of the present application. DETAILED DESCRIPTION

[0064] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0065] As shown in Figure 1 An embodiment of the present application proposes a writing disorder detection method for children based on artificial intelligence, which comprises:

[0066] Obtaining original handwriting data of a target child when completing a designated writing task in a preset monitoring environment, and synchronously collecting environmental parameter data and emotional feature data at the corresponding time, to obtain a synchronous multi-source original data set;

[0067] According to the synchronous multi-source original data set, data preprocessing is performed, the original handwriting data is segmented according to a time window, and the original handwriting data is correlated with the environmental parameter data and emotional feature data of the same period to obtain a segmented and correlated data set;

[0068] The segmented associated data set is input into an identification model trained by historical big data, the handwriting behavior of each segment and the corresponding environmental and emotional characteristics thereof are automatically analyzed, the abnormality in the segmented handwriting behavior is discriminated, and preliminary abnormality labeling data is generated;

[0069] According to the preliminary abnormality labeling data, abnormal behavior attribution processing is performed, the abnormality caused by environmental or emotional influence is excluded, the abnormality not affected by the environment or emotion is marked as suspected disorder behavior, and disorder behavior candidate data set is obtained;

[0070] According to the disorder behavior candidate data set, persistence and periodicity analysis is performed, the disorder behavior appearing multiple times is identified, and a disorder behavior set is formed;

[0071] According to the disorder behavior set, a writing disorder detection report of the target child is generated, and the report content includes the specific type of suspected disorder behavior, the occurrence frequency and the associated context.

[0072] In the embodiment of the present application, by obtaining the original handwriting data of the target child when completing the specified writing task in the preset monitoring environment, and synchronously collecting the environmental parameter data and emotional characteristic data at the corresponding moment, the writing behavior of the child in the real scene and the external influencing factors can be comprehensively reflected. By performing data preprocessing on the synchronous multi-source original data set, the original handwriting data is segmented according to the time window, and is associated with the corresponding environmental parameter data and emotional characteristic data, so that each segment of writing behavior can be analyzed in combination with the environment and emotional state, and the recognition accuracy of the cause of abnormal behavior is effectively improved. For example, in a test, if it is found that the writing abnormality at a moment is accompanied by an increase in environmental noise or emotional fluctuation, the system can automatically associate these external factors with the behavior performance at the data level.

[0073] By inputting the segmented associated data set into the identification model trained by historical big data, the handwriting behavior of each segment and the corresponding environmental and emotional characteristics thereof are automatically analyzed, and the intelligent discrimination of the abnormality of the segmented handwriting behavior is realized. The model can adaptively fuse multiple characteristics, and improve the accuracy of abnormal behavior detection in complex scenes. Taking practical application as an example, when facing a large number of different types of child samples, the model can continuously learn the relationship between various characteristics, and gradually optimize the abnormality detection capability.

[0074] When abnormal behavior is identified, the system automatically excludes the behavior caused by environmental or emotional factors through attribution processing, and only marks the behavior not affected by external factors as suspected disorder behavior, obtaining a disorder behavior candidate dataset. Such a process can reduce false positives and accurately identify behaviors related to writing disorders. For example, a child repeatedly interrupts writing and has disordered stroke structure in a quiet environment, and the system can effectively distinguish such behavior from occasional abnormalities affected by the environment.

[0075] Further analysis of the disorder behavior candidate dataset for persistence and periodicity can identify multiple persistent disorder behaviors and form a disorder behavior set. This can provide a comprehensive understanding of the behavior patterns and development trends of children's writing disorders, and provide objective and systematic data support for subsequent intervention. For example, the system can find that a child's handwriting shakes every certain period of time, prompting parents or teachers to pay attention to the child's attention fluctuation problem.

[0076] Finally, the system automatically generates a detection report containing suspected disorder behavior specific types, occurrence frequency, and associated context based on the disorder behavior set, providing fine-grained and personalized diagnostic evidence and intervention suggestions for professionals. This detection method not only improves the scientificity and reliability of writing disorder identification, but also enables real-time feedback and dynamic tracking in daily teaching and rehabilitation training, significantly improving the value and effectiveness of practical applications.

[0077] The system obtains the original handwriting data of the target child when completing the specified writing task in the preset monitoring environment, and synchronously collects the environmental parameter data and emotional feature data at the corresponding time, to obtain a synchronous multi-source original dataset, which specifically includes:

[0078] First, the system configures a digital writing board or a handwriting input device for each target child to collect handwriting trajectory information generated when completing a specified writing task (such as copying, dictation, fill-in-the-blank, etc.). The handwriting trajectory information includes stroke order, writing speed, stroke pressure, spatial trajectory, and timestamp, etc. At the same time, during the writing task, the system collects environmental parameter data through environmental sensors configured in the classroom or learning space, including noise decibel, illumination, temperature and humidity, seat space changes, etc. Synchronous collection also includes emotional feature data, which can be obtained through a facial expression recognition module, a speech emotion analysis module, or a special physiological monitoring device (such as a skin conductance sensor, a heart rate band, etc.), to real-time perceive and record the emotional state, emotional fluctuation or stress indicators exhibited by the target child during the writing process.

[0079] All the above collected data are marked by a unified timestamp to realize accurate alignment and synchronization of data, and to ensure that the handwriting behavior at each moment can be recorded in correspondence with the environmental state and emotional performance at that time. Finally, the system automatically integrates the handwriting data, environmental parameter data and emotional feature data into a multi-source original data set, and provides full-quantity, synchronized original data basis for subsequent data preprocessing and intelligent analysis.

[0080] Among them, according to the obstacle behavior set, the writing obstacle detection report of the target child is generated, and the report content includes the specific type, occurrence frequency and associated context of the suspected obstacle behavior, specifically including:

[0081] After the system completes the identification and induction of the obstacle behavior set, for each type of obstacle behavior in the set, according to the information structuring principle, the time period, performance characteristics, attribution results and associated information with the external environment and emotional state of each item are analyzed. First, the system names and classifies each suspected obstacle behavior, such as irregular letter structure, abnormal stroke order, sudden change in writing speed, etc. Then, the occurrence frequency of each type of obstacle behavior during the entire monitoring period is counted, including the total number of occurrences, distribution in each monitoring period, duration and interval period.

[0082] For each suspected obstacle behavior, the system further queries the corresponding environmental parameters and emotional characteristics when the behavior occurs, combines the synchronization information at the time of data collection, and automatically induces the typical associated context when the behavior occurs. For example, at each occurrence of the obstacle behavior, the system records the noise level, illumination, emotional fluctuation state of the child, etc. Finally, all the above information is summarized in the detection report to form a result list or descriptive analysis containing specific behavior types, occurrence frequencies and main associated contexts, providing detailed and traceable scientific data basis for subsequent intervention, training and school-home communication.

