A Method and System for Intelligent Assessment and Injury Risk Warning of Fitness Movements Based on Movement Phase Perception

Abstract

This invention discloses a method and system for intelligent assessment and injury risk warning of fitness movements based on motion phase perception, belonging to the field of sports biomechanics. It includes: collecting user motion images, basic health physiological data, and musculoskeletal geometric parameters to construct an assessment dataset; identifying skeletal points, dividing the movement execution phase, and extracting the temporal sequence of key joint angles; constructing a collaborative pattern dictionary grouped by movement type and execution phase; dynamically invoking time-varying collaborative constraints using a soft gating mechanism; constructing a combined loss function by fusing inverse dynamic physical consistency terms; and achieving real-time torque prediction of multiple joints based on a spatiotemporal graph neural network. It also involves calculating joint contact forces and soft tissue stress distribution by simplifying the musculoskeletal model, constructing a cumulative injury factor and a dynamic safety threshold, and combining a contrastive learning motion quality embedding network to generate a motion correctness score and an individualized injury risk assessment value. This invention solves the problems existing in the prior art.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent assessment and injury risk warning of fitness movements based on motion phase perception, belonging to the field of sports biomechanics technology. Background Technology

[0002] With the popularization of the concept of fitness for all, home fitness and smart fitness equipment have developed rapidly. Improper exercise movements are the core cause of sports injuries. According to statistics from sports medicine, more than 80% of home fitness injuries are caused by joint torque exceeding the physiological safety range or soft tissue stress exceeding the limit during the execution of movements. These potential risks cannot be fully identified by the appearance of the movements alone.

[0003] In existing technologies, intelligent fitness movement assessment solutions mainly fall into two categories. The first category is computer vision-based movement pattern recognition solutions, which extract skeletal point trajectories and angle deviations from user motion images to determine the correctness of movements. These solutions can only identify obvious deviations in the appearance of movements and cannot delve into the biomechanical aspects of joints. They are unable to predict the hidden risk of injury such as "compliant movement pattern but joint torque overload" or "abnormal joint contact force leading to cartilage wear," and their adaptability to individualized physiological characteristics and injury risks is insufficient. The second category is inverse dynamics-based musculoskeletal model calculation solutions, which rely on high-precision motion capture equipment and ground reaction force data. The modeling is complex and computation is time-consuming, making it difficult to achieve real-time calculations on ordinary smart terminals. Moreover, existing solutions are mostly aimed at clinical gait analysis and specific joint disease scenarios, and are not adapted to the diverse, dynamic, and non-periodic movement scenarios in mass fitness.

[0004] In recent years, existing technologies have proposed joint torque prediction methods based on general cooperative constraints. These methods utilize non-negative matrix factorization to extract static cooperative pattern matrices and embed them into the loss function of deep learning models to improve prediction accuracy and physiological plausibility. However, these technologies primarily target single hip joint torque prediction in periodic gait scenarios, employing a static cooperative pattern shared across the entire sequence. This fails to adapt to the non-periodic characteristics of significant changes in muscle recruitment strategies during the eccentric, transitional, and concentric phases of fitness movements. Furthermore, these technologies do not incorporate physical consistency constraints of the human kinetic chain. When performing individualized inference for users of different body types, the prediction results may exhibit physiologically unreasonable jumps that violate the torque balance relationship of the lower limb closed kinetic chain. In addition, these technologies only stop at torque prediction and have not established a comprehensive evaluation system from joint torque to joint contact force and soft tissue stress distribution. Moreover, they have not formed a complete closed-loop system for fitness scenarios by integrating movement pattern recognition and real-time risk warning.

[0005] For example, Chinese patent application CN121747898A discloses a hip joint torque prediction method based on general cooperative constraints. It extracts a static cooperative pattern matrix through non-negative matrix factorization and embeds it into the loss function of a deep learning model to improve the accuracy and physiological rationality of hip joint torque prediction. However, this technology has the following shortcomings: (1) It adopts a static cooperative pattern shared by the whole sequence, which is only for single hip joint torque prediction in periodic gait scenarios and cannot adapt to the non-periodic characteristics of significant changes in muscle recruitment strategies during the eccentric, transition and concentric phases of fitness movements; (2) It does not introduce physical consistency constraints of the human kinetic chain. When performing individualized reasoning for users of different body types, the prediction results may show physiologically unreasonable jumps that violate the torque balance relationship of the lower limb closed kinetic chain; (3) It only stops at torque prediction and does not establish a deeper evaluation system from joint torque to joint contact force and soft tissue stress distribution. Chinese patent application CN122004838A discloses a method and system for predicting joint torque based on graph neural networks. It constructs static and dynamic graphs using IMU sensor data to extract spatial features and combines them with gated loop units for temporal modeling. However, this method does not introduce a collaborative constraint mechanism for motion phase perception, nor does it establish a complete chain for physical consistency verification and damage risk assessment.

[0006] In summary, existing technologies lack a comprehensive fitness movement assessment solution that integrates computer vision motion recognition, dynamic collaborative constraints based on motion stage perception, inverse dynamic physical consistency verification, and assessment of joint contact force and cumulative damage factors. This makes it impossible to achieve in-depth assessment from the compliance of movement appearance to the biomechanical safety of tissues, and it is difficult to meet the personalized, highly accurate, and forward-looking safety guidance needs in home fitness scenarios. Summary of the Invention

[0007] The purpose of this invention is to propose a method and system for intelligent assessment and injury risk warning of fitness movements based on movement stage perception. This addresses the problems in existing technologies, such as the inability of static collaborative models to adapt to the phased mechanical changes of fitness movements, the lack of physical consistency in individualized reasoning of pure data-driven models, and the inability of risk assessment to reflect the cumulative state of soft tissue stress by only focusing on torque thresholds. This invention aims to achieve high-precision assessment and forward-looking injury risk warning of fitness movements.

[0008] The intelligent assessment and injury risk warning method for fitness movements based on motion phase perception described in this invention includes the following steps: S1: Acquire motion data of user fitness movements and user-specific physiological data, and at the same time acquire standard musculoskeletal dynamics dataset; S2: Perform motion recognition on the motion data of the user's fitness movements, determine the type of fitness movement performed by the user, divide the movement execution stages according to kinematic characteristics, and generate motion type recognition results and motion stage sequences; S3: Based on the standard musculoskeletal dynamics dataset, construct an action-stage specific cooperative pattern dictionary according to action type and action execution stage; dynamically call the corresponding stage cooperative pattern matrix according to the action type identification result and action stage sequence to generate time-varying cooperative constraint terms; construct a combined loss function of fusion data fitting term, the time-varying cooperative constraint term and inverse dynamic physical consistency term, train the joint torque prediction model, and output the real-time torque prediction results of multiple joints during user action execution; S4: Based on the real-time torque prediction results and the user's individualized physiological data, calculate the joint biomechanical load parameters, construct the user's personalized dynamic safety threshold, generate the action correctness score and individualized injury risk assessment value, and perform risk level stratification according to the individualized injury risk assessment value to generate risk grading results. S5: Based on the risk classification results, implement differentiated real-time early warning strategies and generate personalized action correction guidance data.

