A two-person sliding dynamic collaborative network modeling and collaborative error analysis method

By synchronously collecting multi-source data and modeling heterogeneous graph networks, the problems of single data, inaccurate modeling, and fuzzy errors in the collaborative analysis of two-person skating were solved. This achieved the accuracy of collaborative measurement and the relevance of training schemes, thereby improving training efficiency and effectiveness.

CN123069249APending Publication Date: 2026-11-10SHENYANG SPORT UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611115436.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-27
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies for collaborative analysis and training in pairs figure skating suffer from problems such as limited data collection dimensions, inaccurate collaborative modeling, vague error analysis, and reliance on experience in training guidance, leading to inaccurate collaborative measurement and low training efficiency.

Method used

By employing a method of multi-source collaborative synchronous data acquisition, dual-person exclusive collaborative feature extraction, dynamic collaborative heterogeneous graph network modeling, and collaborative error quantification and attribution, a heterogeneous graph neural network is constructed by integrating human skeleton posture, ice skate-ice surface force field, and dual-person eye contact interaction data. This enables accurate quantification and visual localization of collaboration score and error type, and establishes a dynamic training mechanism for error-task matching.

Benefits of technology

It has achieved improved accuracy in collaborative quantification, precise identification of error types and locations, enhanced targeting and adaptability of training programs, formed an efficient dynamic training closed loop, and improved training effectiveness and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069249A_ABST
    Figure CN123069249A_ABST
Patent Text Reader

Abstract

The application discloses a pair of people sliding dynamic cooperative network modeling and cooperative error analysis method, including the following steps: S1: multi-source cooperative data synchronous acquisition, collecting human skeleton posture data, ice skate-ice surface force field data and pair of people visual line interaction data in the pair of people sliding movement process, adopting hardware triggering and software calibration double mechanism to realize time synchronization of three types of data; S2: pair of people exclusive cooperative feature extraction; S3: dynamic cooperative heterogeneous graph network modeling; S4: cooperative error quantification attribution; S5: error-task matching dynamic training.The beneficial effects of the application are that the technical scheme fuses three types of data of human skeleton posture, ice skate-ice surface force field and pair of people visual line interaction, fills the blank of only collecting single posture data in the prior art, completely captures the physical interaction and visual interaction logic of cooperative action, and provides a data basis for accurate cooperative analysis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent sports training technology, and in particular to a method for modeling dynamic collaborative networks and analyzing collaborative errors in pairs figure skating. Background Technology

[0002] As a winter sport that demands a high degree of coordination, pairs figure skating's scoring hinges on the synchronization, consistency, and artistic expression of the skaters' movements. However, current technologies have many limitations in the analysis and training of pair figure skating coordination.

[0003] The data collection is limited to a single dimension, relying heavily on visual sensors to collect human posture data. This only reflects the trajectory of limb movements and lacks data on the force interaction between the ice skates and the ice surface, as well as the interaction data between the two people's gazes. Therefore, it cannot capture the physical root and interaction logic of coordinated movements.

[0004] The collaborative modeling method has limitations. It analyzes collaboration by modeling single-person actions and simply superimposing them. It evaluates the quality of collaboration by calculating the similarity of two-person postures, but does not consider implicit collaborative relationships such as force transmission and rhythm coupling, resulting in inaccurate quantification of collaboration.

[0005] Error analysis is vague and general, only able to output an overall conclusion of low coordination, and cannot pinpoint the error type, specific location, or time series interval. Coaches need to rely on experience to make judgments, which is inefficient.

[0006] Training guidance relies on experience and lacks a precise matching mechanism between error types and training tasks. Adjustments to training plans depend on the coach's subjective experience, making it impossible to achieve targeted error correction and closed-loop optimization of effect feedback.

[0007] Therefore, there is an urgent need for a comprehensive technical solution covering data acquisition, collaborative modeling, error diagnosis, and training optimization to address the problems of inaccurate collaborative analysis and lack of basis for training guidance in existing technologies. Summary of the Invention

[0008] The main technical problem solved by this invention is to provide a method for modeling dynamic cooperative networks and analyzing cooperative errors in two-person skating, thereby solving one or more of the problems mentioned above in the prior art.

