A method for fusion analysis of strength training data and generation of personalized training programs

By combining dynamic correlation fusion and individual causal models with digital twin verification, personalized strength training programs are generated, which solves the problems of insufficient accuracy of multimodal data fusion and lack of causal constraints in training programs, and achieves accurate perception of training status and safety and reliability of the program.

CN123116148APending Publication Date: 2026-11-17HUNAN UNIV OF SCI & ENG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611319479.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing strength training data analysis, multimodal data fusion cannot adapt to the dynamic changes of the human body, resulting in insufficient accuracy in the representation of training states. The training scheme generation lacks causal constraints and adaptability, cannot effectively eliminate interference from confounding factors, and the generated scheme lacks interpretability and security.

Method used

By using an adaptive fusion mechanism based on dynamic correlation, combined with an individual training causal model and a musculoskeletal digital twin, a personalized strength training program is generated. The dynamic correlation between modalities is used for data fusion to identify the causal effects between training intervention parameters and outcome variables, and simulation prediction and safety verification are performed to form a closed-loop adaptive mechanism.

Benefits of technology

It improves the accuracy of training state representation and the safety and effectiveness of personalized solutions, enabling dynamic adjustment of solutions to adapt to changes in the user's physiological state, providing interpretable causal decision-making basis, and ensuring the long-term safety and individual fit of the solution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116148A_ABST
    Figure CN123116148A_ABST
Patent Text Reader

Abstract

This application relates to the field of intelligent sports science and data processing technology, and provides a method for the fusion analysis of strength training data and the generation of personalized training programs. The method includes: acquiring user multimodal training data and adaptively fusion-generating individual training state representations based on the dynamic correlation between modalities; identifying and ranking the causal effects between training intervention parameters and outcome variables through an individual training causal model to generate candidate programs; obtaining physiological dynamics and safety constraints through simulation prediction of candidate programs based on a musculoskeletal digital twin; generating personalized training programs by combining simulation results under causal effect constraints; and updating digital twin parameters and causal edge weights based on actual data after execution to form a closed-loop adaptive mechanism. This application improves the accuracy of training state representations and the safety, effectiveness, and individual adaptability of program generation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of intelligent sports science and data processing technology, specifically to a method for the fusion analysis of strength training data and the generation of personalized training programs. Background Technology

[0002] Currently, in the field of strength training data analysis and program generation, multimodal data acquisition combined with machine learning models is commonly used to assess user training status and generate training suggestions. However, existing solutions often employ fixed weights or simple concatenation strategies when fusing multimodal data, making it difficult to adapt to the dynamic correlations between different modalities under varying training stages and physiological states, resulting in insufficient accuracy in representing training status. Furthermore, in the training program generation stage, reliance on correlational statistical models or black-box predictive models fails to effectively eliminate the interference of confounding factors such as sleep, nutrition, and psychological factors. This makes it difficult to accurately identify the true causal relationship between training intervention parameters and strength gain, fatigue accumulation, and injury risk, leading to a lack of interpretability and individual adaptability in the generated training programs. Moreover, the inability to calibrate model parameters online based on actual user feedback makes it difficult to guarantee the safety and long-term effectiveness of the programs. Summary of the Invention

[0003] To address the issues of insufficient accuracy in multimodal training data fusion and the lack of causal constraints and adaptive closed-loop mechanisms in training scheme generation in existing technologies, this application proposes a method for the fusion analysis of strength training data and the generation of personalized training schemes. This method achieves accurate state perception based on dynamic correlation, reliable decision constraints based on causal effects, and secure verification and online optimization based on digital twin simulation.

[0004] To achieve the above objectives, this application adopts the following technical solution: The system acquires multimodal training data from users, adaptively fuses the multimodal training data based on the dynamic correlation between modalities, and generates an individual training state representation. Based on the individual training state representation and historical training data, it identifies the direct and indirect causal effects between training intervention parameters and training result variables through an individual training causal model, and ranks the training intervention parameters according to the direct and indirect causal effects to obtain a ranking result. Based on the ranking result, it generates multiple candidate training schemes within a preset safety constraint range. It performs simulation prediction on the multiple candidate training schemes to obtain the physiological dynamic prediction results and safety constraints corresponding to each candidate training scheme. The simulation prediction is performed based on a pre-constructed musculoskeletal digital twin. Under the constraint of the causal effects, and based on the physiological dynamic prediction results and the safety constraints, it generates a personalized strength training scheme. After the user executes the personalized strength training scheme, it collects actual training data and recovery data, updates the model parameters of the musculoskeletal digital twin based on the actual training data and recovery data, and adjusts the causal edge weights in the individual training causal model based on the actual strength growth and fatigue recovery results to form a closed-loop adaptive mechanism.

