Method and system for recognizing risk of sports injury based on deep learning

Abstract

This invention provides a deep learning-based method and system for identifying sports injury risks, relating to the field of sports health monitoring technology. The method includes: collecting exercise data and extracting movement trajectory and load response features, concatenating them to obtain an initial feature matrix; converting the initial feature matrix into a graph structure for information aggregation and causal convolution processing to generate a stage state vector and comparing it with a standard template to obtain a deviation metric; identifying abnormal time periods and extracting features; matching action patterns using a capsule network to obtain matching confidence; constructing a state transition matrix for evolutionary inference to obtain risk accumulation weights; obtaining a global risk assessment value through Bayesian inference and generating a risk level identifier and injury risk report.

Description

Technical Field

[0001] This invention relates to the field of sports health monitoring technology, and in particular to a method and system for identifying sports injury risks based on deep learning. Background Technology

[0002] In the fields of sports science and health monitoring, using sensor data to predict and assess the risk of sports injuries has become an important research direction. Current technologies typically rely on the acquisition and analysis of kinematic data and physiological signals, aiming to provide early warnings of potential injuries by identifying abnormal movement patterns or excessive physiological load. Conventional approaches focus on detecting obvious anomalies or overload moments from the overall sequence.

[0003] However, sports injuries are often not caused by a single momentary abnormality, but rather by the repeated accumulation of incorrect movement patterns and the continuous action of the body's compensatory mechanisms. Existing methods mostly process features through static splicing or shallow fusion, which makes it difficult to characterize the complex dynamic coupling relationships between various joints and muscle groups in the body, as well as the evolution of these dynamic coupling relationships over time. This results in an inability to effectively capture subtle but cumulative non-obvious abnormal patterns, potentially missing the risk of progressive injury.

[0004] Meanwhile, existing technologies typically classify risks as isolated events based on current or recent data fragments, lacking modeling of the dynamic evolution of risk states. This results in risk assessment results lacking temporal consistency and causal interpretability, limiting the lead time and accuracy of early warnings. Summary of the Invention

[0005] This invention provides a method and system for identifying sports injury risks based on deep learning, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides a deep learning-based method for identifying sports injury risks, comprising:

[0007] Collect motion sequence data and physiological response data of the moving object, extract the motion trajectory features corresponding to the motion sequence data, extract the load response features corresponding to the physiological response data, and concatenate them with the motion trajectory features to obtain an initial feature matrix;

[0008] The initial feature matrix is ​​transformed into a graph structure to construct a kinematic constraint topology. Multi-hop information aggregation is performed on the kinematic constraint topology through a message passing mechanism to obtain node representations. Causal convolution is used to generate the stage state vectors corresponding to the node representations. The similarity between the stage state vectors and the standard action template is calculated to obtain the deviation metric.

[0009] Based on the deviation metric value, abnormal time periods are identified, and the node representations corresponding to the abnormal time periods are extracted to obtain the features of the region of interest. The features of the region of interest are matched in the posture space using the capsule network dynamic routing mechanism and the preset action pattern library to obtain the matching confidence. Based on the matching confidence, a Markov state transition matrix is ​​constructed and combined with the stage state vector to perform state evolution reasoning to obtain the risk accumulation weight.

[0010] A global risk assessment value is obtained through Bayesian inference based on the risk accumulation weight. A risk level identifier is generated by comparing the global risk assessment value with a safety threshold. A damage risk report is generated by combining the risk level identifier with the deviation metric value.

[0011] In one alternative implementation,

[0012] The process involves collecting motion sequence data and physiological response data of the moving object, extracting the motion trajectory features corresponding to the motion sequence data, extracting the load response features corresponding to the physiological response data, and concatenating these features with the motion trajectory features to obtain an initial feature matrix, including:

[0013] A continuous frame image sequence of a moving object is acquired and joint position information is extracted. The acceleration and angular velocity signals of the moving object are collected and transformed into a coordinate system to obtain motion velocity information, which is then combined with the joint position information to obtain a motion data sequence. The acquisition device acquires the heart rate signal and electromyography signal of the moving object and combines them to obtain the physiological response data.

[0014] Based on the joint position information in the motion sequence data, a skeleton topology is constructed and an attitude descriptor is calculated. The motion velocity information is then subjected to time window sliding sampling to determine the acceleration sequence, which is then time-aligned and spliced ​​with the attitude descriptor to obtain the motion trajectory features.

[0015] The heart rate signal in the physiological response data is subjected to spectral decomposition and the frequency band energy distribution is extracted to obtain the heart rate frequency domain features. The envelope corresponding to the electromyographic signal is extracted and the slope of the envelope is calculated to obtain the muscle activation mode. The load response features are obtained by combining the heart rate frequency domain features.

[0016] Based on the skeleton topology, the main motion links are determined, the coordination index between the attitude descriptors of adjacent joints in the main motion links is calculated, and the motion trajectory features are weighted and modulated to obtain the modulated trajectory features. The modulated trajectory features and the load response features are concatenated according to the time dimension to obtain the initial feature matrix.

[0017] In one alternative implementation,

[0018] The initial feature matrix is ​​transformed into a graph structure to construct a kinematically constrained topology. Multi-hop information aggregation is then performed on this kinematically constrained topology using a message passing mechanism to obtain node representations, including:

[0019] Based on the initial feature matrix, kinematic attribute types and temporal position indices are determined. Based on the kinematic attribute types, biomechanical coupling strength is calculated to obtain an attribute adjacency matrix. Based on the temporal position indices, motion continuity strength is calculated to obtain a temporal adjacency matrix, which is then fused with the attribute adjacency matrix using tensors to obtain a kinematic constraint topology.

[0020] The initial feature matrix is ​​mapped to the initial embedding vector of the node in the kinematic constraint topology. The set of neighbor nodes of each node in the kinematic constraint topology is extracted and the initial embedding vector of each neighbor node is linearly projected to obtain the neighbor message set. Based on the initial embedding vector of the current node and the neighbor message set, the aggregate message vector is calculated and concatenated with the initial embedding vector of the current node and nonlinearly transformed to obtain the single-hop embedding vector.

[0021] The shortest path length distribution is calculated based on the kinematic constraints and topology. The information propagation attenuation coefficient is determined based on the shortest path length distribution and weighted and modulated with the single-hop embedding vector to obtain an iterative embedding vector. The initial embedding vector is updated based on the iterative embedding vector and message passing is repeated to obtain a multi-hop embedding vector sequence. The cosine similarity of adjacent embedding vectors in the multi-hop embedding vector sequence is calculated and a convergence criterion is constructed. The termination embedding vector is determined based on the convergence criterion and a preset convergence threshold and is used as a node representation.

[0022] In one alternative implementation,

[0023] The node representation generates a stage state vector through causal convolution. The similarity between the stage state vector and the standard action template is calculated to obtain a deviation metric, including:

[0024] Temporal dimension information is extracted from the node representation and a temporal index mapping table is constructed. The node representation is temporally sorted based on the temporal index mapping table to obtain a temporally arranged node representation. The temporally arranged node representation is segmented according to a preset window length to obtain multiple temporal segments. A causal convolution kernel is applied to each temporal segment to perform convolution operation to obtain the segment convolution result.

[0025] The convolution results of the segments are nonlinearly transformed and the peak response position of each temporal segment is extracted. Based on the peak response position, the motion stage boundaries are divided to obtain multiple motion stages. The convolution results of the segments in each motion stage are weighted and averaged to obtain the stage feature representation. The feature dimension is then transformed to obtain the stage state vector.

[0026] The standard motion template matching the current motion type is retrieved from the preset standard motion template library and decomposed into multiple standard stage templates. The dot product between the stage state vector and the corresponding standard stage template is calculated and the amplitude is normalized to obtain the stage similarity. The stage similarity is then weighted and summed by stage weights and subjected to numerical inversion transformation to obtain the deviation metric value.

[0027] In one alternative implementation,

[0028] Based on the deviation metric value, abnormal time periods are identified. The node representations corresponding to the abnormal time periods are extracted to obtain features of the region of interest. These features of the region of interest are then matched in pose space using a capsule network dynamic routing mechanism and a preset action pattern library to obtain matching confidence, including:

[0029] The deviation metric is decomposed by wavelet multi-scale to obtain the deviation component. The local energy density of the deviation component is calculated and the location of energy abrupt change is identified to obtain the set of abnormal candidate points. A time series interval graph is constructed based on the set of abnormal candidate points and connected component analysis is performed on the time series interval graph to obtain the abnormal time period. The node data corresponding to the abnormal time period is extracted from the node representation to obtain the features of the region of interest.

[0030] The feature dimension of the region of interest is expanded and a query vector is constructed. Standard action pattern features are extracted from the action pattern library and a pattern feature matrix is ​​constructed. A candidate pattern feature set is filtered based on the vector distance between the query vector and the pattern feature matrix.

[0031] The features of the region of interest and the set of candidate pattern features are input into the capsule layer of the capsule network for dynamic routing iteration to obtain a query capsule representation and a set of candidate capsule representations. A query pose matrix is ​​extracted based on the query capsule representation, and a set of candidate pose matrices is extracted based on the set of candidate capsule representations. The matrix similarity between the candidate pose matrix and the query pose matrix is ​​calculated to obtain a pose similarity sequence. The collaborative matching score is calculated by combining the query capsule representation and the set of candidate capsule representations to obtain the matching confidence.

