An athlete action biomechanics evaluation method, system, device and medium based on deep learning

Abstract

The application relates to an athlete action biomechanics evaluation method, system, device and medium based on deep learning, the method comprising: obtaining a motion signal sequence by time synchronization and cycle segmentation of a multi-modal technical action signal of an athlete, generating a joint representation tensor by kinematics and dynamics feature fusion, obtaining a bottom feature map by spatiotemporal coding of fused individual morphological parameters, performing cross-attention decoding based on a pre-constructed biomechanics semantic knowledge space to generate a semantic state matrix and a concept activation sequence with a concept node as a query vector, generating an evaluation result along a causal topological structure, and outputting an interpretable biomechanics evaluation report through natural language generation. The method can improve the interpretability, causal tracing ability and training guidance value of the athlete action biomechanics evaluation.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and sports biomechanics, and in particular relates to a method, system, device and medium for evaluating the biomechanics of athlete movements based on deep learning. Background Technology

[0002] With the deepening penetration of artificial intelligence technology into the field of sports science, deep learning-based athlete motion analysis has become an important technical support for the scientification of sports training. In recent years, deep learning has made significant progress in human posture estimation, motion recognition, and motion quality assessment. Deep models, represented by convolutional neural networks, temporal neural networks, and Transformer architectures, can automatically extract motion features from multimodal data such as videos, inertial measurement units, electromyography signals, and force tables, enabling quantitative descriptions of athletes' technical movements.

[0003] The research direction combining deep learning and biomechanical analysis is rapidly developing. Researchers are attempting to use biomechanical features such as joint angles, angular velocities, and torques as inputs or intermediate features of deep neural networks to improve the accuracy of motion recognition and motion quality assessment. Some works further embed physical constraints into the neural network training process, such as using the Newton-Euler equations to impose constraints on joint torque predictions, or using biomechanical models to post-process and verify kinematic parameters. These studies demonstrate that the introduction of domain knowledge can effectively improve the performance of deep models in motion analysis tasks.

[0004] However, current deep learning-based biomechanical evaluation methods for athletes suffer from the following problems: First, there is a fundamental semantic gap between the internal representation of deep models and biomechanical semantics. The distributed numerical vectors learned by the network are difficult to correspond to clear conceptual systems such as joint range of motion and torque generation efficiency. Existing visualization methods can only show the areas that the model focuses on, and cannot reveal what kind of movement laws the model understands. Second, existing methods cannot distinguish between statistical correlation and biomechanical causal relationship. The model is easily affected by spurious correlations such as clothing color and background features, and often misjudges outcome features as root causes, making it difficult to trace the deep-seated causes. Third, there is a lack of effective integration mechanism between domain expert knowledge and neural network numerical calculation paradigm. Existing solutions either completely abandon knowledge dependence and are purely data-driven, or solidify knowledge into rigid feature engineering rules. Fourth, interpretability generation generally adopts a post-interpretation separation paradigm, which lacks deep coupling with the model evaluation process, making it difficult to guarantee the accuracy and fidelity of the interpretation. Fifth, existing systems lack a unified underlying semantic framework that represents the common biomechanical evaluation principles of different sports, resulting in poor model reusability, difficulty in knowledge transfer across sports, and a lack of consistency in interpretation output. Summary of the Invention

[0005] Based on this, it is necessary to provide a deep learning-based method, system, device, and medium for evaluating the biomechanics of athlete movements that can eliminate the semantic gap between the internal representation of deep models and biomechanical semantics, distinguish between statistical correlation and biomechanical causal association, achieve deep integration of domain expert knowledge and neural networks, deeply couple interpretability generation and evaluation process, and have a unified biomechanical semantic framework to address the above-mentioned technical problems.

[0006] Firstly, this application provides a deep learning-based method for evaluating the biomechanics of athlete movements, including:

[0007] S1. Acquire the multimodal technical motion signals of the athlete, and perform time synchronization and motion cycle segmentation on the multimodal technical motion signals to obtain the motion signal time series;

[0008] S2. Multi-dimensional biomechanical features are extracted and fused from the time series of motion signals to generate a kinematic-dynamic joint representation tensor. The kinematic-dynamic joint representation tensor is used to characterize the relationship between the spatiotemporal motion morphology and dynamic force characteristics of the athlete during the corresponding action cycle.

[0009] S3. Obtain the individual morphological parameters of the athlete, encode the spatiotemporal features of the individual morphological parameters and the kinematic-dynamic joint representation tensor, and generate the underlying spatiotemporal feature map;

[0010] S4. Based on the pre-constructed biomechanical semantic knowledge space, the concept nodes in the biomechanical semantic knowledge space are used as query vectors to perform cross-attention decoding and fusion on the underlying spatiotemporal feature map, generating a semantic state matrix corresponding to each concept node, as well as a sequence of concept activation vectors of each concept node on the time axis.

[0011] S5. Based on the causal topological structure of the biomechanical semantic knowledge space, perform anomaly tracing analysis on the semantic state matrix and concept activation vector sequence to generate biomechanical evaluation results.

[0012] S6. Perform natural language processing on the biomechanical evaluation results to obtain an interpretable biomechanical evaluation report.

[0013] Secondly, this application also provides a deep learning-based biomechanical evaluation system for athlete movements, including:

[0014] The time series generation module is used to acquire the multimodal technical motion signals of athletes, and to perform time synchronization and motion cycle segmentation on the multimodal technical motion signals to obtain the motion signal time series;

[0015] The joint representation construction module is used to extract and fuse multi-dimensional biomechanical features from motion signal time series to generate a kinematic-dynamic joint representation tensor. The kinematic-dynamic joint representation tensor is used to represent the relationship between the spatiotemporal motion morphology and dynamic force characteristics of an athlete during the corresponding action cycle.

[0016] The spatiotemporal feature encoding module is used to obtain the individual morphological parameters of athletes, encode the individual morphological parameters and the kinematic-dynamic joint representation tensor in a spatiotemporal manner, and generate the underlying spatiotemporal feature map.

[0017] The concept activation module is used to perform cross-attention decoding and fusion on the underlying spatiotemporal feature map based on the pre-constructed biomechanical semantic knowledge space, using concept nodes in the biomechanical semantic knowledge space as query vectors, to generate a semantic state matrix corresponding to each concept node, as well as a sequence of concept activation vectors for each concept node on the time axis.

[0018] The causal attribution module is used to perform anomaly attribution analysis on semantic state matrices and concept activation vector sequences based on the causal topology of the biomechanical semantic knowledge space, and generate biomechanical evaluation results.

[0019] The report generation module is used to process the biomechanical evaluation results using natural language to generate an interpretable biomechanical evaluation report.