[0083] In a preferred embodiment of the present application, according to the synchronized multi-source original data set, data preprocessing is performed, the original handwriting data is segmented according to a time window, and it is associated with the environmental parameter data and emotional feature data in the same period to obtain a segmented associated data set, including:

[0084] According to the original handwriting data, the initial time window is pre-segmented according to the preset initial time window to obtain an initial segmented handwriting data set;

[0085] For each initial segmented handwriting data, based on the behavior feature points, when any behavior feature point is detected, the position of the behavior feature point is taken as a new segmentation starting point, the segmentation boundary is dynamically adjusted, and a dynamic segmented handwriting data set driven by behavior feature points is formed, the behavior feature points include handwriting speed change, pressure mutation, trajectory turning point;

[0086] The dynamic segmented handwriting data set is associated with the environmental parameter data and the emotional feature data of the corresponding time period in segments to obtain a segmented associated data set.

[0087] In the embodiment of the present application, the original handwriting data is pre-segmented according to a preset initial time window, and in each initial segmented handwriting data, behavior feature points such as handwriting speed change, pressure mutation, and trajectory turning are analyzed, so that a processing mode of dynamically adjusting the segmentation boundary is realized. This processing mode can accurately capture the subtle behavior changes of children in the writing process, combines ordinary time segmentation with actual writing behavior characteristics, greatly improves the sensitivity and accuracy of anomaly detection. For example, when the child's writing speed suddenly increases or the pressure abnormally fluctuates, the system can automatically adjust the segmentation to ensure that such abnormalities are identified in time and are not easily missed by the traditional fixed window method. In addition, the density clustering method is used to merge the feature points that appear continuously in a short time, and by eliminating the too short segments, the false judgment caused by noise and other invalid signals is further reduced, the robustness of data processing is improved, and the dynamic segmented handwriting data set more accurately reflects the actual writing behavior.

[0088] The dynamic segmented handwriting data set is associated with the environmental parameter data and the emotional feature data of the corresponding time period in segments to obtain a segmented associated data set, and specifically includes:

[0089] Firstly, the system sets an independent time range for each segmented data in the dynamic segmented handwriting data set, ensuring that each segmented handwriting behavior has a clear start and end time. Subsequently, the system automatically retrieves the environmental parameter data and emotional feature data corresponding to the time period of each segmented handwriting data according to the time range covered by the segmented handwriting data. In order to ensure accurate matching of data, the system uses time stamp comparison to associate environmental parameter data (such as noise level, illumination, etc.) and emotional feature data (such as facial expression recognition results, voice emotional state, etc.) with corresponding segmented handwriting data. If a segment spans multiple environmental or emotional sampling points, the system can use methods such as average value, maximum value, or most recent value to normalize the environmental and emotional data, ensuring that each segmented handwriting data corresponds to unique environmental and emotional information. Through the above steps, the handwriting behavior, environmental state, and emotional features are bound in segments in the time dimension to generate a structured segmented associated data set, providing a standardized, multi-source fusion data basis for subsequent intelligent analysis.

[0090] In a preferred embodiment of the present application, the segmented associated data set is input into an identification model trained by historical big data to automatically analyze the handwriting behavior of each segment and its corresponding environmental and emotional features, identify the abnormalities in the segmented handwriting behavior, and generate preliminary abnormality annotation data, including:

[0091] For each segment correlation data, a multi-source input feature group containing handwriting behavior feature vector, environment parameter feature vector and emotion feature vector is constructed, and the multi-source input feature group is taken as input data of the recognition model;

[0092] The recognition model automatically adjusts the weight coefficients of the handwriting behavior features, the environment parameter features and the emotion features according to the multi-source input feature group, adopts an adaptive weight distribution manner, and makes the influence factors of each feature category dynamically change when abnormal behaviors are discriminated in different scenarios.

[0093] The recognition model outputs an abnormal probability score and an abnormal type label for each segment correlation data, the abnormal probability score is used to quantify the confidence of the segment handwriting behavior being abnormal, and the abnormal type label is used to distinguish environment-affected type, emotion-affected type or suspected disorder behavior.

[0094] According to the abnormal probability score and the abnormal type label, preliminary abnormal labeling data is generated.

[0095] In the embodiment of the present application, the segment correlation data set is input into the recognition model trained by historical big data, a multi-source input feature group containing handwriting behavior feature vector, environment parameter feature vector and emotion feature vector is constructed, so that the model can comprehensively analyze the handwriting behavior of each segment and its correlation with external factors. Through adaptive adjustment of the weight coefficients of each feature category by the recognition model, the model can dynamically optimize the basis for abnormal behavior discrimination according to the actual data performance in different application scenarios, thereby improving the generalization ability and personalized recognition ability of the model. For example, in some children, the influence of emotional changes on writing behavior abnormalities is greater, and the model can automatically increase the weight of emotion features in the discrimination process, thereby improving the detection accuracy. The model outputs an abnormal probability score and an abnormal type label, which can quantitatively determine the confidence and attribution type of abnormal behavior, and provides a scientific basis for subsequent data processing and intervention decision-making.

[0096] The recognition model outputs an abnormal probability score and an abnormal type label for each segment correlation data, and specifically includes:

[0097] Firstly, the system inputs each segment correlation data into an intelligent recognition model trained by historical big data. For each segment correlation data, the model will extract handwriting behavior features, environment parameter features and emotion features, and standardize these features. The model uses existing training experience to automatically judge whether the segment handwriting behavior has abnormal performance in combination with the actual feature distribution of the input data. Specifically, the model will evaluate the difference between each segment behavior and normal behavior samples, and quantify this difference as an abnormal probability score. The abnormal probability score reflects the possibility of the current segment handwriting behavior being abnormal, and is usually expressed in percentage or score form.

[0098] On the basis of the abnormal probability score, the model further assigns an abnormal type label to each segmented behavior according to the performance of each feature category and the comprehensive analysis result. The abnormal type label is used to distinguish the attribution source of the abnormal behavior, such as environmental influence type, emotion influence type, or suspected disorder behavior, etc. The model combines the previously set attribution rules and data performance to output a label that best represents the nature of the current behavior abnormality. Through this process, the system can generate analysis results containing abnormal probability scores and abnormal type labels for each segmented data, providing a scientific basis for subsequent attribution processing and disorder behavior identification.