[0032] In one alternative implementation,

[0033] Based on the matching confidence, a Markov state transition matrix is ​​constructed, and state evolution reasoning is performed in conjunction with the stage state vector to obtain the risk accumulation weights, including:

[0034] The matching confidence is discretized into a numerical interval to obtain a discrete state label set. The temporal transition frequency between each state label in the discrete state label set is counted and a transition frequency matrix is ​​constructed. The transition frequency matrix is ​​row normalized to obtain the Markov state transition matrix. The diagonal elements are extracted from the Markov state transition matrix to construct a self-transition probability vector and risk-sensitive states are identified based on the self-transition probability vector.

[0035] The stage state vectors are arranged into a state sequence according to the time sequence and the initial stage state vector is extracted. The initial stage state vector is then subjected to matrix operation with the Markov state transition matrix to obtain a single-step evolution state vector. Based on the single-step evolution state vector and the Markov state transition matrix, iterative matrix operation is performed to obtain a multi-step evolution state vector sequence. The component values ​​of each evolution state vector in the corresponding dimension of the risk-sensitive state are extracted from the multi-step evolution state vector sequence to obtain the risk evolution trajectory.

[0036] The cumulative risk amount is obtained by accumulating the risk evolution trajectory along the time dimension. The cumulative risk amount is combined with the vector magnitude of the state vector of each stage in the state sequence to obtain the stage risk contribution value. The stage risk contribution value is then subjected to exponential normalization transformation to obtain the cumulative risk weight.

[0037] In one alternative implementation,

[0038] A global risk assessment value is obtained through Bayesian inference based on the risk accumulation weight. A risk level identifier is generated by comparing the global risk assessment value with a safety threshold. A damage risk report is generated by combining the risk level identifier with the deviation metric value, including:

[0039] Extract the weight values ​​corresponding to each stage from the risk accumulation weight and construct a prior probability distribution. Use the weight values ​​as observation evidence to perform a Bayesian update on the prior probability distribution to obtain a posterior probability distribution. Extract the expected value from the posterior probability distribution and combine it with the weight values ​​to calculate the global risk assessment value.

[0040] A multi-level security threshold sequence is initialized and compared with the global risk assessment value at each level. The risk level code is determined based on the threshold interval into which the global risk assessment value falls. The confidence interval boundary value is extracted from the posterior probability distribution and the degree of dispersion is calculated to obtain the uncertainty index. The risk level code and the uncertainty index are combined to obtain the risk level identifier.

[0041] The deviation amplitude sequence for each time point is extracted from the deviation metric. The deviation extraction window length is determined based on the risk level identifier, and key deviation segments are extracted from the deviation amplitude sequence. The causal correlation strength between the key deviation segments and the global risk assessment value is calculated to obtain the contribution weight. The key deviation segments are sorted based on the contribution weight to generate a deviation tracing chain. The risk level identifier and the deviation tracing chain are encapsulated and the statistical feature parameters of the posterior probability distribution are added to obtain the damage risk report.

[0042] A second aspect of the present invention provides a deep learning-based sports injury risk identification system, comprising:

[0043] The data acquisition unit is used to acquire motion sequence data and physiological response data of the moving object, extract the motion trajectory features corresponding to the motion sequence data, extract the load response features corresponding to the physiological response data, and concatenate them with the motion trajectory features to obtain an initial feature matrix.

[0044] The graph structure transformation unit is used to perform graph structure transformation on the initial feature matrix to construct a kinematic constraint topology, perform multi-hop information aggregation on the kinematic constraint topology through a message passing mechanism to obtain node representations, generate the stage state vectors corresponding to the node representations through causal convolution, and calculate the similarity between the stage state vectors and the standard action template to obtain a deviation metric.

[0045] An anomaly identification unit is used to identify abnormal time periods based on the deviation metric value, extract the node representation corresponding to the abnormal time period to obtain the features of the region of interest, perform pose space matching on the features of the region of interest through the capsule network dynamic routing mechanism and the preset action pattern library to obtain the matching confidence, construct a Markov state transition matrix based on the matching confidence and perform state evolution reasoning in combination with the stage state vector to obtain the risk accumulation weight.

[0046] The risk assessment unit is used to obtain a global risk assessment value through Bayesian inference based on the risk accumulation weight, generate a risk level identifier by comparing the global risk assessment value with a safety threshold, and generate a damage risk report by combining the risk level identifier with the deviation metric value.

[0047] A third aspect of the present invention provides an electronic device, comprising:

[0048] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0049] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0050] In this invention, initial features are converted into a graph structure and multi-hop aggregation is performed using a message passing mechanism, effectively modeling the kinematic constraints between human joints and enhancing the representational power of the features. Causal convolution is used to process node representations to generate stage state vectors, accurately depicting the temporal evolution of motion states. Quantitative measurement of motion deviations is achieved through similarity calculation with standard templates. By identifying abnormal time periods and extracting corresponding node features, the analysis focus can be concentrated on potential risk areas, improving the targeting of detection. The dynamic routing mechanism of capsule networks and pose matching with a pre-set action pattern library enhance the discrimination and generalization capabilities for complex action patterns, resulting in more reliable matching confidence. Bayesian inference is applied to integrate risk accumulation weights to obtain a global risk assessment value, fully considering the uncertainty of historical states and current evidence, making the final assessment result probabilistically interpretable and more robust. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the deep learning-based sports injury risk identification method according to an embodiment of the present invention.

[0052] Figure 2 This is a flowchart of the abnormal behavior capsule network matching process for the deep learning-based sports injury risk identification method according to an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Figure 1 This is a flowchart illustrating the deep learning-based sports injury risk identification method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] Collect motion sequence data and physiological response data of the moving object, extract the motion trajectory features corresponding to the motion sequence data, extract the load response features corresponding to the physiological response data, and concatenate them with the motion trajectory features to obtain an initial feature matrix;

[0057] The initial feature matrix is ​​transformed into a graph structure to construct a kinematic constraint topology. Multi-hop information aggregation is performed on the kinematic constraint topology through a message passing mechanism to obtain node representations. Causal convolution is used to generate the stage state vectors corresponding to the node representations. The similarity between the stage state vectors and the standard action template is calculated to obtain the deviation metric.

[0058] Based on the deviation metric value, abnormal time periods are identified, and the node representations corresponding to the abnormal time periods are extracted to obtain the features of the region of interest. The features of the region of interest are matched in the posture space using the capsule network dynamic routing mechanism and the preset action pattern library to obtain the matching confidence. Based on the matching confidence, a Markov state transition matrix is ​​constructed and combined with the stage state vector to perform state evolution reasoning to obtain the risk accumulation weight.

[0059] A global risk assessment value is obtained through Bayesian inference based on the risk accumulation weight. A risk level identifier is generated by comparing the global risk assessment value with a safety threshold. A damage risk report is generated by combining the risk level identifier with the deviation metric value.

[0060] In one alternative implementation,

[0061] The process involves collecting motion sequence data and physiological response data of the moving object, extracting the motion trajectory features corresponding to the motion sequence data, extracting the load response features corresponding to the physiological response data, and concatenating these features with the motion trajectory features to obtain an initial feature matrix, including:

[0062] A continuous frame image sequence of a moving object is acquired and joint position information is extracted. The acceleration and angular velocity signals of the moving object are collected and transformed into a coordinate system to obtain motion velocity information, which is then combined with the joint position information to obtain a motion data sequence. The acquisition device acquires the heart rate signal and electromyography signal of the moving object and combines them to obtain the physiological response data.

[0063] Based on the joint position information in the motion sequence data, a skeleton topology is constructed and an attitude descriptor is calculated. The motion velocity information is then subjected to time window sliding sampling to determine the acceleration sequence, which is then time-aligned and spliced ​​with the attitude descriptor to obtain the motion trajectory features.

[0064] The heart rate signal in the physiological response data is subjected to spectral decomposition and the frequency band energy distribution is extracted to obtain the heart rate frequency domain features. The envelope corresponding to the electromyographic signal is extracted and the slope of the envelope is calculated to obtain the muscle activation mode. The load response features are obtained by combining the heart rate frequency domain features.

[0065] Based on the skeleton topology, the main motion links are determined, the coordination index between the attitude descriptors of adjacent joints in the main motion links is calculated, and the motion trajectory features are weighted and modulated to obtain the modulated trajectory features. The modulated trajectory features and the load response features are concatenated according to the time dimension to obtain the initial feature matrix.

[0066] The acquisition equipment typically includes a depth camera, an inertial measurement unit (IMU), and a physiological signal acquisition module. The depth camera continuously captures images of the moving object at a frame rate of 30 frames per second (fps) or 60 fps, forming a continuous sequence of images. By applying a deep learning-based pose estimation algorithm to this sequence, key joints of the human body, such as the shoulder, elbow, wrist, hip, knee, and ankle joints, can be detected and located. Each joint has coordinates in three-dimensional space, which constitute the joint position information. Simultaneously, an IMU worn on the wrist, waist, or ankle of the moving object acquires three-axis acceleration and angular velocity signals in real time, with a sampling frequency typically set between 100 Hz and 200 Hz.

[0067] Because the initial coordinate system of the inertial measurement unit (IMU) has a rotational deviation from the global coordinate system, coordinate system transformation is required to convert the acceleration and angular velocity signals from the sensor coordinate system to the global coordinate system. Coordinate system transformation is achieved by constructing a rotation matrix, which is calculated based on the quaternion representation of the initial attitude. The converted acceleration signal is integrated to obtain velocity information, while the angular velocity signal reflects the joint's rotational rate. The velocity and angular velocity signals are synchronized according to timestamps to form motion velocity information. The motion velocity information and joint position information are aligned on the time axis and then combined to obtain a motion data sequence. The data at each moment in the motion data sequence includes the joint's spatial position and motion velocity, providing basic data for subsequent motion trajectory analysis.

