Badminton action evaluation training method and system based on power chain multi-modal timing

Abstract

The application provides a badminton action evaluation training method and system based on a power chain multimodal time sequence, and belongs to the field of human-computer interaction and intelligent training assistance. The method comprises the following steps: collecting multimodal motion data of a measured object during the execution of a badminton hitting action; and pre-processing and time aligning the multimodal time sequence data to obtain an action time sequence segment with a fixed time length. The system comprises a multimodal data collection module, a power chain feature learning module, an action evaluation module and a training guidance module. The application aims to solve the problems of low accuracy of the existing hitting action evaluation method and the inability to quantitatively explain the key force link and time sequence deviation and form targeted training guidance.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction and intelligent training assistance, and in particular to a badminton action evaluation and training method and system based on kinetic chain multimodal temporal sequence. Background Technology

[0002] Badminton is one of the most widely played sports globally. Due to its combination of high explosive power and precise hand control, motion analysis faces significant challenges. Traditional motion analysis systems primarily rely on the subjective observation of coaches. This method is limited by the coach's personal experience and struggles to capture the rapidly changing biomechanical characteristics during high-speed movements, such as subtle adjustments in racket angle at the moment of impact, hidden power generation, and millisecond-level footwork reactions. Furthermore, the scarcity and high cost of coaching resources prevent many athletes, especially beginners, from receiving timely and personalized feedback, leading to the accumulation of technical errors and an increased risk of injury.

[0003] With the continuous development of human motion sensing technology, human posture estimation and wearable sensor technologies have gradually matured. Compared to vision-based motion analysis methods, wearable sensors can be directly attached to the surface of the human body or sports equipment, continuously collecting kinematic and physiological signals during human movement. They are also less affected by changes in lighting conditions, occlusion, and shooting angles, thus offering significant advantages in high-frequency motion information acquisition and internal biomechanical feature acquisition. Therefore, researching methods for objectively evaluating motion quality and providing personalized training guidance based on wearable sensor data, focusing on basic badminton hitting motions, has become a crucial research area.

[0004] In recent years, deep learning technology has been increasingly applied to badminton motion analysis and is considered to have significant application potential. Existing methods typically rely on multi-source wearable sensor data, utilizing deep models to automatically learn and evaluate motion characteristics during the sport. Compared to rule-based or manually generated methods, this approach can more effectively characterize complex dynamic motion properties. In specific applications, some techniques improve the ability to identify differences in hitting motions and skill levels by incorporating angular velocity, electromyography (EMG) signals, or pressure distribution information. Other methods model key features of hitting motions by analyzing muscle activation timing or upper and lower limb coordination. Beyond motion quality assessment, deep learning is also used in badminton training instruction. Some existing systems combine wearable sensors to visualize force patterns during exercise or provide feedback to athletes through touch and vision. Other methods identify motion deviations and provide improvement suggestions by comparing user actions with pre-established expert models. Furthermore, some technologies attempt to construct ideal or modified reference movements using generative models to assist in training instruction.

[0005] Despite the significant advancements made by deep learning in badminton motion assessment and training guidance, the following challenges remain: First, existing models struggle to accurately model the force transmission process during a shot. While deep learning can extract macroscopic features of motion, its ability to capture the dynamic details of force transmission during a shot remains insufficient, resulting in a lack of precise characterization in motion evaluation.

[0006] Second, badminton movements have a strong temporal dependency, but existing models often fail to adequately consider these temporal relationships. Ignoring the temporal sequence of movements leads to insufficient accuracy in capturing motion details and dynamic changes, affecting the accuracy of movement assessment and training guidance.

[0007] Third, existing deep learning models face challenges in terms of interpretability, especially in how to provide personalized feedback to athletes and provide clear evidence in the spatial and temporal dimensions of movement, which remains an urgent problem to be solved.