[0009] To solve the above-mentioned technical problems, the present invention adopts a technical solution as follows: a method for modeling and analyzing the collaborative error of a dynamic collaborative network for two skaters, the innovation of which lies in the following steps: S1: Multi-source collaborative data synchronous acquisition: Collect human skeleton posture data, blade-ice surface force field data, and pair eye-to-eye interaction data during pairs skating. A dual mechanism of hardware triggering and software calibration is used to achieve time synchronization of these three types of data. S2: Pair-specific collaborative feature extraction: Based on the multi-source data synchronized in step S1, extract force collaboration features, rhythm collaboration features, and spatial collaboration features related to the pair's interaction dimensions. S3: Dynamic collaborative heterogeneous graph network modeling: Construct a heterogeneous graph neural network containing multiple types of nodes and dual-type edges, outputting collaboration score and error probability. S4: Collaborative error quantification and attribution: Based on feature deviation and collaboration, determine the error type using a classification model, quantify the degree of deviation, and visualize its location using heatmaps and time-series curves. S5: Dynamic training with error-task matching: Match training tasks according to error type, adjust tasks every 5 minutes based on improvement rate, and optimize weights weekly based on collaboration, forming a dynamic training closed loop.

[0010] In some implementations, in step S1: The human skeleton posture data consists of the three-dimensional coordinates, angular velocity or acceleration of 26 key joints for each of the two individuals, acquired through high-speed motion capture at ≥1000 frames / second. The ice skate-ice surface force field data, which includes the pressure distribution and center of force coordinates of the two ice skates, was acquired through a flexible piezoelectric sensor array with ≥1024 elements. The data on two-person eye contact interaction includes the coordinates of the eye contact point and the duration of the interaction, which are acquired through an infrared eye-tracking camera with a frequency of ≥60Hz. The software calibration uses a dynamic time warping algorithm to ensure that the time synchronization error is less than 10ms.

[0011] In some implementations, in step S2: Force synergy characteristics include peak time difference of blade pressure. θ, the angle between the lines connecting the centers of force, and η, the efficiency of lifting force transmission; Rhythmic coordination features include time differences between key action nodes. , limb angular velocity correlation r, movement cycle similarity γ; Spatial coordination features include the range of center of gravity distance fluctuation σ, the offset of lifting space d, the offset difference of rotation axis d, and the duration of line-of-sight interaction T.

[0012] In some implementations, the heterogeneous graph neural network modeling in step S3 includes: Construct a set of 61 nodes of various types, including 52 human joint nodes, 2 blade center of force nodes, 1 line-of-sight interaction node, and 1 global center of gravity node; Design individual edges that connect the internal joints of the same player and collaborative edges that connect the interaction nodes of two players. The edge weights are dynamically calculated based on real-time collaborative features. The model is trained using a weighted total loss function that combines synergy prediction loss and error type classification loss, and outputs a synergy score of 0-100 and a probability distribution of four types of errors.

[0013] In some implementations, in step S4: The random forest model was used to determine four types of errors: force coordination error, rhythm coordination error, spatial coordination error, and interaction coordination error. The error is further divided into moderate error and severe error, and the deviation degree index e and the comprehensive deviation index are calculated. ; Collaborative error heatmaps are obtained through nodal error contribution. Visual positioning is achieved by marking out the threshold range with color and time-series deviation curve.

[0014] In some implementations, in step S5: Error improvement rate The threshold is 20%, and the training task is adjusted accordingly. The weight of the basic task is calculated weekly based on the collaboration score s. High-intensity task weight Optimize the proportion of training tasks.

[0015] In some implementations, the force synergy characteristics are calculated as follows: ,in , These are the moments when skater A and skater B reached their peak skate pressure, respectively. ,in , For the coordinates of the ice skate's center of force, , The coordinates of the body's center of gravity; ,in The difference in pressure on the skate blades before and after the lift. The difference in pressure on the ice skates before and after the lifter lifts the ice.