[0009] Preferably, the motion data of the user's fitness movements in step S1 includes motion image data of the user's fitness movements; the user's individualized physiological data includes the user's basic health physiological data and the user's musculoskeletal geometric parameters; the musculoskeletal geometric parameters include at least the lower limb segment length, the relative position of the joint center, and the estimated value of body mass distribution.

[0010] Preferably, the division of the action execution stages based on kinematic characteristics in step S2 includes: for various fitness movements such as squats, deadlifts, lunges, bench presses, and bent-over rows, the movement cycle is divided into at least three stages based on biomechanical inflection points: eccentric stage, transition stage, and concentric stage; the division is based on the key joint angle threshold and the change in angular velocity direction.

[0011] Preferably, the step S3 of constructing the action-stage specific cooperative pattern dictionary includes: independently performing nonnegative matrix decomposition on the multi-joint torque time-series data within each action and stage. ,in Indicates the action type index. Indicates the stage index. For action stage The collaborative mode matrix, This is the corresponding activation coefficient matrix; The number of cooperative patterns for each action and each stage is determined based on the interpretable variance ratio, and a cooperative pattern dictionary with a hierarchical index structure is constructed. .

[0012] Preferably, the dynamic invocation of the corresponding stage's collaborative mode matrix in step S3 includes: Based on the joint angle time series data extracted in step S2, calculate the current time. Membership function values ​​relative to each action phase The membership function value The key joint angles and angular velocities are determined using a Gaussian mixture model or fuzzy logic. Generate time-varying cooperative constraint matrix ,in ; The time-varying cooperative constraint term To predict the torque relative to the time-varying cooperative constraint matrix The L2 norm of the reconstruction residual in the collaborative subspace is calculated as follows: ; Where: is time. The predicted torque vector, These are the optimal reconstruction coefficients.

[0013] Preferably, the calculation method for the inverse dynamics physical consistency term in step S3 is as follows: A simplified inverse dynamics model of the lower limb closed kinetic chain is constructed based on user musculoskeletal geometry parameters, and the theoretical torque of each joint is calculated based on joint angle time-series data. ; Calculate the predicted torque With theoretical torque The residuals are determined, and a lower limb sagittal plane moment balance constraint is introduced: the hip joint moment, knee joint moment and ankle joint moment satisfy a closed chain transmission relationship; The inverse dynamic physical consistency term is: ; Where: represents the torque balance residual term of the closed chain. This is the balance coefficient.

[0014] Preferably, the joint torque prediction model in step S3 is a spatiotemporal graph neural network. The spatiotemporal graph neural network constructs a graph structure based on the topology of the human skeleton, with nodes corresponding to the hip, knee, ankle, shoulder, elbow, and wrist joints, and edges corresponding to bone connection segments. The network layer includes a spatiotemporal graph convolutional layer and a gated recurrent layer. The spatiotemporal graph convolutional layer is used to aggregate the temporal features of biomechanically adjacent joints, and the gated recurrent layer is used to capture the temporal dependencies of the action phase.

[0015] Preferably, the joint biomechanical load parameters in step S4 include joint contact force and soft tissue stress distribution; calculating the joint contact force and soft tissue stress distribution includes: estimating the tension of major muscle groups using a simplified muscle force line model based on the user's musculoskeletal geometry parameters and real-time torque prediction results; and calculating the joint contact force according to the relationship between the muscle group tension and joint geometry. Tensile stress on ligaments / tendons ; Establish cumulative damage factors: ; in: For dynamic security thresholds, Used to quantify the time-cumulative effect of overload stress.

[0016] Preferably, the generation of the action correctness score in step S4 includes: mapping the joint angle temporal data of the standard action template and the user action data to the same embedding space, learning the action quality representation through a dual-branch encoder, using the standard action embedding as the anchor point and the user action embedding as a positive or negative sample, and optimizing with a triplet loss function to make the action correctness score monotonically mapped to the distance in the embedding space; the individualized injury risk assessment value is generated through a deep association model, which uses a Transformer-based temporal encoder, taking the joint contact force temporal sequence, soft tissue stress temporal sequence, action angle deviation temporal sequence and user physiological risk factors as inputs, and capturing the cross-temporal dependencies of multi-source features through a multi-head self-attention mechanism to output the individualized injury risk assessment value.

[0017] The intelligent fitness movement assessment and injury risk warning system based on movement phase perception described in this invention is used to execute the intelligent fitness movement assessment and injury risk warning method based on movement phase perception as described above. The system includes: The data acquisition module is used to acquire motion data of user fitness movements, user-specific physiological data, and standard musculoskeletal dynamics datasets; The motion recognition and stage segmentation module is used to recognize motion data and segment the motion execution stages. The joint torque prediction module is used to construct a motion-stage specific cooperative pattern dictionary, generate time-varying cooperative constraints by dynamically calling them, and train the joint torque prediction model based on the combined loss function that integrates inverse dynamic physical consistency terms, and output the real-time torque prediction results of multiple joints. The assessment and risk grading module is used to calculate joint biomechanical load parameters, generate movement correctness scores and individualized injury risk assessment values, and stratify risk levels. The early warning and guidance generation module is used to execute differentiated early warning strategies and generate personalized action correction guidance data.

[0018] Compared with existing technologies, the intelligent assessment and injury risk warning method and system for fitness movements based on motion phase perception of the present invention exhibits the following beneficial effects in terms of technical performance and practical application: (1) A dynamic collaborative constraint mechanism for action stage perception was proposed. Through grouped non-negative matrix decomposition and soft gating dynamic invocation, the technical bottleneck of static collaborative mode in the existing technology cannot adapt to the stage changes of fitness movements was solved, and the predictive smoothness and physiological rationality at the stage transition were significantly improved.

[0019] (2) The physical consistency term of inverse dynamics is embedded in the loss function and combined with the user's musculoskeletal geometric parameters to realize the physical self-consistency verification during individualized reasoning, filling the technical gap of the disconnect between "data-driven prediction" and "biomechanical physical laws" in the existing technology.