[0099] In a preferred embodiment of the present application, according to the preliminary abnormal annotation data, abnormal behavior attribution processing is performed, and the abnormal behaviors caused by environmental or emotional influence are excluded, and the abnormal behaviors not affected by environmental or emotional influence are marked as suspected disorder behaviors, to obtain a disorder behavior candidate data set, including:

[0100] According to the preliminary abnormal annotation data, the environmental parameter data and the emotional feature data corresponding to each segmented handwriting behavior are analyzed for correlation, and a correlation score R between the abnormal behavior and the environmental or emotional change is calculated, wherein the correlation score R is used to measure the influence degree of environmental or emotional factors on the abnormal behavior.

[0101] According to the correlation score R, a plurality of attribution level thresholds are set, and the abnormal behavior is divided into strong correlation type, medium correlation type and weak correlation type attribution labels, respectively corresponding to R>T1, T2≤R≤T1, R

[0102] The abnormal behaviors with strong correlation type and medium correlation type attribution are marked as environmental or emotional influence type abnormal behaviors, respectively, and the abnormal behaviors with weak correlation type attribution are marked as suspected disorder behaviors, to obtain a disorder behavior candidate data set with multiple labels.

[0103] For each abnormal behavior in the disorder behavior candidate data set, the distribution of the attribution label in the historical sample is counted to form attribution distribution data.

[0104] In the embodiment of the present application, by performing attribution processing on the preliminary abnormality labeled data, the correlation scores between each segmented handwriting behavior and the environmental parameter data and the emotional feature data are calculated, and the attribution level threshold is set according to the correlation scores, so as to divide the abnormal behaviors into strong correlation type, medium correlation type and weak correlation type, thereby realizing the multi-level labeled abnormal behavior candidate data set. The processing procedure can accurately distinguish which abnormal behavior is mainly caused by environmental or emotional factors, and which abnormal behavior is more likely to be related to the child's own writing disorder, thereby significantly reducing the false judgment and omission. For example, if the segmented behavior abnormality coincides with the increase of classroom noise, the system can automatically determine that it is an environmental influence type, thereby avoiding invalid intervention. Further, the distribution of the attribution label in the historical sample is counted to form the attribution distribution data, which provides a data basis for individualized sensitive factor analysis and subsequent intervention suggestions, thereby realizing the accurate auxiliary diagnosis for each child.

[0105] In the embodiment of the present application, by performing attribution processing on the preliminary abnormality labeled data, the correlation scores between each segmented handwriting behavior and the environmental parameter data and the emotional feature data are calculated, and the attribution level threshold is set according to the correlation scores, so as to divide the abnormal behaviors into strong correlation type, medium correlation type and weak correlation type, thereby realizing the multi-level labeled abnormal behavior candidate data set. The processing procedure can accurately distinguish which abnormal behavior is mainly caused by environmental or emotional factors, and which abnormal behavior is more likely to be related to the child's own writing disorder, thereby significantly reducing the false judgment and omission. For example, if the segmented behavior abnormality coincides with the increase of classroom noise, the system can automatically determine that it is an environmental influence type, thereby avoiding invalid intervention. Further, the distribution of the attribution label in the historical sample is counted to form the attribution distribution data, which provides a data basis for individualized sensitive factor analysis and subsequent intervention suggestions, thereby realizing the accurate auxiliary diagnosis for each child.

[0106] After obtaining the preliminary abnormality labeled data, the system first extracts the environmental parameter data and the emotional feature data corresponding to each segmented handwriting behavior. Next, the system analyzes the environmental and emotional data change trend of the segmented behavior in the current detection process and the historical detection process, and compares it with the abnormal performance of the segmented handwriting behavior. In the correlation analysis process, the system will count the change amplitude of the environmental parameters (such as noise, illumination, etc.) and the emotional features (such as expression tension, heart rate increase, etc.) before and after the occurrence of the segmented behavior abnormality, and combine the probability of the simultaneous occurrence of environmental or emotional changes and abnormal behaviors in the historical data to calculate the correlation degree between them.

[0107] This calculation process can use methods such as statistical correlation method, joint occurrence frequency comparison or normalized score, etc. Finally, the system will generate a correlation score for each segmented handwriting behavior, which reflects the relationship between the behavior abnormality and the environmental change or emotional fluctuation. The higher the score, the more likely the abnormal behavior is caused by environmental or emotional changes, and vice versa, which means that it has a lower correlation with external factors and is more likely to belong to the child's own writing disorder.

[0108] The setting method of the attribution level threshold comprises:

[0109] To achieve fine attribution of abnormal behaviors, the system pre-sets several attribution level thresholds based on a large amount of historical data and expert experience, which are used to classify the correlation scores of abnormal behaviors with environmental and emotional factors. Specifically, first, the correlation scores of abnormal behaviors with known attribution relationships in multiple historical samples are statistically analyzed, and according to the distribution characteristics of the scores, boundary values that can effectively distinguish strong, medium and weak correlations are selected as level thresholds.

[0110] For example, the system can classify samples with scores higher than a certain proportion (such as 80%) as strong correlation type, samples with scores in the middle interval (such as 50%-80%) as medium correlation type, and samples with scores lower than another proportion (such as 50%) as weak correlation type according to the percentile distribution of the correlation scores in all historical samples. The level thresholds can also be adjusted by professionals according to actual needs with reference to actual intervention experience and specific application scenarios in the field. Once the attribution level thresholds are set, they serve as the basis for subsequent analysis and attribution label division, achieving scientific classification and fine attribution of each abnormal behavior.

[0111] In a preferred embodiment of the present application, based on the candidate data set of obstacle behaviors, continuity and periodicity analysis is performed to identify obstacle behaviors that occur continuously multiple times to form an obstacle behavior set, including:

[0112] For each suspected obstacle behavior, based on a pre-set multi-scale time window, the occurrence frequency and duration of the behavior at different time scales are counted to obtain multi-scale behavior time series statistical data;

[0113] According to the multi-scale behavior time series statistical data, a behavior sequence clustering method is used to perform clustering analysis on the time series distribution of suspected obstacle behaviors of the same category during the entire writing task process to identify obstacle behavior patterns with high frequency recurrence or periodicity characteristics;

[0114] According to the obstacle behavior patterns, the adaptive trigger threshold of the obstacle behavior is automatically adjusted. When the suspected obstacle behavior occurs continuously or for a duration exceeding the adaptive trigger threshold within the pre-set multi-scale time window, it is determined to be a continuous or periodic obstacle behavior and is included in the obstacle behavior set;

[0115] An obstacle behavior set containing obstacle behavior types, occurrence time periods and periodicity statistical information is generated.

