Learning ability dynamic evaluation method and system based on writing track of touch pen
By constructing a cluster of writing behaviors and using the median vector and Euclidean distance to calculate sequence deviation, the problem of difficulty in distinguishing between unqualified writing trajectories and abnormal rhythms in existing technologies is solved, thus achieving a more accurate assessment of learning ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 北京爱宾果科技有限公司
- Filing Date
- 2026-04-15
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to distinguish and assess two different types of problems in students' writing trajectories: unqualified trajectories and abnormal rhythms. In particular, existing methods are unable to sensitively capture issues such as severe trajectory deviations or localized rhythmic abnormalities.
By constructing a cluster of writing behaviors, standard writing behavior sequences are generated using unsupervised clustering and median vectors. The sequence deviation is calculated by combining Euclidean distance to determine whether the writing trajectory is qualified, and local anomalies are identified by rhythm ratio.
It improves the effectiveness of learning ability assessment, avoids misjudgments caused by distorted standards, and can more accurately reflect the degree of deviation between the target object and normal behavior, thereby improving the accuracy of assessment and the ability to identify subtle anomalies.
Smart Images
Figure CN122045902A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of learning assessment, specifically a dynamic assessment method for learning ability based on stylus writing trajectory. Background Technology
[0002] In digital learning environments, styluses are widely used for data collection during student responses. Through built-in sensors, they record behavioral parameters such as coordinates, touch pressure, and tilt angle at a fixed frequency, generating raw behavioral data arranged chronologically. This data can reconstruct the writing trajectory, supporting retrospective analysis of the response process, and has become an important means of obtaining procedural behavioral evidence in intelligent education systems.
[0003] Existing methods typically process the aforementioned data holistically, such as calculating the average writing speed and total pause duration, or directly comparing the geometric distance between the final trajectory and the standard answer, using this as a basis for judgment. However, such methods have significant limitations: when the target student's writing trajectory deviates significantly from the normal range due to complete inability to answer, their rhythmic characteristics (such as local acceleration or pauses) are no longer diagnostically meaningful; conversely, when the trajectory is basically in line with the standard but has local rhythmic abnormalities (such as a slow start or skipped steps at the end), the overall indicators often fail to capture them sensitively, making it difficult for existing technologies to assess the two different types of problems: unqualified trajectory and rhythmic abnormalities. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a dynamic evaluation method and system for learning ability based on stylus writing trajectory, which solves the technical problems mentioned in the background by introducing secondary learning evaluation.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] In a first aspect, the present invention discloses a dynamic evaluation method for learning ability based on stylus writing trajectory, comprising the following steps:
[0007] S1. Within the target group, select N sample objects with object IDs;
[0008] S2. Construct N writing behavior sequences for N sample objects;
[0009] Among them, the writing behavior sequence represents the node behavior vector arranged in M sampling timestamps, and the correlation between adjacent sampling timestamps represents the writing rhythm ratio of multidimensional writing behavior changes per unit time.
[0010] S3. Embed and encode N writing behavior sequences to obtain N sequence representation vectors; where each sequence representation vector is associated with the object ID of the sample object.
[0011] S4. Perform unsupervised clustering on the N sequence representation vectors to obtain K clusters of writing behaviors;
[0012] S5. Construct the standard writing behavior sequences for each of the K writing behavior clusters;
[0013] S6. Obtain the actual writing behavior sequence of the target object, and determine the most similar behavior cluster of the target object from K writing behavior clusters based on the actual writing behavior sequence;
[0014] S7. Extract the standard writing behavior sequence of the most similar behavior cluster and calculate the sequence deviation between it and the real writing behavior sequence;
[0015] S8. If the sequence deviation is greater than the first threshold, the writing trajectory of the target object is deemed unqualified; otherwise, the writing rhythm ratio of the target object is extracted for learning anomaly label evaluation.
[0016] In some specific embodiments, N writing behavior sequences for N sample objects are constructed, including:
[0017] S2-1. Select any target sample object from N sample objects;
[0018] S2-2, Locate the target sample object on the timeline and use the stylus to write M sampling timestamps;
[0019] S2-3. Based on the time order of the M sample timestamps, generate M node numbers arranged in a time sequence.
[0020] S2-4. Obtain the writing behavior parameters of M node numbers; where the writing behavior parameters include at least: the coordinates of the stylus at the sampling timestamp, the touch pressure, and the touch tilt angle;
[0021] S2-5. Perform dimensionless processing on the writing behavior parameters and concatenate them into a node behavior vector of the target sample object at the node index.
[0022] S2-6. Arrange the M node behavior vectors based on the node index to generate the writing behavior sequence;
[0023] S2-7. In the writing behavior sequence, associate the sampling timestamps of the M node behavior vectors.
[0024] S2-8. Based on the sampling timestamps corresponding to the writing behavior sequence and its node behavior vectors, calculate the M-1 writing rhythm ratios between adjacent sampling timestamps; where the writing rhythm ratio is inherited from the node number of the previous sampling timestamp.
[0025] In some specific embodiments, based on the writing behavior sequence and its associated sampling timestamps, M-1 writing rhythm ratios between adjacent sampling timestamps are calculated, including:
[0026] S2-8-1. Select the behavior vectors of adjacent nodes in the writing behavior sequence one pair at a time;
[0027] S2-8-2, Calculate the Euclidean distance between the behavior vectors of adjacent nodes;
[0028] S2-8-3. Based on the behavior vectors of adjacent nodes, obtain their respective associated adjacent sampling timestamps;
[0029] S2-8-4. Calculate the writing duration between adjacent sampling timestamps;
[0030] S2-8-5. Traverse the writing behavior sequence and perform a ratio operation between the Euclidean distance and the writing duration of adjacent node behavior vectors to obtain M-1 writing rhythm ratios.