[0016] In some implementations, the rhythmic coordination features are calculated as follows: ,in , These are the key moments when contestants A and B complete their synchronized actions; ,in , For the corresponding limb angular velocity sequence; ,in , These represent the cycle duration for contestants A and B to complete the same periodic movement, respectively.

[0017] In some implementations, spatial cooperative features are calculated as follows: ,in Let be the distance between the centers of gravity of the two people at time i. This is the average distance between the centers of gravity; ,in( , , ( ) is the ideal center of gravity coordinate of the person being lifted, ( , , () represents the actual center of gravity coordinates; ,in( , ), ( , ( ) represent the rotation axis coordinates of contestants A and B, respectively;

[0018] in The time interval during which the eyes of the two people meet.

[0019] In some implementations, the model architecture of heterogeneous graph neural networks includes: Input layer: Initializes the 61 nodes into an initial feature matrix, and inputs the edge types and weight matrices. ; Graph convolutional layer: A type-aware message passing mechanism is used to update node features, and two graph convolutional layers are set to avoid overfitting; Fully connected layer: Global average pooling is performed on the output of the graph convolutional layer to obtain the global collaborative feature vector; Multi-task output layer: via The activation function outputs a synergy score from 0 to 100. The activation function outputs four types of error probability distributions.

[0020] The beneficial effects of this invention are: Comprehensive data dimensions: It integrates three types of data: human skeleton posture, ice skate-ice surface force field and two-person eye contact interaction, filling the gap of existing technologies that only collect single posture data, and fully capturing the physical and visual interaction logic of coordinated actions, providing a data foundation for accurate collaborative analysis.

[0021] Precise and efficient modeling: By transforming the implicit two-person collaborative relationship into a computable network structure through heterogeneous graph networks, this method differs from the coarse approach of traditional single-person modeling and overlay, significantly improving the accuracy of collaborative measurement and making collaborative quality assessment more in line with actual motion patterns.

[0022] Precise error diagnosis: Breaking through the limitations of existing technologies that can only judge low coordination and cannot locate errors, it achieves full-dimensional precise positioning of error type, severity, spatial location and time interval, making errors explainable and traceable, and providing a clear basis for targeted error correction.

[0023] Scientific training guidance: Establish a precise matching mechanism between error types and training tasks, combine real-time feedback and periodic optimization to form a dynamic training closed loop, reduce reliance on the coach's subjective experience, make training programs more targeted and adaptable, and improve the efficiency and effectiveness of collaborative training. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is an overall flowchart of the method for modeling and analyzing collaborative errors in a dynamic collaborative network for two-person skating, as described in this invention.

[0025] Figure 2 This is a schematic diagram of multi-source data synchronous acquisition for a method of dynamic collaborative network modeling and collaborative error analysis for two-person skating, as proposed in this invention.

[0026] Figure 3 This is a schematic diagram of the node and edge structure of the HGNN model for the two-person skating dynamic cooperative network modeling and cooperative error analysis method of the present invention.

[0027] Figure 4 This is a logic diagram for the quantification and attribution of collaborative error in a dynamic collaborative network modeling and collaborative error analysis method for two-person skating, as described in this invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0029] like Figures 1 to 4 As shown, the embodiments of the present invention include: a specific implementation method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters is as follows.

[0030] I. Synchronous Acquisition of Multi-Source Collaborative Data 1.1 Data Acquisition Content and Equipment This step collects three types of core data, overcoming the limitation of only collecting single-person posture data.

[0031] Human skeleton posture data: The three-dimensional coordinates, angular velocity or acceleration of 26 key joints of each of the two people are obtained through a high-speed motion capture system with a sampling frequency of ≥1000 frames / second and joint coordinate accuracy of ≤0.1mm, accurately capturing the details of limb movement.