[0116] In the embodiments of the present application, by continuously and periodically analyzing the candidate data set of obstacle behaviors, the frequency and duration of each suspected obstacle behavior at different time scales can be obtained, and multi-scale behavior time series statistical data can be formed. By combining the behavior sequence clustering method, the time series distribution of suspected obstacle behaviors of the same category in the entire writing task process can be clustered and analyzed, and obstacle behavior patterns with high frequency recurrence or periodicity can be effectively identified. This processing method can comprehensively reveal the repeatability and regularity of children's writing obstacle behaviors. For example, if a child continuously appears a disorderly character structure every certain period of time, the system can automatically detect the periodicity of the abnormality. The system further adjusts the adaptive trigger threshold of the obstacle behavior according to the obstacle behavior pattern, accurately determines the persistence and periodicity of the abnormal behavior in the multi-scale time window, and significantly improves the recognition ability of complex behavior patterns. Finally, the generated obstacle behavior set not only contains the type and time period of the obstacle behavior, but also carries periodicity statistical information, providing scientific, continuous and personalized intervention reference for teachers and parents.

[0117] The setting method of the multi-scale time window specifically includes:

[0118] In order to comprehensively analyze the distribution and regularity of obstacle behaviors at different time scales, the system sets multiple time windows with different lengths when counting each suspected obstacle behavior, which are used to capture short-term, long-term and periodic abnormalities. Specifically, the system pre-divides several time windows according to the total duration of the actual writing task and the interval distribution of the historical obstacle behaviors, for example, windows with minutes, hours or days as units. For each suspected obstacle behavior, the system respectively counts its frequency and duration in each time window, thereby obtaining multi-scale behavior time series statistical data.

[0119] The setting method of the time window can be flexibly adjusted according to the actual monitoring scene. For example, a shorter window (such as 1-5 minutes) can be used for short-time high-frequency obstacle behaviors, and a longer window (such as 15 minutes, 30 minutes or longer) can be used for low-frequency or periodic obstacle behaviors. In addition, the system can dynamically optimize the division of the time window according to the historical data, so that the analysis process can adapt to the behavior rhythm and abnormal characteristics of different children, ensuring the effectiveness and comprehensiveness of multi-scale analysis.

[0120] The automatic adjustment of the adaptive trigger threshold of the obstacle behavior according to the obstacle behavior pattern specifically includes:

[0121] After the system completes the identification and clustering analysis of the obstacle behavior patterns, it analyzes the statistical results such as the occurrence frequency, duration and periodicity intensity of each obstacle behavior. Combined with the distribution characteristics of different types of obstacle behaviors in each time window, the system dynamically adjusts the trigger threshold of each type of obstacle behavior. The adjustment method is: for obstacle behaviors that frequently occur or have obvious periodicity, the system can appropriately lower the trigger threshold, so that the detection sensitivity is improved, and such abnormalities can be found in time; for occasional or more interfering obstacle behaviors, the threshold can be appropriately increased to reduce false positives.

[0122] In the specific implementation process, the system compares the normal fluctuation range of each type of obstacle behavior in historical monitoring with the currently monitored behavior patterns, and automatically optimizes the threshold setting according to the behavior occurrence trend, repeatability and intervention demand. In this way, without manual repeated adjustment, the system can adapt to the behavior changes of different children at different stages, realize accurate identification and real-time dynamic management of obstacle behaviors.

[0123] In a preferred embodiment of the present application, for each initial segmented handwriting data, based on the behavior feature points, when any behavior feature point is detected, the position of the behavior feature point is taken as a new segmentation starting point, the segmentation boundary is dynamically adjusted, and a behavior feature point driven dynamic segmented handwriting data set is formed, including:

[0124] For each initial segmented handwriting data, the average speed, average pressure and trajectory angle change rate in the segment are calculated respectively;

[0125] In the initial time window range, the speed change value, pressure change value and angle change value between adjacent data points are calculated, and when any one of the speed change value, pressure change value or angle change value exceeds the corresponding preset dynamic threshold, the current data point is determined as a behavior feature point;

[0126] When multiple behavior feature points are continuously detected within a preset time threshold, a density clustering method is used to combine adjacent behavior feature points into a segmentation starting point;

[0127] The noise filtering is performed on the segmented data set segmented by the behavior feature points, and the abnormal short segments with a segment length less than a set segment length threshold are removed, to obtain a dynamic segmented handwriting data set.

[0128] In the embodiments of the present application, for each initial segmented handwriting data, based on the behavior feature points, when any behavior feature point such as handwriting speed change, pressure mutation, trajectory turning point is detected, the position of the behavior feature point is taken as a new segmentation starting point, the segmentation boundary is dynamically adjusted, and the real behavior change in the writing process of children is effectively reflected. By calculating the speed average, pressure average and trajectory angle change rate of each segment respectively, the system can distinguish the subtle differences of writing behavior more carefully. Within the initial time window range, further using the comparison of the speed change value, the pressure change value and the angle change value with the dynamic threshold, the precise detection of sudden abnormalities is realized. If multiple feature points are continuously detected within a preset time threshold, the adjacent feature points are merged by using the density clustering method to prevent the data redundancy caused by over-dense segmentation. The segmented data set is filtered for noise, and the abnormal short segments are removed, which improves the robustness of data processing. For example, in the process of children practicing writing, the system can distinguish between habitual hesitation and real writing disorders, and significantly improve the representation ability of the dynamic segmented handwriting data set to the actual behavior.

[0129] When multiple behavior feature points are continuously detected within a preset time threshold, the adjacent behavior feature points are merged into a segmentation starting point by using the density clustering method, specifically including:

[0130] During the generation of the dynamic segmented handwriting data set, the system continuously monitors the distribution of behavior feature points at each time in each initial segmented handwriting data. If the system continuously detects multiple behavior feature points within a preset time threshold (for example, between a few hundred milliseconds and a few seconds), it usually means that these feature points are likely to belong to the same writing behavior change or the same type of continuous abnormal performance. In order to avoid over-dense segmentation or data redundancy caused by taking these densely appearing feature points as multiple segmentation starting points, the system analyzes the time interval and relative distribution density of these feature points.

[0131] Combining the density clustering idea in the prior art, the system classifies these time-close and densely-distributed behavior feature points into the same cluster. In the cluster, the system selects the starting position or the average time point of these feature points as a new segmentation starting point, and merges them into a single segment. This processing can reduce the excessive increase in the number of segments caused by excessive sensitivity, and effectively improve the representativeness of the segments and the efficiency and accuracy of subsequent data analysis.

[0132] The segmented data set segmented by the behavior feature points is filtered for noise, and the abnormal short segments with a length less than a set segmentation length threshold are removed to obtain the dynamic segmented handwriting data set, specifically including:

[0133] When the system completes the dynamic segmentation of handwriting data based on behavioral feature points, some segments may be abnormally short due to incidental minor fluctuations in writing movements or noise of the sensor itself, and cannot truly reflect effective writing behavior. In order to ensure the actual analysis value of the segmented data, the system will perform length screening on all segmented data sets.