[0031] In some specific embodiments, standard writing behavior sequences are constructed for each of the K writing behavior clusters, including:
[0032] S5-1. Based on the object ID inherited from the sequence representation vector, anchor G sample objects for each writing behavior cluster; where G is the number of sample objects within the cluster.
[0033] S5-2. Obtain the learning scores of the G sample objects;
[0034] S5-3. Based on the learning scores of the G sample objects, select J sample objects whose learning scores are greater than the high score threshold from high to low.
[0035] S5-4. Extract J writing behavior sequences corresponding to J sample objects;
[0036] S5-5. Align the J writing behavior sequences according to the node index of their node behavior vectors to generate an instantaneous vector set with M node indices; where the instantaneous vector set represents the J node behavior vectors that share the same node index.
[0037] S5-6. Based on the behavior vectors of the J nodes, construct the median vector of the instantaneous vector set;
[0038] S5-7. Traverse the instantaneous vector set of M node indices, and repeat the construction of the median vector until the median vector of M node indices is obtained.
[0039] S5-8. Arrange the median vectors of the M node indices in order to construct the standard writing behavior sequence of the writing behavior cluster;
[0040] S5-9. Traverse the K writing behavior clusters and repeat the construction of the standard writing behavior sequence until the standard writing behavior sequence of each of the K writing behavior clusters is obtained.
[0041] In some specific embodiments, the median vector of the instantaneous vector set is constructed based on the J node behavior vectors, including:
[0042] S5-6-1. Traverse the current dimension of the node behavior vectors, extract the writing behavior parameters of J node behavior vectors in the instantaneous vector set in the current dimension, and mark them as J parameters of the same dimension.
[0043] S5-6-2. Sort the J parameters of the same dimension in ascending order of their numerical values;
[0044] S5-6-3. If J is odd, select the writing behavior parameter located at the (J+1) / 2th position after ascending sorting and mark it as the median parameter of the current dimension.
[0045] S5-6-4. If J is even, calculate the average of the writing behavior parameters at the J / 2 and (J / 2+1)th positions after ascending order, and mark it as the median parameter of the current dimension.
[0046] S5-6-5. Repeat the marking of the median parameters until the median parameters for each dimension are obtained;
[0047] S5-6-6. Concatenate the median parameters of each dimension in the original dimension order to form the median vector of the instantaneous vector set.
[0048] In some specific embodiments, based on the actual writing behavior sequence in K writing behavior clusters, the most similar behavior cluster of the target object is determined, including:
[0049] S6-1. Encode the real writing behavior sequence into a real representation vector;
[0050] S6-2. Calculate the K similarities between the true representation vector and the cluster center vectors of the K writing behavior clusters;
[0051] S6-3. Select the maximum similarity among the K similarities, and locate the most similar behavior cluster of the target object based on the maximum similarity.
[0052] In some specific embodiments, the learning anomaly labels of the target object are evaluated, including:
[0053] S8-1. Based on the actual writing behavior sequence of the target object, calculate the M-1 actual writing rhythm ratios of the target object; where the actual writing rhythm ratio is inherited from the node number of the previous sampling timestamp.
[0054] S8-2. Based on the most similar behavior cluster of the target object, extract M-1 standard writing rhythm ratios of its standard writing behavior sequence; where the standard writing rhythm ratio is inherited from the node number of the presampled timestamp.
[0055] S8-3. Pair the M-1 real writing rhythm ratios with the M-1 standard writing rhythm ratios according to their shared node numbers and calculate the difference to obtain the M-1 rhythm ratio deviations.
[0056] S8-4. Among M-1 rhythm ratio deviations, identify Q abnormal deviations that exceed the second threshold;
[0057] S8-5. Extract P abnormal sampling timestamps corresponding to the Q abnormal deviation; where P≤Q;
[0058] S8-6. Based on P abnormal sampling timestamps, identify the abnormal clustering period;
[0059] S8-7. Based on the mapping labels pre-stored in the label database during the abnormal clustering period, evaluate the learning abnormal labels of the target object.
[0060] In some specific embodiments, based on P abnormal sampling timestamps, abnormal clustering periods are identified, including:
[0061] S8-6-1. Arrange the P abnormal sampling timestamps in ascending order of time.
[0062] S8-6-2. Select two adjacent abnormal sampling timestamps in ascending order one by one to obtain P-1 pairs of adjacent timestamps.
[0063] S8-6-3. Calculate the time interval between each pair of adjacent abnormal sampling timestamps to obtain P-1 time intervals;
[0064] S8-6-4. Compare P-1 time intervals with the preset time adjacency threshold T;
[0065] S8-6-5. Based on the comparison results of P-1 time intervals with the time adjacency threshold T, group the P abnormal sampling timestamps into abnormal time periods.
[0066] The abnormal time period grouping is as follows: if the time interval between any two adjacent abnormal sampling timestamps is less than or equal to the time adjacency threshold T, then the consecutive abnormal sampling timestamps are merged into the same abnormal cluster time period; wherein, the abnormal cluster time period is defined by the earliest and latest sampling timestamps contained therein.
[0067] This invention provides a dynamic evaluation method for learning ability based on stylus writing trajectory, which has the following beneficial effects:
[0068] This invention, when constructing standard writing behavior sequences for each writing behavior cluster, extracts node behavior vectors from high-scoring samples within the cluster and generates a median vector for each dimension's parameters using the median method. This effectively suppresses occasional abnormal writing behaviors in individual high-scoring samples (such as sudden pressure changes or trajectory jitter caused by brief distraction), making the standard sequence more representative of the stable behavior patterns of outstanding students in that cluster. Furthermore, the median vectors of each node are arranged sequentially to form a standard writing behavior sequence, and the Euclidean distance is calculated point-by-point with the actual writing behavior sequence to obtain the sequence deviation. If this deviation is greater than a first threshold, the writing trajectory is deemed unqualified; otherwise, local anomalies are identified based on the deviation between the actual writing rhythm ratio and the standard writing rhythm ratio. Therefore, because the standard sequence itself is resistant to interference due to the median vector, the sequence deviation can more accurately reflect the degree of deviation between the target object and normal behavior, avoiding misjudgments caused by standard distortion and improving the effectiveness of learning ability assessment.