[0008] Therefore, there is an urgent need for a badminton motion evaluation method and system that can integrate multimodal motion data, has an explicit kinetic chain diagram structure and temporal modeling capability, and can output interpretable bias and personalized training feedback. Summary of the Invention

[0009] This invention provides a badminton motion evaluation and training method and system based on kinetic chain multimodal timing to solve the above-mentioned problems in the prior art. The badminton motion evaluation and training method based on kinetic chain multimodal timing includes the following steps: S1: Collect multimodal motion data of the test subject during the badminton hitting action. The multimodal motion data includes at least one or more of joint posture data, electromyographic signal data, and plantar pressure data. The multimodal motion data is synchronously recorded through a unified time reference to form multimodal time-series data corresponding to the hitting action. S2: Preprocess and time-align the multimodal timing data to obtain action timing segments of fixed length; S3: Based on the power transmission relationship in the human hitting action, the action sequence segment is modeled into a kinetic chain, which is divided into multiple kinetic chain nodes arranged in the order of force exertion from bottom to top. The corresponding multimodal feature data of each kinetic chain node is extracted, and a unified dimension kinetic chain node feature representation is obtained through node feature mapping. S4: Construct a chain-like power chain graph structure that reflects the power transmission sequence of the human body based on the power chain nodes, and input the chain-like power chain graph into a chain graph convolutional neural network for spatial feature extraction to obtain a power chain spatial feature representation; S5: Input the spatial feature representation of the kinetic chain into the temporal modeling network, model the temporal features of the hitting action through a self-attention mechanism, and output the action-level feature representation that characterizes the overall temporal features of the hitting action; S6: Input the action-level feature representation into the classification and evaluation network, and output the action quality evaluation result corresponding to the hitting action; S7: Perform interpretability analysis on the motion quality assessment results to obtain the contribution distribution of the power chain node dimension and time dimension; S8: Based on the contribution distribution, perform deviation diagnosis on the hitting action to be evaluated and generate corresponding training guidance information.

[0010] Preferably, the multimodal motion data includes at least two of the following: joint posture data, upper limb electromyography (EMG) signal data, lower limb EMG signal data, and plantar pressure data.

[0011] Preferably, step S2 includes: deduplicating and removing anomalies from the time information of each modality data; constructing a unified target time axis and resampling each modality data; performing noise suppression processing on each modality data respectively; performing numerical normalization on each processed modality data; and clipping the action interval of the multimodal time series data based on the start and end annotation information of the hitting action.

[0012] Preferably, the kinetic chain nodes include at least the plantar support node, the lower limb muscle group node, the core trunk node, the upper arm node, the forearm node, and the wrist end node.

[0013] Preferably, node feature mapping uses linear mapping and nonlinear activation functions to map node features of different dimensions into node feature vectors of a unified dimension.

[0014] Preferably, the chain-based graph convolutional neural network performs layer-by-layer propagation and aggregation of the features of the kinetic chain nodes through multi-layer graph convolution operations to extract the spatial correlation features between the kinetic chain nodes.

[0015] Preferably, the temporal modeling network is a self-attention mechanism-based temporal modeling network, which models the temporal features of the hitting action by establishing the correlation between different time frames.

[0016] Preferably, the action-level feature representation is composed of a global temporal feature vector output by the temporal modeling network, which is used to characterize the overall temporal characteristics of the hitting action.

[0017] Preferably, the interpretability analysis in step S7 includes: determining the output value corresponding to the target evaluation category as the interpretable target; calculating the contribution of the input features to the interpretable target; and propagating the contribution layer by layer along the model network hierarchy through a hierarchical correlation propagation method to obtain the contribution distribution of the power chain node dimension and the time dimension.

[0018] Preferably, the deviation diagnosis in step S8 includes: comparing the kinetic chain node features of the action to be evaluated with the pre-established standard action feature data, calculating the degree of deviation of each kinetic chain node; and marking the corresponding abnormal node when the deviation exceeds a preset threshold.

[0019] Preferably, the deviation diagnosis further includes: performing a combined analysis of the power chain nodes according to the preset power chain structure, calculating the chain-level deviation, and outputting chain-level abnormal information when the chain-level deviation exceeds a preset threshold.

[0020] Preferably, the deviation diagnosis further includes: in the time dimension, performing time-series deviation analysis on key time phases with high contribution, and outputting abnormal information related to the force exertion sequence or action rhythm.