[0020] (3) A multi-joint torque prediction framework based on spatiotemporal graph neural network was constructed. The temporal coupling of biomechanical adjacent joints was explicitly modeled using the topological structure of human skeleton. Only joint angles and musculoskeletal geometric parameters are needed to achieve high-precision and high-consistency real-time torque prediction. No additional data such as ground reaction force is required. It can be directly deployed on ordinary smart terminals.

[0021] (4) A comprehensive assessment system was established, encompassing joint torque, joint contact force, soft tissue stress distribution, and cumulative injury factors. By combining comparative learning motion quality embedding with the Transformer deep association model, a fusion assessment of motion morphology and tissue biomechanical safety was achieved, significantly improving the accuracy and foresight of injury risk assessment.

[0022] (5) The solution forms a complete technical closed loop of "data collection - stage identification - dynamic torque prediction - contact force / stress assessment - risk classification - closed-loop guidance", which can be directly applied to various products such as smart fitness APP, smart fitness mirror, and home fitness equipment, and has high industrial application value. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall steps of an intelligent fitness movement assessment and injury risk warning method based on dynamic collaborative constraints and physical consistency with motion phase perception, according to the present invention. Figure 2 This is a schematic diagram of the fitness movement recognition and kinematic parameter extraction process in step S2 of the present invention; Figure 3 This is a schematic diagram of the action-stage specific cooperative pattern dictionary construction and soft gating dynamic invocation process in step S3 of the present invention; Figure 4 This is a schematic diagram of the construction process of the combined loss function that incorporates the inverse dynamics physical consistency term in step S3 of the present invention; Figure 5 This is a schematic diagram of the motion assessment and risk classification process based on joint contact force and cumulative damage factors in step S4 of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] Example 1: like Figure 1 As shown, this embodiment uses a user performing a weighted squat at home as an example to illustrate the complete execution flow of the method of the present invention. The specific steps are as follows: Step S1: Data Acquisition and Preprocessing Step S11: The user captures continuous images of the squatting motion using the front-facing camera of their mobile phone. The capture frame rate is set to 30fps, and the capture period covers the complete motion cycle of the user from standing up, squatting down to the lowest point, and then standing up to return to the starting position. A total of 3 consecutive squatting motion images are captured to generate user motion image data.

[0026] Step S12: Collect the user's basic health and physiological data, including: height 175cm, weight 80kg, age 30, male, no history of sports injury, exercise 3 times a week, and moderate training level, to form the user's basic health and physiological data.

[0027] Step S13: Collect the user's musculoskeletal geometric parameters, including: thigh length 0.45m, calf length 0.38m, foot mass as a percentage of total body mass approximately 1.4%, and thigh mass as a percentage of total body mass approximately 10.5%. These parameters are estimated based on anthropometry regression equations after the user inputs their height and weight, or obtained through simple measurement of the relative distance between skeletal points in the image.

[0028] Step S14: Collect kinematic data of the standard squat movement and the corresponding gold standard data of hip, knee and ankle joint flexion / extension torque, adduction / abduction torque and rotation torque. The data comes from the motion collection of professional athletes and the simulation calculation results of the OpenSim musculoskeletal model to form a standard musculoskeletal dynamics dataset.

[0029] Step S15: Perform Gaussian filtering for noise reduction, BodyPix network background segmentation, and brightness normalization preprocessing on the motion image data; perform Butterworth low-pass filtering, Z-score normalization, and temporal alignment preprocessing on the joint angle and torque data; standardize and encode the user's musculoskeletal geometric parameters and physiological data; construct a fitness movement evaluation dataset, in which 70% of the data is divided into a training set and 30% into a test set.

[0030] like Figure 2 As shown, step S2: Fitness movement recognition and kinematic parameter extraction: Step S21: Using the MediaPipeBlazePose lightweight pose estimation algorithm, skeletal point recognition is performed on the preprocessed motion image data to locate key body nodes of the user's shoulder, elbow, wrist, hip, knee, and ankle, and generate three-dimensional spatial coordinate temporal data of each node in consecutive frames.

[0031] Step S22: Based on the three-dimensional spatial coordinate temporal data of key nodes, calculate the joint angles of the hip-knee-ankle kinematic chain segment, and extract the temporal data of the flexion / extension angles of the hip, knee, and ankle joints during the squatting motion cycle.

[0032] Step S23: Match the user's joint angle time series data with the standard parameter set of movements such as squat, deadlift, and lunge in the standard movement library, calculate the time series similarity through the Dynamic Time Warping (DTW) algorithm, and finally identify the user's current movement as a weighted squat with a matching confidence of 0.96, and generate the movement type recognition result.

[0033] Step S24: Divide the movement execution phases based on changes in knee flexion angle and angular velocity. Set knee flexion angle thresholds: when the knee flexion angle is less than 20 degrees and the absolute value of the angular velocity is less than 5 degrees / second, it is determined to be the standing starting position; when the angular velocity is negative (knee flexion increases) and the flexion angle is in the range of 20 to 100 degrees, it is determined to be the eccentric phase; when the angular velocity changes from negative to positive or the flexion angle is in the range of 100 to 110 degrees, it is determined to be the bottom transition phase; when the angular velocity is positive (knee extension) and the flexion angle decreases from 110 degrees to 20 degrees, it is determined to be the concentric phase. Generate a sequence of movement phases and mark the phase to which each moment belongs.

[0034] Step S3: Joint torque prediction based on dynamic cooperative constraints and physical consistency constraints of motion phase perception: like Figure 3 As shown, step S31: The multi-dimensional torque time-series data of the hip, knee, and ankle joints in the standard squat exercise dataset are grouped according to the eccentric phase, bottom transition phase, and concentric phase. Non-negative matrix decomposition is then performed on each group of data. ; in, Corresponding to the centrifugation stage, Corresponding to the bottom transition phase, Corresponding to the centripetal phase. Calculate the explained variance ratio for each group, when the number of cooperative patterns in the centrifugal phase... The explained variance ratio reached 93% during the bottom transition phase. When it reaches 95%, the centripetal phase The success rate reached 94%, based on which a collaborative pattern dictionary for the squatting motion was constructed. .

[0035] Step S32: Calculate the stage membership degree at each moment based on the user action stage sequence generated in step S24. For example, when the user's knee flexion angle is 85 degrees and the angular velocity is -25 degrees / second (downward movement), the eccentric stage membership degree... Membership degree of the bottom transition phase Centripetal phase membership Generate time-varying collaborative constraint matrix: ; like Figure 4 As shown, step S33: Construct the combined loss function: ; Among them, data fitting term The mean square error between the predicted torque and the actual torque is used; the balance coefficient is employed. , .