[0020] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0021] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0022] The aforementioned method, system, device, and medium for evaluating athlete movement biomechanics based on deep learning acquire multimodal technical movement signals from athletes and perform time synchronization and movement cycle segmentation to obtain a time series of movement signals. Multi-dimensional biomechanical feature extraction and fusion of the movement signal time series generate a kinematic-dynamic joint representation tensor that characterizes the spatiotemporal movement morphology and dynamic force characteristics, laying a multi-source data foundation for comprehensively depicting the biomechanical state of athletes' movements. By acquiring individual morphological parameters of athletes and performing spatiotemporal feature encoding on these parameters and the kinematic-dynamic joint representation tensor, a low-level spatiotemporal feature map is generated, integrating individual anatomical differences into the movement representation and providing an individualized feature foundation for subsequent semantic alignment. Furthermore, by pre-constructing a biomechanical semantic knowledge space including a concept ontology, a set of semantic token vectors, and an inter-concept causal graph, and using concept nodes in the biomechanical semantic knowledge space as query vectors, the low-level spatiotemporal feature map is cross-referenced. Attention-based decoding and fusion to generate semantic state matrices and concept activation vector sequences establish a one-to-one correspondence between the distributed numerical representations learned by deep models and explicit biomechanical conceptual systems such as joint range of motion, torque generation efficiency, and kinetic chain transmission timing. This achieves native alignment between the model's internal reasoning logic and biomechanical semantics, fundamentally eliminating the semantic gap. Anomaly tracing analysis of the semantic state matrix and concept activation vector sequences based on a causal topological structure of the biomechanical semantic knowledge space generates biomechanical evaluation results. This distinguishes between statistical correlation and biomechanical causal association, tracing back from observed movement deviations along the causal graph to the root cause node and quantifying the attribution contribution of each cause to the abnormal results, avoiding misjudging outcome features as root causes. Multi-granularity natural language generation processing of the biomechanical evaluation results yields an interpretable biomechanical evaluation report, deeply integrating the evaluation process and interpretation generation into the same cognitive process, ensuring the fidelity and professionalism of the interpretation content. This method enhances the interpretability, causal tracing ability, and training guidance value of athlete movement biomechanical evaluation. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating a deep learning-based biomechanical evaluation method for athlete movement, provided as an exemplary embodiment of this application;

[0025] Figure 2A flowchart illustrating a method for generating a low-level spatiotemporal feature map, provided as an exemplary embodiment of this application;

[0026] Figure 3 This is a schematic diagram of the structure of a deep learning-based biomechanical evaluation system for athlete movement, provided as an exemplary embodiment of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] In one embodiment, such as Figure 1 As shown, a deep learning-based biomechanical evaluation method for athlete movements is provided. This embodiment illustrates the application of this method to an intelligent evaluation terminal. It is understood that this method can also be applied to an intelligent evaluation server, and further to a system including both an intelligent evaluation terminal and an intelligent evaluation server, and is implemented through the interaction between the intelligent evaluation terminal and the intelligent evaluation server. In this embodiment, the method includes the following steps:

[0029] S1. Acquire the multimodal technical motion signals of the athlete, and perform time synchronization and motion cycle segmentation on the multimodal technical motion signals to obtain the motion signal time series.

[0030] Specifically, the intelligent evaluation terminal can connect to corresponding acquisition devices through a preset multi-source data acquisition adapter interface to acquire multimodal technical movement signals of athletes. Using a unified hardware trigger clock as the time reference, it timestamps the multimodal technical movement signals, eliminating transmission delays and sampling frequency deviations between different acquisition devices, and completing time synchronization processing. Furthermore, the intelligent evaluation terminal can identify the start and end points of a complete movement based on the extreme points of the movement's timing characteristics, and uniformly segment the synchronized multimodal technical movement signals into single-movement cycles and multi-movement cycles to obtain a time sequence of motion signals with time alignment and complete cycles.

[0031] Optionally, the athlete's multimodal technical movement signal can be spatial motion modal data and dynamic modal data containing the athlete's complete technical movement process. The spatial motion modal data may include, but is not limited to, optical motion capture data, monocular or multi-view video acquisition data, and attitude timing data acquired by IMU (Inertial Measurement Unit). The dynamic modal data may include, but is not limited to, ground reaction force data, plantar pressure distribution data, and surface electromyography signal data acquired by a three-dimensional force table.

[0032] Optionally, the motion signal time series can be used to characterize the motion technique signals of an athlete that change continuously over time in each acquisition dimension throughout the complete motion cycle.

[0033] S2. Multi-dimensional biomechanical features are extracted and fused from the time series of motion signals to generate a kinematic-dynamic joint representation tensor.

[0034] Specifically, the intelligent evaluation terminal can perform joint spatial coordinate calculation and temporal smoothing on the spatial motion modal data of motion signal time series to generate a continuous temporal kinematic parameter sequence. Furthermore, the intelligent evaluation terminal can perform adaptive filtering and denoising and baseline calibration on the dynamic modal data of motion signal time series to extract core dynamic features. Based on a unified time axis, the intelligent evaluation terminal can complete the spatiotemporal dimension alignment and feature concatenation of the kinematic parameter sequence and core dynamic features, and after tensor organization, generate a joint kinematic-dynamic representation tensor.

[0035] Optionally, the core dynamic features may include, but are not limited to, contact forces and force loading rates.

[0036] Optionally, the kinematic-dynamic joint characterization tensor can be used to characterize the relationship between the spatiotemporal motion morphology and dynamic force characteristics of an athlete during a corresponding action cycle.

[0037] S3. Obtain the individual morphological parameters of the athlete, encode the spatiotemporal features of the individual morphological parameters and the kinematic-dynamic joint representation tensor, and generate the underlying spatiotemporal feature map.

[0038] Specifically, the intelligent evaluation terminal can acquire the athlete's individual morphological parameters through a preset parameter input interface. Further, the intelligent evaluation terminal can standardize and nonlinearly map the athlete's individual morphological parameters, transforming the discrete morphological parameters into fixed-dimensional morphological embedding vectors. The intelligent evaluation terminal can then sequentially concatenate these morphological embedding vectors with the temporal position features of the kinematic-dynamic joint representation tensor along the feature dimension, achieving a deep fusion of individual anatomical features and biomechanical representations. Furthermore, the intelligent evaluation terminal can extract spatiotemporal context features from the enhanced motion representation tensor obtained after concatenation and fusion, and after dimensional mapping and residual feature transfer, generate a low-level spatiotemporal feature map.

[0039] Optionally, the athlete's individual morphological parameters can be parameters of the athlete's individual anatomical characteristics, including but not limited to height, weight, limb length, joint range of motion, body fat percentage, and limb circumference.

[0040] Optionally, the underlying spatiotemporal feature map can be used to characterize the spatiotemporal and biomechanical features of motion that incorporate individual anatomical differences in athletes.

[0041] S4. Based on the pre-constructed biomechanical semantic knowledge space, the concept nodes in the biomechanical semantic knowledge space are used as query vectors. The underlying spatiotemporal feature map is cross-attention decoded and fused to generate a semantic state matrix corresponding to each concept node, as well as a sequence of concept activation vectors of each concept node on the time axis.

[0042] Specifically, the intelligent evaluation terminal can access a pre-constructed biomechanical semantic knowledge space and extract semantic query vectors corresponding to each concept node within that space. Further, the intelligent evaluation terminal can use these semantic query vectors as query vectors in the decoding process, and the underlying spatiotemporal feature maps as key and value vectors, respectively. It then completes feature decoding through a cross-attention mechanism, calculates the correlation weights between each concept node and the underlying spatiotemporal features, and generates a semantic state matrix corresponding to each concept node after weighted fusion and semantic mapping. Furthermore, the intelligent evaluation terminal can calculate the activation level of each concept node at different temporal positions along the time axis, generating a sequence of concept activation vectors corresponding to the semantic state matrix.