[0134] Specifically, the system presets a minimum segment length threshold according to the actual requirements of the writing task and historical experience. For each segment, if its duration or data point number is lower than the threshold, the system determines it as a noise segment or an invalid segment, and removes it from the data set. Only the segments with lengths reaching or exceeding the threshold are retained as effective objects for subsequent writing behavior analysis. In this way, the system can significantly improve the quality of the dynamic segmented handwriting data set, reduce false positives, remove invalid information, and ensure that the analysis results are more reliable and have practical application significance.

[0135] In a preferred embodiment of the present application, the recognition model automatically adjusts the weight coefficients of handwriting behavior features, environmental parameter features and emotion features according to the multi-source input feature group, adopts an adaptive weight distribution method, and dynamically changes the influence factors of each feature category when discriminating abnormal behavior in different scenarios, including:

[0136] According to the correlation of the handwriting behavior feature vector, the environmental parameter feature vector and the emotion feature vector to the abnormal behavior discrimination result in the historical sample data, the corresponding weight coefficients are dynamically updated 、 、 ;

[0137] The following weight dynamic updating formula is adopted:

[0138] ;

[0139] Wherein, is the weight coefficient of the i-th feature, is the mean score of the correlation of the feature to the abnormal behavior discrimination in the historical sample, i∈{b,e,c}, representing handwriting behavior, environmental parameter and emotion feature, respectively;

[0140] The recognition model adjusts the correlation score of each feature weight in real time based on the discrimination accuracy of each new sample after analysis, forming a dynamic adaptive optimization mechanism.

[0141] In the embodiments of the present application, the recognition model can automatically adjust the weight coefficients of handwriting behavior features, environmental parameter features and emotion features for a multi-source input feature group, adopts an adaptive weight distribution method, and dynamically adjusts the influence factors of different feature categories in abnormal behavior discrimination with the accumulation of historical data and the change of actual discrimination accuracy. By statistically analyzing the correlation of each feature vector in the historical samples, the model can dynamically update the weight coefficients, and reasonably distribute the weights according to the following weight dynamic updating formula. Whenever new sample data is input, the model can adjust the correlation score in real time according to the discrimination effect of the sample, forming a dynamic adaptive optimization mechanism. This mechanism effectively improves the adaptability and detection accuracy of the detection system under diversified scenarios and individual differences. For example, in some children, environmental noise has a more significant impact on writing abnormalities, and the model can automatically increase the weight of environmental features, thereby achieving more accurate intelligent discrimination and personalized detection.

[0142] wherein the mean score of the correlation of the feature in the historical samples with the abnormal behavior discrimination, specifically includes:

[0143] During long-term monitoring and historical data accumulation, the system statistically analyzes the correlation between each type of input feature (including handwriting behavior features, environmental parameter features and emotion features) and the discrimination result. Specifically, for the abnormal behaviors in the historical samples that have been identified and labeled, the system calculates the role and contribution of each type of feature in the abnormal behavior discrimination. This process can be achieved by analyzing the synchronicity, cooperativity or correlation between feature changes and abnormal behavior occurrence. For example, if a certain type of feature (such as emotional fluctuations) frequently appears synchronously with the abnormal discrimination result, it indicates that the correlation between the feature and the abnormal discrimination is high. After the system aggregates the correlation statistics of each feature, it takes the mean value as the overall correlation score of the current feature category in the historical samples. This mean score is used in the subsequent weight distribution step, reflecting the actual contribution of each type of feature in the abnormal discrimination, and ensuring the scientificity and individualization of the model weight adjustment.

[0144] wherein the recognition model adjusts the correlation score of each feature weight in real time based on the discrimination accuracy of the sample after each new sample analysis, forms a dynamic adaptive optimization mechanism, and specifically includes:

[0145] The system records the accuracy of the current model in discriminating the sample each time it processes a new piece of segmented correlation data. If the discrimination result is consistent with the actual situation, it is considered that the current feature weight setting is reasonable. If the discrimination result deviates, the system traces the actual correspondence between the sample feature value and the abnormal label, and analyzes the feature categories that cause the discrimination error. Based on the analysis result, the system fine-tunes the correlation score: for the feature categories that contribute more in this discrimination, increase their correlation score and weight; for the feature categories that contribute less or mislead, appropriately reduce their correlation score and weight. Through the continuous input of new samples and discrimination accuracy feedback, the system continuously adjusts the weights of various features, so that the model can adapt to the behavior characteristics of different children and environmental changes, thereby maintaining or improving the overall detection accuracy. This dynamic adaptive optimization mechanism does not require human intervention and can automatically adapt to new data and new scenarios, realizing the continuous evolution and optimization of the intelligent model.

[0146] In a preferred embodiment of the present application, the attribution distribution data is used to identify model adaptive adjustment and individualized intervention, specifically including:

[0147] According to the attribution label distribution of each abnormal behavior in the historical sample in the abnormal behavior candidate data set, the occurrence frequency of each attribution type in the target child historical data is counted, and an individualized attribution frequency vector is generated;

[0148] According to the individualized attribution frequency vector, the attribution type with the highest frequency of occurrence in the target child historical data is identified, the weight coefficient of the feature category corresponding to the attribution type is set to a first preset weight, and the weight coefficients of other feature categories are set to a second preset weight and a third preset weight respectively, the first preset weight is greater than the second preset weight, and the second preset weight is greater than the third preset weight;

[0149] When the identification model discriminates a new sample, the first preset weight, the second preset weight and the third preset weight are used to weight each feature category, the first preset weight is used to weight the high-sensitive feature category, and the second preset weight and the third preset weight are used to weight the medium-sensitive feature category and the low-sensitive feature category respectively.

[0150] In the embodiments of the present application, by using the attribution distribution data to adaptively adjust the feature weight coefficients of the identification model, the sensitivity and accuracy of the model in individualized detection are improved. First, the system determines the sensitivity levels of different feature categories according to the occurrence frequency of each attribution type in the target child historical data, and divides the weight coefficients into multiple levels accordingly. The feature category corresponding to the attribution type with the highest frequency of occurrence is set to the highest weight, so that the model pays more attention to the features that have a major impact on the child during the discrimination process, ensuring that the key behavior characteristics are fully reflected.

[0151] Through hierarchical weighting of weights, the system can effectively enhance the response capability of the model to high-sensitive features, reduce the interference of secondary or irrelevant features on abnormality discrimination, and thus reduce the misjudgment or omission caused by individual differences. For example, if a child's emotional fluctuation is the main abnormality attribution, the model will automatically increase the weight of emotional features, making it more sensitive to emotional changes in new sample discrimination. Conversely, for children who are less affected by environmental factors, the system automatically reduces the weight of environmental features to avoid misleading the detection results by invalid features.