[0036] Time-varying collaborative constraint terms The calculation is as follows: ; The inverse dynamics physical consistency term is calculated as follows: Based on the user's thigh length of 0.45m, calf length of 0.38m, and lower limb mass distribution, the theoretical knee joint torque and theoretical hip joint torque are calculated using a simplified inverse dynamics model (Newton-Euler method). Introducing the sagittal moment balance residual term: ; Wherein is the segmental inertial torque compensation term for the lower limbs; then: ; in .

[0037] Step S34: Using joint angle temporal data and user musculoskeletal geometric parameters as input, and joint torque gold standard data as output, construct a spatiotemporal graph neural network. The graph structure node set is as follows: The edges are defined based on the connection relationships of the human skeleton, and the adjacency matrix elements are initialized using biomechanical coupling strength. The network consists of three spatiotemporal graph convolutional layers and two bidirectional gated recurrent units, used to aggregate spatially adjacent joint features and capture temporal dependencies. The Adam optimizer is used to minimize the combined loss function, and the number of iterations is set to 150 epochs. After training, a joint torque prediction model is obtained.

[0038] Step S35: Input the time-series data of hip, knee, and ankle joint angles and musculoskeletal geometric parameters extracted in step S2 into the trained model, and output the real-time flexion torque prediction results of the hip, knee, and ankle joints during the user's squatting motion. Among them, the peak knee flexion torque at the lowest point of the squat is 1.82 N·m / kg, the peak hip flexion torque is 1.45 N·m / kg, and the peak ankle dorsiflexion torque is 0.68 N·m / kg.

[0039] Step S4: Intelligent motion assessment and injury risk classification based on joint contact force and cumulative injury factors; like Figure 5 As shown, in step S41: based on the user's thigh length of 0.45m, weight of 80kg, and the torque prediction results output in step S35, the quadriceps muscle tension and hamstring muscle tension are estimated using a simplified musculoskeletal model. Based on the patellar tendon force line angle and the geometric relationship of the tibial plateau, the knee joint contact force is calculated. (BW is a multiple of body weight, approximately 2080N), patellar tendon tensile stress MPa, anterior cruciate ligament stress MPa.

[0040] Step S42: Based on the user's BMI of 26.1, no history of injury, and moderate training level, construct dynamic safety thresholds: knee flexion torque threshold 1.9 N·m / kg, knee contact force threshold 3.0 BW, and patellar tendon stress threshold 15.0 MPa.

[0041] Step S43: Calculate torque deviation: (Not exceeding limits); Contact force deviation: (Not exceeding the limit); Patellar tendon stress did not exceed the threshold.

[0042] The user's joint angle time-series data is mapped to a standard squat movement template using a contrastive learning motion quality embedding network. A dual-branch encoder extracts the user's motion embedding vector and the standard template embedding vector separately. Calculate the Euclidean distance The action correctness score generated based on the distance-score mapping function is 93 points (out of 100).

[0043] Step S44: Using a pre-trained Transformer deep association model, input the joint contact force time series, patellar tendon stress time series, ACL stress time series, motion angle deviation sequence, BMI, no injury history label, and training level encoding. A multi-head self-attention mechanism is used to capture cross-time-step feature dependencies, and the cumulative injury factor is calculated. ; The individualized injury risk assessment value was 0.18.

[0044] Step S45: Based on the risk assessment value of 0.18, which falls within the 0-0.3 range, a low-risk risk classification result is generated.

[0045] Step S5: Generation of tiered early warnings and personalized guidance; Step S51: For the low-risk classification result, implement the standard prompting strategy. Extract the optimizable parameters of the hip flexion angle and knee flexion angle in the user's movement, and generate progressive movement optimization suggestions: The quality of this squat movement is excellent, and the knee torque and contact force are within the safe range. It is recommended that in the next training session, the hip flexion angle be appropriately reduced by about 2 degrees to maintain greater trunk stability, and the knee flexion angle deviation be gradually controlled within 2 degrees to further reduce the cumulative load on the patellar tendon.

[0046] Step S52: Simultaneously record the angle deviation, peak torque, contact force, action correctness score, and cumulative injury factor of the user's current action, generate progress tracking data, compare it with the user's historical training data, and show the trend of action progress.

[0047] Step S53: Record the user's behavioral response after receiving guidance. If the user corrects the movement according to the suggestions during subsequent training and the deviation of the hip joint forward tilt angle decreases, then dynamically adjust the focus of subsequent guidance content and fine-tune the dynamic safety threshold to form a closed-loop optimization.

[0048] Example 2 This embodiment addresses a high-risk scenario involving a user's lunge motion, illustrating the risk warning and intervention process of this invention. The core steps are as follows: Steps S1-S2: Collect motion imaging data of the user's lunge movement. The user's basic physiological data are: height 170cm, weight 75kg, age 35, history of right knee anterior cruciate ligament injury, and fitness beginner; the user's musculoskeletal geometry parameters are: thigh length 0.42m, calf length 0.36m. Timing data of hip, knee, and ankle joint angles are extracted through posture recognition, and the movement type is identified as a lunge. Based on the knee flexion angle and angular velocity, the movement cycle is divided into three phases: the right knee joint experiences the eccentric phase (flexion angle 20° → 95°), the bottom transition phase (flexion angle 95° → 105°), and the concentric phase (flexion angle 105° → 20°).

[0049] Step S3: Generate time-varying cooperative constraints using an action-stage specific cooperative pattern dictionary and a soft gating mechanism. During the bottom transition phase, the predicted peak value of the right knee adduction moment is 1.15 N·m / kg. Through inverse dynamics physical consistency verification: based on the user's lower limb segment length and mass distribution, the theoretical knee adduction moment is calculated to be 1.08 N·m / kg. The residual between the predicted and theoretical values ​​is 6.5%, which is less than the set threshold of 10%, thus the prediction result is deemed physically consistent.

[0050] Step S4: Based on the user's history of ACL injury in the right knee, the general safety threshold for knee adduction torque (0.8 N·m / kg) is personalized to 0.55 N·m / kg; the general threshold for knee contact force is adjusted to 2.2 BW. The user's peak torque exceeds the safety threshold by 109%, with an excess duration of 0.9 s. The right knee contact force is calculated using a simplified musculoskeletal model. Exceeding the threshold by 27.3%; medial collateral ligament stress MPa, approaching the high-risk threshold. Cumulative damage factor. The individualized injury risk assessment value is 0.88, falling within the 0.7-1.0 range, resulting in a high-risk risk classification result.