[0032] Ice skate-ice surface force field data: Pressure distribution and center of force coordinates of the two ice skates are obtained through a flexible piezoelectric sensor array containing ≥1024 sensing units, pressure accuracy ≤0.1kPa, and sampling frequency ≥500Hz, reflecting the force interaction between the ice skate and the ice surface.

[0033] Two-person eye contact interaction data: coordinates of the eye contact point and duration of the interaction, are acquired by an infrared eye-tracking camera with a sampling frequency of ≥60Hz to capture two-person visual interaction information.

[0034] 1.2 Data Synchronization Mechanism Data alignment is achieved through a combination of hardware triggering and software calibration to ensure spatiotemporal consistency of the data.

[0035] Hardware triggering: By synchronizing pulses through the motion capture system, the three types of acquisition devices are forced to start simultaneously, avoiding initial time deviation.

[0036] Software calibration: The Dynamic Time Warping (DTW) algorithm is used to align the timestamps of each device, ensuring that the time synchronization error is less than 10ms.

[0037] II. Feature Extraction for Two-Person Collaborative Teams Based on synchronized multi-source data, three types of exclusive collaborative features reflecting the dimensions of two-person interaction are extracted.

[0038] 2.1 Force Synergy Characteristics (Reflecting the quality of force transmission and interaction between the ice skate and the ice surface) Peak pressure time difference of ice skate : ,in , These are the moments when skaters A and B reach peak pressure on their skate blades, respectively; the ideal threshold is 50ms, and anything greater than 100ms is considered a force transmission delay error.

[0039] Angle between the lines connecting the centers of force : ,in , For the coordinates of the ice skate's center of force, , The coordinates are the body's center of gravity; the ideal threshold is 15°, and a value greater than 30° is considered an error due to uneven distribution of falling ice force.

[0040] Lifting force transmission efficiency : ,in The difference in pressure on the ice skates before and after the lift. This represents the difference in ice skate pressure before and after the lifter lifts the ice skate; the ideal threshold is 0.85, and a value below 0.7 is considered an inefficient error in the transmission of lifting force.

[0041] 2.2 Rhythmic Coordination Characteristics (Reflecting the Timing Synchronization of Two Persons' Movements) Time difference of key action nodes : ,in , These are the moments when contestants A and B complete the key nodes of their synchronized actions, respectively; the ideal threshold is 100ms, and when it exceeds 200ms, it is judged as a delay error in the synchronized action node.

[0042] Limb angular velocity correlation r: ,in , The sequence of angular velocities for the corresponding limbs of athletes A and B. , These are their average values. The ideal threshold is 0.8; if it is below 0.5, it is considered an error in limb rhythm synchronization.

[0043] similarity of action cycles : ,in , These represent the cycle duration for contestants A and B to complete the same periodic movement, respectively. The ideal threshold is... When γ is not within the interval, it is determined to be an error in the motion cycle imbalance.

[0044] 2.3 Spatial Coordination Features (Reflecting the degree of spatial position matching between two people) Center of gravity distance fluctuation range : ,in Let be the distance between the centers of gravity of the two people at time i. The average center of gravity distance is 0.2m; the ideal threshold is 0.2m. When it is greater than 0.3m, it is judged as an excessive error in spatial position fluctuation.

[0045] Lifting space offset : ,in( , , ( ) represents the ideal center of gravity coordinates of the person being lifted. , , The actual center of gravity coordinates are shown in the figure. The ideal threshold is 0.1m. When the value is greater than 0.2m, it is determined to be an offset error in the lifting space.

[0046] Rotation axis offset difference : ,in( , ), ( , ( ) represent the rotation axis coordinates of athletes A and B, and the two-dimensional coordinates on the ice surface, calculated by fitting the center of gravity trajectory; the ideal threshold is 0.05m, when If the error is greater than 0.1m, it is determined to be a misalignment error of the rotating shaft center.

[0047] Duration of eye contact :

[0048] in The time interval for eye contact between two people is determined by an infrared eye-tracking camera. The determination criteria are that the angle between the eye directions is ≤15° and the point of eye contact is within the other person's torso area. The ideal threshold is 0.3s. When it is less than 0.1s, it is determined to be an error of missing eye contact.