[0069] Secondly, the present invention discloses a dynamic evaluation system for learning ability based on stylus writing trajectory, used to execute the dynamic evaluation method for learning ability based on stylus writing trajectory described in the first aspect, the system comprising:
[0070] The sample object selection module is used to select N sample objects with object IDs within the target group.
[0071] The sample sequence construction module is used to construct N writing behavior sequences for N sample objects;
[0072] Among them, the writing behavior sequence represents the node behavior vector arranged in M sampling timestamps, and the correlation between adjacent sampling timestamps represents the writing rhythm ratio of multidimensional writing behavior changes per unit time.
[0073] The embedding vector encoding module is used to embed and encode N sequences of writing actions to obtain N sequence representation vectors; where each sequence representation vector is associated with the object ID of the sample object.
[0074] The behavior clustering module is used to perform unsupervised clustering of N sequence representation vectors to obtain K writing behavior clusters;
[0075] The standard sequence construction module is used to construct the standard writing behavior sequences for each of the K writing behavior clusters.
[0076] The similarity cluster determination module is used to obtain the actual writing behavior sequence of the target object, and determine the most similar behavior cluster of the target object from K writing behavior clusters based on the actual writing behavior sequence;
[0077] The sequence deviation calculation module is used to extract the standard writing behavior sequence of the most similar behavior cluster and calculate the sequence deviation between it and the real writing behavior sequence.
[0078] The secondary deviation evaluation module is used to determine whether the writing trajectory of the target object is unqualified if the sequence deviation is greater than the first threshold; otherwise, the writing rhythm ratio of the target object is extracted for learning anomaly label evaluation.
[0079] Compared with the prior art, the beneficial effects of the learning ability dynamic evaluation system based on stylus writing trajectory of the present invention are the same as those of the above-mentioned learning ability dynamic evaluation method based on stylus writing trajectory, so they will not be repeated here. Attached Figure Description
[0080] Figure 1 This is a flowchart illustrating the dynamic evaluation method for learning ability based on stylus writing trajectory according to the present invention.
[0081] Figure 2 This invention is a flowchart illustrating one embodiment of the writing rhythm ratio;
[0082] Figure 3 This invention provides a schematic diagram of the calculation process for another embodiment of the writing rhythm ratio.
[0083] Figure 4 This invention is a schematic diagram illustrating the process of constructing the median vector;
[0084] Figure 5 This is a structural block diagram of the dynamic learning ability assessment system based on stylus writing trajectory of the present invention. Detailed Implementation
[0085] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0086] Example 1: Please refer to Figures 1 to 4 This invention provides a dynamic assessment method for learning ability based on stylus writing trajectory, comprising the following steps:
[0087] S1. Within the target group, select N sample objects with object IDs;
[0088] Here, the sample object refers to the student who participated in the same learning task and completed the corresponding questions; the object ID refers to the number used to identify the student.
[0089] S2. Construct N writing behavior sequences for N sample objects;
[0090] Among them, the writing behavior sequence represents the node behavior vector arranged in M sampling timestamps, and the correlation between adjacent sampling timestamps represents the writing rhythm ratio of multidimensional writing behavior changes per unit time.
[0091] S3. Embed and encode N writing behavior sequences to obtain N sequence representation vectors; where each sequence representation vector is associated with the object ID of the sample object.
[0092] Specifically, in this embodiment, the embedding encoding is implemented in the following way:
[0093] Each sequence of writing actions is input into a one-dimensional convolutional neural network, which contains convolutional layers, max pooling layers, and fully connected layers in sequence.
[0094] The convolutional layer uses a one-dimensional convolutional kernel with a size of 3 and a stride of 1 to perform sliding calculations on the sequence in the time dimension and output local feature maps;
[0095] The max pooling layer downsamples the local feature map to obtain the pooled feature map.
[0096] Finally, the pooled feature map is flattened and mapped to a 128-dimensional vector through a fully connected layer, which serves as the sequence representation vector for the sample object.
[0097] S4. Perform unsupervised clustering on the N sequence representation vectors to obtain K clusters of writing behaviors;
[0098] In this embodiment, the unsupervised clustering uses the K-means algorithm, and the number of clusters K is selected between 2 and 10 to make the writing behavior sequences within the same cluster as similar as possible, and the sequences between different clusters as different as possible.
[0099] S5. Construct the standard writing behavior sequences for each of the K writing behavior clusters;
[0100] S6. Obtain the actual writing behavior sequence of the target object, and determine the most similar behavior cluster of the target object from K writing behavior clusters based on the actual writing behavior sequence;
[0101] S7. Extract the standard writing behavior sequence of the most similar behavior cluster and calculate the sequence deviation between it and the real writing behavior sequence;
[0102] Specifically, in this embodiment, the formula for calculating the sequence bias is:
[0103] ;
[0104] in:
[0105] D represents the sequence deviation of the target object; the larger the value, the more significant the deviation of its writing behavior from the standard pattern.
[0106] M represents the total number of nodes in the writing sequence (i.e., the number of sampling timestamps), used for normalization to eliminate the influence of question length;
[0107] This represents the node behavior vector of the i-th node in the actual writing behavior sequence of the target object;
[0108] The node behavior vector representing the i-th node in the standard written behavior sequence of the most similar behavior cluster;
[0109] It should be noted that the sequence deviation calculation in this embodiment preserves the temporal order, because the temporal structure of the writing behavior directly reflects the task execution process. Even if two sequences contain similar local behaviors, if their order of occurrence is different (such as normal progression and backtracking jump), they are considered different patterns and should produce a higher deviation.