[0051] Step S5: For high-risk classification results, immediately implement a mandatory warning strategy. A voice prompt will say, "Your right knee joint valgus is severely excessive. Both knee joint contact force and ligament stress have exceeded safe limits, posing an extremely high risk of anterior cruciate ligament re-injury. Please stop the current movement immediately." Simultaneously, the right knee joint will be highlighted in red on the screen, along with the stress concentration area (medial collateral ligament and patellar trajectory). High-risk deviations will be categorized as technical errors (knee valgus) combined with pre-existing injury-related limiting errors. Professional guidance data will be generated: "Continue performing the current weighted lunge movement. It is recommended to replace it with a weightless wall squat. Based on stress distribution analysis, there is significant stress concentration in the medial collateral ligament area of ​​your right knee. It is recommended to prioritize strengthening the vastus medialis and gluteus medius muscles to improve knee joint dynamic stability, and then gradually resume lunge training. If you experience knee discomfort, please consult a sports rehabilitation physician promptly." User feedback will be recorded simultaneously. If the user continues to perform high-risk movements, mandatory warnings will continue to be issued until the user stops the movement.

[0052] Experimental verification and comparative analysis: To comprehensively evaluate the performance differences between the method of this invention and existing technologies, a multimodal dataset was constructed, encompassing five common fitness movements: squat, deadlift, lunge, bench press, and bent-over row. The dataset was collected from 326 subjects of different ages (18-55 years old) and training levels (beginner / intermediate / advanced), with gold standard data from professional athletes included as a reference. Three existing technology comparison schemes were set up: Scheme A is a torque prediction method based on full-sequence static cooperative constraints (using a single cooperative mode matrix shared throughout the entire cycle); Scheme B is a purely data-driven spatiotemporal graph neural network method (without cooperative constraints or physical consistency terms); Scheme C is a traditional risk assessment method based on torque threshold and DTW morphological alignment. The comparative experimental results are reported below from four dimensions: joint torque prediction accuracy, physical consistency, injury risk identification accuracy, and system real-time performance.

[0053] (I) Comparison of joint torque prediction accuracy In the joint torque prediction task, the gold standard torque calculated by OpenSim musculoskeletal model simulation was used as a reference, and the root mean square error (RMSE) and transition jump amplitude (TJA) were used as evaluation indicators. As shown in Table 1, the method of the present invention (hereinafter referred to as "the present invention") is significantly better than the three comparative schemes in torque prediction accuracy for the three key joints of hip, knee and ankle.

[0054] Table 1. Comparison of RMSE for multi-joint torque prediction using different methods (unit: N·m / kg)

[0055] As shown in Table 1, compared to the static collaborative constraint method of Scheme A, this invention, through an action-stage specific collaborative pattern dictionary and a soft-gating dynamic invocation mechanism, reduces the predicted RMSE of the knee joint torque from 0.42 to 0.23 N·m / kg, a reduction of 45.2%; the predicted RMSE of the hip joint torque from 0.38 to 0.21 N·m / kg, a reduction of 44.7%; and the predicted RMSE of the ankle joint torque from 0.29 to 0.17 N·m / kg, a reduction of 41.4%. Regarding the smoothness of stage transitions, this invention reduces the predicted jump amplitude at the transition point from the eccentric stage to the bottom stage by 68.3% compared to Scheme B, effectively solving the inherent defect of predicted jumps at stage transitions in the static collaborative mode. This result demonstrates that the dynamic collaborative constraint mechanism can accurately adapt to the significant differences in muscle recruitment strategies at different stages of fitness movements, significantly improving the continuity and physiological rationality of torque prediction.

[0056] (II) Physical Consistency Verification Comparison To verify the effect of the inverse dynamics physical consistency term on improving the physical consistency of the prediction results, the Closed-chain Torque Balance Residual (CTBR) was introduced as a physical consistency evaluation index, and the relative deviation between the predicted torque and the simplified inverse dynamics theoretical torque was calculated. The experimental results are shown in Table 2.

[0057] Table 2 Comparison of Physical Consistency Indicators for Different Methods

[0058] As shown in Table 2, compared to the purely data-driven approach B, the introduction of the inverse dynamics physical consistency term in this invention reduces the average CTBR of predicted torque from 12.7% to 4.1%, improving physical consistency by 67.7%; the proportion of CTBR exceeding the 10% threshold decreases from 23.6% to 3.8%, reducing the abnormal prediction rate by 83.9%. In cross-user generalization tests (using data from 68 new subjects who did not participate in training), the generalization RMSE of this invention is 0.25 N·m / kg, a reduction of 39.0% compared to approach A and 51.9% compared to approach B. These results fully demonstrate that the inverse dynamics physical consistency term can effectively suppress the violation of torque balance relationships and abnormal jumps that may occur in the purely data-driven model when performing individualized inference for users of different body types, ensuring that the prediction results always conform to the basic laws of human biomechanics.

[0059] (III) Comparison of Accuracy Rates of Damage Risk Identification In the injury risk assessment task, the motion data of 326 subjects were independently labeled with risk levels (low / medium / high) by three professional sports rehabilitation physicians as the gold standard. The risk identification performance of the method of this invention and scheme C (traditional torque threshold + DTW method) was compared. The evaluation indicators included high-risk identification accuracy (HA), overall classification accuracy (OA), false positive rate (FPR), and false negative rate (FNR). The experimental results are shown in Table 3.

[0060] Table 3 Comparison of Damage Risk Identification Performance of Different Methods

[0061] As shown in Table 3, the in-depth assessment system based on joint contact force and cumulative injury factors of this invention achieved an accuracy rate of 94.6% in identifying high-risk movements, an improvement of 18.4 percentage points compared to 76.2% in scheme C; the overall classification accuracy reached 95.3%, an improvement of 16.8 percentage points compared to scheme C; the false alarm rate decreased from 19.8% to 6.3%, a reduction of 68.2%; and the false negative rate decreased from 15.3% to 4.8%, a reduction of 68.6%. Particularly noteworthy is that in a subgroup of 42 subjects with a history of sports injuries, the high-risk identification accuracy of this invention reached 97.1%, while scheme C only achieved 71.4%, an improvement of 25.7 percentage points. This result fully demonstrates that by deepening the assessment dimensions from a single torque threshold to joint contact force, soft tissue stress distribution, and cumulative injury factors, this invention can more accurately capture individualized hidden injury risks, especially demonstrating a higher risk prospective identification capability for individuals with a history of injury.