[0021] Preferably, the training guidance information is generated based on abnormal nodes, abnormal chains, and time-related abnormal results, and is used to guide the tested object to perform targeted action training.

[0022] This invention provides a badminton action evaluation and training system based on kinetic chain multimodal timing, which uses any of the badminton action evaluation and training methods based on kinetic chain multimodal timing described above.

[0023] Preferably, the badminton action assessment and training system based on kinetic chain multimodal temporal sequence includes the following modules: a multimodal data acquisition module: synchronously acquiring athlete joint posture, electromyographic signals, and plantar pressure data; a kinetic chain feature learning module: constructing deep features of the kinetic chain of the hitting action based on chain-based topological graph convolution and self-attention temporal network; an action assessment module: automatically assessing the athlete's current action level; and a training guidance module: generating personalized training guidance based on high-level benchmark models and interpretable analysis results, combined with a language model.

[0024] The present invention has the following beneficial effects: This invention provides a badminton action evaluation and training method and system based on multimodal temporal sequence of kinetic chain. Based on the concept of human kinetic chain modeling, the basic badminton hitting action is represented as a chain diagram structure with a clear force transmission path. Combined with the temporal modeling process using multimodal sensor data, the method achieves a joint characterization of key force-generating components and their temporal characteristics in the hitting action. This allows for accurate evaluation of the athlete's action quality and identification of deviations during action execution, generating targeted training guidance information. This effectively overcomes the problems of insufficient accuracy in hitting action evaluation, difficulty in quantifying and analyzing key force-generating components, and lack of targeted training feedback in existing technologies. Attached Figure Description

[0025] Figure 1This is a flowchart of the badminton motion evaluation and training method and system based on kinetic chain multimodal timing of the present invention; Figure 2 This is a schematic diagram of the kinetic chain structure of the key force-generating nodes in the human body according to the present invention; Figure 3 This is a schematic diagram of the architecture of the dynamic chain diagram structure and multimodal time series analysis model described in this invention; Figure 4 This is a user interface demonstration diagram of the badminton basic movement assessment and training guidance system of the present invention.

[0026] Among them: 1, node 1; 2, node 2; 3, node 3; 4, node 4; 5, node 5; 6, node 6; 7, node 7. Detailed Implementation

[0027] The following detailed description, with reference to specific embodiments, illustrates a method for evaluating and guiding the quality of badminton hitting actions based on a kinetic chain graph structure and multimodal temporal analysis, as proposed in this invention. Those skilled in the art should understand that equivalent substitutions for specific parameters or implementation methods made without departing from the overall concept of this invention should fall within the scope of protection of this invention.

[0028] According to one embodiment of this application, such as Figure 2 As shown, node 1 is the wrist; node 2 is the forearm; node 3 is the upper arm; node 4 is the dominant leg; node 5 is the right foot; node 6 is the left foot; and node 7 is the hip, torso, and shoulder.

[0029] In this embodiment, the method is designed for badminton hitting actions (such as the forehand high clear). It follows a process of multimodal acquisition, preprocessing alignment, kinetic chain node modeling, chain graph convolutional space modeling, temporal Transformer modeling, classification evaluation, LRP interpretability analysis, and benchmark comparison to output the hitting action quality evaluation results, the kinetic chain node and temporal dimension contribution distribution, and generate training guidance information accordingly.

[0030] S1: Collect multimodal motion data of the tested object during the badminton hitting action.

[0031] S11: The subject wears an inertial motion capture device to collect joint posture data. In this embodiment, the PerceptionNeuronStudio motion capture kit, containing 17 IMUs, can be used to output the local Euler angles of each joint (3-dimensional angles per joint), with a sampling frequency set to 62.4 Hz. The joint posture data is used to characterize the changes in upper limb and trunk posture during the hitting action.

[0032] S12: Acquire upper limb electromyography (EMG) signal data. In this embodiment, gForcePro+ EMG armbands are worn on the forearm and upper arm on the racket side to acquire 8-channel dry electrode EMG data; the sampling frequency for the forearm can be set to 76Hz, and the sampling frequency for the upper arm can be set to 162Hz, with Unix timestamp writing enabled. The upper limb EMG signals are used to characterize the intensity and timing of upper limb muscle activation during the hitting process.