[0043] Optionally, the pre-constructed biomechanical semantic knowledge space can be built based on prior knowledge of motion biomechanics. The biomechanical semantic knowledge space can be a standardized knowledge system that includes a set of biomechanical concepts, a causal topological structure between concepts, and semantic query vectors corresponding to each concept node.

[0044] Optionally, the set of biomechanical concepts may include, but is not limited to, joint range of motion, torque generation efficiency, kinetic chain transmission timing, and motion rhythm stability.

[0045] Optionally, the semantic state matrix can be used to characterize the feature expression state of the corresponding concept node within the complete action cycle, and to quantify the degree of matching between action features and corresponding biomechanical concepts.

[0046] Optionally, the concept activation vector sequence of each concept node on the time axis can be used to characterize the activation degree and change pattern of the corresponding biomechanical concept at different temporal stages of the action cycle, reflecting the temporal performance of the biomechanical characteristics corresponding to the concept during the action process.

[0047] S5. Based on the causal topological structure of the biomechanical semantic knowledge space, perform anomaly tracing analysis on the semantic state matrix and concept activation vector sequence to generate biomechanical evaluation results.

[0048] Specifically, the intelligent evaluation terminal can extract a predefined causal topology from the biomechanical semantic knowledge space, obtaining the directed causal connections and complete causal transmission paths between various concept nodes. Furthermore, based on the causal topology, the intelligent evaluation terminal can identify abnormal concept nodes whose activation levels deviate from the standard range through a preset standard threshold interval. The intelligent evaluation terminal can then follow the reverse causal path of the causal topology, performing layer-by-layer attribution from the abnormal result node to the causal precursor node, quantifying the contribution of each precursor node to the abnormal result, locating the root cause node of the abnormal action, and integrating the results to generate a biomechanical evaluation result.

[0049] Preferably, the contribution can be calculated using the following formula:

[0050]

[0051] In the formula, Indicates the root cause node For abnormal result nodes causal attribution contribution Indicates the root cause node To the abnormal result node The set of all directed causal paths, Represents the set of all directed causal paths. A single directed causal path in the middle. Represents a single directed causal path Up, from the predecessor node Point to the successor node The directed causal edge, Represents a single directed causal path directed edges on The causal strength parameter, based on the predefined causal topology of the biomechanical semantic knowledge space, can be used to quantify the degree of influence of predecessor nodes on successor nodes.

[0052] Optionally, the biomechanical evaluation results can be used to characterize the degree of deviation between the biomechanical characteristics of an athlete's movements and the standard movement model, the location results of abnormal biomechanical characteristics, the causal tracing path of the abnormal causes, and the quantification results of the contribution of each cause.

[0053] S6. Perform natural language processing on the biomechanical evaluation results to obtain an interpretable biomechanical evaluation report.

[0054] Specifically, the intelligent evaluation terminal can perform structured analysis on abnormal nodes, causal paths, and contribution quantification data in biomechanical evaluation results to extract core evaluation elements. Based on a pre-set sports biomechanics terminology database and standardized report generation templates, the intelligent evaluation terminal can convert structured evaluation data into natural language text conforming to sports training and biomechanics professional standards, match corresponding training improvement guidance content with abnormal features, complete multi-granular text generation and content integration, and generate an interpretable biomechanical evaluation report.

[0055] Optionally, the core evaluation elements may be quantitative evaluation indicators of the overall quality of the athlete's movements, abnormal biomechanical concept nodes that deviate from the standard movement model, the temporal occurrence stage corresponding to the abnormal nodes, the causal transmission path of the abnormal causes, the contribution ratio of each causal node to the abnormal result, and the training optimization direction that matches the abnormal characteristics.

[0056] Optionally, the preset sports biomechanics terminology database can be a standardized set of terms built on authoritative academic norms in the field of sports biomechanics at home and abroad and general standards in the sports training industry. The sports biomechanics terminology database may include, but is not limited to, professional terms in human anatomy, kinesiology, dynamics, and sports training, and form a one-to-one correspondence with the concept node terms in the biomechanics semantic knowledge space.

[0057] Optionally, the training improvement guidance content can be personalized training content matched with abnormal biomechanical characteristics and the root cause obtained by tracing the source. The training improvement guidance content may include, but is not limited to, strength activation and strengthening training programs for corresponding muscle groups, flexibility training programs to improve joint range of motion, patterned training programs to correct movement timing deviations, synergistic training programs to optimize kinetic chain transmission efficiency, and targeted control training content to avoid the risk of sports injuries.

[0058] Optionally, an interpretable biomechanical evaluation report can be a standardized professional report that combines biomechanical expertise, causal interpretability, and practical training guidance. An interpretable biomechanical evaluation report may include, but is not limited to, an overall quality score of the athlete's movements, sub-evaluations of biomechanical characteristics across various dimensions, precise localization of movement abnormalities, causal analysis of the causes of these abnormalities, directions for movement optimization, and personalized training improvement suggestions.

[0059] In the aforementioned deep learning-based biomechanical evaluation method for athlete movements, the intelligent evaluation terminal acquires multimodal technical movement signals from athletes and performs time synchronization and movement cycle segmentation to obtain a time series of movement signals. Multi-dimensional biomechanical feature extraction and fusion of the time series of movement signals generate a kinematic-dynamic joint representation tensor that characterizes the spatiotemporal movement morphology and dynamic force characteristics, laying a multi-source data foundation for comprehensively depicting the biomechanical state of athletes' movements. By acquiring individual morphological parameters of athletes and performing spatiotemporal feature encoding on these parameters and the kinematic-dynamic joint representation tensor, a low-level spatiotemporal feature map is generated, integrating individual anatomical differences into the movement representation and providing an individualized feature foundation for subsequent semantic alignment. Finally, a pre-constructed biomechanical semantic knowledge space, including a concept ontology, a set of semantic token vectors, and an inter-concept causal graph, is used, with concept nodes in the biomechanical semantic knowledge space serving as query vectors, to perform cross-annotation on the low-level spatiotemporal feature map. This method decodes and fuses semantic state matrices and concept activation vector sequences, establishing a one-to-one correspondence between the distributed numerical representations learned by deep models and explicit biomechanical conceptual systems such as joint range of motion, torque generation efficiency, and kinetic chain transmission timing. This achieves native alignment between the model's internal reasoning logic and biomechanical semantics, fundamentally eliminating the semantic gap. By performing anomaly tracing analysis on the semantic state matrix and concept activation vector sequences based on the causal topology of the biomechanical semantic knowledge space, it generates biomechanical evaluation results. This distinguishes between statistical correlation and biomechanical causal association, tracing back from observed movement deviations along the causal graph to the root cause node and quantifying the attribution contribution of each cause to the abnormal results, avoiding misjudging outcome features as root causes. Multi-granularity natural language generation processing of the biomechanical evaluation results yields an interpretable biomechanical evaluation report, deeply integrating the evaluation process and interpretation generation into the same cognitive process, ensuring the fidelity and professionalism of the interpretation content. This method enhances the interpretability, causal tracing ability, and training guidance value of athlete movement biomechanical evaluation.