[0068] Physiological response data is acquired using wearable physiological signal acquisition devices. Heart rate signals are obtained via a heart rate strap worn on the chest or wrist using a photoplethysmography (PPG) sensor, with a sampling frequency set at 250 Hz. Electromyography (EMG) signals are acquired via surface EMG sensors attached to the target muscle groups; for lower limb movements, sensors are typically placed in the quadriceps, hamstrings, and gastrocnemius muscles, with a sampling frequency set at 1000 Hz. The acquisition device synchronizes the heart rate and EMG signals in time and transmits the data to the processing unit via a wireless transmission module. The heart rate and EMG signals are combined sequentially over time to form physiological response data.

[0069] Based on the joint position information in motion sequence data, the connections between joints are defined according to human anatomy. For example, the shoulder and elbow joints connect to form the upper arm skeleton, and the elbow and wrist joints connect to form the forearm skeleton. A skeletal topology is constructed using these connections, which can be represented by an undirected graph where nodes represent joints and edges represent bone connections. Based on this skeletal topology, a pose descriptor for each joint is calculated. The pose descriptor includes the joint's three-dimensional coordinates, the bone vector formed with adjacent joints, the length of the bone vector, and the angle between the bone vectors. The length of the bone vector reflects the degree of limb extension, and the angle between the bone vectors reflects the flexion or extension state of the joint.

[0070] Motion velocity information is sampled using a sliding time window to extract an acceleration sequence reflecting changes in motion intensity. The time window length is set to 0.5 to 2 seconds, and the sliding step size is set to 50% of the window length. Within each time window, statistical characteristics of the acceleration signal are calculated, including mean, standard deviation, maximum, and minimum values. These statistical characteristics are combined to form the acceleration sequence. The acceleration sequence reflects the changes in motion intensity and motion pattern of the moving object over different time periods. The acceleration sequence is temporally aligned with the attitude descriptor to ensure that each sampling point in the acceleration sequence corresponds to a timestamp in the attitude descriptor. Temporal alignment is achieved using linear interpolation or nearest neighbor interpolation methods. The aligned acceleration sequence and attitude descriptor are then concatenated according to feature dimensions to form motion trajectory features. The motion trajectory features are a multi-dimensional time series containing information such as joint spatial position, posture angle, and motion acceleration.

[0071] The heart rate signal in the physiological response data was preprocessed, including the removal of baseline drift and high-frequency noise. Baseline drift was removed using a high-pass filter with a cutoff frequency set to 0.5 Hz. High-frequency noise was removed using a low-pass filter with a cutoff frequency set to 40 Hz. The preprocessed heart rate signal was then subjected to spectral decomposition using a Fast Fourier Transform (FFT) to obtain its frequency domain representation. After spectral decomposition, the frequency domain was divided into a very low frequency band (0.003 Hz to 0.04 Hz), a low frequency band (0.04 Hz to 0.15 Hz), and a high frequency band (0.15 Hz to 0.4 Hz). The energy within each frequency band was calculated by integrating the square of the spectral amplitude within that band. The combination of the energy distributions of the three frequency bands forms the heart rate frequency domain characteristics, which reflect the regulatory state of the autonomic nervous system in response to exercise load.

[0072] Electromyography (EMG) signals are filtered to remove motion artifacts and power frequency interference using a bandpass filter with a passband range of 20Hz to 500Hz. The filtered EMG signals are then subjected to full-wave rectification and low-pass filtering to obtain an envelope, with the low-pass filter's cutoff frequency set to 6Hz. The envelope reflects the intensity changes of muscle activation. The slope is calculated on the envelope using differential calculations between adjacent sampling points. A positive slope indicates increased muscle activation intensity, while a negative slope indicates decreased muscle activation intensity. The envelope slope is segmented and statistically analyzed within a time window, and the mean, variance, and peak value of the slope are extracted. These statistics are combined to form a muscle activation pattern. The muscle activation pattern reflects the coordination patterns and fatigue states of different muscle groups during exercise. The muscle activation pattern is aligned with heart rate frequency domain features along the time dimension, and load response features are obtained through weighted summation or feature concatenation. The weights in the weighted summation are determined based on the physiological characteristics of different exercise types; for example, heart rate features have a higher weight in aerobic exercise, while EMG features have a higher weight in strength training.

[0073] Based on skeletal topology, the dominant joint links in specific movements are identified. For example, in a squat, the lower limb link consisting of the hip, knee, and ankle joints is the primary motion link. In a throwing motion, the upper limb link consisting of the shoulder, elbow, and wrist joints is the primary motion link. The primary motion links are determined by analyzing the amplitude and velocity changes of each joint in the motion data sequence; links with significant changes in amplitude and velocity are identified as primary motion links. Within the primary motion links, a coordination index is calculated between the posture descriptors of adjacent joints. The coordination index is obtained by calculating the consistency of motion direction and the velocity ratio of adjacent joints. Motion direction consistency is calculated using the cosine similarity of the velocity vectors of adjacent joints; a cosine similarity close to 1 indicates a high degree of consistency in motion direction. The velocity ratio is calculated using the ratio of the velocity amplitudes of adjacent joints; a ratio close to a certain ideal value indicates good coordination. The coordination index is normalized to the range of 0 to 1 and used as a weighting coefficient to weighted modulate the motion trajectory features. Weighted modulation is achieved by multiplying the feature vector of the corresponding joint in the motion trajectory features with the coordination index to obtain the modulated trajectory features. Modulation trajectory features reinforce well-coordinated motion patterns and suppress anomalous patterns with poor coordination.

[0074] The modulation trajectory features and load response features are concatenated along the time dimension. The temporal resolution of the modulation trajectory features and the load response features may differ, requiring resampling or interpolation to align their time axes. After alignment, at each time point, the modulation trajectory feature vector and the load response feature vector are concatenated along the feature dimension to form a comprehensive feature vector containing both kinematic and physiological information. The comprehensive feature vectors from all time points are arranged in chronological order to form the initial feature matrix. The number of rows in the initial feature matrix corresponds to the number of time sampling points, and the number of columns corresponds to the dimension of the comprehensive feature. The initial feature matrix provides a complete data foundation for subsequent graph structure transformation and risk identification.

[0075] In one alternative implementation,

[0076] The initial feature matrix is ​​transformed into a graph structure to construct a kinematically constrained topology. Multi-hop information aggregation is then performed on this kinematically constrained topology using a message passing mechanism to obtain node representations, including:

[0077] Based on the initial feature matrix, kinematic attribute types and temporal position indices are determined. Based on the kinematic attribute types, biomechanical coupling strength is calculated to obtain an attribute adjacency matrix. Based on the temporal position indices, motion continuity strength is calculated to obtain a temporal adjacency matrix, which is then fused with the attribute adjacency matrix using tensors to obtain a kinematic constraint topology.

[0078] The initial feature matrix is ​​mapped to the initial embedding vector of the node in the kinematic constraint topology. The set of neighbor nodes of each node in the kinematic constraint topology is extracted and the initial embedding vector of each neighbor node is linearly projected to obtain the neighbor message set. Based on the initial embedding vector of the current node and the neighbor message set, the aggregate message vector is calculated and concatenated with the initial embedding vector of the current node and nonlinearly transformed to obtain the single-hop embedding vector.

[0079] The shortest path length distribution is calculated based on the kinematic constraints and topology. The information propagation attenuation coefficient is determined based on the shortest path length distribution and weighted and modulated with the single-hop embedding vector to obtain an iterative embedding vector. The initial embedding vector is updated based on the iterative embedding vector and message passing is repeated to obtain a multi-hop embedding vector sequence. The cosine similarity of adjacent embedding vectors in the multi-hop embedding vector sequence is calculated and a convergence criterion is constructed. The termination embedding vector is determined based on the convergence criterion and a preset convergence threshold and is used as a node representation.

[0080] The initial feature matrix is ​​reorganized into a topological structure that reflects the intrinsic kinematic constraints. The initial feature matrix contains information in multiple dimensions, with each feature vector corresponding to a specific kinematic measurement point or physiological monitoring point. These feature points are used as nodes in a graph structure, and the connections between these nodes need to be determined to reflect the true kinematic constraints.

[0081] Each feature vector in the initial feature matrix is ​​categorized and labeled according to its source. For example, joint angle features, acceleration features, and electromyographic signal features correspond to different kinematic attribute types. Each attribute type is represented by a unique identifier and recorded in the attribute label vector. The temporal position index is sorted according to the acquisition timestamp, and each feature point is assigned a position number in the time series to form a temporal index vector.

[0082] The biomechanical coupling strength is calculated based on the physical correlation between different kinematic attributes. For example, there is a strong coupling relationship between the knee joint angle and the quadriceps electromyography signal, while the coupling strength with the distal ankle joint is weaker. For attribute types... and The biomechanical coupling strength of the two nodes Obtained by looking up a predefined biomechanical coupling matrix, which is constructed based on knowledge of kinesiology. Attribute adjacency matrix. elements Equal to the biomechanical coupling strength of the corresponding node If there is no direct mechanical relationship between the attribute types of two nodes, the element is zero.

[0083] The motion continuity intensity is calculated based on the continuity between adjacent or close feature points at temporal locations, where the temporal location indices are respectively... and The two nodes, their motion continuity strength It is obtained by calculating the reciprocal of the time interval and then normalizing it. When the time interval... When the interval is less than a preset threshold, the continuity strength is high; when the interval exceeds a preset window length, the continuity strength decays to zero. Temporal Adjacency Matrix elements Equal to the motion continuity strength of the corresponding node .