[0033] S13: Acquire lower limb electromyography (EMG) signal data. In this embodiment, EMG electrodes are attached to the rectus femoris, vastus medialis, vastus lateralis, and biceps femoris muscles of the supporting leg, and connected to the CGXAIM host to acquire 4-channel lower limb EMG data at a sampling frequency of approximately 502 Hz, with timestamp synchronization enabled. The lower limb EMG signals are used to characterize muscle group activity during the lower limb exertion and support phases.

[0034] S14: Collect plantar pressure data. In this embodiment, left and right MotionOpenGo smart insoles are used, outputting 16-dimensional pressure sensing unit data and COP trajectory for each insole, with a sampling frequency set to 53Hz. The plantar pressure data is used to characterize the support phase and center of gravity transfer features.

[0035] S15: Synchronous Recording. Each device simultaneously writes sensor data with a Unix timestamp upon starting recording, forming continuous multimodal time-series data covering the entire process of preparation, backswing, acceleration, hitting, and follow-through. The unified timestamp is used to ensure the alignment and comparability of different modal data in the time dimension.

[0036] S2: Preprocess and time-align the multimodal raw time series data collected in step S1.

[0037] S21: Timestamp deduplication and anomaly removal. For each modal time series... Perform deduplication: ;in, This is used to remove duplicate timestamps and return the corresponding index idx, where x' is the deduplicated feature sequence. Furthermore, abnormal sampling points with abnormal time intervals are removed to ensure the time axis is monotonically continuous, thus facilitating subsequent resampling and interpolation calculations.

[0038] S22: Construct a unified target time axis and resample. Estimate the actual sampling rate for each time series. , and in the action range [ , Internally constructed equally spaced target time axis: ; In this embodiment, the resampling rate is... Action window Therefore, N=150 frames. The target timeline is used to form a unified set of sampling points. To achieve multimodal timing alignment.

[0039] Then, monotonic interpolation (e.g., piecewise linear interpolation) is applied to each mode to obtain the resampled sequence: ; in Used to map the feature x' at the original sampling point t' to the target time axis This allows for the acquisition of synchronized multimodal sequences.

[0040] S23: Noise suppression processing.

[0041] S231: Upper limb gForce EMG. The absolute value of the raw data was taken to form the envelope, and a 4th-order Butterworth low-pass filter was applied with a cutoff frequency of 5Hz. ;in This represents the Butterworth low-pass filter operator. The cutoff frequency, The sampling rate is denoted as . The processing is used to suppress high-frequency noise and extract the electromyographic envelope trend.

[0042] S232: Lower limb CGXEMG. First, take the absolute value; then, remove artifacts according to the jump condition (e.g., differential threshold difff=80, interval time=5), and set outliers to NaN before median interpolation; subsequently, use a low-pass filter with a cutoff frequency of 20Hz. The processing is used to remove motion artifacts and spike interference, and to maintain the smoothness and comparability of electromyographic activation sequences.

[0043] S233: Plantar pressure. The absolute value is then processed using a low-pass filter with a cutoff frequency of 5Hz to suppress high-frequency artifacts from sole vibration. This processing is used to obtain a stable support pressure variation curve.

[0044] S24: Normalization. Scale different modalities according to their physical dimensions and empirical ranges: for example, upper limb EMG is linearly scaled and translated according to the empirical amplitude range [0, 300], lower limb EMG is normalized and translated using half-range normalization, and plantar pressure is normalized and translated using global range normalization; joint Euler angles are linearly scaled from their domain [-180, 180] to [-1, 1]. The normalization is used to eliminate the differences in the dimensions of different modalities and improve the numerical stability of network training and inference.

[0045] S25: Action range clipping. Read the external annotations (start, stop), and take [[] on each modality. , The intervals are resampled to a fixed frame length of 150 to obtain a fixed-length action sequence segment. The cropping is used to ensure that different samples are consistent in the time length dimension.