[0060] In one embodiment, such as Figure 2 As shown, a flowchart illustrating a method for generating a low-level spatiotemporal feature map is provided. S3 may include:

[0061] S31. Map individual morphological parameters to morphological embedding vectors, and concatenate the morphological embedding vectors with the spatiotemporal positions of the kinematic-dynamic joint representation tensor in the feature dimension to obtain the enhanced motion representation tensor.

[0062] For example, the intelligent evaluation terminal can perform dimensionless standardization on the acquired individual morphological parameters to eliminate dimensional and numerical range differences between different types of parameters. The intelligent evaluation terminal can perform nonlinear mapping using a multilayer perceptron to transform discrete scalar morphological parameters into dense morphological embedding vectors of fixed dimensions. Furthermore, the intelligent evaluation terminal can extend the dense morphological embedding vectors to a spatiotemporal dimension consistent with the kinematic-dynamic joint representation tensor through a broadcast operation. It then performs bit-by-bit concatenation with the temporal and spatial features of the tensor along the feature channel dimension, completing feature fusion to generate an enhanced motion representation tensor.

[0063] Preferably, the expression for the morphological embedding vector can be:

[0064]

[0065]

[0066] In the formula, This represents a scalar vector composed of the original individual's morphological parameters. This represents the statistical mean of the morphological parameter training set. This represents the statistical variance of the morphological parameter training set. This represents a dimensionless standardized vector of individual morphological parameters, used to eliminate differences in dimensions and numerical ranges between different types of parameters. Indicates training parameters A multilayer perceptron mapping network is used to perform nonlinear embedding mapping of standardized morphological parameters. This represents a fixed-dimensional morphological embedding vector used to characterize the individual anatomical features of an athlete.

[0067] Optionally, the enhanced motion representation tensor can be used to characterize spatiotemporal motion dynamics states that incorporate individual anatomical differences.

[0068] S32. Extract spatiotemporal context features from the enhanced motion representation tensor to generate a motion feature sequence.

[0069] For example, the intelligent evaluation terminal can extract local contextual features of the action temporal dimension in the enhanced motion representation tensor through a temporal convolutional network, capturing the correlation of action changes and local motion patterns between adjacent temporal positions. The intelligent evaluation terminal can perform global dependency modeling on the local contextual features output by the temporal convolution, capturing long-distance spatiotemporal correlation features within the complete action cycle, and simultaneously performing deep extraction of action spatial structure features and temporal evolution features. After feature fusion and dimensionality normalization, a motion feature sequence matching the temporal length of the action cycle is generated.

[0070] Preferably, the intelligent evaluation terminal can complete global spatiotemporal dependency modeling through a multi-head self-attention mechanism, and the calculation formula is as follows:

[0071]

[0072]

[0073] In the formula, This represents the output features of multi-head self-attention. Represents local contextual features, Indicates the first The computational output of each attention head. This represents the predefined total number of attention heads. Indicates the first The query vector corresponding to each attention head. Indicates the first The key vector corresponding to each attention head. Indicates the first The value vector corresponding to each attention head This indicates a feature concatenation operation, used to fuse the output features of multiple attention heads. The linear mapping weight matrix represents the multi-head attention output, used to map the concatenated features to the target output dimension.

[0074] Optionally, the motion feature sequence can be used to characterize the high-order kinematics features of each temporal stage within the motion cycle, preserving the local temporal variation patterns of the motion and the global spatiotemporal context information, while taking into account the representational ability of motion detail features and overall motion patterns.

[0075] S33. Perform dimensional mapping and residual feature transfer on the motion feature sequence to generate the underlying spatiotemporal feature map.

[0076] For example, the intelligent evaluation terminal can perform dimension adaptation mapping on the motion feature sequence, transforming the motion feature sequence to a preset target feature dimension space to match the input specifications of the subsequent semantic decoding stage. The intelligent evaluation terminal can add and fuse the high-order motion features obtained by dimension mapping with the original motion feature sequence that has not been dimension mapped. While extracting the high-order motion features adapted to the target dimension, it fully retains the effective spatiotemporal information carried in the original features. After feature integration, a low-level spatiotemporal feature map is generated.

[0077] In this embodiment, the intelligent evaluation terminal maps individual morphological parameters into morphological embedding vectors, and extracts spatiotemporal context features by splicing and fusing them with the kinematic-dynamic joint representation tensor. Then, through dimensional mapping and residual feature transfer, it generates a low-level spatiotemporal feature map, which can deeply integrate the individual anatomical differences of athletes into the spatiotemporal mechanical representation of movements, providing a feature foundation for subsequent semantic alignment that takes into account both individualized features and the fidelity of movement details.

[0078] In one embodiment, S33 may include:

[0079] S331. Normalize the motion feature sequence to obtain a normalized motion feature sequence.

[0080] For example, the intelligent evaluation terminal can calculate the statistical mean and variance of a single motion feature sequence in the feature dimension, and perform linear transformation and standardization calibration on the motion feature sequence based on the obtained mean and variance to eliminate the distribution offset and dimensional differences of features at different time positions, reduce the risk of gradient instability in the subsequent feature mapping process, and obtain a normalized motion feature sequence with stable distribution.

[0081] Optionally, the normalized motion feature sequence can be used to characterize the temporal motion features of actions after distribution standardization, fully preserving the spatiotemporal correlation information carried by the original features.

[0082] S332. Project the normalized motion feature sequence onto the predefined target feature dimension to obtain the main path projection feature.

[0083] For example, the intelligent evaluation terminal can retrieve a predefined target feature dimension, which is used to ensure dimensional consistency in feature transmission throughout the entire process. Furthermore, the intelligent evaluation terminal can use a fully connected layer with a nonlinear activation function to perform nonlinear transformation and dimensional projection on the normalized motion feature sequence, transforming the original dimension of the normalized motion feature sequence to the predefined target feature dimension, thus completing high-order feature abstraction and dimensional adaptation to obtain the main path projection feature.

[0084] Optionally, the predefined target feature dimension can be a fixed feature dimension that is adapted to the input feature dimension of the subsequent cross-attention decoding stage and the dimension of the semantic query vector corresponding to each concept node in the biomechanical semantic knowledge space. The target feature dimension can be adapted to the size of the biomechanical concept set, the length of the action sequence, and the dimensional scale of the multimodal features to ensure that the features can complete accurate attention weight matching calculation with the semantic query vector in the subsequent semantic decoding process.

[0085] Optionally, the main path projection features are used to characterize higher-order motion features after nonlinear mapping.

[0086] S333. Perform residual fusion on the normalized motion feature sequence and the main path projection feature to obtain the underlying spatiotemporal feature map.

[0087] For example, the intelligent evaluation terminal can perform identity mapping on the normalized motion feature sequence to ensure that the processed feature dimension and the target feature dimension of the main path projection feature are completely matched. The intelligent evaluation terminal can then add and fuse the dimension-matched normalized motion feature sequence with the main path projection feature element by element, preserving the effective spatiotemporal details of the original features while incorporating higher-order abstract motion features. After nonlinear activation processing, the fused features generate a low-level spatiotemporal feature map.