[0084] When performing tensor fusion of the attribute adjacency matrix and the temporal adjacency matrix, a weighted summation method is used. Let the attribute weight coefficients be... The time-series weighting coefficient is ,satisfy The fused kinematic constraint topological adjacency matrix Calculated as This adjacency matrix preserves both biomechanical constraints and reflects the temporal continuity of motion. By adjusting... and The value of controls the relative importance of the two constraints. In practical applications, It is usually set to 0.6. Set to 0.4 to highlight the dominant role of biomechanical constraints.

[0085] When mapping the initial feature matrix to the initial embedding vectors of the nodes, each row or column of the initial feature matrix (depending on the feature arrangement) is projected through a linear transformation layer. The dimension of the weight matrix of the linear transformation layer is determined according to the dimension of the target embedding space, typically projecting the original feature dimensions to a 128-dimensional or 256-dimensional embedding space. The projected vectors are then batch normalized to obtain the initial embedding vectors. The superscript indicates the iteration number, and the subscript indicates the node index.

[0086] Extract the set of neighboring nodes for each node, and then perform a topological adjacency matrix based on kinematic constraints. Non-zero elements in the table determine neighbor relationships. For nodes... Its set of neighboring nodes Includes all that satisfy nodes Iterate through the set, for each neighbor node... initial embedding vector Perform a linear projection. The linear projection is achieved through a learnable weight matrix. To achieve this, calculate the neighbor messages. . Node All neighbor messages are collected to form a neighbor message set. .

[0087] The aggregated message vector is calculated based on the initial embedding vector of the current node and the set of neighbor messages, using an attention-weighted mechanism. For the current node... Calculate its initial embedding vector Messages to each neighbor The correlation score between the two values ​​is used to calculate the attention weights by performing a dot product operation and then normalizing the correlation score using softmax. Aggregate message vectors The calculation is a weighted sum of all neighbor messages, i.e. This aggregated message vector combines information passed from neighboring nodes, with weights reflecting the importance of different neighbors.

[0088] The aggregated message vector is concatenated with the initial embedding vector of the current node by directly linking the two vectors along the feature dimension, resulting in a concatenated vector. The dimension of the concatenated vector is the sum of the dimensions of the two input vectors. A nonlinear transformation is applied to the concatenated vector using a fully connected layer and an activation function. The fully connected layer maps the concatenated vector dimension back to the embedding space dimension, and the activation function is either ReLU or LeakyReLU. The resulting nonlinear transformation yields a single-hop embedding vector. This indicates that after one hop of message passing, the node... The update indicates that...

[0089] Calculate the shortest path length distribution. For any two nodes in the kinematically constrained topology, calculate the shortest path hop count between them using a breadth-first search algorithm. Count the shortest path lengths for all node pairs and construct a path length distribution histogram. The path length distribution histogram reflects the overall connectivity of the graph and the reachability characteristics between nodes.

[0090] The information propagation attenuation coefficient is determined based on the shortest path length distribution, considering the attenuation law of information propagation on the graph. The longer the path, the greater the loss during information transmission. For a path length of... The information propagation attenuation coefficient of the node pair. It can be designed in an exponential decay form, for example... ,in This is the attenuation rate hyperparameter. Piecewise linear attenuation or adaptive attenuation based on path length distribution can also be used. When weighting the attenuation coefficient with the single-hop embedding vector, the node... single-hop embedding vector Based on its positional characteristics in the graph and the path length to the information source node, an appropriate attenuation coefficient is applied. The modulated vector is obtained. .

[0091] The initial embedding vector is updated based on the modulated vector, and the modulated single-hop embedding vector is used as the new node representation, replacing the original initial embedding vector. The node's embedding vector has already incorporated information from its one-hop neighbors. The message passing process is repeated, and on the [number]th [time / phase]... In the iteration, using the The embedding vector obtained in the next iteration As the representation of the current node, extract the embedding vectors of neighboring nodes, calculate neighbor messages, aggregate and update them to obtain the th node. Embedding vector of the next iteration As the number of iterations increases, the node's embedding vector gradually incorporates information from more distant neighbors, forming a multi-hop embedding vector sequence. .

[0092] Calculate the cosine similarity of adjacent embedding vectors in a multi-hop embedding vector sequence, for the . Second and third The cosine similarity is calculated from the embedding vectors in the next iteration. This similarity reflects the magnitude of change in node representation between two consecutive iterations. When the similarity is close to 1, it indicates that the node representation tends to stabilize, and the information gain from further iterations is limited.

[0093] A convergence criterion can be constructed, which can be the average cosine similarity of adjacent embedding vectors as the global convergence criterion, or the minimum similarity of all nodes as a conservative criterion. The convergence criterion is compared with a preset convergence threshold. When the criterion exceeds the threshold (e.g., 0.99), the message passing process is considered to have converged. At this point, iteration stops, and the current embedding vector is... As the terminating embedding vector, it integrates the node's own features and the structural information of its multi-hop neighbors, fully reflecting the global association relationships in the kinematically constrained topology, and serves as the node representation output for subsequent processing.

[0094] In one alternative implementation,

[0095] The node representation generates a stage state vector through causal convolution. The similarity between the stage state vector and the standard action template is calculated to obtain a deviation metric, including:

[0096] Temporal dimension information is extracted from the node representation and a temporal index mapping table is constructed. The node representation is temporally sorted based on the temporal index mapping table to obtain a temporally arranged node representation. The temporally arranged node representation is segmented according to a preset window length to obtain multiple temporal segments. A causal convolution kernel is applied to each temporal segment to perform convolution operation to obtain the segment convolution result.

[0097] The convolution results of the segments are nonlinearly transformed and the peak response position of each temporal segment is extracted. Based on the peak response position, the motion stage boundaries are divided to obtain multiple motion stages. The convolution results of the segments in each motion stage are weighted and averaged to obtain the stage feature representation. The feature dimension is then transformed to obtain the stage state vector.

[0098] The standard motion template matching the current motion type is retrieved from the preset standard motion template library and decomposed into multiple standard stage templates. The dot product between the stage state vector and the corresponding standard stage template is calculated and the amplitude is normalized to obtain the stage similarity. The stage similarity is then weighted and summed by stage weights and subjected to numerical inversion transformation to obtain the deviation metric value.

[0099] After obtaining the node representations, temporal modeling is performed to capture dynamic change patterns during motion. A node representation is typically a three-dimensional tensor with dimensions of [missing information]. ,in Indicates batch size, Indicates the number of nodes. This represents the node feature dimension. For temporal analysis, the temporal dimension information must first be extracted from this tensor. Specifically, during motion sequence data acquisition, each timestamp generates a set of node representations; therefore, a temporal index mapping table needs to be constructed to record the timestamp information corresponding to each node representation. This mapping table is stored in dictionary form, with the node index as the key and the corresponding timestamp sequence as the value. Through this mapping table, the node representations originally organized in spatial dimension can be rearranged according to chronological order, forming a temporally arranged node representation, whose dimension is transformed into… ,in Indicates the timing length.

[0100] To divide the temporally ordered node representation into stages, a sliding window mechanism is needed. (Preset window length) It is usually set according to the characteristics of the specific type of sport, such as running. It can be set to 30 frames to capture a complete gait cycle; for squat movements, It can be set to 20 frames to cover the complete squatting and standing process. Based on this window length, the temporal arrangement of node representations is segmented along the temporal dimension using a non-overlapping segmentation method, resulting in... A time series segment, in which The dimension of each time segment is... .

[0101] A causal convolution kernel is applied to each time segment for convolution. The core characteristic of causal convolution is that its receptive field only covers the current time step and previous time steps, without introducing future information, thus ensuring the causality of the prediction. The parameter matrix of the causal convolution kernel has dimensions of... ,in Indicates the kernel size. Indicates the input feature dimension. This indicates the output feature dimension. In practical applications, It is usually set to 3 or 5. With node feature dimensions same, Set to 128 or 256. During convolution operations, for the first time segment... At each time step, its convolutional output is from the previous... The data is obtained by weighted summation of features from each time step, ensuring that data from future time steps is not accessed. After convolution, each time segment produces a segment convolution result with dimensions of [dimension number missing]. .

[0102] A nonlinear transformation is applied to the convolutional results of the segments, typically using the ReLU or GELU activation functions, to introduce nonlinear characteristics and enhance the model's expressive power. After the nonlinear transformation, the peak response locations of each time segment need to be extracted. Specifically, the norm of the segment convolutional results is calculated in both the node and feature dimensions to obtain the response intensity value at each time step, with dimensions of [missing information]. Find local maxima in the sequence, i.e., points that satisfy the condition. and Location ,in Indicates the first The response intensity values ​​at each time step. These local maxima correspond to key turning points in the motion, such as the moment of transition from squatting to standing, or the moment of transition from forward arm swing to backward arm swing.

[0103] Motion phase boundaries are defined based on peak response locations. The time interval between two adjacent peak response locations is defined as a motion phase. If a motion phase is detected within a certain time segment... If there are peak response locations, then the segment is divided into The motion sequence consists of several motion stages. For the entire motion sequence, the motion stages of all time segments are merged to obtain a complete motion stage sequence. In practice, to avoid over-segmentation, a minimum stage length threshold is set. The value is typically between 5 and 10 frames. If the length of a certain motion phase is less than... If so, then merge it with the adjacent stage.