[0152] Overall, this scheme can flexibly optimize the model structure according to the behavior and attribution characteristics of children, realize the individualization and precision of writing disorder detection, and significantly improve the adaptive ability and practical application effect of the system to diversified disorder performance.

[0153] The setting method of the first preset weight, the second preset weight and the third preset weight specifically includes:

[0154] After the system performs attribution distribution statistics on the historical samples of the target child, it will set a clear numerical range for each weight coefficient according to the importance of each feature category in abnormal behavior discrimination, combined with a large amount of historical detection data and artificial experience. In specific operation, the system first determines the feature category with the highest frequency in the attribution distribution of the target child, and sets its weight coefficient as the first preset weight. The numerical value of the first preset weight is generally set as the maximum value among all feature weight coefficients, ensuring that this feature category has the most significant influence in subsequent model discrimination.

[0155] For feature categories with a medium frequency in the attribution distribution, the system sets their weight coefficients as the second preset weight, with a numerical value between the first preset weight and the third preset weight, reflecting their moderate influence on the discrimination result. The remaining feature categories with the lowest frequency are assigned the third preset weight, which is the minimum value among all weights, and only has an impact on the discrimination result in a few cases.

[0156] The specific numerical value of the weight coefficient can be adjusted in combination with the field standard, the performance of the validation set of model training, and the actual discrimination effect. For example, before actual system deployment, a weight level template can be pre-set according to the statistical analysis results of previous large sample data, and the weight levels can be fine-tuned and optimized in the subsequent use process through dynamic feedback of model discrimination accuracy. Through this multi-level, data-driven weight setting method, the recognition model can ensure good individual adaptation and detection performance under different children and different attribution modes.

[0157] In a preferred embodiment of the present application, according to multi-scale behavior timing statistical data, a behavior sequence clustering method is used to cluster analyze the timing distribution of suspected disorder behaviors in the same category in the whole writing task process, and identify disorder behavior patterns with high frequency recurrence or periodicity characteristics, including:

[0158] According to multi-scale behavior timing statistical data, for each suspected disorder behavior, the dynamic time warping algorithm is used to calculate the timing similarity in different time periods, and the behavior sequences with timing similarity higher than the preset timing similarity threshold are divided into the same behavior cluster, and the repeatedly appearing disorder behavior patterns are identified;

[0159] For the disorder behavior sequences contained in each behavior cluster, the sliding autocorrelation analysis method is used to calculate the cycle length and cycle strength of each disorder behavior sequence, wherein the cycle length is the mean of the behavior interval, and the cycle strength is the maximum autocorrelation value obtained by the sliding autocorrelation analysis method;

[0160] When the occurrence frequency or cycle strength of the disorder behaviors in a behavior cluster continuously exceeds the corresponding preset mode threshold, the behavior cluster is determined as a disorder behavior pattern with periodicity characteristics, and the corresponding cycle length and cycle strength are output as the statistical results of the periodicity characteristics.

[0161] In the embodiment of the present application, through clustering analysis of multi-scale behavior timing statistical data, high frequency recurrence and periodicity characteristics of suspected disorder behaviors in the same category in the whole writing task process can be effectively identified. Specifically, the system uses the dynamic time warping algorithm to calculate the timing similarity of each suspected disorder behavior in different time periods, and the behavior sequences with similarity higher than the threshold are classified into the same cluster. This not only can aggregate abnormal behaviors showing high repeatability, but also can accurately mine disorder behavior patterns hidden in big data.

[0162] Further, for each behavior cluster, the system quantitatively analyzes the periodicity of the disorder behaviors in the cluster by the sliding autocorrelation analysis method. The output of statistical indicators such as cycle length and cycle strength enables the system to specifically reflect the repetition frequency and its significance of the disorder behavior pattern. When the occurrence frequency or cycle strength of the disorder behaviors in a cluster continuously exceeds the preset mode threshold, the system can automatically determine that the cluster is a disorder behavior pattern with high frequency recurrence or periodicity characteristics, and output the specific statistical characteristics as the quantitative results of the periodic disorder pattern.

[0163] This scheme not only improves the automatic recognition ability of complex obstacle behavior patterns, but also provides accurate data support for subsequent personalized intervention. For example, if a child has a fixed interval of dysgraphia in the writing process, the system can accurately identify the periodic behavior pattern and quantify the periodic characteristics, providing scientific basis for educators and parents to develop targeted intervention strategies, significantly enhancing the practicality and intelligence level of the writing disorder detection system.

[0164] Among them, according to the multi-scale behavior time sequence statistical data, the time sequence similarity of each suspected obstacle behavior in different time periods is calculated by using the dynamic time warping algorithm, which specifically includes:

[0165] After obtaining the multi-scale behavior time sequence statistical data, the system first extracts the performance sequence of each suspected obstacle behavior in different time periods during the entire writing task. Then, the dynamic time warping algorithm is used to compare these sequences pairwise, and the similarity of behaviors in different time periods is evaluated by comparing the change trend of each behavior sequence in the time dimension. This algorithm can stretch, align and compare the behavior patterns with different speeds or intervals on the time axis, so as to judge whether the overall shape of different sequences is similar. If the change patterns of two behavior sequences are highly consistent, even if their actual time points are different, the algorithm can still determine that they are high-similarity behaviors. The system classifies behavior sequences with a time sequence similarity higher than a preset threshold into the same behavior cluster, and finally realizes the classification and identification of high-frequency recurring or repetitive obstacle behaviors.

[0166] Among them, for each obstacle behavior sequence included in the behavior cluster, the sliding autocorrelation analysis method is used to calculate the period length and period intensity of each obstacle behavior sequence, which specifically includes:

[0167] After the behavior cluster is determined, the system performs sliding autocorrelation analysis on the obstacle behavior sequences in each cluster. The specific operation is that the system traverses the time sequence of the obstacle behavior by moving the analysis window, compares the sequences at different times with themselves, and calculates the similarity of the behavior sequences at different time delays. By analyzing the repeated matching of the behavior sequence itself at different intervals, the system can identify the periodic repetition pattern existing in it. The determination of the period length depends on finding the time interval that makes the sequence most consistent with itself, which is usually manifested as the average time of the behavior interval; the period intensity is reflected as the consistency degree of the behavior sequence under this period, that is, the significance of the repeated appearance of the sequence. Finally, the system outputs the period length and period intensity of the obstacle behavior sequence in each cluster, providing data support for the fine analysis and intervention of periodic obstacle behavior patterns.