[0110] Since the standard writing behavior sequence and the actual writing behavior sequence are aligned in terms of node number (i.e., the i-th node corresponds to the same stage of answering the question), the average value of the node-by-node Euclidean distance can effectively characterize the overall deviation: when the deviation of each node is generally large, the sequence deviation is high; when most nodes are similar, the sequence deviation is low.
[0111] S8. If the sequence deviation is greater than the first threshold, the writing trajectory of the target object is deemed unqualified; otherwise, the writing rhythm ratio of the target object is extracted for learning anomaly label evaluation.
[0112] Specifically, the first threshold represents the 75th percentile of the target group sequence deviation; because the 75th percentile can cover most of the normal writing behavior samples in the target group, while effectively separating the tail abnormal individuals with large deviations, this percentile is used as the threshold for determining whether the writing trajectory is qualified.
[0113] This invention uses sequence deviation to determine whether the writing trajectory of the target object deviates significantly from the normal range (i.e., "unqualified"). If the deviation is too large, it is directly judged as unqualified without further analysis of rhythm details. Only when the trajectory is basically qualified will it enter the second evaluation of rhythm ratio deviation analysis to identify more subtle learning status problems (such as difficulty in starting to solve problems, stuttering in step transitions, or lack of answer verification).
[0114] Furthermore, secondary assessment can effectively reduce the misjudgment rate. For example, if a student's trajectory deviates due to physiological reasons in their hands but their rhythm is normal, the system can classify it as "abnormal trajectory" rather than "abnormal learning ability," avoiding incorrect attribution of learning status.
[0115] In this embodiment, step S2 specifically includes:
[0116] S2-1. Select any target sample object from N sample objects;
[0117] S2-2, Locate the target sample object on the timeline and use the stylus to write M sampling timestamps;
[0118] S2-3. Based on the time order of the M sample timestamps, generate M node numbers arranged in a time sequence.
[0119] S2-4. Obtain the writing behavior parameters of M node numbers; where the writing behavior parameters include at least: the coordinates of the stylus at the sampling timestamp, the touch pressure, and the touch tilt angle;
[0120] S2-5. Perform dimensionless processing on the writing behavior parameters and concatenate them into a node behavior vector of the target sample object at the node index.
[0121] Specifically, dimensionless processing can be achieved using the Min-Max normalization method, which linearly maps each writing behavior parameter to the [0,1] interval.
[0122] S2-6. Arrange the M node behavior vectors based on the node index to generate the writing behavior sequence;
[0123] S2-7. In the writing behavior sequence, associate the sampling timestamps of the M node behavior vectors.
[0124] S2-8. Based on the sampling timestamps corresponding to the writing behavior sequence and its node behavior vectors, calculate the M-1 writing rhythm ratios between adjacent sampling timestamps; where the writing rhythm ratio is inherited from the node number of the previous sampling timestamp.
[0125] It should be noted that the writing behavior sequence itself consists only of M node behavior vectors arranged in chronological order, and does not contain any explicit writing rhythm ratio.
[0126] Furthermore, the writing rhythm ratio is not only based on the calculation of coordinate displacement, but also integrates multi-dimensional behavioral parameters such as pressure and tilt angle, so as to achieve sensitive capture of abnormal behavioral patterns such as hesitation, pause, repeated correction or unstable handwriting during the writing process; it is used to characterize the rate of multi-dimensional behavioral changes between adjacent sampling points.
[0127] Specifically, in this embodiment, step S2-1 further includes:
[0128] S2-8-1. Select the behavior vectors of adjacent nodes in the writing behavior sequence one pair at a time;
[0129] S2-8-2, Calculate the Euclidean distance between the behavior vectors of adjacent nodes;
[0130] S2-8-3. Based on the behavior vectors of adjacent nodes, obtain their respective associated adjacent sampling timestamps;
[0131] S2-8-4. Calculate the writing duration between adjacent sampling timestamps;
[0132] S2-8-5. Traverse the writing behavior sequence and perform a ratio operation between the Euclidean distance and the writing duration of adjacent node behavior vectors to obtain M-1 writing rhythm ratios.
[0133] In this embodiment, the writing rhythm ratio is calculated by integrating multi-dimensional behavioral parameters, which can more comprehensively reflect the overall writing dynamics per unit time. Compared with the traditional speed index based solely on coordinate displacement, this method is more sensitive to subtle anomalies such as sudden pressure changes and tilt angle fluctuations, thereby improving the ability to identify behaviors such as hesitation, pauses, or unstable control.
[0134] In this embodiment, step S5 specifically includes:
[0135] S5-1. Based on the object ID inherited from the sequence representation vector, anchor G sample objects for each writing behavior cluster; where G is the number of sample objects within the cluster.
[0136] S5-2. Obtain the learning scores of the G sample objects;
[0137] It should be noted that the learning performance described in this embodiment represents objective indicators related to learning ability, rather than writing quality; the data source for learning performance is teacher scoring or automatic scoring by the system, etc.
[0138] S5-3. Based on the learning scores of the G sample objects, select J sample objects whose learning scores are greater than the high score threshold from high to low.
[0139] Specifically, in this embodiment, the high score threshold is set as the 70th percentile of the learning performance within the cluster;
[0140] This embodiment sets the high-score threshold at the 70th percentile of learning performance within the cluster, aiming to balance the representativeness and purity of the standard sequence. If the threshold is too high (e.g., 90th percentile), the number of high-scoring samples will be too small (especially in small clusters), making them susceptible to noise interference from individual samples; if the threshold is too low (e.g., 50th percentile), a large number of average-performing samples will be mixed in, weakening the typicality of the "excellent behavior pattern." Experiments show that the 70th percentile can guarantee J≥5 in most class sizes, meeting the statistical requirements.