[0049] III. Modeling of Dynamic Collaborative Heterogeneous Graph Networks A heterogeneous graph neural network (HGNN) with multiple types of nodes and dual types of edges is constructed to achieve accurate modeling of the collaborative relationship between two people.

[0050] 3.1 Construction of Multiple Node Types Based on the collaborative needs of two-person skating, a set of 61 nodes in 4 categories was constructed, and all nodes are associated with the interaction dimension features of the two people.

[0051] There are 52 human joint nodes, corresponding to 26 key joints for contestants A and B respectively. The feature vector of each node is as follows: , , In three-dimensional coordinates, ω is the angular velocity.

[0052] There are two blade center of force nodes, corresponding to the blade centers of force for skaters A and B respectively. The feature vector of each node is... , , For the force center in two-dimensional coordinates, This represents the peak pressure.

[0053] Eye contact interaction node: 1, corresponding to the point where the eyes of both people meet, and the feature vector is: , , Let T be the three-dimensional coordinates of the interaction point, and let T be the duration of the interaction.

[0054] Global centroid node: 1, corresponding to the overall centroid of the two people, with the following feature vector: , , The three-dimensional coordinates of the global center of gravity The standard deviation of the center of gravity fluctuation is calculated by weighting the coordinates of human joint nodes according to their body weight percentage.

[0055] 3.2 Design of Two Types of Edges and Calculation of Dynamic Weights The design includes two types of edges: individual edges and collaborative edges. The edge weights are dynamically calculated based on real-time collaborative features.

[0056] Individual edges: 48 in total, connecting only adjacent joint nodes of the same athlete, such as athlete A's left shoulder. left elbow Left hand, right hip right knee Right ankle, 24 items each for contestants A and B, 12 items for upper limbs and 12 items for lower limbs; the weighting formula is: ,in , Let i and j be the angular velocity vectors, respectively, with weights ranging from [0,1]. This indicates that the joint movements are continuous. It indicates stiffness in movement.

[0057] Collaborative edges: 53 in total, divided into 3 categories.

[0058] For two athletes, there are 26 corresponding joint edges, connecting the same-named joint nodes of athlete A and athlete B, associated with the limb angular velocity correlation feature r; the weight calculation formula is as follows: ,in Let Pearson correlation coefficient be the angular velocity sequence of athlete A joint i and athlete B joint j. , .

[0059] Double skate center of force edge: 1 line, connecting the center of force nodes of skaters A and B, associated with the angle feature θ of the line connecting the centers of force. The weight calculation formula is: ,in Let p be the angle between the line connecting skater A's blade center of force and q be the line connecting the centers of gravity of both skaters. , .

[0060] Gaze-Global Center of Gravity Connection Edges: 26, connecting gaze interaction nodes with the human joint nodes of players A and B, intersecting with the gaze interaction duration feature T; weight calculation formula is: ,in The duration of eye contact between two people is recorded at the eye-interaction node v. , .

[0061] 3.3 HGNN Model Architecture Construction Construct a model architecture consisting of an input layer, a graph convolutional layer, a fully connected layer, and a multi-task output layer.

[0062] Input layer: The initial features of 61 nodes are matrixed to obtain the feature matrix. d represents the average feature dimension of the nodes, which is approximately 5.2. The input edge type and weight matrix are also considered. .

[0063] Graph convolutional layer: Node features are updated using a type-aware message passing mechanism, with the update formula as follows: in A set of edge types, Let i be the set of neighboring nodes connected to node i via edge of type t. For connection weights, The feature transformation matrix, For bias terms, (.) represents the ReLU activation function; the number of convolutional layers is set to... This ensures that features are fully integrated and avoids overfitting.

[0064] Fully connected layer: The final node feature matrix output by the graph convolutional layer. Perform global average pooling to obtain the global collaborative feature vector. The pooling formula is: ;in The average pooling coefficient ensures that the feature vector dimension is independent of the number of nodes.

[0065] Multi-task output layer: contains two parallel branches.