[0168] The embodiments of the present application also provide a child writing obstacle detection system based on artificial intelligence, which comprises:

[0169] a data collection module configured to acquire original handwriting data of a target child when the target child completes a designated writing task in a preset monitoring environment, and synchronously collect environmental parameter data and emotional feature data at a corresponding time, to obtain a synchronous multi-source original data set;

[0170] a data preprocessing module configured to perform data preprocessing according to the synchronous multi-source original data set, segment the original handwriting data according to a time window, associate the original handwriting data with the environmental parameter data and the emotional feature data of the same period, and obtain a segmented associated data set;

[0171] an abnormality discrimination module configured to input the segmented associated data set into a recognition model trained by historical big data, automatically analyze handwriting behavior of each segment and corresponding environmental and emotional features, discriminate abnormalities in the segmented handwriting behavior, and generate preliminary abnormality labeling data;

[0172] an abnormality attribution module configured to perform abnormality behavior attribution processing according to the preliminary abnormality labeling data, exclude abnormalities caused by environmental or emotional influences, mark abnormalities not affected by environmental or emotional influences as suspected disorder behaviors, and obtain a disorder behavior candidate data set;

[0173] a behavior analysis module configured to perform continuity and periodicity analysis according to the disorder behavior candidate data set, identify disorder behaviors that appear continuously multiple times, and form a disorder behavior set;

[0174] a report generation module configured to generate a writing disorder detection report of the target child according to the disorder behavior set, the report content including specific types of suspected disorder behaviors, occurrence frequencies and associated contexts.

[0175] It should be noted that the system corresponds to the above method, and all implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0176] Embodiments of the present application also provide a computing device, comprising a processor and a memory storing a computer program, when the computer program is run by the processor, the method described above is executed. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0177] Embodiments of the present application also provide a computer readable storage medium storing instructions, when the instructions are run on a computer, the computer executes the method described above. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0178] The above is the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the principles described in the present application, can also be made several improvements and refinements, these improvements and refinements should also be considered the scope of protection of the present application.