[0141] It should be noted that 70% is the preferred implementation value, and can be dynamically adjusted according to the difficulty of the task or the distribution of the group in actual application. For example, it can be increased to 75% in high-discrimination tests and reduced to 65% in basic exercises, but the core principle remains unchanged: select sample subjects whose performance is significantly better than the average level of the group.
[0142] S5-4. Extract J writing behavior sequences corresponding to J sample objects;
[0143] S5-5. Align the J writing behavior sequences according to the node index of their node behavior vectors to generate an instantaneous vector set with M node indices; where the instantaneous vector set represents the J node behavior vectors that share the same node index.
[0144] S5-6. Based on the behavior vectors of the J nodes, construct the median vector of the instantaneous vector set;
[0145] S5-7. Traverse the instantaneous vector set of M node indices, and repeat the construction of the median vector until the median vector of M node indices is obtained.
[0146] S5-8. Arrange the median vectors of the M node indices in order to construct the standard writing behavior sequence of the writing behavior cluster;
[0147] S5-9. Traverse the K writing behavior clusters and repeat the construction of the standard writing behavior sequence until the standard writing behavior sequence of each of the K writing behavior clusters is obtained.
[0148] In this embodiment, high-scoring samples within the cluster are selected to construct a standard writing behavior sequence. This is based on the fact that academic performance objectively reflects a student's mastery of knowledge and fluency in task execution. High-scoring students typically exhibit a more stable rhythm, fewer hesitations and pauses, and a more standardized writing trajectory. Therefore, using their behavioral patterns as a "standard" can effectively represent the standard writing norms at this ability level.
[0149] Specifically, in this embodiment, step S5-6 further includes:
[0150] S5-6-1. Traverse the current dimension of the node behavior vectors, extract the writing behavior parameters of J node behavior vectors in the instantaneous vector set in the current dimension, and mark them as J parameters of the same dimension.
[0151] S5-6-2. Sort the J parameters of the same dimension in ascending order of their numerical values;
[0152] S5-6-3. If J is odd, select the writing behavior parameter located at the (J+1) / 2th position after ascending sorting and mark it as the median parameter of the current dimension.
[0153] S5-6-4. If J is even, calculate the average of the writing behavior parameters at the J / 2 and (J / 2+1)th positions after ascending order, and mark it as the median parameter of the current dimension.
[0154] S5-6-5. Repeat the marking of the median parameters until the median parameters for each dimension are obtained;
[0155] S5-6-6. Concatenate the median parameters of each dimension in the original dimension order to form the median vector of the instantaneous vector set.
[0156] Specifically, the median vector is used to suppress abnormal writing behaviors that may exist in individual high-scoring samples (such as handwriting shakiness caused by accidental distraction). The median is not sensitive to outliers and can more robustly reflect the common behavioral characteristics of the high-scoring group, thereby improving the representativeness of the standard writing behavior sequence.
[0157] In this embodiment, step S6 specifically includes:
[0158] S6-1. Encode the real writing behavior sequence into a real representation vector;
[0159] S6-2. Calculate the K similarities between the true representation vector and the cluster center vectors of the K writing behavior clusters;
[0160] Specifically, the similarity is preferably represented as the reciprocal or negative value of the Euclidean distance, with a smaller distance indicating a higher similarity.
[0161] It should be noted that the purpose of calculating similarity in step S6-2 is to determine the most similar behavior cluster from the K clusters of writing behavior, thereby providing a comparison benchmark for sequence bias; therefore, the similarity (reciprocal of Euclidean distance) calculation itself is not used to directly evaluate the writing quality of the target object.
[0162] The sequence deviation calculation in step S7 is to compare the node behavior vectors of the real writing behavior sequence with the standard writing behavior sequence corresponding to the most similar behavior cluster node by node to determine whether the writing trajectory is qualified as a whole.
[0163] In other words, in this embodiment, the actual writing behavior sequence participates in two vector distance calculations with completely different purposes: a similarity calculation to determine the most similar behavior cluster and a calculation to determine the sequence deviation.
[0164] S6-3. Select the maximum similarity among the K similarities, and locate the most similar behavior cluster of the target object based on the maximum similarity.
[0165] In this embodiment, after matching the target object to the most similar behavior cluster, both the sequence bias calculation and rhythm ratio analysis use the standard writing behavior sequence corresponding to that cluster as the comparison benchmark. Since different clusters correspond to different task types or ability levels (such as geometric proof vs. Chinese character writing), this matching ensures that the comparison is performed in the same task context, avoiding biased judgments caused by task differences.
[0166] In this embodiment, step S8 specifically includes:
[0167] S8-1. Based on the actual writing behavior sequence of the target object, calculate the M-1 actual writing rhythm ratios of the target object; where the actual writing rhythm ratio is inherited from the node number of the previous sampling timestamp.
[0168] S8-2. Based on the most similar behavior cluster of the target object, extract M-1 standard writing rhythm ratios of its standard writing behavior sequence; where the standard writing rhythm ratio is inherited from the node number of the presampled timestamp.
[0169] S8-3. Pair the M-1 real writing rhythm ratios with the M-1 standard writing rhythm ratios according to their shared node numbers and calculate the difference to obtain the M-1 rhythm ratio deviations.
[0170] S8-4. Among M-1 rhythm ratio deviations, identify Q abnormal deviations that exceed the second threshold;
[0171] Specifically, the second threshold is set as twice the standard deviation of M-1 standard writing rhythm ratios. The setting rule is based on the 3σ principle, that is, under the assumption of normal distribution, about 95% of the rhythm ratio data fall within the range of mean ± twice the standard deviation, and deviations outside this range are considered significant anomalies.