[0046] S26: Multimodal concatenation. Concatenates modal segments from the same trial along the channel dimension to form a multimodal input tensor. Where T=150. The tensor X is used for subsequent feature extraction and structured modeling of the kinetic chain nodes.

[0047] S3: Based on the power transmission relationship in the human hitting action, the timing segment of the hitting action is divided into multiple kinetic chain nodes.

[0048] S31: Kinetic Chain Node Division. Based on the bottom-up force application sequence, the hitting action is divided into chain-like kinetic chain nodes, including at least: foot support node, lower limb muscle group node, core trunk node, upper arm node, forearm node, and wrist distal node; in this embodiment, it is further refined into a seven-node chain topology (consistent with the illustration). The chain topology is used to characterize the sequential relationship of power transmission from the lower limbs through the trunk to the upper limb distal nodes.

[0049] S32: Node Feature Extraction. Let the multimodal input be... For the channel range corresponding to the v-th node Child selection characteristics: ;in The eigenvector representing time step t, Indicates in the channel set Subvectors on the kinetic chain. The node features are used to characterize the motion and physiological state of each kinetic chain node in the time dimension.

[0050] S33: Node Feature Mapping (Uniform Dimension). Since the original dimensions of each node are inconsistent, a linear embedding is applied to each node, mapping it to a uniform embedding dimension. =64: ;in and These are the learnable parameters of the node feature mapping layer. The activation function is non-linear, with GELU being the preferred choice. The node feature mapping is implemented by node_tokenizer, which is used to unify heterogeneous modal features into a comparable representation space.

[0051] S34: Form the node temporal tensor. Embed the nodes and stack them in node order: ; Therefore, we get: The tensor H is used as an input feature for the graph convolutional network.

[0052] S4: Based on the multiple power chain nodes, construct a chain-like power chain diagram structure that reflects the power transmission sequence of the human body.

[0053] S41: Construct a linked adjacency matrix. Construct a seven-node linked topological adjacency matrix. The propagation matrix is ​​obtained by setting adjacent node connections to 1 and all others to 0, and then performing symmetric normalization. Where I is the identity matrix and D is the degree matrix. Used to implement neighborhood propagation and normalized aggregation in graph convolution.

[0054] S42: Chained GCN spatial feature extraction. For each time frame t, perform graph convolutional propagation: ;in , This is the convolution weight matrix for the first layer graph. The activation function is nonlinear, with GELU being the preferred choice. The propagation is used to aggregate information from adjacent kinetic chain nodes layer by layer and extract spatial correlation features.

[0055] S43: Network layer and dimension settings. In this embodiment, the chain graph convolutional layer has 2 layers, and the feature dimension of each layer is 64. Dropout is introduced after each layer with a dropout rate of 0.1 for regularization, which is used to reduce the risk of overfitting and improve generalization ability.

[0056] S44: Obtain the spatial feature representation of the kinetic chain. After two layers of propagation, obtain the spatially enhanced features: ; And form: The S is used as the input to the time series modeling network.

[0057] S5: Input the spatial feature representation of the kinetic chain into the trained temporal Transformer network.

[0058] S51: Frame-level aggregation. The node features of each frame are pooled along the node dimension (e.g., average pooling) to obtain a frame-level representation: The pooling is used to convert the node-level spatial representation into a frame-level temporal token representation to adapt to the sequence input format of the Transformer.

[0059] S52: Linear Projection and Position Encoding. Projecting g(t) onto the Transformer input dimension. =128, and superimposed with sine position code (maximum length 500): ; in , Here, p(t) represents the linear projection parameter, and p(t) represents the position code. The position code is used to introduce temporal sequence information.

[0060] S53: Introduce CLS notation. Learnable CLS vectors... The CLS markers are concatenated to the beginning of the sequence to obtain an input sequence of length T+1. The CLS markers are used to aggregate global timing information and form an action-level representation.

[0061] S54: Self-attention calculation. For each Transformer layer, multi-head self-attention is calculated: ; in: The self-attention is used to establish dependencies between different time frames, thereby extracting action rhythm and temporal structure features.

[0062] S55: Network layer and head count settings. In this embodiment, the Transformer encoder has 2 layers, each with 128 hidden dimensions, and 4 attention heads. These settings aim to achieve a balance between computational complexity and modeling capability.