[0088] In this embodiment, the intelligent evaluation terminal eliminates distribution offset by normalizing the motion feature sequence, projects the normalized features onto the target feature dimension that matches the semantic query vector, and performs residual fusion between the normalized features and the main path projection features. This enables the high-order semantic adaptation to be completed while preserving the original motion details, providing a stable, dimensionally aligned, and lossless underlying spatiotemporal feature map for subsequent cross-attention decoding.

[0089] In one embodiment, the construction of the biomechanical semantic knowledge space in S4 may include:

[0090] S41. Obtain prior biomechanical knowledge information, extract biomechanical concepts from the prior biomechanical knowledge information, and construct a set of biomechanical concepts by using each biomechanical concept as a concept node.

[0091] For example, the intelligent evaluation terminal can acquire structured and unstructured biomechanical prior knowledge from professional theories, technical movement standards, common movement error patterns, and injury risk assessment rules in the field of sports biomechanics. It then performs concept extraction and deduplication categorization on this information, eliminating redundant expressions and retaining core evaluation elements directly applicable to movement evaluation. Furthermore, the intelligent evaluation terminal can define each independent evaluation element as a concept node, and aggregate and organize all concept nodes to generate a unified and standardized set of biomechanical concepts.

[0092] Optionally, the biomechanical concepts in the prior biomechanical knowledge information can be specialized concepts used to describe motion quality, mechanical state, kinetic chain transmission relationships, and risk characteristics. Biomechanical concepts may include, but are not limited to, joint range of motion, torque generation efficiency, motion timing consistency, and muscle group synergy.

[0093] Optionally, the set of biomechanical concepts can be used to characterize the entire professional conceptual system upon which the biomechanical evaluation of an athlete's movements relies, providing a standardized node foundation for subsequent knowledge graph construction.

[0094] S42. Based on the causal logic between various biomechanical concepts, directed connection edges are constructed between the concept nodes in the biomechanical concept set to generate a biomechanical semantic graph containing a causal topology.

[0095] For example, the intelligent evaluation terminal can determine the influence relationships and transmission directions between different biomechanical concepts based on the physical laws and technical transmission logic of motion formation in sports biomechanics. Following a cause-to-effect approach, it can establish directed connections between concept nodes with direct or indirect influence relationships. Furthermore, the intelligent evaluation terminal can perform overall structured organization of all biomechanical concept nodes and directed connections to generate a causal topology that reflects the influence paths between biomechanical concepts, thus obtaining a biomechanical semantic graph.

[0096] Optionally, the causal logic between various biomechanical concepts can be objective physical logic such as the generation and transmission relationship of movement deviations, the constraint relationship between muscle group function and joint movement, and the correspondence between abnormal mechanical parameters and technical defects.

[0097] Optionally, the biomechanical semantic map can be used to characterize the influence paths and transmission relationships between various biomechanical concepts, providing a reasonable topological basis for subsequent anomaly tracing.

[0098] S43. Perform semantic feature embedding processing on each concept node in the biomechanical semantic graph to generate a semantic query vector corresponding to each concept node, and combine the semantic query vector and the biomechanical semantic graph to obtain the biomechanical semantic knowledge space.

[0099] For example, the intelligent evaluation terminal can perform semantic vectorization mapping on each concept node in the biomechanical semantic graph, converting the textual professional concepts into dense numerical feature vectors of fixed dimensions, so that each vector can carry the professional semantic information of the corresponding concept. Furthermore, the intelligent evaluation terminal can uniformly store and manage the vectors corresponding to all concept nodes, and integrate the vectors corresponding to all concept nodes with the constructed biomechanical semantic graph to generate a biomechanical semantic knowledge space that can be directly used for feature decoding.

[0100] Optionally, the semantic query vector can be used to represent the professional semantic connotation of the corresponding biomechanical concept, and serve as a matching benchmark in the subsequent cross-attention decoding process to achieve the alignment of the underlying action features and the high-level semantic concepts.

[0101] In this embodiment, the intelligent evaluation terminal extracts concepts from biomechanical prior knowledge information to construct a concept set, establishes directed connection edges based on causal logic to generate a semantic graph, and generates semantic query vectors by semantically vectorizing each concept node and combining them with the graph. This enables the construction of a unified biomechanical semantic knowledge space that integrates the concept system, causal topology, and semantic representation, providing a standardized and reasonable semantic anchoring foundation for subsequent cross-attention decoding.

[0102] In one embodiment, S2 may include:

[0103] S21. Perform motion state analysis on the spatial motion modal data to generate a sequence of kinematic parameters.

[0104] For example, the intelligent evaluation terminal can preprocess the acquired spatial motion modal data, which may originate from acquisition terminals such as optical motion capture systems, monocular or multi-view video acquisition devices, and IMUs. Preprocessing may include outlier removal, temporal interpolation, and sliding window smoothing to eliminate environmental noise, missing data, and jitter bias generated during acquisition. Furthermore, the intelligent evaluation terminal can perform joint spatial coordinate calculations on the preprocessed spatial motion modal data to obtain the temporal position data of each joint in a unified world coordinate system. Through differential operations and joint coordinate system transformation, temporal parameters are calculated. The intelligent evaluation terminal can then normalize the data according to the motion cycle time axis to generate a kinematic parameter sequence.

[0105] Optionally, the motion signal time series may include, but is not limited to, spatial motion mode data and dynamic mode data.

[0106] Optionally, the sequence of kinematic parameters can be used to characterize the spatiotemporal motion pattern of an athlete during a corresponding motion cycle.

[0107] S22. Perform inverse dynamics calculation on the kinematic parameter sequence and dynamic modal data to generate a dynamic parameter sequence.

[0108] For example, the intelligent evaluation terminal can use a sequence of kinematic parameters as input for motion state, combine it with the athlete's individual morphological parameters to construct a suitable multi-rigid-body dynamic model of the human body, use measured force data from the dynamic modal data as boundary constraints, and deduce the core dynamic temporal parameters within the motion cycle through an inverse dynamics solution process. Furthermore, the intelligent evaluation terminal can iteratively calibrate the deduced results based on preset physical motion equations to eliminate error accumulation and deviations during the solution process, obtaining a dynamic parameter sequence that is strictly aligned with the motion cycle temporal sequence.

[0109] Optionally, the core dynamic timing parameters may include, but are not limited to, the torque and power of each joint during the motion cycle.

[0110] Optionally, the sequence of dynamic parameters can be used to characterize the dynamic force, work efficiency, and energy transfer characteristics of each joint of the mobilizer during the corresponding action cycle.

[0111] S23. The kinematic parameter sequence and the dynamic parameter sequence are spliced ​​and tensorized in the spatiotemporal dimension to generate a joint kinematic-dynamic tensor.