Claims

1. A method for detecting a writing disorder of a child based on artificial intelligence, the method comprising: The method comprises: obtaining original handwriting data of a target child when completing a designated writing task in a preset monitoring environment, and synchronously collecting environmental parameter data and emotional feature data at the corresponding moment to obtain a synchronous multi-source original data set; According to the synchronous multi-source original data set, data preprocessing is performed, the original handwriting data is segmented according to a time window, and the original handwriting data is associated with the environmental parameter data and the emotional feature data in the same period to obtain a segmented associated data set, comprising: According to the original handwriting data, pre-segmentation is performed according to a preset initial time window to obtain an initial segmented handwriting data set; For each initial segmented handwriting data, analyze based on behavior feature points, when any behavior feature point is detected, the position of the behavior feature point is taken as a new segmentation starting point, the segmentation boundary is dynamically adjusted, and a dynamic segmented handwriting data set driven by behavior feature points is formed, the behavior feature points include handwriting speed change, pressure mutation and trajectory turning; The dynamic segmented handwriting data set is associated with the environmental parameter data and the emotional feature data in the corresponding time period in segments to obtain a segmented associated data set; The segmented associated data set is input into a recognition model trained by historical big data, and the handwriting behavior of each segment and the corresponding environment and emotional features are automatically analyzed, the abnormality in the segmented handwriting behavior is discriminated, and preliminary abnormality labeling data is generated; According to the preliminary abnormality labeling data, abnormal behavior attribution processing is performed, the abnormality caused by environmental or emotional influence is excluded, and the abnormality not affected by the environment or emotion is marked as suspected disorder behavior to obtain a disorder behavior candidate data set, comprising: According to the preliminary abnormality labeling data, the environmental parameter data and the emotional feature data corresponding to each segmented handwriting behavior are analyzed for correlation, and a correlation score R between the abnormal behavior and the environmental or emotional change is calculated, the correlation score R is used to measure the influence degree of environmental or emotional factors on the abnormal behavior; According to the correlation score R, set multiple attribution level thresholds, divide the abnormal behavior into strong correlation type, medium correlation type and weak correlation type attribution labels, respectively corresponding to R>T1, T2≤R≤T1, R For the abnormal behaviors attributed to strong correlation type and medium correlation type, mark them as environmental or emotional influence type abnormalities respectively; for the abnormal behavior attributed to weak correlation type, mark it as suspected disorder behavior to obtain a disorder behavior candidate data set with multiple labels; For each abnormal behavior in the disorder behavior candidate data set, the distribution of the attribution label in the historical sample is counted to form an attribution distribution data; According to the disorder behavior candidate data set, perform continuity and periodicity analysis, identify the disorder behaviors that appear multiple times continuously, and form a disorder behavior set; According to the disorder behavior set, a writing disorder detection report of the target child is generated, and the report content includes the specific type, occurrence frequency and associated context of the suspected disorder behavior; According to the synchronous multi-source original data set, data preprocessing is performed, the original handwriting data is segmented according to a time window, and the original handwriting data is associated with the environmental parameter data and the emotional feature data in the same period to obtain a segmented associated data set According to the preliminary anomaly labeling data, abnormal behavior attribution processing is performed, and the abnormal behaviors caused by environmental or emotional influences are excluded, and the abnormal behaviors not affected by the environment or emotions are marked as suspected disorder behaviors to obtain a disorder behavior candidate data set. 2.The AI-based child writing disorder detection method of claim 1, wherein, The segmented correlation data set is input into the recognition model trained by historical big data, and the handwriting behavior of each segment and its corresponding environmental and emotional features are automatically analyzed to determine the abnormality in the segmented handwriting behavior, and generate preliminary anomaly labeling data, including: For each segmented correlation data, a multi-source input feature group containing handwriting behavior feature vector, environmental parameter feature vector and emotional feature vector is constructed, and the multi-source input feature group is used as the input data of the recognition model; The recognition model automatically adjusts the weight coefficients of handwriting behavior features, environmental parameter features and emotional features according to the multi-source input feature group, and uses an adaptive weight distribution method to make the influence factors of each feature category dynamically change when determining abnormal behaviors in different scenarios; The recognition model outputs an abnormal probability score and an abnormal type label for each segmented correlation data, wherein the abnormal probability score is used to quantify the confidence of the segmented handwriting behavior as abnormal, and the abnormal type label is used to distinguish environmental impact type, emotional impact type or suspected disorder behavior; According to the abnormal probability score and the abnormal type label, preliminary anomaly labeling data is generated. 3.The AI-based child writing disorder detection method of claim 1, wherein, According to the disorder behavior candidate data set, continuity and periodicity analysis is performed to identify disorder behaviors that occur multiple times, and a disorder behavior set is formed, including: For each suspected disorder behavior, the occurrence frequency and duration of the behavior in different time scales are counted based on a pre-set multi-scale time window to obtain multi-scale behavior time series statistical data; According to the multi-scale behavior time series statistical data, a behavior sequence clustering method is used to cluster analyze the time series distribution of suspected disorder behaviors of the same category in the entire writing task process to identify disorder behavior patterns with high frequency recurrence or periodicity characteristics; According to the disorder behavior pattern, the adaptive trigger threshold of the disorder behavior is automatically adjusted, and when the suspected disorder behavior occurs continuously or continuously for more than the adaptive trigger threshold in the pre-set multi-scale time window, it is determined as a continuous or periodic disorder behavior and included in the disorder behavior set; A disorder behavior set containing disorder behavior type, occurrence time period and periodicity statistical information is generated. 4.The AI-based child writing disorder detection method of claim 1, wherein For each initial segmented handwriting data, the behavior feature points are analyzed, and when any behavior feature point is detected, the position of the behavior feature point is taken as a new segment starting point, and the segment boundary is dynamically adjusted to form a dynamic segmented handwriting data set driven by behavior feature points, including: For each initial segmented handwriting data, the average handwriting speed, average pressure and trajectory angle change rate in the segment are calculated respectively; In the initial time window range, the speed change value, pressure change value and angle change value between adjacent data points are calculated, and when any one of the speed change value, pressure change value or angle change value exceeds the corresponding pre-set dynamic threshold, the current data point is determined as a behavior feature point; When a plurality of behavior feature points are continuously detected within a preset time threshold, a density clustering method is used to combine adjacent behavior feature points into a segment starting point; Noise filtering is performed on the segment data set segmented by the behavior feature points, and abnormal short segments with a segment length lower than a set segment length threshold are removed, to obtain a dynamic segmented handwriting data set. 5.The AI-based child writing disorder detection method of claim 2, wherein, The recognition model automatically adjusts the weight coefficients of the handwriting behavior features, environmental parameter features, and emotional features according to the multi-source input feature groups, and uses an adaptive weight distribution method to dynamically change the influence factors of each feature category when discriminating abnormal behaviors in different scenarios, including: The handwriting behavior feature vector, the environment parameter feature vector and the emotion feature vector are respectively dynamically updated according to the correlation of the abnormal behavior discrimination results in the historical sample data , , ; The following weight dynamic updating formula is used:    ， wherein, is a weight coefficient of the i-th feature, is a mean score of the relevance of the feature to the abnormal behavior discrimination in the historical samples, i∈{b,e,c}, representing the handwriting behavior, the environmental parameter and the emotional feature, respectively. After analyzing each new sample, the recognition model adjusts the relevance scores of the feature weights in real time based on the discrimination accuracy of the sample, forming a dynamic adaptive optimization mechanism. 6.The AI-based child writing disorder detection method of claim 1, wherein, The attribution distribution data are used for adaptive adjustment of the recognition model and individualized intervention, specifically including: According to the attribution label distribution of each abnormal behavior in the historical sample in the candidate data set of the abnormal behavior, the occurrence frequency of each attribution type in the historical data of the target child is counted, and an individualized attribution frequency vector is generated; According to the individualized attribution frequency vector, the attribution type with the highest occurrence frequency in the historical data of the target child is identified, the weight coefficient of the corresponding feature category is set as a first preset weight, and the weight coefficients of other feature categories are set as a second preset weight and a third preset weight, respectively, the first preset weight being greater than the second preset weight, and the second preset weight being greater than the third preset weight; When discriminating a new sample, the first preset weight, the second preset weight, and the third preset weight are used to weight each feature category, the first preset weight is used to weight the high-sensitivity feature category, and the second preset weight and the third preset weight are used to weight the medium-sensitivity feature category and the low-sensitivity feature category, respectively. 7.The AI-based child writing disorder detection method of claim 3, wherein, According to the multi-scale behavior time sequence statistical data, a behavior sequence clustering method is used to cluster analyze the time sequence distribution of suspected abnormal behaviors of the same category in the entire writing task process, and identify abnormal behavior patterns with high-frequency recurrence or periodic characteristics, including: According to the multi-scale behavior time sequence statistical data, for each suspected abnormal behavior, the time sequence similarity of the behavior in different time periods is calculated by using a dynamic time warping algorithm, and behavior sequences with a time sequence similarity higher than a preset time sequence similarity threshold are divided into the same behavior cluster, to identify repeated abnormal behavior patterns; For the abnormal behavior sequences included in each behavior cluster, a sliding autocorrelation analysis method is used to calculate the cycle length and cycle strength of each abnormal behavior sequence, wherein the cycle length is the mean of the behavior interval, and the cycle strength is the maximum autocorrelation value obtained by the sliding autocorrelation analysis method; When the occurrence frequency or cycle strength of the abnormal behaviors in a behavior cluster continuously exceeds the corresponding preset pattern threshold, the behavior cluster is determined as an abnormal behavior pattern with periodic characteristics, and the cycle length and cycle strength thereof are output as the statistical results of the periodic characteristics.

8. An artificial intelligence-based children's writing disorder detection system, characterized by, The system is applied to the method of any one of claims 1 to 7, and the system comprises: A data acquisition module is configured to acquire original handwriting data of a target child when the target child completes a designated writing task in a preset monitoring environment, and synchronously acquire environmental parameter data and emotional feature data at a corresponding time, to obtain a synchronous multi-source original data set; A data preprocessing module is configured to perform data preprocessing according to the synchronous multi-source original data set, segment the original handwriting data according to a time window, associate the original handwriting data with the environmental parameter data and the emotional feature data in the same period, and obtain a segmented associated data set; An abnormality discrimination module is configured to input the segmented associated data set into a recognition model trained by historical big data, automatically analyze handwriting behavior of each segment and corresponding environmental and emotional features, discriminate abnormalities in the segmented handwriting behavior, and generate preliminary abnormality labeling data; An abnormality attribution module is configured to perform abnormality behavior attribution processing according to the preliminary abnormality labeling data, exclude abnormalities caused by environmental or emotional influences, mark abnormalities not affected by environmental or emotional influences as suspected disorder behaviors, and obtain a disorder behavior candidate data set; A behavior analysis module is configured to perform continuity and periodicity analysis according to the disorder behavior candidate data set, identify disorder behaviors that occur continuously multiple times, and form a disorder behavior set; A report generation module is configured to generate a writing disorder detection report of the target child according to the disorder behavior set, and the report content includes specific types of suspected disorder behaviors, occurrence frequencies and associated contexts.

Images