[0066] Synergy scoring branch: ,in This is the weight matrix. For bias terms, The function maps the output to Multiply the range by 100 to scale the score range to 0-100 points.

[0067] Error type prediction branch: ,in This is the weight matrix. For bias terms, The function ensures that the output is a probability distribution of four types of coordinated errors (force coordinated error, rhythm coordinated error, spatial coordinated error, and interaction coordinated error), satisfying... .

[0068] 3.4 Multi-task model training Training dataset construction: Standard movement data (including lifts, synchronized jumps, spirals, and other typical coordinated movements) of 50 pairs of top figure skaters (top 20 pairs in the world) were used. Each data set was labeled with two types of labels: one is the coordination quality score (1-10 points) from 3 international judges, scaled to 10-100 points as the coordination degree label. Secondly, the labels for the presence or absence of four types of cooperative errors are obtained using one-hot encoding. , This indicates that the nth sample has a k-th type of error. This indicates that it does not exist.

[0069] Loss function design: A weighted summation multi-task loss function is adopted, and the total loss function is... ,in Weighting coefficients: Synergy Prediction Loss Mean squared error (MSE) loss is used. N=50 is the number of training samples.

[0070] Error type classification loss : Using cross-entropy loss, .

[0071] Training parameter settings: The Adam optimizer is used, and the initial learning rate is set to... The training iterations consist of 100 epochs, with the learning rate decaying to 0.5 times the previous epoch every 20 epochs; L2 regularization is used, and the weight decay coefficient is set to... After training, the model must meet the performance indicators of predicting the degree of collaboration with no more than 0.5 points and the accuracy of error type determination with no less than 92%.

[0072] IV. Quantitative Attribution of Collaborative Error Based on feature bias and synergy, error type determination, degree quantification, and visual positioning are achieved.

[0073] 4.1 Error Type Determination and Degree Quantification A random forest model was used to identify four types of collaborative errors. For errors identified as existing, they were further categorized into moderate and severe errors based on their characteristic bias values. A bias severity index e and a comprehensive bias index were then calculated. .

[0074] Lifting force delay: moderate Severe cases are: Uneven distribution of falling ice force: moderate Severe Synchronization action node latency: Moderate Severe Movement cycle imbalance: moderate Severe Lifting space offset: moderate Severe Rotation axis misalignment: moderate Severe Lack of eye contact: moderate Severe Excessive limb occlusion: moderate Severe ; Comprehensive Deviation Index E k The calculation formula is: ,in Let be the deviation index of the i-th sub-error under the k-th type of error. For weights.

[0075] 4.2 Error Visualization and Positioning Collaborative error heatmap: Calculate the contribution of node errors ,in Let i be the attention weight of node i in the HGNN model. (The deviation index of the associated error for node i), nodes are labeled by color: ≥0.7 is marked in red (severe error node), 0.3≤ Points with a value <0.7 are marked in yellow (moderate error nodes). Nodes with a value <0.3 are marked in green (no significant error nodes).

[0076] Time series deviation curve: characteristics of time series Wait, calculate the deviation value for each frame. ,draw The curve is marked with shaded areas exceeding the threshold: The interval (moderate threshold) is marked with a yellow shading. The range of (severe threshold) is marked with red shading, such as... of =100ms =200ms.

[0077] 4.3 Error Diagnosis Report An error diagnosis report should include at least: a list of error types, such as force coordination error - delay in lifting force transmission, the deviation index (e-value) of each error, a heatmap of coordination error including node error annotations, a time-series deviation curve including the threshold interval, and an analysis of the causes of the error, such as... =250ms causes a delay in the transmission of lifting force.

[0078] V. Dynamic Training of Error-Task Matching Establish a precise matching mechanism between error types and training tasks to form a dynamic optimization closed loop of analysis, training, and feedback.

[0079] 5.1 Training Task Matching and Real-time Adjustment Based on the error type, a specific training task is matched from the preset training task library to generate an initial training plan; the error improvement rate is calculated every 5 minutes during training. Ebefore and Eafter represent the error deviation indices before and after training, respectively. If E ≥ 20%, the current task continues; if E < 20%, the alternative task is switched.