[0104] A weighted average is applied to the convolutional results of segments within each motion phase to extract representative features for that phase. The weighting coefficients are dynamically determined based on the time step's position within the phase, with time steps in the middle of the phase receiving higher weights and those at the phase boundaries receiving lower weights. The weighted average is calculated as follows: ,in Indicates the length of this stage. Indicates the first [number]th ... The convolution results at each time step This represents the corresponding weight coefficient, which satisfies Weighting coefficients A Gaussian distribution is typically used to maximize the weight of the central location of each stage. The weighted average yields the stage feature representation, with dimensions of [missing value]. .

[0105] To transform the stage feature representation into a stage state vector, a feature dimension transformation is required. This is achieved by using a fully connected layer to transform the dimension from... Transform into ,in The dimension representing the state vector is typically set to 64 or 128. During the dimension transformation process, global average pooling is first performed along the node dimension, resulting in a dimension of... The tensor is then mapped through a fully connected layer to... Dimension. After this transformation, each motion stage corresponds to a stage state vector, which compactly expresses the motion characteristics of that stage.

[0106] The system retrieves a standard movement template matching the current movement type from a pre-built standard movement template library. This library is a pre-built database storing reference templates for various standardized movements. Each template contains a sequence of phase state vectors for the movement in its standard execution. The retrieval process uses an index query based on the movement type label; for example, if the current movement type is squat, the system retrieves the standard squat movement template. The retrieved standard movement template is then broken down into multiple standard phase templates, each corresponding to a phase in the standard movement execution process, with a dimension of [dimension not specified]. .

[0107] Calculate the dot product between the stage state vector and the corresponding standard stage template. For the ... There are several motion stages, and the stage state vector is: The corresponding standard stage template is The vector inner product is calculated as follows: To eliminate the influence of vector magnitude, magnitude normalization is required, calculated as follows: ,in This represents the Euclidean norm of a vector. The normalized value. The range of values ​​is The value is 1 when the two vectors are in the same direction and negative 1 when they are in opposite directions. Mapped to Intervals are used to obtain stage similarity. .

[0108] The similarity of each stage is summed using stage-weighted methods. Different stages of a movement have varying degrees of impact on the overall motion quality. For example, in a jump, the takeoff and landing stages are more important than the air stage; therefore, different weights need to be assigned to each stage. Stage Weights It can be set based on expert experience or learned automatically through statistical analysis of historical damage data. The weighted summation is calculated as follows: ,in Represents the total number of motion stages, satisfying To obtain the deviation metric, a numerical inversion transformation is needed to perform a numerical similarity transformation on the weighted similarity. The calculation method is as follows: The range of values ​​for this deviation metric is: A larger value indicates a greater deviation between the actual action and the standard action, and a higher potential risk of injury. In practical applications, the deviation measurement value can also be scaled, for example, multiplied by 100 to convert it into a percentage for easier understanding and display.

[0109] Figure 2 This is a flowchart of the abnormal behavior capsule network matching process for the deep learning-based sports injury risk identification method according to an embodiment of the present invention.

[0110] In one alternative implementation,

[0111] Based on the deviation metric value, abnormal time periods are identified. The node representations corresponding to the abnormal time periods are extracted to obtain features of the region of interest. These features of the region of interest are then matched in pose space using a capsule network dynamic routing mechanism and a preset action pattern library to obtain matching confidence, including:

[0112] The deviation metric is decomposed by wavelet multi-scale to obtain the deviation component. The local energy density of the deviation component is calculated and the location of energy abrupt change is identified to obtain the set of abnormal candidate points. A time series interval graph is constructed based on the set of abnormal candidate points and connected component analysis is performed on the time series interval graph to obtain the abnormal time period. The node data corresponding to the abnormal time period is extracted from the node representation to obtain the features of the region of interest.

[0113] The feature dimension of the region of interest is expanded and a query vector is constructed. Standard action pattern features are extracted from the action pattern library and a pattern feature matrix is ​​constructed. A candidate pattern feature set is filtered based on the vector distance between the query vector and the pattern feature matrix.

[0114] The features of the region of interest and the set of candidate pattern features are input into the capsule layer of the capsule network for dynamic routing iteration to obtain a query capsule representation and a set of candidate capsule representations. A query pose matrix is ​​extracted based on the query capsule representation, and a set of candidate pose matrices is extracted based on the set of candidate capsule representations. The matrix similarity between the candidate pose matrix and the query pose matrix is ​​calculated to obtain a pose similarity sequence. The collaborative matching score is calculated by combining the query capsule representation and the set of candidate capsule representations to obtain the matching confidence.

[0115] After obtaining the deviation metric, in-depth analysis of the deviation information is required to locate potential risk periods. Wavelet multi-scale decomposition is used to process the deviation metric, selecting the Daubechies wavelet as the mother wavelet function and setting the decomposition level to four to obtain deviation components at different scales. Wavelet decomposition can decompose the deviation signal into different frequency components; high-frequency components reflect local abrupt changes, while low-frequency components reflect the overall trend. The local energy density is calculated for the deviation component at each scale. Within a sliding window with a time window length of 50 sampling points, the sum of squares of the signal within the window is calculated as the energy index. By setting an energy change rate threshold where the energy ratio between adjacent windows exceeds 2.5, locations where significant energy jumps are identified, and these locations are marked as anomaly candidate points. The anomaly candidate points identified at all scales are summarized, and duplicate points with a time interval of less than 10 sampling points are removed to form an anomaly candidate point set.

[0116] A time-series interval graph is constructed based on a set of anomaly candidate points. Each anomaly candidate point is treated as a node in the graph. When the time interval between two candidate points is less than a preset connection distance threshold of 100 sampling points, an edge connection is established between the corresponding nodes. Connectivity analysis is performed on the constructed time-series interval graph. A depth-first search algorithm is used to traverse the graph structure and identify all connected subgraphs. Each connected subgraph represents a potential anomalous time period. The timestamps corresponding to all nodes in the connected subgraph are extracted. The minimum timestamp is taken as the starting point of the anomalous time period, and the maximum timestamp as the ending point. A buffer is formed by extending 20 sampling points before the starting point and 20 sampling points after the ending point to form a complete definition of the anomalous time period. When multiple anomalous time periods are identified, the interval between adjacent time periods is checked. If the interval is less than 30 sampling points, they are merged.

[0117] Node data corresponding to abnormal time periods are extracted from the node representation, which is a three-dimensional tensor with dimensions of time steps × number of nodes × feature dimension. For each abnormal time period, slices are performed along the time dimension based on the start and end time indices to extract the feature vectors of all nodes within that time period. Since the lengths of different abnormal time periods may vary, a time interpolation method is used to uniformly adjust the node data of all abnormal time periods to a fixed length of 80 time steps. Cubic spline interpolation is chosen as the interpolation method to maintain the smoothness and continuity of the data. The interpolated node data is flattened along both the node and feature dimensions to obtain a feature matrix of the region of interest with dimensions of 80 × total number of features.

[0118] The feature dimensions of the region of interest are expanded by using a fully connected layer to map the original feature dimensions from the total number of features to 512 dimensions. The GELU activation function is used to enhance non-linear expressiveness. Layer normalization is applied to the expanded features to eliminate the dimensional differences between different feature dimensions. The normalized features are then compressed in time using an attention pooling mechanism. The attention weights are obtained by performing a dot product between the learnable parameter matrix and the features, followed by softmax normalization. The weighted sum is then used to obtain a query vector with a fixed dimension of 512.

[0119] Standard motion pattern features are extracted from a pre-defined motion pattern library. This library stores standard motion patterns for various motion types, each containing key pose sequences and corresponding feature descriptions. Each standard motion pattern in the library has been preprocessed, with a unified feature dimension of 512, consistent with the query vector dimension. All standard motion pattern features are stacked row-wise to construct a pattern feature matrix, with a matrix dimension of the total number of patterns × 512. The cosine distance between the query vector and each row of vectors in the pattern feature matrix is ​​calculated, defined as 1 minus the normalized inner product of the two vectors. All patterns are sorted in ascending order of cosine distance, and the top 15 patterns with the smallest distances are selected as candidate patterns. Their corresponding feature vectors are extracted to form a candidate pattern feature set.

[0120] The set of features from the region of interest and the set of candidate patterns are input into a capsule network for pose space matching. The input layer of the capsule network first extracts local patterns from the region of interest features and each candidate pattern feature through convolutional layers. The convolutional kernel size is 3×3, the stride is 1, and the number of output channels is 64. The output of the convolutional layers is then reshaped into primary capsule layers, each with a dimension of 8. The length of the capsule vector is normalized to between 0 and 1 using a squeezing function. The capsule layers employ a dynamic routing iteration mechanism for information transmission, with three iterations. In each iteration, the coupling coefficient between the primary and advanced capsules is calculated. The coupling coefficient is initialized with a uniform distribution and updated during iteration based on the consistency between the predicted vectors of the primary capsules and the outputs of the advanced capsules. The predicted vectors are obtained by linearly transforming the primary capsule vectors using a learnable weight matrix. The input to the advanced capsule is the weighted sum of the predicted vectors of all primary capsules, with the weights being the coupling coefficients. The output is non-linearly activated using a squeezing function. After dynamic routing iteration, the query capsule representation corresponding to the region of interest features (16×16 dimension) and the candidate capsule representation corresponding to each candidate pattern feature are obtained, forming a set of candidate capsule representations.