[0172] S8-5. Extract P abnormal sampling timestamps corresponding to the Q abnormal deviation; where P≤Q;
[0173] Since adjacent tempo ratios share the same sampling timestamp, a single timestamp may correspond to two tempo ratio deviations, therefore P≤Q≤2P.
[0174] S8-6. Based on P abnormal sampling timestamps, identify the abnormal clustering period;
[0175] S8-7. Based on the mapping labels pre-stored in the label database during the abnormal clustering period, evaluate the learning abnormal labels of the target object.
[0176] The tag database is pre-stored in the form of structured data tables. Each record contains four fields: subject type, learning chapter, abnormal clustering time period, and corresponding learning abnormal tag. For example, a record can be represented as: (subject = mathematics, chapter = solving algebraic equations, time period = last 20%, tag = missing answer review). During matching, the table is queried to obtain unique tags based on the target object's task metadata (subject, chapter) and the detected abnormal time period.
[0177] All mapping rules were developed by domain experts based on long-term teaching observations. For example, “abnormal starting segment of geometry proof problems” often corresponds to students’ inability to quickly extract known conditions and target conclusions, so it is defined as “difficulty in starting the problem-solving process”.
[0178] It should be noted that different periods of abnormal clustering correspond to different learning chapters and learning abnormal tags in the tag database;
[0179] For example:
[0180] If the abnormal clustering period is located at the first 10% of the question (corresponding to the task understanding and planning stage), and the chapter is "Geometric Proof", then the tag is "Difficulty in Starting the Problem Solving".
[0181] If the anomalies are concentrated in areas with complex strokes (30%–50% of node numbers), and the chapter is “Chinese Character Writing”, then the tag is “Unstable Character Shape Control”.
[0182] If the anomaly occurs at the transition point between multiple steps (such as formula substitution or unit conversion), and the chapter is "Physics Calculation Problems", then the tag will be "Step Conversion Stuttering";
[0183] If anomalies are clustered in the last 20% of the questions (corresponding to the answer organization and review stage), and the chapter is "Solving Algebraic Equations", then the tag will be "Answer Review Missing".
[0184] Specifically, in this embodiment, step S8-6 further includes:
[0185] S8-6-1. Arrange the P abnormal sampling timestamps in ascending order of time.
[0186] S8-6-2. Select two adjacent abnormal sampling timestamps in ascending order one by one to obtain P-1 pairs of adjacent timestamps.
[0187] S8-6-3. Calculate the time interval between each pair of adjacent abnormal sampling timestamps to obtain P-1 time intervals;
[0188] S8-6-4. Compare P-1 time intervals with the preset time adjacency threshold T;
[0189] Specifically, the temporal adjacency threshold T is used to define the temporal continuity of abnormal behavior. Considering that the typical sampling frequency of a stylus is 100–200Hz (i.e., a sampling period of 5–10ms), a single cognitive hesitation or writing pause typically lasts 20–100ms. Therefore, setting T to 5 times the sampling period (e.g., T=25ms when the sampling period is 5ms) can effectively aggregate multiple abnormal points within the same behavioral event into a meaningful clustered period, while avoiding erroneous merging across events.
[0190] S8-6-5. Based on the comparison results of P-1 time intervals with the time adjacency threshold T, group the P abnormal sampling timestamps into abnormal time periods.
[0191] The abnormal time period grouping is as follows: if the time interval between any two adjacent abnormal sampling timestamps is less than or equal to the time adjacency threshold T, then the consecutive abnormal sampling timestamps are merged into the same abnormal cluster time period; wherein, the abnormal cluster time period is defined by the earliest and latest sampling timestamps contained therein.
[0192] For example:
[0193] Assume the abnormal sampling timestamp P=5, and the timestamps are sorted as: t1, t2, t3, t4, t5;
[0194] The adjacent time intervals are: t1-t2=10ms, t2-t3=30ms, t3-t4=8ms, t4-t5=12ms;
[0195] Set the time adjacency threshold T=20ms.
[0196] Since the interval between t1 and t2 (10ms) is less than or equal to the time adjacency threshold T, they belong to the same clustering period.
[0197] The interval between t2 and t3 (30ms) is greater than the time adjacency threshold T, so a break occurs here, and t3 is not merged with the previous group;
[0198] t3 and t4 (8ms ≤ time adjacency threshold T) and t4 and t5 (12ms ≤ time adjacency threshold T) both satisfy the adjacency condition, therefore t3, t4, and t5 are classified as another clustering period.
[0199] The final output consists of two abnormal clustering periods: [t1,t2] and [t3,t5], which represent two time-continuous and internally closely related abnormal behavior intervals.
[0200] Example 2: See Figure 5The technical solution of Embodiment 2 differs from Embodiment 1 in that this embodiment of the invention also provides a dynamic evaluation system for learning ability based on the writing trajectory of a stylus. This system is used to implement the above-described method embodiments, and details already described will not be repeated. The terms "module," "unit," and "subunit" used below refer to combinations of software and / or hardware that achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0201] Figure 5 This is a structural block diagram of the dynamic assessment system for learning ability based on stylus writing trajectory of the present invention. The system includes:
[0202] The sample object selection module is used to select N sample objects with object IDs within the target group.
[0203] The sample sequence construction module is used to construct N writing behavior sequences for N sample objects;
[0204] Among them, the writing behavior sequence represents the node behavior vector arranged in M sampling timestamps, and the correlation between adjacent sampling timestamps represents the writing rhythm ratio of multidimensional writing behavior changes per unit time.
[0205] The embedding vector encoding module is used to embed and encode N sequences of writing actions to obtain N sequence representation vectors; where each sequence representation vector is associated with the object ID of the sample object.