[0063] S56: Output action-level temporal feature representation. The CLS vector of the final layer output sequence is taken as the action-level feature. The action-level feature h is used as input to the classification evaluation network.

[0064] S6: Input the action-level feature representation into the trained multilayer perceptron classification network, and output the action quality evaluation result corresponding to the hitting action.

[0065] S61: Input the action-level feature h into the multilayer perceptron classifier. The classifier head contains multiple fully connected layers, each with 64 neurons, using GELU as the activation function and Dropout (0.1) for nonlinear discriminative mapping of the action-level feature.

[0066] S62: The output layer dimension is set to 3, corresponding to the scores of the 3 technical skill categories. The category scores are used to characterize the discrimination strength of different quality levels.

[0067] S63: Softmax Probability and Prediction Categories: Where pc is the predicted probability of category c. The output category is defined. The Softmax function is used to normalize the category scores into a probability distribution.

[0068] S64: Training Loss: ;in This represents the one-hot encoding of the true label. The loss function is used to guide model parameter training and improve classification accuracy.

[0069] S7: Perform interpretability analysis on the motion quality evaluation results output in step S6 to obtain the contribution distribution of the power chain node dimension and time dimension.

[0070] S71: Determine the interpretation target. Select the output value corresponding to the target evaluation category as the interpretation target, such as the component corresponding to the target category index c* in logits, which will serve as the starting point for relevance backpropagation.

[0071] S72: Gradient and Input Contribution. The gradient of the input feature x is calculated and multiplied by the input to obtain the local contribution. ;in This is the output function for the target category. The contribution value is used to quantify the direction and intensity of the influence of each input feature on the target output.

[0072] S73: Layer-by-layer propagation (LRP). The output z of a certain layer is stabilized (by introducing a small constant ε): And calculate the correlation ratio coefficient: The correlation is backpropagated layer by layer along the network hierarchy of the evaluation model to obtain the correlation value of each node at each time step. The backpropagation is used to allocate the output correlation to the input space, realizing the interpretation of nodes and the time dimension.

[0073] S74: Calculation of Time-Dimensional Contribution. The total time-step contribution is obtained by summing the cls_relevance output of the Transformer at each time step, and the relative contribution coefficient is calculated. ;in This represents the CLS correlation contribution at time step t. The αt is used to obtain the time-dimensional contribution distribution.

[0074] S75: Aggregation of node-dimensional, time-dimensional, and two-dimensional matrix. The correlation values ​​are aggregated by node and time to obtain the contribution of each power chain node, the contribution of each time step, and the node-time two-dimensional contribution matrix, which serve as the basis for subsequent deviation localization and diagnosis.

[0075] S8: Based on the contribution distribution obtained in step S7, perform deviation diagnosis on the action to be evaluated and generate training guidance information.

[0076] S81: Obtain standard movement feature data. The standard movement feature data is constructed from a sample of high-level athletes and includes reference statistical characteristics (mean, standard deviation, etc.) of each kinetic chain node at each time step, which is used as a diagnostic comparison benchmark.

[0077] S82: Node Deviation Calculation and Abnormal Node Marking. The Z-score is calculated by comparing the characteristics of the action node to be evaluated with the benchmark mean and standard deviation. ; in and These represent the mean and standard deviation of the standard action at node v and time step t, respectively. When When its aggregate value exceeds a preset threshold (e.g., 2.0), the corresponding abnormal node is marked to locate the source of the key deviation.

[0078] S83: Chain Deviation Score. The chain deviation score is obtained by averaging the Z-scores of the nodes within the chain, organizing the nodes into at least one of the following: lower limb chain, core chain, and upper limb chain. Where N is the number of nodes in the chain. This represents the aggregate deviation of node v within the critical time period. If the chain deviation score exceeds a preset threshold, a chain-level anomaly is output to indicate an overall problem in the power transmission link.