[0062] (iv) Comparison of consistency in motion quality scoring In the movement quality rating task, the gold standard (out of 100) was the manual rating of the participants' movement quality by three professional sports rehabilitation physicians. The consistency of the ratings was compared between the contrastive learning movement quality embedding network of this invention and the traditional DTW temporal alignment method in scheme C. Pearson correlation coefficient, Spearman rank correlation coefficient, and mean absolute error (MAE) were used as evaluation indicators. The experimental results are shown in Table 4.

[0063] Table 4. Comparison of consistency in motion quality scoring using different methods

[0064] As shown in Table 4, the Pearson correlation coefficient between the action correctness scores generated by the comparative learning action quality embedding network and the expert human scores of this invention reached 0.91 (p<0.001), an improvement of 26.4% compared to 0.72 in scheme C; the Spearman rank correlation coefficient reached 0.89, an improvement of 29.0% compared to 0.69 in scheme C; and the mean absolute error of the scores decreased from 8.6 points to 3.2 points, a reduction of 62.8%. These results indicate that this invention, by mapping action morphology evaluation to an embedding space anchored by standard actions, can more accurately quantify action execution quality and avoid the scoring bias caused by the oversensitivity of the DTW method to temporal deformation.

[0065] (v) Real-time performance testing of the system To verify the feasibility of deploying the method of this invention on ordinary smart terminals, end-to-end inference latency tests were conducted on three mainstream mobile processor platforms. The tests covered the complete inference chain, including motion image skeleton point recognition, action type and stage recognition, multi-joint torque prediction, joint contact force and stress calculation, injury risk assessment, and early warning guidance generation. The experimental results are shown in Table 5.

[0066] Table 5 End-to-end inference latency tests on different platforms

[0067] As shown in Table 5, the end-to-end inference latency of the method of the present invention is only 38ms on smartphones equipped with flagship mobile processors, corresponding to a frame rate of 26.3fps, which meets the real-time processing requirements under 30fps input; on mid-range smartphones, the latency is 52ms, corresponding to 19.2fps, which can meet the requirements of near real-time scenarios; on smart fitness mirror devices, the latency is 71ms. Although it does not reach the full real-time standard, smooth real-time feedback can be achieved through inter-frame interpolation strategies. The above results show that the method of the present invention does not rely on high-precision motion capture equipment and force tables or other dedicated hardware. It only requires the camera of an ordinary smart terminal to realize the real-time calculation of the entire chain from motion recognition to injury risk assessment, and has broad industrial deployment conditions and high cost-effectiveness advantages.

[0068] (vi) Comprehensive experimental conclusions Based on the results of the five sets of comparative experiments, the method of this invention has achieved a comprehensive and significant improvement over existing technologies: In terms of joint torque prediction accuracy, the dynamic collaborative constraint mechanism reduces the average RMSE of multi-joint torque prediction by 43.8% and the stage switching jump amplitude by 68.3%; in terms of physical consistency, the inverse dynamics physical consistency term reduces the closed-chain torque balance residual by 67.7% and the cross-user generalization error by 51.9%; in terms of injury risk identification, the in-depth assessment system improves the accuracy of high-risk identification by 18.4 percentage points, reduces the false alarm rate by 68.2%, and reduces the false negative rate by 68.6%; in terms of motion quality scoring, the contrastive learning embedding network increases the correlation coefficient of expert scoring consistency from 0.72 to 0.91; and in terms of system deployment, the overall end-to-end inference latency is as low as 38ms, enabling real-time calculation on ordinary smart terminals. The above experimental data fully demonstrate the creative contribution and significant beneficial effects of the technical solution of this invention. It solves the key technical bottlenecks in the prior art, such as insufficient adaptability of static collaborative mode, lack of physical consistency of pure data-driven model, and single risk assessment dimension. It can provide home fitness users with personalized, highly accurate, and forward-looking intelligent motion assessment and injury risk warning services.

[0069] Example 3 The intelligent fitness movement assessment and injury risk warning system based on movement phase perception described in this invention is used to execute the intelligent fitness movement assessment and injury risk warning method based on movement phase perception as described in Example 1. The system includes: Data acquisition module The data acquisition module is used to acquire motion data of user fitness movements, individualized physiological data of users, and standard musculoskeletal dynamics datasets. This module consists of three sub-units: The motion capture subunit uses the camera of a smart terminal to capture continuous video data of the user performing fitness movements in real time, with a frame rate of no less than 30fps, covering the complete cycle of the user's movements. This subunit supports various terminal forms such as front-facing cameras of mobile phones, cameras of smart fitness mirrors, and tablet cameras, without relying on professional motion capture equipment.

[0070] The physiological data entry subunit is used to acquire users' basic health physiological data and musculoskeletal geometric parameters. Users input basic information such as height, weight, age, gender, history of sports injuries, and training level through an interactive interface. Based on the height and weight data input by the user, the system automatically estimates the lower limb segment length, relative position of joint centers, and estimated body mass distribution based on anthropometry regression equations, or obtains these values ​​through simple measurement of the relative distances between skeletal points in images.

[0071] The standard data storage subunit is used to store pre-collected standard musculoskeletal dynamics datasets of common fitness movements. The datasets include kinematic data of common fitness movements such as squats, deadlifts, lunges, bench presses, and bent-over rows, as well as corresponding gold standard data of multidimensional torques of the hip, knee, ankle, shoulder, elbow, and wrist joints. The data comes from the motion collection of professional athletes and the simulation calculation results of the OpenSim musculoskeletal model.

[0072] The data acquisition module also integrates data preprocessing functions, including Gaussian filtering for noise reduction, background segmentation, and brightness normalization preprocessing for motion image data; Butterworth low-pass filtering, Z-score normalization, and temporal alignment preprocessing for joint angle and torque data; and standardized encoding of user musculoskeletal geometric parameters and physiological data.

[0073] Action recognition and stage segmentation module: The motion recognition and stage segmentation module is connected to the data acquisition module and is used to recognize motion data and segment the motion execution stages. This module includes three sub-units: The skeletal point recognition subunit uses the MediaPipe BlazePose lightweight pose estimation algorithm to perform skeletal point recognition on the preprocessed motion image data, locate key body nodes of the user's shoulder, elbow, wrist, hip, knee, and ankle, and generate three-dimensional spatial coordinate temporal data of each node in consecutive frames.