[0080] 5.2 Periodic Optimization of Task Weights Adjust task weights weekly based on collaboration score s: Calculate base task weights High-intensity task weight When the coordination score s≥85, the proportion of high-intensity tasks is increased; when s<60, the proportion of basic tasks is increased, making the training program more targeted and adaptable.

[0081] The beneficial effects of this invention are as follows: Comprehensive data dimensions: It integrates three types of data: human skeleton posture, ice skate-ice surface force field and two-person eye contact interaction, filling the gap of existing technologies that only collect single posture data, and fully capturing the physical and visual interaction logic of coordinated actions, providing a data foundation for accurate collaborative analysis.

[0082] Precise and efficient modeling: By transforming the implicit two-person collaborative relationship into a computable network structure through heterogeneous graph networks, this method differs from the coarse approach of traditional single-person modeling and overlay, significantly improving the accuracy of collaborative measurement and making collaborative quality assessment more in line with actual motion patterns.

[0083] Precise error diagnosis: Breaking through the limitations of existing technologies that can only judge low coordination and cannot locate errors, it achieves full-dimensional precise positioning of error type, severity, spatial location and time interval, making errors explainable and traceable, and providing a clear basis for targeted error correction.

[0084] Scientific training guidance: Establish a precise matching mechanism between error types and training tasks, combine real-time feedback and periodic optimization to form a dynamic training closed loop, reduce reliance on the coach's subjective experience, make training programs more targeted and adaptable, and improve the efficiency and effectiveness of collaborative training.

[0085] VI. Examples A pair from a national pairs figure skating team was selected as the experimental subjects, and the method of this invention was applied to conduct collaborative training for 8 weeks: Data acquisition: Three types of data are collected simultaneously using a high-speed motion capture system, a flexible piezoelectric sensor array, and an infrared line-of-sight tracking camera to ensure that the time synchronization error is less than 10ms.

[0086] Feature extraction and modeling: Extract force coordination, rhythm coordination and spatial coordination features, construct and train an HGNN model, the model coordination degree prediction error is 0.4 points, and the error type identification accuracy is 94%.

[0087] Error diagnosis and training optimization: In the early stages of training, a delay in the transmission of lifting force was diagnosed in this combination. Errors such as 220ms (e=1) and asynchronous limb rhythm (r=0.45, e=1) were addressed by matching targeted training tasks. During training, the task was adjusted every 5 minutes, and the task weights were optimized weekly.

[0088] Training results: After 8 weeks, the synergy score of this combination improved from 62 to 88, and the delay error in lifting force transmission was eliminated. =45ms), significantly improved limb rhythm synchronization (r=0.88), and significantly improved training efficiency and quality.

[0089] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters, characterized in that: Includes the following steps: S1: Multi-source collaborative data synchronous acquisition: Collect human skeleton posture data, blade-ice surface force field data, and pair eye-to-eye interaction data during pairs skating. A dual mechanism of hardware triggering and software calibration is used to achieve time synchronization of these three types of data. S2: Pair-specific collaborative feature extraction: Based on the multi-source data synchronized in step S1, extract force collaboration features, rhythm collaboration features, and spatial collaboration features related to the pair's interaction dimensions. S3: Dynamic collaborative heterogeneous graph network modeling: Construct a heterogeneous graph neural network containing multiple types of nodes and dual-type edges, outputting collaboration score and error probability. S4: Collaborative error quantification and attribution: Based on feature deviation and collaboration, determine the error type using a classification model, quantify the degree of deviation, and visualize its location using heatmaps and time-series curves. S5: Dynamic training with error-task matching: Match training tasks according to error type, adjust tasks every 5 minutes based on improvement rate, and optimize weights weekly based on collaboration, forming a dynamic training closed loop.

2. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 1, characterized in that: In step S1: The human skeleton posture data consists of the three-dimensional coordinates, angular velocity or acceleration of 26 key joints for each of the two individuals, acquired through high-speed motion capture at ≥1000 frames / second. The ice skate-ice surface force field data, which includes the pressure distribution and center of force coordinates of the two ice skates, was acquired through a flexible piezoelectric sensor array with ≥1024 elements. The data on two-person eye contact interaction includes the coordinates of the eye contact point and the duration of the interaction, which are acquired through an infrared eye-tracking camera with a frequency of ≥60Hz. The software calibration uses a dynamic time warping algorithm to ensure that the time synchronization error is less than 10ms.

3. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 1, characterized in that: In step S2: Force synergy characteristics include peak time difference of blade pressure. θ, the angle between the lines connecting the centers of force, and η, the efficiency of lifting force transmission; Rhythmic coordination features include time differences between key action nodes. , limb angular velocity correlation r, movement cycle similarity γ; Spatial coordination features include the range of center of gravity distance fluctuation σ, the offset of lifting space d, the offset difference of rotation axis d, and the duration of line-of-sight interaction T.

4. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 1, characterized in that: Step S3, heterogeneous graph neural network modeling, includes: Construct a set of 61 nodes of various types, including 52 human joint nodes, 2 blade center of force nodes, 1 line-of-sight interaction node, and 1 global center of gravity node; Design individual edges that connect the internal joints of the same player and collaborative edges that connect the interaction nodes of two players. The edge weights are dynamically calculated based on real-time collaborative features. The model is trained using a weighted total loss function that combines synergy prediction loss and error type classification loss, and outputs a synergy score of 0-100 and a probability distribution of four types of errors.

5. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 1, characterized in that: In step S4: The random forest model was used to determine four types of errors: force coordination error, rhythm coordination error, spatial coordination error, and interaction coordination error. The error is further divided into moderate error and severe error, and the deviation degree index e and the comprehensive deviation index are calculated. ; Collaborative error heatmaps are obtained by analyzing the contribution of nodal errors. Visual positioning is achieved by marking out the threshold range with color and time-series deviation curve.

6. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 1, characterized in that: In step S5: Error improvement rate The threshold is 20%, and the training task is adjusted accordingly. The weight of the basic task is calculated weekly based on the collaboration score s. High-intensity task weight Optimize the proportion of training tasks.

7. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 3, characterized in that: The calculation method for force synergy characteristics is as follows: ,in , These are the moments when skater A and skater B reached their peak skate pressure, respectively. ,in , Let's define the coordinates of the ice skate's center of force. , The coordinates of the body's center of gravity; ,in The difference in pressure on the skate blades before and after the lift. The difference in pressure on the ice skates before and after the lifter lifts the ice.

8. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 3, characterized in that: The rhythmic coordination feature is calculated as follows: ,in , These are the key moments when contestants A and B complete their synchronized actions; ,in , For the corresponding limb angular velocity sequence; ,in , These represent the cycle duration for contestants A and B to complete the same periodic movement, respectively.

9. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 3, characterized in that: The calculation method for spatial collaborative features is as follows: ,in Let be the distance between the centers of gravity of the two people at time i. This is the average distance between the centers of gravity; ,in( , , ( ) is the ideal center of gravity coordinate of the person being lifted, ( , , () represents the actual center of gravity coordinates; ,in( , ), ( , ( ) represent the rotation axis coordinates of contestants A and B, respectively; in The time interval during which the eyes of the two people meet.

10. The method for modeling and analyzing collaborative errors in a dynamic collaborative network for two skaters according to claim 4, characterized in that: The model architecture of heterogeneous graph neural networks includes: Input layer: Initializes the feature matrix of 61 nodes, and inputs the edge types and weight matrices. ; Graph convolutional layer: A type-aware message passing mechanism is used to update node features, and two graph convolutional layers are set to avoid overfitting; Fully connected layer: Global average pooling is performed on the output of the graph convolutional layer to obtain the global collaborative feature vector; Multi-task output layer: via The activation function outputs a synergy score from 0 to 100. The activation function outputs four types of error probability distributions.