[0112] For example, the intelligent evaluation terminal can use a unified timeline of the complete action cycle as a benchmark to strictly align the kinematic parameter sequence and the dynamic parameter sequence in the temporal dimension. This ensures that the kinematic parameter sequence and the dynamic parameter sequence correspond to the same stage of action execution at the same temporal position, eliminating temporal offset deviations between the kinematic parameter sequence and the dynamic parameter sequence. Furthermore, the intelligent evaluation terminal can perform position-by-position concatenation of the temporally aligned kinematic parameter sequence and the dynamic parameter sequence in the feature channel dimension to obtain a multidimensional feature sequence that integrates kinematic spatiotemporal morphological features and intrinsic mechanical features of dynamics.

[0113] Furthermore, the intelligent evaluation terminal can perform dimensional normalization and tensor quantization on the spliced ​​multidimensional feature sequence, converting it into a multidimensional tensor that conforms to the subsequent feature encoding input specifications, namely the kinematic-dynamic joint representation tensor.

[0114] In this embodiment, the intelligent evaluation terminal generates a kinematic parameter sequence by parsing the spatial motion modal data, generates a dynamic parameter sequence by inverse dynamic calculation of the kinematic parameter sequence and dynamic modal data, and splices and tensors the kinematic parameter sequence and dynamic parameter sequence in the spatiotemporal dimension. This enables the integration of the spatiotemporal motion morphology and dynamic force characteristics of the athlete's movements into a unified joint representation tensor, providing a complete input foundation of both kinematic and dynamic information for subsequent individualized feature encoding.

[0115] In one embodiment, S22 may include:

[0116] S221. Perform feature mapping on the kinematic parameter sequence and dynamic modal data to generate an initial predicted joint torque vector.

[0117] For example, the intelligent evaluation terminal can strictly align the kinematic parameter sequence and dynamic modal data along the temporal dimension, ensuring a one-to-one correspondence between the kinematic state parameters and the measured force parameters at the same moment. The intelligent evaluation terminal can extract the joint angles, angular velocities, angular accelerations, and measured force vectors corresponding to the temporally aligned kinematic parameter sequence and dynamic modal data at each moment, and then perform feature concatenation on these parameters. The intelligent evaluation terminal can input the concatenated fused features into a multilayer perceptron mapping network, complete the associated feature mapping through the nonlinear transformation of the multilayer perceptron mapping network, output the initial predicted joint torque vector at the corresponding moment, and integrate them along the action cycle time axis to obtain a complete temporal sequence of initial predicted joint torque vectors.

[0118] Optionally, the initial predicted joint torque vector can be used to characterize the initial value of the joint driving torque calculated based on the real-time kinematic state of the action and the measured boundary force conditions, providing an initial solution benchmark for subsequent iterative correction based on physical constraints.

[0119] Preferably, the expression for the initial predicted joint moment vector can be:

[0120]

[0121] In the formula, for The initial predicted joint torque vector at time t. For trainable parameters Multilayer perceptron mapping network, for Joint angles in the kinematic parameter sequence at any given time. for Angular velocity in the kinematic parameter sequence at time t. for Angular acceleration in the kinematic parameter sequence at time t. for The measured force vectors of the foot or body surface in the dynamic modal data at any given time. This is a feature splicing operation.

[0122] S222. Substitute the initial predicted joint moment vector into the preset physical motion equation for forward deduction to generate theoretical physical state values, and calculate the physical residual between the theoretical physical state values ​​and the actual measured values ​​in the dynamic modal data.

[0123] For example, the intelligent evaluation terminal can retrieve preset physical motion equations, substitute the initial predicted joint torque vectors at each moment into the physical motion equations, and perform forward dynamic deduction by combining the kinematic parameter sequence and athlete's individual morphological parameters at the corresponding moment to obtain the theoretical contact force vector at the corresponding moment, i.e., the theoretical physical state value. Further, the intelligent evaluation terminal can calculate the fitting residual between the theoretical physical state value and the measured force data, and the regularized residual between the predicted joint torque and the initial predicted joint torque. After weighted summing of the fitting residual and the regularized residual, the physical residual is obtained, completing the physical constraint verification and residual calculation.

[0124] Optionally, the physical motion equation can be a human multi-rigid-body dynamics equation constructed based on the principles of Newton-Euler dynamics. The physical motion equation can be adapted by combining individual morphological parameters of athletes to describe the objective physical mapping relationship between human joint movement and joint torque and end force.

[0125] Optionally, the theoretical physical state value can be used to characterize the theoretical contact force vector at the end of the motion obtained by forward deduction based on the predicted joint torque, reflecting the theoretical state of the interaction between the human body and the environment under the current joint torque.

[0126] Preferably, the expression for the physical residual can be:

[0127]

[0128] In the formula, For physical residuals, for The theoretical physical state value at time t. for The measured force vectors of the foot or body surface in the dynamic modal data at any given time. The square of the L2 norm. The preset regularization weight coefficients, express The predicted joint torque vector to be solved at any given time. express The initial predicted joint torque vector at time t.

[0129] S223. Based on the physical residual, iteratively correct the initial predicted joint torque vector until the physical residual meets the preset convergence condition, output the corrected joint torque vector sequence, and generate the dynamic parameter sequence.

[0130] For example, the intelligent evaluation terminal can take minimizing the physical residual as the optimization objective, and iteratively correct the initial predicted joint moment vector based on a preset iterative update rule. During each iteration, the intelligent evaluation terminal can calculate the gradient of the physical residual with respect to the current iterative joint moment vector, update the value of the joint moment vector along the gradient descent direction, and then resubstitute it into the preset physical motion equation to calculate the new physical residual. This iterative optimization continues until the physical residual meets the preset convergence condition. The iteration stops, and the corrected full-time sequence of joint moment vectors is output, generating a dynamic parameter sequence.

[0131] Optionally, the preset convergence condition can be that the physical residual is less than or equal to a preset convergence threshold, or that the number of iterations reaches a preset maximum number of iterations, in order to ensure the accuracy of the solution while taking into account the computational efficiency of the model.

[0132] Preferably, the update formula for iterative correction can be:

[0133]

[0134] In the formula, For the first During the next iteration The joint torque vector at time t, The preset learning rate, This represents the gradient of the physical residual with respect to the joint torque in the current iteration. This represents the number of iterations.

[0135] In this embodiment, the intelligent evaluation terminal generates an initial predicted joint moment vector by performing correlation feature mapping on the kinematic parameter sequence and dynamic modal data, substitutes the initial predicted joint moment vector into the physical motion equation for forward deduction and calculates the physical residual, and performs gradient descent iterative correction on the initial predicted joint moment vector based on the physical residual. This enables the deep integration of data-driven preliminary estimation of joint moment with Newton-Euler rigid body dynamic constraints, and outputs a high-precision dynamic parameter sequence that fits the measured force data and strictly satisfies the physical consistency verification.