[0074] The action type recognition subunit calculates the angle time-series data of key joints such as hip, knee, ankle, shoulder, elbow, and wrist based on the three-dimensional spatial coordinate time-series data of key nodes. It then uses the Dynamic Time Warping (DTW) algorithm to match the user's joint angle time-series data with the action standard parameter set in the standard action library, calculates the time-series similarity, identifies the type of fitness action currently being performed by the user, and generates the action type recognition result.

[0075] The movement phase is divided into sub-units. The movement execution phase is automatically divided according to the key joint angle threshold and the change of angular velocity direction. For various fitness movements such as squat, deadlift, lunge, bench press and bent row, the movement cycle is divided into eccentric phase, transition phase and concentric phase according to the biomechanical inflection point, generating a movement phase sequence and marking the phase to which each moment belongs.

[0076] Joint torque prediction module: The joint torque prediction module is connected to both the data acquisition module and the motion recognition and stage segmentation module. It constructs a motion-stage-specific cooperative pattern dictionary, generates time-varying cooperative constraints through dynamic invocation, and trains a joint torque prediction model based on a combined loss function that incorporates inverse dynamic physical consistency terms. The module outputs real-time torque prediction results for multiple joints. This module comprises four sub-units: The co-operational pattern dictionary construction subunit groups the standard musculoskeletal dynamics dataset by action type and action execution stage, performs non-negative matrix decomposition on the multi-joint torque time series data of each group, determines the number of co-operational patterns for each action and each stage based on the interpretable variance ratio, and constructs a co-operational pattern dictionary with a hierarchical index structure.

[0077] The soft-gated dynamic calling sub-unit calculates the membership function value of the current moment relative to each action stage according to the action stage sequence. The membership degree is determined by Gaussian mixture model or fuzzy logic based on the key joint angle and angular velocity. The cooperative mode matrix of each stage is weighted and fused to generate a time-varying cooperative constraint matrix. The time-varying cooperative constraint term is calculated as the reconstruction residual L2 norm of the predicted torque relative to the time-varying cooperative constraint matrix in the cooperative subspace.

[0078] The physical consistency verification subunit constructs a simplified inverse dynamics model of the lower limb closed kinetic chain based on the user's musculoskeletal geometric parameters, calculates the theoretical torque of each joint, introduces the lower limb sagittal plane torque balance constraint, and generates inverse dynamics physical consistency terms.

[0079] The model training and inference subunit takes temporal data of joint angles and user musculoskeletal geometric parameters as input and gold standard data of joint torques as output to construct and train a joint torque prediction model based on a spatiotemporal graph neural network. The spatiotemporal graph neural network constructs a graph structure based on the topology of the human skeleton, with nodes corresponding to hip, knee, ankle, shoulder, elbow, and wrist joints and edges corresponding to bone connection segments. The network layers include spatiotemporal graph convolutional layers and gated recurrent layers. The trained model is deployed on a smart terminal to receive temporal data of key joint angles and musculoskeletal geometric parameters from the user and output real-time torque prediction results of multiple joints during the execution of user actions.

[0080] Assessment and Risk Classification Module: The assessment and risk grading module is connected to the joint torque prediction module. It is used to calculate joint biomechanical load parameters, generate movement accuracy scores and individualized injury risk assessment values, and stratify risk levels. This module includes four sub-units: The biomechanical load calculation subunit estimates the tension of major muscle groups based on real-time torque prediction results and user musculoskeletal geometry parameters using a simplified muscle force line model. It calculates joint contact force and soft tissue stress distribution based on the relationship between muscle tension and joint geometry, and establishes a cumulative damage factor to quantify the time-cumulative effect of overload stress.

[0081] The dynamic safety threshold construction subunit, based on joint contact force, soft tissue stress distribution and real-time torque prediction results, combined with user physiological data and past injury history, constructs a user-specific dynamic joint torque safety threshold, and corrects the threshold for users with a past injury history.

[0082] The motion quality assessment subunit maps the joint angle time-series data of the standard motion template and the user motion data to the same embedding space. It learns the motion quality representation through a dual-branch encoder, using the standard motion embedding as the anchor point and the user motion embedding as the positive or negative sample. It uses a triplet loss function for optimization, so that the motion correctness score has a monotonic mapping relationship with the distance of the embedding space.

[0083] The injury risk assessment subunit uses a Transformer-based temporal encoder to construct a deep correlation model. It takes the temporal sequence of joint contact force, soft tissue stress, movement angle deviation, and user physiological risk factors as inputs. It captures the cross-temporal dependencies of multi-source features through a multi-head self-attention mechanism, outputs individualized injury risk assessment values, and classifies them into low, medium, and high risk levels according to preset risk thresholds.

[0084] Early warning and guidance generation module: The early warning and guidance generation module is connected to the assessment and risk grading module to execute differentiated early warning strategies and generate personalized action correction guidance data. This module includes three sub-units: The tiered early warning subunit executes differentiated real-time early warning strategies based on the risk classification results: for low-risk results, a regular prompting strategy is executed, providing progressive action optimization suggestions; for medium-risk results, an enhanced prompting strategy is executed, providing real-time corrective reminders through voice and visual annotations; for high-risk results, a mandatory early warning strategy is executed, through voice prompts to force a stop, visual red markings as warnings, and highlighting of stress concentration areas. If the user continues to perform high-risk actions, mandatory early warnings will continue to be issued until the user stops performing the action.

[0085] Personalized guidance generation sub-units combine movement angle deviation characteristics, torque abnormality characteristics, and soft tissue stress distribution characteristics to generate stratified personalized movement correction guidance and rehabilitation training suggestions data; for high-risk classification results, stress concentration areas are identified based on soft tissue stress distribution characteristics to generate targeted local muscle strengthening and joint stability training programs.

[0086] The data recording and closed-loop optimization subunit synchronously records the angle deviation, peak torque, contact force, action correctness score, and cumulative injury factor of each user action, generating progress tracking data for comparison with historical training data; it also records the user's behavioral feedback after receiving guidance, dynamically adjusts the focus of the guidance content and fine-tunes the dynamic safety threshold based on the user's subsequent action correction in training, forming a closed-loop optimization mechanism.

[0087] The five modules are connected sequentially. The output of the data acquisition module is connected to the input of the motion recognition and stage division module; the output of the motion recognition and stage division module is connected to the input of the joint torque prediction module; the output of the joint torque prediction module is connected to the input of the assessment and risk grading module; and the output of the assessment and risk grading module is connected to the input of the early warning and guidance generation module, forming a complete processing chain from data collection to risk early warning and guidance. The system is deployed on smart terminal devices to meet real-time processing requirements.