[0079] S84: Timing Deviation in Critical Time Phases. Based on the time contribution distribution obtained in step S7, select the critical phases with higher contributions and calculate the time difference between the action to be evaluated and the standard action at critical events (e.g., peak moments): ; when When the time exceeds a preset threshold (e.g., 0.1s), output timing anomalies related to rhythm or force application sequence to indicate deviations in movement rhythm.

[0080] S85: Training guidance information generation. Based on abnormal nodes, abnormal chains, and time anomalies, training guidance information is generated and output. The training guidance information may include key training areas corresponding to abnormal nodes, chain-level collaborative effort suggestions, and rhythm adjustment suggestions for key time stages. It can also be displayed in a visualization interface as a node heatmap, time contribution curve, and deviation location results corresponding to the contribution distribution.

[0081] According to one embodiment of this application, Figure 4 This is a user interface demonstration diagram of the badminton basic movement assessment and training guidance system of the present invention. The blank area in the lower right corner of the diagram displays information such as the current skill level, kinetic chain status, and kinetic chain segment analysis. Specific information is illustrated in the following examples: [Coach's Technical Feedback | Badminton Hitting Technique] 1. Current technological level and power chain situation The athlete is currently at the beginner level and has significant room for improvement in technique. There is insufficient power generation in the overall kinetic chain; core rotation and power transfer are not smooth enough, affecting hitting efficiency. There is a slight problem with the forearm whipping motion, which may affect shot control.

[0082] Areas requiring improvement include: hip rotation (slight deviation), shoulder roll (significant deviation), and shoulder joint rotation (large deviation). Overall footwork is stable, but weight transfer can be further optimized.

[0083] 2. Segmentation analysis of the kinetic chain (in sequence) • Footwork: Overall performance was good, with a slight deviation in foot height. Adjusting this will help improve control of the racket face angle when hitting the ball.

[0084] • Lower limbs (Leg): There is a significant discrepancy between the force exertion and initiation of the lower limbs, requiring increased involvement of leg drive and rotation.

[0085] • Core: There is a slight deviation, and the core is not fully involved, which affects the upward transmission of power.

[0086] • Upper Arm: Overall function is normal; minor adjustments can further improve the efficiency of force generation from proximal to distal ends.

[0087] • Forearm: There is a slight deviation in power generation. Targeted tips can improve the driving effect of the shot.

[0088] • Wrist: Normal, no obvious problems found.

[0089] 3. Training Recommendations 1. Core activation training • Key point: Hip and spine rotation • Objective: To complete 5–7 high and low ball hits while maintaining a stable racket face angle. 2. Forefoot control training • Key points: Foot placement and weight transfer • Goal: Complete 8–10 shots while maintaining footwork alignment. 3. Forearm strength and timing training • Key point: Forearm power combined with wrist movement • Goal: Complete 6–8 controlled wrist extension hits, reducing power lag.

Claims

1. A badminton motion evaluation and training method based on kinetic chain multimodal timing, characterized in that, Includes the following steps: S1: Collect multimodal motion data of the test subject during the badminton hitting action. The multimodal motion data includes at least one or more of joint posture data, electromyographic signal data, and plantar pressure data. The multimodal motion data is synchronously recorded through a unified time reference to form multimodal time-series data corresponding to the hitting action. S2: Preprocess and time-align the multimodal timing data to obtain action timing segments of fixed length; S3: Based on the power transmission relationship in the human hitting action, the action sequence segment is modeled into a kinetic chain, which is divided into multiple kinetic chain nodes arranged in the order of force exertion from bottom to top. The corresponding multimodal feature data of each kinetic chain node is extracted, and a unified dimension kinetic chain node feature representation is obtained through node feature mapping. S4: Construct a chain-like power chain graph structure that reflects the power transmission sequence of the human body based on the power chain nodes, and input the chain-like power chain graph into a chain graph convolutional neural network for spatial feature extraction to obtain a power chain spatial feature representation; S5: Input the spatial feature representation of the kinetic chain into the temporal modeling network, model the temporal features of the hitting action through a self-attention mechanism, and output the action-level feature representation that characterizes the overall temporal features of the hitting action; S6: Input the action-level feature representation into the classification and evaluation network, and output the action quality evaluation result corresponding to the hitting action; S7: Perform interpretability analysis on the motion quality assessment results to obtain the contribution distribution of the power chain node dimension and time dimension; S8: Based on the contribution distribution, perform deviation diagnosis on the hitting action to be evaluated and generate corresponding training guidance information.