[0136] In the aforementioned method, system, device, and medium for biomechanical evaluation of athlete movements based on deep learning, the intelligent evaluation terminal obtains a motion signal time series by performing time synchronization and periodic segmentation on multimodal motion signals. Through motion state analysis and iterative inverse dynamics calculation based on physical constraints, a kinematic-dynamic joint representation tensor is generated. Individual morphological parameters are then fused for spatiotemporal feature encoding to generate a low-level spatiotemporal feature map. Based on a pre-constructed biomechanical semantic knowledge space containing concept sets, causal topological structures, and semantic query vectors, the low-level spatiotemporal feature map is cross-attention decoded to generate a semantic state matrix and a sequence of concept activation vectors. The causal topological structure is then traced back to generate the biomechanical evaluation result. Finally, an interpretable biomechanical evaluation report is output after natural language generation processing. This technical solution constructs a biomechanical semantic knowledge space and uses concept nodes as query vectors to achieve deep representation and native alignment of biomechanical semantics, fundamentally eliminating the semantic gap. Through physical constraint inverse dynamic iteration and causal topological reverse tracing, statistical correlation is upgraded to causal association. This effectively solves the core technical problems of semantic gap, spurious correlation interference, lack of domain knowledge integration, and unfaithful ex post facto interpretation in existing technologies, improving the interpretability, causal traceability, and training guidance value of the evaluation.

[0137] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0138] Based on the same inventive concept, this application also provides a system for implementing the aforementioned deep learning-based biomechanical evaluation method for athlete movement. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more deep learning-based biomechanical evaluation system embodiments provided below can be found in the above-described limitations of the deep learning-based biomechanical evaluation method for athlete movement, and will not be repeated here.

[0139] In one exemplary embodiment, such as Figure 3 As shown, a deep learning-based biomechanical evaluation system for athlete movements 70 is provided, comprising:

[0140] The time series generation module 71 can be used to acquire multimodal technical motion signals of athletes, and to perform time synchronization and motion cycle segmentation on the multimodal technical motion signals to obtain a motion signal time series;

[0141] The joint representation construction module 72 can be used to extract and fuse multi-dimensional biomechanical features from motion signal time series to generate a kinematic-dynamic joint representation tensor. The kinematic-dynamic joint representation tensor is used to represent the relationship between the spatiotemporal motion morphology and dynamic force characteristics of an athlete during the corresponding action cycle.

[0142] The spatiotemporal feature encoding module 73 can be used to obtain the individual morphological parameters of athletes, perform spatiotemporal feature encoding on the individual morphological parameters and the kinematic-dynamic joint representation tensor, and generate the underlying spatiotemporal feature map;

[0143] The concept activation module 74 can be used to perform cross-attention decoding and fusion on the underlying spatiotemporal feature map based on the pre-constructed biomechanical semantic knowledge space, using the concept nodes in the biomechanical semantic knowledge space as query vectors, to generate a semantic state matrix corresponding to each concept node, as well as a sequence of concept activation vectors for each concept node on the time axis.

[0144] The causal attribution module 75 can be used to perform anomaly attribution analysis on semantic state matrices and concept activation vector sequences based on the causal topology of biomechanical semantic knowledge space, and generate biomechanical evaluation results.

[0145] The report generation module 76 can be used to perform natural language generation processing on the biomechanical evaluation results to obtain an interpretable biomechanical evaluation report.

[0146] In one embodiment, the spatiotemporal feature encoding module includes:

[0147] The morphological embedding fusion unit can be used to map individual morphological parameters into morphological embedding vectors, and then concatenate the morphological embedding vectors with the spatiotemporal positions of the kinematic-dynamic joint representation tensor in the feature dimension to obtain the enhanced motion representation tensor; wherein, the enhanced motion representation tensor is used to represent the spatiotemporal motion mechanical state incorporating individual anatomical differences.

[0148] The spatiotemporal feature extraction unit can be used to extract spatiotemporal context features from the enhanced motion representation tensor to generate a motion feature sequence;

[0149] The mapping residual transfer unit can be used to perform dimensional mapping and residual feature transfer on motion feature sequences to generate a low-level spatiotemporal feature map.

[0150] In one embodiment, the mapping residual transfer unit includes:

[0151] The normalization subunit can be used to normalize motion feature sequences to obtain normalized motion feature sequences.

[0152] The dimension projection subunit can be used to project a normalized motion feature sequence onto a predefined target feature dimension to obtain the main path projection feature; wherein, the main path projection feature is used to characterize the higher-order motion features after nonlinear mapping.

[0153] The residual fusion subunit can be used to perform residual fusion on the normalized motion feature sequence and the main path projection feature to obtain the underlying spatiotemporal feature map.

[0154] In one embodiment, the concept activation module can also be used for:

[0155] Obtain prior knowledge information in biomechanics, extract biomechanical concepts from the prior knowledge information in biomechanics, and construct a set of biomechanical concepts by using each biomechanical concept as a concept node.

[0156] Based on the causal logic between various biomechanical concepts, directed connection edges are constructed between concept nodes in the biomechanical concept set to generate a biomechanical semantic graph containing causal topology.

[0157] Semantic feature embedding is performed on each concept node in the biomechanical semantic graph to generate a semantic query vector corresponding to each concept node. The semantic query vector and the biomechanical semantic graph are combined to obtain the biomechanical semantic knowledge space.

[0158] In one embodiment, the joint characterization construction module includes:

[0159] The kinematic analysis unit can be used to analyze the motion state of spatial motion modal data and generate a kinematic parameter sequence; the kinematic parameter sequence is used to characterize the spatiotemporal motion pattern of an athlete within the corresponding action cycle;

[0160] The inverse dynamics calculation unit can be used to perform inverse dynamics calculations on kinematic parameter sequences and dynamic modal data to generate dynamic parameter sequences;

[0161] The spatiotemporal tensor fusion unit can be used to perform feature splicing and tensor organization of kinematic parameter sequences and dynamic parameter sequences in the spatiotemporal dimension to generate a joint kinematic-dynamic representation tensor.

[0162] In one embodiment, the spatiotemporal tensor fusion unit includes:

[0163] The torque vector generation subunit can be used to perform feature mapping on kinematic parameter sequences and dynamic modal data to generate initial predicted joint torque vectors;

[0164] The expression for the initial predicted joint moment vector is as follows:

[0165]

[0166] In the formula, for The initial predicted joint torque vector at time t. For trainable parameters Multilayer perceptron mapping network, for Joint angles in the kinematic parameter sequence at any given time. for Angular velocity in the kinematic parameter sequence at time t. for Angular acceleration in the kinematic parameter sequence at time t. for The measured force vectors of the foot or body surface in the dynamic modal data at any given time. For feature splicing operations;

[0167] The physical residual calculation subunit can be used to substitute the initial predicted joint moment vector into the preset physical motion equation for forward deduction, generate theoretical physical state values, and calculate the physical residual between the theoretical physical state values ​​and the actual measured values ​​in the dynamic modal data.

[0168] The expression for the physical residual is:

[0169]

[0170] In the formula, For physical residuals, for The theoretical physical state value at time t. for The measured force vectors of the foot or body surface in the dynamic modal data at any given time. The square of the L2 norm. The preset regularization weight coefficients, express The predicted joint torque vector to be solved at any given time. express The initial predicted joint torque vector at time t;

[0171] The iterative correction subunit can be used to iteratively correct the initial predicted joint moment vector based on the physical residual until the physical residual meets the preset convergence condition, and output the corrected joint moment vector sequence to generate the dynamic parameter sequence.