[0005] The above-mentioned scheme overcomes the shortcomings of traditional fixed-weight fusion, which cannot capture the complex dynamic state of the human body, by adopting an adaptive fusion mechanism based on the dynamic correlation between modalities, thus improving the accuracy of training state representation. By identifying direct and indirect causal effects and ranking training intervention parameters through an individual training causal model, it effectively eliminates the interference of non-causal confounding factors and provides an interpretable causal basis for scheme generation. Combined with simulation prediction using a pre-constructed musculoskeletal digital twin, it introduces physical safety constraints and physiological dynamic verification before generating the scheme. After the user executes the scheme, it uses actual data to update the parameters of the digital twin model and the causal edge weights, forming a closed-loop adaptive mechanism of "perception-decision-verification-update", which significantly improves the safety, effectiveness and individual adaptability of personalized strength training schemes.

[0006] As one implementation method, the temporal feature vectors corresponding to each modality in the multimodal training data are mapped to a common semantic space; an attention weight matrix between different modalities is calculated based on the temporal feature vectors, and the attention weight matrix is ​​scaled and dot-product calculated according to the dynamic correlation between modalities; the temporal feature vectors are weighted and fused based on the attention weight matrix to generate the individual training state representation.

[0007] As one implementation method, the temporal feature vectors corresponding to each modality are mapped to query vectors, key vectors, and value vectors through linear transformations, respectively; multiple parallel scaling dot products are calculated on the query vectors and key vectors using a multi-head attention mechanism to obtain multiple attention weight matrices; the value vectors are weighted and summed based on the multiple attention weight matrices to obtain fused features; the fused features are adaptively weighted using a gating mechanism, and the fused features are input into a fully connected layer for dimensionality reduction to obtain the individual training state representation.

[0008] As one implementation method, the individual training state representation and the historical training data are combined to form an individual longitudinal dataset; an initial causal graph is constructed based on the individual longitudinal dataset and prior knowledge in the field of sports science, wherein the prior knowledge in the field of sports science is used to constrain the direction of the causal edges in the initial causal graph; path coefficients are estimated based on the initial causal graph, and the direct and indirect causal effects between the training intervention parameters and the training outcome variables are identified.

[0009] As one implementation method, a constraint-based causal discovery algorithm is used to construct the initial causal graph. The training intervention parameters include training action selection, load intensity, training volume, rest time between sets, recovery strategy, and training frequency. The training outcome variables include strength gain, fatigue state, and injury risk. Structural equation modeling is used to estimate the path coefficients in the initial causal graph to calculate the direct causal effect of the training intervention parameters on the training outcome variables. Based on the mediating variables in the initial causal graph, the indirect causal effect of the training intervention parameters on the training outcome variables through the mediating variables is calculated. The direct causal effect and the indirect causal effect are added together to obtain the total causal effect, and the training intervention parameters are ranked based on the magnitude of the total causal effect.

[0010] As one implementation method, the generation of a personalized strength training program under the constraint of the causal effect and based on the physiological dynamic prediction results and the safety constraints includes at least one of the following methods: generating the personalized strength training program based on knowledge graph retrieval and a large language model; generating the personalized strength training program based on a reinforcement learning model.

[0011] As one implementation method, the generation of the personalized strength training plan based on knowledge graph retrieval and a large language model includes: converting user multi-source data into semantic tags, and retrieving relevant subgraphs in a pre-constructed strength training knowledge graph based on the semantic tags; converting the retrieved relevant subgraphs into natural language prompt text; inputting the natural language prompt text, the semantic tags, the physiological dynamic prediction results, and the individual training state representation into a pre-trained large language model, and generating a draft training plan through the pre-trained large language model, wherein the draft training plan conforms to the physiological dynamic prediction results and incorporates prior knowledge from the strength training knowledge graph; during the decoding process of the pre-trained large language model, An embedding vector constraint is introduced for knowledge graph entities. This constraint is achieved by applying cosine similarity between the vector representations of the knowledge graph entities and the decoded vectors. The token sequences generated by the pre-trained large language model are restricted through constraint decoding. This constraint decoding is achieved by setting allowed and prohibited token sets, ensuring that the output training scheme draft satisfies safety rules and individual injury taboos. A rule validator is set up to perform safety rule verification on the training scheme draft. These safety rules include load safety range, action taboos, fatigue state adaptability, and action combination safety. If the training scheme draft fails the safety rule verification, it is returned to be regenerated until it passes the verification, resulting in the personalized strength training scheme.