[0121] The query pose matrix is ​​extracted based on the query capsule representation by reorganizing the elements of the capsule vectors into a matrix form. The 16×16 capsule representation is decomposed into four sub-capsule groups, each containing 4×16 capsule vectors. Each group of capsule vectors is reshaped to obtain a 4×4 pose sub-matrix. These four pose sub-matrices are then concatenated spatially to form an 8×8 query pose matrix. The same method is applied to the candidate capsule representation set, extracting the corresponding candidate pose matrix for each candidate pattern, forming a candidate pose matrix set. The matrix similarity between the query pose matrix and each candidate pose matrix is ​​calculated, using the Frobenius norm to measure the matrix difference. Similarity is defined as the negative exponential function of the Frobenius norm of the difference between two matrices, normalized to the interval 0 to 1. The similarity values ​​calculated for all candidate patterns are arranged sequentially to form a pose similarity sequence.

[0122] The co-matching score is calculated by combining the query capsule representation and the candidate capsule representation set. The query capsule representation is flattened into a 256-dimensional vector, and each candidate capsule representation is also flattened into a 256-dimensional vector. The element-wise product of the query vector and each candidate vector is calculated to obtain the interaction feature vector. The interaction feature vector is then non-linearly mapped through a two-layer fully connected network. The first layer has an output dimension of 128 and uses the ReLU activation function, while the second layer has an output dimension of 1 and uses the Sigmoid activation function. The output value represents the degree of co-matching between the candidate pattern and the query feature. A co-matching score sequence is calculated for all candidate patterns. This sequence is then weighted and fused with the pose similarity sequence, with weight coefficients set to 0.6 and 0.4 respectively. The weighted sum is then used to obtain the comprehensive matching score. The score corresponding to the candidate pattern with the highest comprehensive matching score is selected as the final matching confidence. This confidence value reflects the degree of matching between the motion posture and the standard action pattern within the current abnormal time period. A higher confidence indicates a smaller deviation and lower risk, while a lower confidence indicates a larger posture deviation and a higher risk of motion injury.

[0123] In one alternative implementation,

[0124] Based on the matching confidence, a Markov state transition matrix is ​​constructed, and state evolution reasoning is performed in conjunction with the stage state vector to obtain the risk accumulation weights, including:

[0125] The matching confidence is discretized into a numerical interval to obtain a discrete state label set. The temporal transition frequency between each state label in the discrete state label set is counted and a transition frequency matrix is ​​constructed. The transition frequency matrix is ​​row normalized to obtain the Markov state transition matrix. The diagonal elements are extracted from the Markov state transition matrix to construct a self-transition probability vector and risk-sensitive states are identified based on the self-transition probability vector.

[0126] The stage state vectors are arranged into a state sequence according to the time sequence and the initial stage state vector is extracted. The initial stage state vector is then subjected to matrix operation with the Markov state transition matrix to obtain a single-step evolution state vector. Based on the single-step evolution state vector and the Markov state transition matrix, iterative matrix operation is performed to obtain a multi-step evolution state vector sequence. The component values ​​of each evolution state vector in the corresponding dimension of the risk-sensitive state are extracted from the multi-step evolution state vector sequence to obtain the risk evolution trajectory.

[0127] The cumulative risk amount is obtained by accumulating the risk evolution trajectory along the time dimension. The cumulative risk amount is combined with the vector magnitude of the state vector of each stage in the state sequence to obtain the stage risk contribution value. The stage risk contribution value is then subjected to exponential normalization transformation to obtain the cumulative risk weight.

[0128] After obtaining the matching confidence score, it is necessary to quantitatively model the risk accumulation during the motion process using the Markov state transition matrix. The matching confidence score is typically a continuous value, ranging from 0 to 1. To adapt to the discrete state requirements of the Markov model, a numerical interval discretization mapping strategy is adopted. Specifically, the interval [0, 1] is divided into several non-overlapping sub-intervals, for example, five intervals: [0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), and [0.8, 1.0], corresponding to the five discrete state labels: "high risk," "medium-high risk," "medium risk," "medium-low risk," and "low risk," respectively. The matching confidence score at each time step is mapped to the corresponding state label according to its numerical interval, thus obtaining a set of discrete state labels. This set records the risk state classification at each time step during the motion process in chronological order.

[0129] After obtaining the set of discrete state labels, it is necessary to statistically analyze the transition patterns between states at adjacent time points. This involves iterating through adjacent state label pairs in the set and recording the transitions from state to state. Transition to state Construct a transition frequency matrix based on the frequency of the transitions. , where matrix elements Indicates from state Transition to state The number of times. The dimension of the transition frequency matrix is... ,in This represents the total number of discrete state labels. To convert the frequency matrix into a probability matrix, row normalization is performed on the transition frequency matrix, i.e., calculating... The Markov state transition matrix is ​​obtained. Each element of the matrix Indicates from state Transition to state The probability is such that the sum of the elements in each row equals 1, which conforms to the basic properties of a probability distribution.

[0130] Diagonal elements of the Markov state transition matrix Representing state The self-transition probability remains unchanged. The self-transition probability vector is constructed by extracting the diagonal elements. A high self-transfer probability indicates strong stability in a state, making it prone to prolonged stays once entered. In risk assessment scenarios, a high self-transfer probability in a high-risk state suggests strong "stickiness," meaning it's difficult to quickly recover to a safe state after entering a high-risk state; such states are identified as risk-sensitive states. A self-transfer probability threshold can be set... (For example, 0.6) will satisfy Furthermore, states with higher risk levels are marked as risk-sensitive states.

[0131] After identifying risk-sensitive states, state evolution reasoning is performed using stage state vectors. The stage state vectors are temporal feature representations generated by a causal convolutional network, arranged in chronological order to form a state sequence. ,in Indicates the first The stage state vector for each time period. Extract the initial stage state vector. This is then combined with the Markov state transition matrix through matrix operations. It's important to note that the stage state vector is typically represented in a high-dimensional feature space, while the Markov state transition matrix is ​​a low-dimensional matrix constructed based on discrete states. To achieve this combination, the stage state vector is first mapped through a fully connected layer to a matrix corresponding to the number of discrete states. The same dimensions yield the state distribution vector. Each element of this vector represents the probability distribution of the current discrete state.

[0132] Single-step evolution through computation Implementation, in which This represents the state distribution vector after one state transition. Continue with iterative matrix operations, the... The evolution state vector of step is A multi-step evolution state vector sequence is obtained through multiple iterations. ,in This is the preset maximum number of evolution steps. This sequence describes the evolution of the state distribution over time, starting from the initial state and under the influence of Markov transition laws.

[0133] Extract risk evolution trajectories from a multi-step evolutionary state vector sequence. For each evolutionary state vector... Extract the component values ​​of the dimension corresponding to the risk-sensitive state. Assume the index set corresponding to the risk-sensitive state is... Then in the first The risk component of the step is ,in express The Each component is a risk component. Arranging the risk components of all evolutionary steps in chronological order yields the risk evolution trajectory. This trajectory reflects the cumulative trend of risk states under the influence of state transition laws.

[0134] The cumulative risk amount is obtained by cumulatively calculating the risk evolution trajectory along the time series dimension. This value comprehensively reflects the degree of accumulation of risk states over multiple time steps starting from the initial state. The greater the accumulated risk, the higher the probability that the system will enter or remain in a high-risk state during its future evolution based on the current state.

[0135] To assess the contribution of each stage to the overall risk, the cumulative risk is combined with the vector magnitude of the state vectors at each stage in the state sequence. The vector magnitude of the stage state vector is calculated... This amplitude reflects the overall activation intensity of the motion characteristics during that time period. The stage risk contribution value is calculated using a weighted product approach. ,in This is the time decay coefficient, which assigns smaller weights to stages that are farther away from the current time. A common form is... This is the attenuation rate parameter.

[0136] The risk contribution values ​​at all stages are subjected to exponential normalization to obtain the cumulative risk weight. Exponential normalization is performed using the softmax function. ,in The temperature parameter controls the smoothness of the distribution. When the temperature parameter is small, the normalized weight distribution is more concentrated in stages with higher risk contributions; when the temperature parameter is large, the weight distribution is more uniform. Exponential normalization ensures that the sum of the cumulative risk weights across all stages is 1, highlighting the importance of high-risk contribution stages. The final risk cumulative weight vector is obtained. This provides a quantitative weighting basis for each stage of the subsequent overall risk assessment, enabling the risk assessment to focus on the movement stages that contribute significantly to the risk of injury.

[0137] In one alternative implementation,

[0138] A global risk assessment value is obtained through Bayesian inference based on the risk accumulation weight. A risk level identifier is generated by comparing the global risk assessment value with a safety threshold. A damage risk report is generated by combining the risk level identifier with the deviation metric value, including:

[0139] Extract the weight values ​​corresponding to each stage from the risk accumulation weight and construct a prior probability distribution. Use the weight values ​​as observation evidence to perform a Bayesian update on the prior probability distribution to obtain a posterior probability distribution. Extract the expected value from the posterior probability distribution and combine it with the weight values ​​to calculate the global risk assessment value.

[0140] A multi-level security threshold sequence is initialized and compared with the global risk assessment value at each level. The risk level code is determined based on the threshold interval into which the global risk assessment value falls. The confidence interval boundary value is extracted from the posterior probability distribution and the degree of dispersion is calculated to obtain the uncertainty index. The risk level code and the uncertainty index are combined to obtain the risk level identifier.

[0141] The deviation amplitude sequence for each time point is extracted from the deviation metric. The deviation extraction window length is determined based on the risk level identifier, and key deviation segments are extracted from the deviation amplitude sequence. The causal correlation strength between the key deviation segments and the global risk assessment value is calculated to obtain the contribution weight. The key deviation segments are sorted based on the contribution weight to generate a deviation tracing chain. The risk level identifier and the deviation tracing chain are encapsulated and the statistical feature parameters of the posterior probability distribution are added to obtain the damage risk report.