2. The badminton motion evaluation and training method based on kinetic chain multimodal timing as described in claim 1, characterized in that, The multimodal motion data includes at least two of the following: joint posture data, upper limb electromyography (EMG) signal data, lower limb EMG signal data, and plantar pressure data.

3. The badminton motion evaluation and training method based on kinetic chain multimodal timing as described in claim 1, characterized in that, Step S2 includes: Deduplication and anomaly removal are performed on the temporal information of each modality data; Construct a unified target timeline and resample the data for each modality; Noise suppression processing was performed on each modal data separately; The processed modal data are numerically normalized. Action range clipping is performed on multimodal time series data based on the start and end annotation information of the hitting action.

4. The badminton motion evaluation and training method based on kinetic chain multimodal timing as described in claim 1, characterized in that, The kinetic chain nodes include at least the plantar support node, the lower limb muscle group node, the core trunk node, the upper arm node, the forearm node, and the wrist end node. The node feature mapping uses linear mapping and nonlinear activation functions to map node features of different dimensions into node feature vectors of a unified dimension. The chain-based graph convolutional neural network performs layer-by-layer propagation and aggregation of the features of the power chain nodes through multi-layer graph convolution operations to extract the spatial correlation features between the power chain nodes.

5. The badminton motion evaluation and training method based on kinetic chain multimodal timing as described in claim 1, characterized in that, The temporal modeling network is a self-attention mechanism-based temporal modeling network that models the temporal features of the hitting action by establishing the correlation between different time frames. The action-level feature representation is composed of a global temporal feature vector output by the temporal modeling network, which is used to characterize the overall temporal characteristics of the hitting action.

6. The badminton motion evaluation and training method based on kinetic chain multimodal timing as described in claim 1, characterized in that, The interpretability analysis in step S7 includes: Determine the output value corresponding to the target evaluation category as the explanatory target; Calculate the contribution of the input features to the explanatory objective; The contribution is propagated back along the model network hierarchy layer by layer using a hierarchical correlation propagation method to obtain the contribution distribution in the power chain node dimension and time dimension.

7. The badminton motion evaluation and training method based on kinetic chain multimodal timing as described in claim 1, characterized in that, The deviation diagnosis in step S8 includes: The characteristics of the kinetic chain nodes of the motion to be evaluated are compared with the pre-established standard motion characteristic data, and the degree of deviation of each kinetic chain node is calculated. When the deviation exceeds a preset threshold, the corresponding abnormal node is marked; The deviation diagnosis further includes: The power chain nodes are combined and analyzed according to the preset power chain structure, the chain-level deviation is calculated, and the chain-level abnormal information is output when the chain-level deviation exceeds the preset threshold. The deviation diagnosis further includes: In the time dimension, timing deviation analysis is performed on key time periods with high contribution, and abnormal information related to the force exertion sequence or action rhythm is output.

8. The badminton motion evaluation and training method based on kinetic chain multimodal timing as described in claim 1, characterized in that, The training guidance information is generated based on abnormal nodes, abnormal chains, and time-related abnormal results, and is used to guide the tested object to perform targeted action training.

9. A badminton motion evaluation and training system based on kinetic chain multimodal timing, characterized in that, Use the badminton motion assessment and training method based on kinetic chain multimodal timing as described in any one of claims 1 to 13.

10. The badminton motion evaluation and training system based on kinetic chain multimodal timing according to claim 9, characterized in that, Includes the following modules: Multimodal data acquisition module: Simultaneously acquires athlete joint posture, electromyographic signals, and plantar pressure data; Kinetic Chain Feature Learning Module: Constructs deep features of the kinetic chain of the hitting action based on chain-based topological graph convolution and self-attention temporal network; Movement assessment module: Automatically assesses the athlete's current movement level; Training guidance module: Based on high-level benchmark models and interpretable analysis results, personalized training guidance is generated by combining language models.

Images