[0012] As one implementation method, the acquired multimodal training data is preprocessed, including: clock synchronization, time alignment using interpolation and resampling; noise reduction of inertial measurement unit data and surface electromyography data using wavelet threshold denoising, normalization of heart rate variability data and subjective fatigue level data using Z-score normalization, and imputation of missing data using multiple interpolation methods; the preprocessed multimodal training data is divided into sequence segments of three stages—pre-training, during-training, and post-training—according to time windows for subsequent feature extraction.

[0013] As one implementation method, generating the personalized strength training program based on the reinforcement learning model includes: using the individual training state representation and the digital twin simulation state as the state space of the reinforcement learning model, and using the training action selection, load intensity, training volume, rest time between sets, recovery strategy, and training frequency as the action space of the reinforcement learning model; using strength growth, fatigue recovery, and injury risk as multi-objective reward functions, and employing a proximal policy optimization algorithm or a soft actor critic algorithm to optimize the reinforcement learning model and output a dynamic training strategy; and using the dynamic training strategy as the personalized strength training program.

[0014] As one implementation, the step of updating the model parameters of the musculoskeletal digital twin based on the actual training data and the recovery data, adjusting the causal edge weights in the individual training causal model based on the actual strength growth and fatigue recovery results, and adjusting the objective weight coefficients of the multi-objective reward function includes: using an extended Kalman filter method or a Bayesian optimization method to update the model parameters of the musculoskeletal digital twin based on the actual training data and the recovery data; updating the causal edge weights of the individual training causal model based on the actual strength growth and fatigue recovery results; calculating the reward function prediction error based on the actual strength growth and fatigue recovery results, and adjusting the objective weight coefficients of the multi-objective reward function based on the reward function prediction error.

[0015] The beneficial effects of this application are as follows: By generating individual training state representations through an adaptive fusion mechanism based on dynamic intermodal correlations, the fusion weights of each modality data can be dynamically adjusted according to changes in the training stage and individual physiological state, solving the problem of representation distortion in complex human states under fixed-weight fusion and improving the robustness of state perception; by identifying and ranking the direct and indirect causal effects between training intervention parameters and training result variables through an individual training causal model, the influence of non-causal confounding factors is effectively eliminated, providing an interpretable and reliable causal decision basis for training scheme generation; by combining a pre-constructed musculoskeletal digital twin to simulate and predict candidate schemes, safety constraints and physiological dynamic verification in accordance with biomechanical principles are introduced before scheme generation, avoiding unsafe suggestions that may be generated by purely data-driven models; after the user executes the scheme, the parameters of the digital twin model and the causal edge weights are updated online using actual training and recovery data, forming a continuously optimized closed-loop adaptive mechanism, enabling the system to dynamically evolve with changes in the user's physical state, significantly improving the long-term safety, effectiveness, and individual adaptability of personalized strength training schemes. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for fusing and analyzing strength training data and generating personalized training plans according to an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0019] Embodiment 1 of this application is as follows: like Figure 1 As shown, this embodiment provides a method for the fusion analysis of strength training data and the generation of personalized training programs. This method achieves accurate generation and continuous optimization of personalized strength training programs through the collaborative work of multimodal data perception, causal inference decision-making, digital twin verification, and closed-loop adaptive updating. Specifically, the method includes the following steps: Step S100: Obtain the user's multimodal training data, and adaptively fuse the multimodal training data based on the dynamic correlation between modalities to generate an individual training state representation.

[0020] Specifically, multimodal training data encompasses information reflecting multiple dimensions, including user kinematics, dynamics, physiology, and subjective feelings. For example, multimodal training data may include, but is not limited to, inertial measurement unit data, surface electromyography (SEMG) data, force table data, heart rate variability data, and subjective fatigue level data. In this embodiment, the adaptive fusion mechanism does not employ fixed weighting coefficients or simple feature concatenation, but rather allocates fusion weights in real-time based on the dynamic correlation of each modality's data at the current moment. For instance, when a user is in a state of deep fatigue, the correlation between the spectral characteristics of the SEMG signal and the subjective fatigue level may be significantly enhanced. In this case, the adaptive fusion mechanism automatically increases the contribution of these two modalities in generating the individual training state representation, while relatively reducing the weight of kinematic data that is more susceptible to noise interference. This dynamic correlation-based fusion method effectively overcomes the shortcomings of traditional fixed-weight fusion methods in representing complex nonlinear states of the human body, ensuring that the generated individual training state representation can realistically and sensitively reflect the user's current comprehensive functional level.