[0142] After obtaining the cumulative risk weights, the phased risk indicators need to be transformed into a determinable global assessment result. The weight values ​​corresponding to each phase are extracted from the cumulative risk weights; these weight values ​​reflect the degree of local risk contribution at different time periods or action phases. These weight values ​​are used as the sample space to construct a prior probability distribution, typically using a Beta distribution or a Dirichlet distribution as the conjugate prior to ensure the parsing feasibility of subsequent Bayesian updates. Specifically, during construction, the distribution range of each weight value is statistically analyzed, and the shape parameters of the prior distribution are set based on historical data or expert experience. For example, when using a Beta distribution, parameters are set... and This makes the mean of the prior distribution close to the historical average risk level of the moving object.

[0143] The obtained weight values ​​for each stage are considered as observational evidence, and a Bayesian update is performed on the prior probability distribution. The update process follows Bayes' theorem, multiplying the prior distribution by the likelihood function and then normalizing to obtain the posterior probability distribution. The likelihood function is constructed by statistically analyzing the frequency and magnitude of the weight values ​​for each stage, reflecting the correlation between the observed data and the parameters. When the cumulative risk weights for certain stages are observed to be significantly higher, the likelihood function will concentrate in high-risk areas, causing a corresponding shift in the posterior distribution. After the Bayesian update, the expected value is extracted from the posterior probability distribution. This expected value integrates prior knowledge and current observational evidence, serving as a central estimate of the risk level. Considering the differences in weight values ​​across stages, a weighted average is calculated between the expected value and the weight values ​​for each stage. The weights are determined by the duration or importance coefficient of each stage in the overall process, ultimately yielding the global risk assessment value.

[0144] To translate continuous global risk assessment values ​​into understandable risk levels, a multi-level safety threshold sequence needs to be initialized. This sequence contains several threshold nodes, such as setting three thresholds: low risk upper limit, medium risk upper limit, and high risk upper limit, dividing the risk space into four intervals: safe, caution, warning, and danger. The thresholds are set based on sports medicine research results and statistical analysis of a large number of samples, ensuring a significant correlation between different levels and the actual probability of injury. The global risk assessment value is compared level by level with each threshold, traversing upwards from the lowest threshold. When the global risk assessment value first falls below a certain threshold, its corresponding threshold interval is determined. A risk level code is assigned based on the interval; for example, interval one corresponds to code 0 for safe, interval two to code 1 for caution, interval three to code 2 for warning, and interval four to code 3 for danger.

[0145] Relying solely on a single numerical value for risk classification introduces uncertainty, especially when the overall risk assessment value approaches the threshold boundary. To quantify this uncertainty, confidence interval boundary values ​​are extracted from the posterior probability distribution, typically choosing the upper and lower bounds of the 95% confidence interval. The width or standard deviation of the confidence interval is calculated as a measure of dispersion, yielding an uncertainty index. A large uncertainty index indicates high volatility in the risk assessment and relatively low reliability of the risk classification. Combining the risk level code with the uncertainty index, for example using a binary tuple, with the risk level code occupying the higher digit and the uncertainty index, after normalization or discretization, occupying the lower digit, forms a complete risk level identifier. This identifier contains both risk level information and the confidence level of the judgment, facilitating comprehensive consideration in subsequent decision-making.

[0146] When generating an injury risk report, it is necessary not only to present the risk level but also to trace the specific causes of the risk. The deviation amplitude sequence for each time point is extracted from the deviation metric. These deviation amplitudes reflect the degree of difference between the moving object and the standard action template at each moment. The deviation extraction window length is determined based on the risk level identifier; the higher the risk level, the longer the window length to cover more potential abnormal behaviors. When the risk level identifier is low risk, the window length is set to a shorter value, such as covering the most recent 5 time points; when the risk level identifier is high risk, the window length is extended to cover the most recent 20 time points or longer. From the deviation amplitude sequence, data segments of a corresponding length are truncated backward from the current time point to obtain the key deviation segments.

[0147] To determine the impact of each key deviation segment on the overall risk assessment value, it is necessary to calculate the strength of the causal relationship between them. Granger causality tests or mutual information methods are used to analyze the predictive contribution of historical values ​​of the deviation amplitude sequence to the overall risk assessment value. Specifically, the deviation amplitude at each time point in the key deviation segment is used as the independent variable, and the overall risk assessment value is used as the dependent variable. A regression model or time series model is constructed, and the regression coefficients or causal influence coefficients are extracted as the strength of the causal relationship. The causal relationship strength at all time points is normalized to obtain contribution weights. The larger the contribution weight, the more significant the contributing effect of the deviation behavior at that time point on the overall risk.

[0148] Based on contribution weights, all time points in the key deviation segment are sorted in descending order. The top-ranked nodes (e.g., the first five nodes) are extracted as the primary sources of risk. These nodes are then rearranged chronologically to form a deviation source chain. This chain clearly shows the timeline of risk accumulation, indicating which abnormal actions or physiological responses had a critical impact on the final risk assessment. The deviation source chain includes the deviation magnitude, action type description, and abnormal physiological indicator type for each node, enhancing the report's interpretability.

[0149] After generating the risk level identifier and deviation tracing chain, the information is encapsulated to construct an injury risk report. The risk level identifier is the core field of the report, clearly indicating the current risk status of the participant. The deviation tracing chain serves as supplementary diagnostic information, detailing the deviations at each key time point to facilitate coaches or medical personnel in pinpointing the root cause of the problem. Statistical characteristics of the posterior probability distribution are also included, such as mean, variance, skewness, and kurtosis. These parameters comprehensively characterize the morphological features of the risk distribution. The mean reflects the central tendency of the risk, the variance reflects the volatility of the risk, the skewness reflects the symmetry of the risk distribution, and the kurtosis reflects the tendency for extreme risk events to occur. All information is organized in a structured format, such as JSO7 or XML, to ensure the report can be automatically parsed and visualized by the system.

[0150] Once generated, the injury risk report can be presented in various ways to meet specific needs. For professional users, a detailed report with complete mathematical parameters and statistical indicators is provided; for ordinary athletes, a simplified version is offered, showing only the risk level and key improvement recommendations. Visual charts can be embedded in the report, such as plotting the deviation amplitude sequence as a time-series line graph, marking the location of key deviation segments; and plotting the posterior probability distribution as a probability density curve to visually display the range of uncertainty in the risk assessment. Through multi-level and multi-dimensional information integration, the injury risk report ensures both scientific rigor and practical operability, providing a solid basis for sports injury prevention.

[0151] A second aspect of the present invention provides a deep learning-based sports injury risk identification system, comprising:

[0152] The data acquisition unit is used to acquire motion sequence data and physiological response data of the moving object, extract the motion trajectory features corresponding to the motion sequence data, extract the load response features corresponding to the physiological response data, and concatenate them with the motion trajectory features to obtain an initial feature matrix.

[0153] The graph structure transformation unit is used to perform graph structure transformation on the initial feature matrix to construct a kinematic constraint topology, perform multi-hop information aggregation on the kinematic constraint topology through a message passing mechanism to obtain node representations, generate the stage state vectors corresponding to the node representations through causal convolution, and calculate the similarity between the stage state vectors and the standard action template to obtain a deviation metric.

[0154] An anomaly identification unit is used to identify abnormal time periods based on the deviation metric value, extract the node representation corresponding to the abnormal time period to obtain the features of the region of interest, perform pose space matching on the features of the region of interest through the capsule network dynamic routing mechanism and the preset action pattern library to obtain the matching confidence, construct a Markov state transition matrix based on the matching confidence and perform state evolution reasoning in combination with the stage state vector to obtain the risk accumulation weight.

[0155] The risk assessment unit is used to obtain a global risk assessment value through Bayesian inference based on the risk accumulation weight, generate a risk level identifier by comparing the global risk assessment value with a safety threshold, and generate a damage risk report by combining the risk level identifier with the deviation metric value.

[0156] A third aspect of the present invention provides an electronic device, comprising:

[0157] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0158] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0159] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for identifying sports injury risks, characterized in that, include: Collect motion sequence data and physiological response data of the moving object, extract the motion trajectory features corresponding to the motion sequence data, extract the load response features corresponding to the physiological response data, and concatenate them with the motion trajectory features to obtain an initial feature matrix; The initial feature matrix is ​​transformed into a graph structure to construct a kinematic constraint topology. Multi-hop information aggregation is performed on the kinematic constraint topology through a message passing mechanism to obtain node representations. Causal convolution is used to generate the stage state vectors corresponding to the node representations. The similarity between the stage state vectors and the standard action template is calculated to obtain the deviation metric. Based on the deviation metric value, abnormal time periods are identified, and the node representations corresponding to the abnormal time periods are extracted to obtain the features of the region of interest. The features of the region of interest are matched in the posture space using the capsule network dynamic routing mechanism and the preset action pattern library to obtain the matching confidence. Based on the matching confidence, a Markov state transition matrix is ​​constructed and combined with the stage state vector to perform state evolution reasoning to obtain the risk accumulation weight. A global risk assessment value is obtained through Bayesian inference based on the risk accumulation weight. A risk level identifier is generated by comparing the global risk assessment value with a safety threshold. A damage risk report is generated by combining the risk level identifier with the deviation metric value.