[0088] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception, characterized in that, Includes the following steps: S1: Acquire motion data of user fitness movements and user-specific physiological data, and at the same time acquire standard musculoskeletal dynamics dataset; S2: Perform motion recognition on the motion data of the user's fitness movements, determine the type of fitness movement performed by the user, divide the movement execution stages according to kinematic characteristics, and generate motion type recognition results and motion stage sequences; S3: Based on the standard musculoskeletal dynamics dataset, construct an action-stage specific cooperative pattern dictionary according to action type and action execution stage; dynamically call the corresponding stage cooperative pattern matrix according to the action type identification result and action stage sequence to generate time-varying cooperative constraint terms; construct a combined loss function of fusion data fitting term, the time-varying cooperative constraint term and inverse dynamic physical consistency term, train the joint torque prediction model, and output the real-time torque prediction results of multiple joints during user action execution; S4: Based on the real-time torque prediction results and the user's individualized physiological data, calculate the joint biomechanical load parameters, construct the user's personalized dynamic safety threshold, generate the action correctness score and individualized injury risk assessment value, and perform risk level stratification according to the individualized injury risk assessment value to generate risk grading results. S5: Based on the risk classification results, implement differentiated real-time early warning strategies and generate personalized action correction guidance data.

2. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that, The motion data of the user's fitness movements in step S1 includes motion image data of the user's fitness movements; the user's individualized physiological data includes the user's basic health physiological data and the user's musculoskeletal geometric parameters; the musculoskeletal geometric parameters include at least the lower limb segment length, the relative position of the joint center, and the estimated value of body mass distribution.

3. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that, The step S2, which divides the movement execution stages according to kinematic characteristics, includes: for various fitness movements such as squats, deadlifts, lunges, bench presses, and bent-over rows, the movement cycle is divided into at least three stages based on biomechanical inflection points: eccentric stage, transition stage, and concentric stage; the division is based on the key joint angle thresholds and changes in angular velocity direction.

4. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that: The step S3 of constructing the action-stage specific cooperative pattern dictionary includes: independently performing nonnegative matrix decomposition on the multi-joint torque time-series data within each action and stage. ,in Indicates the action type index. Indicates the stage index. For action stage The collaborative mode matrix, This is the corresponding activation coefficient matrix; The number of cooperative patterns for each action and each stage is determined based on the interpretable variance ratio, and a cooperative pattern dictionary with a hierarchical index structure is constructed. .

5. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that: The dynamic invocation of the corresponding stage's collaborative mode matrix in step S3 includes: Based on the joint angle time series data extracted in step S2, calculate the current time. Membership function values ​​relative to each action phase The membership function value The key joint angles and angular velocities are determined using a Gaussian mixture model or fuzzy logic. Generate time-varying cooperative constraint matrix ,in ; The time-varying cooperative constraint term To predict the torque relative to the time-varying cooperative constraint matrix The L2 norm of the reconstruction residual in the collaborative subspace is calculated as follows: ； Where: is time. The predicted torque vector, These are the optimal reconstruction coefficients.

6. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that: The calculation method for the inverse dynamics physical consistency term in step S3 is as follows: A simplified inverse dynamics model of the lower limb closed kinetic chain is constructed based on user musculoskeletal geometry parameters, and the theoretical torque of each joint is calculated based on joint angle time-series data. ; Calculate the predicted torque With theoretical torque The residuals are determined, and a lower limb sagittal plane moment balance constraint is introduced: the hip joint moment, knee joint moment and ankle joint moment satisfy a closed chain transmission relationship; The inverse dynamic physical consistency term is: ； Where: represents the torque balance residual term of the closed chain. This is the balance coefficient.

7. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that: In step S3, the joint torque prediction model is a spatiotemporal graph neural network. The spatiotemporal graph neural network constructs a graph structure based on the topology of the human skeleton. The nodes correspond to the hip, knee, ankle, shoulder, elbow, and wrist joints, and the edges correspond to the bone connection segments. The network layer includes a spatiotemporal graph convolutional layer and a gated recurrent layer. The spatiotemporal graph convolutional layer is used to aggregate the temporal features of biomechanical adjacent joints, and the gated recurrent layer is used to capture the temporal dependencies of the action phase.

8. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that: The biomechanical load parameters of the joint in step S4 include joint contact force and soft tissue stress distribution. Calculating joint contact force and soft tissue stress distribution includes: estimating the tension of major muscle groups using a simplified muscle force line model based on user musculoskeletal geometry parameters and real-time torque prediction results; and calculating the joint contact force based on the relationship between the muscle group tension and joint geometry. Tensile stress on ligaments / tendons ; Establish cumulative damage factors: ； in: For dynamic security thresholds, Used to quantify the time-cumulative effect of overload stress.

9. The method for intelligent assessment and injury risk warning of fitness movements based on motion phase perception according to claim 1, characterized in that: The generation of the action correctness score in step S4 includes: mapping the joint angle temporal data of the standard action template and the user action data to the same embedding space, learning the action quality representation through a dual-branch encoder, using the standard action embedding as the anchor point and the user action embedding as a positive or negative sample, and optimizing with a triplet loss function to make the action correctness score monotonically mapped to the distance in the embedding space; the individualized injury risk assessment value is generated through a deep association model, which uses a Transformer-based temporal encoder, taking the joint contact force temporal sequence, soft tissue stress temporal sequence, action angle deviation temporal sequence and user physiological risk factors as inputs, and capturing the cross-temporal dependencies of multi-source features through a multi-head self-attention mechanism to output the individualized injury risk assessment value.

10. A fitness movement intelligent assessment and injury risk early warning system based on movement phase perception, characterized in that, The system is used to perform the intelligent assessment and injury risk warning method for fitness movements based on motion phase perception as described in any one of claims 1-9, the system comprising: The data acquisition module is used to acquire motion data of user fitness movements, user-specific physiological data, and standard musculoskeletal dynamics datasets; The motion recognition and stage segmentation module is used to recognize motion data and segment the motion execution stages. The joint torque prediction module is used to construct a motion-stage specific cooperative pattern dictionary, generate time-varying cooperative constraints by dynamically calling them, and train the joint torque prediction model based on the combined loss function that integrates inverse dynamic physical consistency terms, and output the real-time torque prediction results of multiple joints. The assessment and risk grading module is used to calculate joint biomechanical load parameters, generate movement correctness scores and individualized injury risk assessment values, and stratify risk levels. The early warning and guidance generation module is used to execute differentiated early warning strategies and generate personalized action correction guidance data.

Images