[0172] The update formula for iterative correction is as follows:

[0173]

[0174] In the formula, For the first During the next iteration The joint torque vector at time t, The preset learning rate, This represents the gradient of the physical residual with respect to the joint torque in the current iteration. This represents the number of iterations.

[0175] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a deep learning-based biomechanical evaluation method for athlete movements as described above.

[0176] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0177] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0178] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A deep learning-based biomechanical evaluation method for athlete movements, characterized in that, The method includes: S1. Acquire the multimodal technical motion signals of the athlete, and perform time synchronization and motion cycle segmentation on the multimodal technical motion signals to obtain a motion signal time series; S2. Multi-dimensional biomechanical features are extracted and fused from the motion signal time series to generate a kinematic-dynamic joint representation tensor; wherein, the kinematic-dynamic joint representation tensor is used to characterize the spatiotemporal motion morphology and dynamic force characteristics of the athlete in the corresponding action cycle; S3. Obtain the individual morphological parameters of the athlete, encode the individual morphological parameters and the kinematic-dynamic joint representation tensor in a spatiotemporal manner, and generate a low-level spatiotemporal feature map; S4. Based on the pre-constructed biomechanical semantic knowledge space, the concept nodes in the biomechanical semantic knowledge space are used as query vectors to perform cross-attention decoding and fusion on the underlying spatiotemporal feature map, generating a semantic state matrix corresponding to each concept node, and a sequence of concept activation vectors of each concept node on the time axis. S5. Based on the causal topological structure of the biomechanical semantic knowledge space, perform anomaly tracing analysis on the semantic state matrix and the concept activation vector sequence to generate biomechanical evaluation results; S6. Perform natural language generation processing on the biomechanical evaluation results to obtain an interpretable biomechanical evaluation report.

2. The method according to claim 1, characterized in that, The S3 includes: S31. The individual morphological parameters are mapped to morphological embedding vectors, and the morphological embedding vectors are concatenated with the spatiotemporal positions of the kinematic-dynamic joint representation tensor in the feature dimension to obtain the enhanced motion representation tensor; wherein, the enhanced motion representation tensor is used to represent the spatiotemporal motion mechanical state incorporating individual anatomical differences; S32. Extract spatiotemporal context features from the enhanced motion representation tensor to generate a motion feature sequence; S33. Perform dimensional mapping and residual feature transfer on the motion feature sequence to generate the underlying spatiotemporal feature map.

3. The method according to claim 2, characterized in that, S33 includes: S331. Normalize the motion feature sequence to obtain a normalized motion feature sequence; S332. Project the normalized motion feature sequence onto a predefined target feature dimension to obtain the main path projection feature; wherein, the main path projection feature is used to characterize the higher-order motion features after nonlinear mapping. S333. Perform residual fusion on the normalized motion feature sequence and the main path projection feature to obtain the underlying spatiotemporal feature map.

4. The method according to claim 1, characterized in that, The construction method of the biomechanical semantic knowledge space in S4 includes: S41. Obtain prior biomechanical knowledge information, extract biomechanical concepts from the prior biomechanical knowledge information, and construct a set of biomechanical concepts by using each of the biomechanical concepts as concept nodes. S42. Based on the causal interaction logic between the biomechanical concepts, directed connection edges are constructed between the concept nodes in the biomechanical concept set to generate a biomechanical semantic graph containing a causal topology. S43. Perform semantic feature embedding processing on each concept node in the biomechanical semantic graph to generate a semantic query vector corresponding to each concept node, and combine the semantic query vector and the biomechanical semantic graph to obtain the biomechanical semantic knowledge space.

5. The method according to claim 1, characterized in that, S2 includes: S21. Perform motion state analysis on the spatial motion modal data to generate a kinematic parameter sequence; wherein, the kinematic parameter sequence is used to characterize the spatiotemporal motion pattern of the athlete within the corresponding action cycle; the motion signal time series includes spatial motion modal data and dynamic modal data; S22. Perform inverse dynamics calculation on the kinematic parameter sequence and the dynamic mode data to generate a dynamic parameter sequence; S23. The kinematic parameter sequence and the dynamic parameter sequence are spliced ​​and tensorized in the spatiotemporal dimension to generate the kinematic-dynamic joint representation tensor.

6. The method according to claim 5, characterized in that, S22 includes: S221. Perform correlation feature mapping on the kinematic parameter sequence and the dynamic modal data to generate an initial predicted joint torque vector; The expression for the initial predicted joint moment vector is as follows: In the formula, for The initial predicted joint torque vector at time t. For trainable parameters Multilayer perceptron mapping network, for Joint angles in the kinematic parameter sequence at any given time. for Angular velocity in the kinematic parameter sequence at time t. for Angular acceleration in the kinematic parameter sequence at time t. for The measured force vectors of the foot or body surface in the dynamic modal data at any given time. For feature splicing operations; S222. Substitute the initial predicted joint moment vector into the preset physical motion equation for forward deduction to generate theoretical physical state values, and calculate the physical residual between the theoretical physical state values ​​and the actual measured values ​​in the dynamic modal data. The expression for the physical residual is as follows: In the formula, For physical residuals, for The theoretical physical state value at time t. for The measured force vectors of the foot or body surface in the dynamic modal data at any given time. The square of the L2 norm. The preset regularization weight coefficients, express The predicted joint torque vector to be solved at any given time. express The initial predicted joint torque vector at time t; S223. Based on the physical residual, iteratively correct the initial predicted joint moment vector until the physical residual meets the preset convergence condition, output the corrected joint moment vector sequence, and generate the dynamic parameter sequence. The update formula for the iterative correction is as follows: In the formula, For the first During the next iteration The joint torque vector at time t, The preset learning rate, This represents the gradient of the physical residual with respect to the joint torque in the current iteration. This represents the number of iterations.

7. A deep learning-based biomechanical evaluation system for athlete movements, characterized in that, The system includes: The time series generation module is used to acquire the multimodal technical motion signals of athletes, and to perform time synchronization and motion cycle segmentation on the multimodal technical motion signals to obtain a motion signal time series; The joint representation construction module is used to extract and fuse multi-dimensional biomechanical features from the motion signal time series to generate a kinematic-dynamic joint representation tensor; wherein, the kinematic-dynamic joint representation tensor is used to represent the correlation between the spatiotemporal motion morphology and dynamic force characteristics of the athlete in the corresponding action cycle; The spatiotemporal feature encoding module is used to obtain the individual morphological parameters of the athlete, perform spatiotemporal feature encoding on the individual morphological parameters and the kinematic-dynamic joint representation tensor, and generate a low-level spatiotemporal feature map; The concept activation module is used to perform cross-attention decoding and fusion on the underlying spatiotemporal feature map based on a pre-constructed biomechanical semantic knowledge space, using concept nodes in the biomechanical semantic knowledge space as query vectors, to generate a semantic state matrix corresponding to each concept node, and a sequence of concept activation vectors for each concept node on the time axis. The causal attribution module is used to perform anomaly attribution analysis on the semantic state matrix and the concept activation vector sequence based on the causal topology of the biomechanical semantic knowledge space, and generate biomechanical evaluation results. The report generation module is used to perform natural language processing on the biomechanical evaluation results to obtain an interpretable biomechanical evaluation report.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images