2. The method according to claim 1, characterized in that, The process involves collecting motion sequence data and physiological response data of the moving object, extracting the motion trajectory features corresponding to the motion sequence data, extracting the load response features corresponding to the physiological response data, and concatenating these features with the motion trajectory features to obtain an initial feature matrix, including: A continuous frame image sequence of a moving object is acquired and joint position information is extracted. The acceleration and angular velocity signals of the moving object are collected and transformed into a coordinate system to obtain motion velocity information, which is then combined with the joint position information to obtain a motion data sequence. The acquisition device acquires the heart rate signal and electromyography signal of the moving object and combines them to obtain the physiological response data. Based on the joint position information in the motion sequence data, a skeleton topology is constructed and an attitude descriptor is calculated. The motion velocity information is then subjected to time window sliding sampling to determine the acceleration sequence, which is then time-aligned and spliced ​​with the attitude descriptor to obtain the motion trajectory features. The heart rate signal in the physiological response data is subjected to spectral decomposition and the frequency band energy distribution is extracted to obtain the heart rate frequency domain features. The envelope corresponding to the electromyographic signal is extracted and the slope of the envelope is calculated to obtain the muscle activation mode. The load response features are obtained by combining the heart rate frequency domain features. Based on the skeleton topology, the main motion links are determined, the coordination index between the attitude descriptors of adjacent joints in the main motion links is calculated, and the motion trajectory features are weighted and modulated to obtain the modulated trajectory features. The modulated trajectory features and the load response features are concatenated according to the time dimension to obtain the initial feature matrix.

3. The method according to claim 1, characterized in that, The initial feature matrix is ​​transformed into a graph structure to construct a kinematically constrained topology. Multi-hop information aggregation is then performed on this kinematically constrained topology using a message passing mechanism to obtain node representations, including: Based on the initial feature matrix, kinematic attribute types and temporal position indices are determined. Based on the kinematic attribute types, biomechanical coupling strength is calculated to obtain an attribute adjacency matrix. Based on the temporal position indices, motion continuity strength is calculated to obtain a temporal adjacency matrix, which is then fused with the attribute adjacency matrix using tensors to obtain a kinematic constraint topology. The initial feature matrix is ​​mapped to the initial embedding vector of the node in the kinematic constraint topology. The set of neighbor nodes of each node in the kinematic constraint topology is extracted and the initial embedding vector of each neighbor node is linearly projected to obtain the neighbor message set. Based on the initial embedding vector of the current node and the neighbor message set, the aggregate message vector is calculated and concatenated with the initial embedding vector of the current node and nonlinearly transformed to obtain the single-hop embedding vector. The shortest path length distribution is calculated based on the kinematic constraints and topology. The information propagation attenuation coefficient is determined based on the shortest path length distribution and weighted and modulated with the single-hop embedding vector to obtain an iterative embedding vector. The initial embedding vector is updated based on the iterative embedding vector and message passing is repeated to obtain a multi-hop embedding vector sequence. The cosine similarity of adjacent embedding vectors in the multi-hop embedding vector sequence is calculated and a convergence criterion is constructed. The termination embedding vector is determined based on the convergence criterion and a preset convergence threshold and is used as a node representation.

4. The method according to claim 1, characterized in that, The node representation generates a stage state vector through causal convolution. The similarity between the stage state vector and the standard action template is calculated to obtain a deviation metric, including: Temporal dimension information is extracted from the node representation and a temporal index mapping table is constructed. The node representation is temporally sorted based on the temporal index mapping table to obtain a temporally arranged node representation. The temporally arranged node representation is segmented according to a preset window length to obtain multiple temporal segments. A causal convolution kernel is applied to each temporal segment to perform convolution operation to obtain the segment convolution result. The convolution results of the segments are nonlinearly transformed and the peak response position of each temporal segment is extracted. Based on the peak response position, the motion stage boundaries are divided to obtain multiple motion stages. The convolution results of the segments in each motion stage are weighted and averaged to obtain the stage feature representation. The feature dimension is then transformed to obtain the stage state vector. The standard motion template matching the current motion type is retrieved from the preset standard motion template library and decomposed into multiple standard stage templates. The dot product between the stage state vector and the corresponding standard stage template is calculated and the amplitude is normalized to obtain the stage similarity. The stage similarity is then weighted and summed by stage weights and subjected to numerical inversion transformation to obtain the deviation metric value.

5. The method according to claim 1, characterized in that, Based on the deviation metric value, abnormal time periods are identified. The node representations corresponding to the abnormal time periods are extracted to obtain features of the region of interest. These features of the region of interest are then matched in pose space using a capsule network dynamic routing mechanism and a preset action pattern library to obtain matching confidence, including: The deviation metric is decomposed by wavelet multi-scale to obtain the deviation component. The local energy density of the deviation component is calculated and the location of energy abrupt change is identified to obtain the set of abnormal candidate points. A time series interval graph is constructed based on the set of abnormal candidate points and connected component analysis is performed on the time series interval graph to obtain the abnormal time period. The node data corresponding to the abnormal time period is extracted from the node representation to obtain the features of the region of interest. The feature dimension of the region of interest is expanded and a query vector is constructed. Standard action pattern features are extracted from the action pattern library and a pattern feature matrix is ​​constructed. A candidate pattern feature set is filtered based on the vector distance between the query vector and the pattern feature matrix. The features of the region of interest and the set of candidate pattern features are input into the capsule layer of the capsule network for dynamic routing iteration to obtain a query capsule representation and a set of candidate capsule representations. A query pose matrix is ​​extracted based on the query capsule representation, and a set of candidate pose matrices is extracted based on the set of candidate capsule representations. The matrix similarity between the candidate pose matrix and the query pose matrix is ​​calculated to obtain a pose similarity sequence. The collaborative matching score is calculated by combining the query capsule representation and the set of candidate capsule representations to obtain the matching confidence.

6. The method according to claim 1, characterized in that, Based on the matching confidence, a Markov state transition matrix is ​​constructed, and state evolution reasoning is performed in conjunction with the stage state vector to obtain the risk accumulation weights, including: The matching confidence is discretized into a numerical interval to obtain a discrete state label set. The temporal transition frequency between each state label in the discrete state label set is counted and a transition frequency matrix is ​​constructed. The transition frequency matrix is ​​row normalized to obtain the Markov state transition matrix. The diagonal elements are extracted from the Markov state transition matrix to construct a self-transition probability vector and risk-sensitive states are identified based on the self-transition probability vector. The stage state vectors are arranged into a state sequence according to the time sequence and the initial stage state vector is extracted. The initial stage state vector is then subjected to matrix operation with the Markov state transition matrix to obtain a single-step evolution state vector. Based on the single-step evolution state vector and the Markov state transition matrix, iterative matrix operation is performed to obtain a multi-step evolution state vector sequence. The component values ​​of each evolution state vector in the corresponding dimension of the risk-sensitive state are extracted from the multi-step evolution state vector sequence to obtain the risk evolution trajectory. The cumulative risk amount is obtained by accumulating the risk evolution trajectory along the time dimension. The cumulative risk amount is combined with the vector magnitude of the state vector of each stage in the state sequence to obtain the stage risk contribution value. The stage risk contribution value is then subjected to exponential normalization transformation to obtain the cumulative risk weight.

7. The method according to claim 1, characterized in that, A global risk assessment value is obtained through Bayesian inference based on the risk accumulation weight. A risk level identifier is generated by comparing the global risk assessment value with a safety threshold. A damage risk report is generated by combining the risk level identifier with the deviation metric value, including: Extract the weight values ​​corresponding to each stage from the risk accumulation weight and construct a prior probability distribution. Use the weight values ​​as observation evidence to perform a Bayesian update on the prior probability distribution to obtain a posterior probability distribution. Extract the expected value from the posterior probability distribution and combine it with the weight values ​​to calculate the global risk assessment value. A multi-level security threshold sequence is initialized and compared with the global risk assessment value at each level. The risk level code is determined based on the threshold interval into which the global risk assessment value falls. The confidence interval boundary value is extracted from the posterior probability distribution and the degree of dispersion is calculated to obtain the uncertainty index. The risk level code and the uncertainty index are combined to obtain the risk level identifier. The deviation amplitude sequence for each time point is extracted from the deviation metric. The deviation extraction window length is determined based on the risk level identifier, and key deviation segments are extracted from the deviation amplitude sequence. The causal correlation strength between the key deviation segments and the global risk assessment value is calculated to obtain the contribution weight. The key deviation segments are sorted based on the contribution weight to generate a deviation tracing chain. The risk level identifier and the deviation tracing chain are encapsulated and the statistical feature parameters of the posterior probability distribution are added to obtain the damage risk report.

8. A deep learning-based sports injury risk identification system, used to implement the method of any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire motion sequence data and physiological response data of the moving object, extract the motion trajectory features corresponding to the motion sequence data, extract the load response features corresponding to the physiological response data, and concatenate them with the motion trajectory features to obtain an initial feature matrix. The graph structure transformation unit is used to perform graph structure transformation on the initial feature matrix to construct a kinematic constraint topology, perform multi-hop information aggregation on the kinematic constraint topology through a message passing mechanism to obtain node representations, generate the stage state vectors corresponding to the node representations through causal convolution, and calculate the similarity between the stage state vectors and the standard action template to obtain a deviation metric. An anomaly identification unit is used to identify abnormal time periods based on the deviation metric value, extract the node representation corresponding to the abnormal time period to obtain the features of the region of interest, perform pose space matching on the features of the region of interest through the capsule network dynamic routing mechanism and the preset action pattern library to obtain the matching confidence, construct a Markov state transition matrix based on the matching confidence and perform state evolution reasoning in combination with the stage state vector to obtain the risk accumulation weight. The risk assessment unit is used to obtain a global risk assessment value through Bayesian inference based on the risk accumulation weight, generate a risk level identifier by comparing the global risk assessment value with a safety threshold, and generate a damage risk report by combining the risk level identifier with the deviation metric value.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images