[0206] The behavior clustering module is used to perform unsupervised clustering of N sequence representation vectors to obtain K writing behavior clusters;
[0207] The standard sequence construction module is used to construct the standard writing behavior sequences for each of the K writing behavior clusters.
[0208] The similarity cluster determination module is used to obtain the actual writing behavior sequence of the target object, and determine the most similar behavior cluster of the target object from K writing behavior clusters based on the actual writing behavior sequence;
[0209] The sequence deviation calculation module is used to extract the standard writing behavior sequence of the most similar behavior cluster and calculate the sequence deviation between it and the real writing behavior sequence.
[0210] The secondary deviation evaluation module is used to determine whether the writing trajectory of the target object is unqualified if the sequence deviation is greater than the first threshold; otherwise, the writing rhythm ratio of the target object is extracted for learning anomaly label evaluation.
[0211] In the above system, the sample object selection module selects N sample objects with object IDs; the sample sequence construction module constructs N writing behavior sequences; where each writing behavior sequence represents a node behavior vector arranged sequentially by M sampling timestamps, and the correlation between adjacent sampling timestamps represents the writing rhythm ratio of multidimensional writing behavior changes per unit time; the embedding vector encoding module obtains N sequence representation vectors; the behavior cluster clustering module obtains K writing behavior clusters; the standard sequence construction module constructs a standard writing behavior sequence; the similarity cluster determination module determines the most similar behavior cluster of the target object; the sequence deviation calculation module calculates the sequence deviation between the target object and the actual writing behavior sequence; and the secondary deviation evaluation module determines that if the sequence deviation is greater than a first threshold, the writing trajectory of the target object is deemed unqualified; otherwise, the writing rhythm ratio of the target object is extracted for learning anomaly label evaluation, thus solving the problem of incomplete evaluation.
[0212] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A dynamic assessment method for learning ability based on stylus writing trajectory, characterized in that, include: S1. Within the target group, select N sample objects with object IDs; S2. Construct N writing behavior sequences for N sample objects; Among them, the writing behavior sequence represents the node behavior vector arranged in M sampling timestamps, and the correlation between adjacent sampling timestamps represents the writing rhythm ratio of multidimensional writing behavior changes per unit time. S3. Embed and encode N writing behavior sequences to obtain N sequence representation vectors; where each sequence representation vector is associated with the object ID of the sample object. S4. Perform unsupervised clustering on the N sequence representation vectors to obtain K clusters of writing behaviors; S5. Construct the standard writing behavior sequences for each of the K writing behavior clusters; S6. Obtain the actual writing behavior sequence of the target object, and determine the most similar behavior cluster of the target object from K writing behavior clusters based on the actual writing behavior sequence; S7. Extract the standard writing behavior sequence of the most similar behavior cluster and calculate the sequence deviation between it and the real writing behavior sequence; S8. If the sequence deviation is greater than the first threshold, the writing trajectory of the target object is deemed unqualified; otherwise, the writing rhythm ratio of the target object is extracted for learning anomaly label evaluation.
2. The dynamic assessment method for learning ability based on stylus writing trajectory according to claim 1, characterized in that, Construct N sequences of writing behaviors for N sample objects, including: S2-1. Select any target sample object from N sample objects; S2-2, Locate the target sample object on the timeline and use the stylus to write M sampling timestamps; S2-3. Based on the time order of the M sample timestamps, generate M node numbers arranged in a time sequence. S2-4. Obtain the writing behavior parameters of M node numbers; where the writing behavior parameters include at least: the coordinates of the stylus at the sampling timestamp, the touch pressure, and the touch tilt angle; S2-5. Perform dimensionless processing on the writing behavior parameters and concatenate them into a node behavior vector of the target sample object at the node index. S2-6. Arrange the M node behavior vectors based on the node index to generate the writing behavior sequence; S2-7. In the writing behavior sequence, associate the sampling timestamps of the M node behavior vectors. S2-8. Based on the sampling timestamps corresponding to the writing behavior sequence and its node behavior vectors, calculate the M-1 writing rhythm ratios between adjacent sampling timestamps; where the writing rhythm ratio is inherited from the node number of the previous sampling timestamp.
3. The dynamic assessment method for learning ability based on stylus writing trajectory according to claim 2, characterized in that, Based on the writing behavior sequence and its associated sampling timestamps, calculate M-1 writing rhythm ratios between adjacent sampling timestamps, including: S2-8-1. Select the behavior vectors of adjacent nodes in the writing behavior sequence one pair at a time; S2-8-2, Calculate the Euclidean distance between the behavior vectors of adjacent nodes; S2-8-3. Based on the behavior vectors of adjacent nodes, obtain their respective associated adjacent sampling timestamps; S2-8-4. Calculate the writing duration between adjacent sampling timestamps; S2-8-5. Traverse the writing behavior sequence and perform a ratio operation between the Euclidean distance and the writing duration of adjacent node behavior vectors to obtain M-1 writing rhythm ratios.
4. The dynamic assessment method for learning ability based on stylus writing trajectory according to claim 1, characterized in that, Construct the standard writing behavior sequences for each of the K writing behavior clusters, including: S5-1. Based on the object ID inherited from the sequence representation vector, anchor G sample objects for each writing behavior cluster; where G is the number of sample objects within the cluster. S5-2. Obtain the learning scores of the G sample objects; S5-3. Based on the learning scores of the G sample objects, select J sample objects whose learning scores are greater than the high score threshold from high to low. S5-4. Extract J writing behavior sequences corresponding to J sample objects; S5-5. Align the J writing behavior sequences according to the node index of their node behavior vectors to generate an instantaneous vector set with M node indices; where the instantaneous vector set represents the J node behavior vectors that share the same node index. S5-6. Based on the behavior vectors of the J nodes, construct the median vector of the instantaneous vector set; S5-7. Traverse the instantaneous vector set of M node indices, and repeat the construction of the median vector until the median vector of M node indices is obtained. S5-8. Arrange the median vectors of the M node indices in order to construct the standard writing behavior sequence of the writing behavior cluster; S5-9. Traverse the K writing behavior clusters and repeat the construction of the standard writing behavior sequence until the standard writing behavior sequence of each of the K writing behavior clusters is obtained.
5. The dynamic assessment method for learning ability based on stylus writing trajectory according to claim 4, characterized in that, Based on the behavior vectors of J nodes, construct the median vector of the instantaneous vector set, including: S5-6-1. Traverse the current dimension of the node behavior vectors, extract the writing behavior parameters of J node behavior vectors in the instantaneous vector set in the current dimension, and mark them as J parameters of the same dimension. S5-6-2. Sort the J parameters of the same dimension in ascending order of their numerical values; S5-6-3. If J is odd, select the writing behavior parameter located at the (J+1) / 2th position after ascending sorting and mark it as the median parameter of the current dimension. S5-6-4. If J is even, calculate the average of the writing behavior parameters at the J / 2 and (J / 2+1)th positions after ascending order, and mark it as the median parameter of the current dimension. S5-6-5. Repeat the marking of the median parameters until the median parameters for each dimension are obtained; S5-6-6. Concatenate the median parameters of each dimension in the original dimension order to form the median vector of the instantaneous vector set.
6. The dynamic assessment method for learning ability based on stylus writing trajectory according to claim 1, characterized in that, Based on the actual writing behavior sequence, determine the most similar behavior cluster of the target object from K writing behavior clusters, including: S6-1. Encode the real writing behavior sequence into a real representation vector; S6-2. Calculate the K similarities between the true representation vector and the cluster center vectors of the K writing behavior clusters; S6-3. Select the maximum similarity among the K similarities, and locate the most similar behavior cluster of the target object based on the maximum similarity.
7. The dynamic assessment method for learning ability based on stylus writing trajectory according to claim 1, characterized in that, The learning anomalous labels of the target object were evaluated, including: S8-1. Based on the actual writing behavior sequence of the target object, calculate the M-1 actual writing rhythm ratios of the target object; where the actual writing rhythm ratio is inherited from the node number of the pre-sampling timestamp; S8-2. Based on the most similar behavior cluster of the target object, extract M-1 standard writing rhythm ratios of its standard writing behavior sequence; where the standard writing rhythm ratio is inherited from the node number of the presampled timestamp. S8-3. Pair the M-1 real writing rhythm ratios with the M-1 standard writing rhythm ratios according to their shared node numbers and calculate the difference to obtain the M-1 rhythm ratio deviations. S8-4. Among M-1 rhythm ratio deviations, identify Q abnormal deviations that exceed the second threshold; S8-5. Extract P abnormal sampling timestamps corresponding to the Q abnormal deviation; where P≤Q; S8-6. Based on P abnormal sampling timestamps, identify the abnormal clustering period; S8-7. Based on the mapping labels pre-stored in the label database during the abnormal clustering period, evaluate the learning abnormal labels of the target object.
8. The dynamic assessment method for learning ability based on stylus writing trajectory according to claim 7, characterized in that, Based on P abnormal sampling timestamps, abnormal clustering periods are identified, including: S8-6-1. Arrange the P abnormal sampling timestamps in ascending order of time. S8-6-2. Select two adjacent abnormal sampling timestamps in ascending order one by one to obtain P-1 pairs of adjacent timestamps. S8-6-3. Calculate the time interval between each pair of adjacent abnormal sampling timestamps to obtain P-1 time intervals; S8-6-4. Compare P-1 time intervals with the preset time adjacency threshold T; S8-6-5. Based on the comparison results of P-1 time intervals with the time adjacency threshold T, group the P abnormal sampling timestamps into abnormal time periods. The abnormal time period grouping is as follows: if the time interval between any two adjacent abnormal sampling timestamps is less than or equal to the time adjacency threshold T, then the consecutive abnormal sampling timestamps are merged into the same abnormal cluster time period; wherein, the abnormal cluster time period is defined by the earliest and latest sampling timestamps contained therein.
9. A dynamic learning ability assessment system based on stylus writing trajectory, used to execute the dynamic learning ability assessment method based on stylus writing trajectory as described in any one of claims 1 to 8, characterized in that, The system includes: The sample object selection module is used to select N sample objects with object IDs within the target group. The sample sequence construction module is used to construct N writing behavior sequences for N sample objects; Among them, the writing behavior sequence represents the node behavior vector arranged in M sampling timestamps, and the correlation between adjacent sampling timestamps represents the writing rhythm ratio of multidimensional writing behavior changes per unit time. The embedding vector encoding module is used to embed and encode N sequences of writing actions to obtain N sequence representation vectors; where each sequence representation vector is associated with the object ID of the sample object. The behavior clustering module is used to perform unsupervised clustering of N sequence representation vectors to obtain K writing behavior clusters; The standard sequence construction module is used to construct the standard writing behavior sequences for each of the K writing behavior clusters. The similarity cluster determination module is used to obtain the actual writing behavior sequence of the target object, and determine the most similar behavior cluster of the target object from K writing behavior clusters based on the actual writing behavior sequence; The sequence deviation calculation module is used to extract the standard writing behavior sequence of the most similar behavior cluster and calculate the sequence deviation between it and the real writing behavior sequence. The secondary deviation evaluation module is used to determine whether the writing trajectory of the target object is unqualified if the sequence deviation is greater than the first threshold; otherwise, the writing rhythm ratio of the target object is extracted for learning anomaly label evaluation.