[0021] Step S200: Based on individual training state representation and historical training data, the direct and indirect causal effects between training intervention parameters and training outcome variables are identified through an individual training causal model. The training intervention parameters are then ranked according to the direct and indirect causal effects to obtain the ranking results.

[0022] Specifically, the individual training causal model is a structured causal network built on the user's individual longitudinal data. Its core function is to extract true causal relationships from statistical correlations. Training intervention parameters refer to adjustable training control variables, such as training movement selection, load intensity, training volume, rest time between sets, recovery strategy, and training frequency. Training outcome variables refer to the target indicators to be optimized, such as strength gain, fatigue recovery rate, and injury risk probability. During the identification process, the system not only calculates the direct impact of intervention parameters on outcome variables but also quantifies indirect impact paths through mediating variables (such as muscle activation and metabolic stress levels). For example, increasing load intensity may directly promote strength gain but may also indirectly inhibit performance in the next training session by increasing neural fatigue. Through comprehensive evaluation and ranking of direct and indirect causal effects, the system can identify which intervention parameters have the highest marginal benefit for achieving training goals at the current stage. This provides an interpretable prioritization basis for subsequent program generation and effectively eliminates the interference of non-training confounding factors such as sleep quality and nutritional intake on decision-making.

[0023] Step S300: Based on the ranking results, generate multiple candidate training schemes within the preset safety constraints.

[0024] Specifically, the generation of candidate training schemes is not a random search within the entire parameter space, but rather a targeted combination guided by the causal effect ranking in step S200. The preset safety constraint range refers to the safe value range of each training intervention parameter determined based on prior knowledge in sports science and individual user historical data, including but not limited to basic safety boundaries such as the upper limit of load intensity, the threshold of single training capacity, and the minimum rest time between groups. For example, the system prioritizes training intervention parameters with high causal effect rankings as core variables and discretizes them within their safety threshold range to form a set of candidate schemes with high potential returns. As one implementation method, the combination generation process can employ a greedy search strategy based on causal ranking, prioritizing the top N ranked training intervention parameters and orthogonally combining them within the safety threshold range of each parameter; alternatively, a genetic algorithm can be used, using causal effect ranking as the prior weight of the fitness function, iteratively generating a set of candidate schemes through selection, crossover, and mutation operations. This generation strategy based on causal ranking combined with preset safety constraints significantly reduces the search space for ineffective schemes, improves computational efficiency, and simultaneously ensures the theoretical validity and physical safety feasibility of the candidate schemes.

[0025] Step S400: Simulation prediction is performed on multiple candidate training schemes to obtain the physiological dynamic prediction results and safety constraints corresponding to each candidate training scheme. The simulation prediction is performed based on the pre-constructed musculoskeletal digital twin.

[0026] Specifically, the musculoskeletal digital twin is a high-fidelity virtual model built based on the user's anatomical structure, biomechanical characteristics, and historical physiological data. During the simulation prediction phase, the system maps each candidate training scheme to this digital twin for virtual execution, simulating physiological processes such as muscle contraction, joint stress, and energy metabolism. The physiological dynamic prediction results include, but are not limited to, peak joint torque under specific movements, activation sequence of target muscle groups, and the expected accumulation of metabolic waste; safety constraints include hard physical boundaries such as maximum joint load-bearing capacity thresholds, ligament tensile limits, and safe eccentric muscle contraction rates. Through digital twin simulation, the system can identify schemes that, while statistically effective, pose high biomechanical risks before the user actually performs training, achieving a physical-level safety safety safeguard.

[0027] Step S500: Under the constraint of causal effect and based on the physiological dynamic prediction results and safety constraints, a personalized strength training program is generated.

[0028] Specifically, the final personalized strength training program is the product of the intersection of causal effectiveness and physical safety. The system first uses the causal effect ranking determined in step S200 as an optimization direction to ensure the program has the potential to promote strength growth or improve function. Then, it uses the physiological dynamic prediction results and safety constraints obtained in step S400 as a filter to eliminate or correct all parameter configurations that violate biomechanical safety boundaries. For example, if a load combination with high causal effect causes the knee joint shear force to exceed the safety threshold in the simulation, the system will automatically reduce the load or replace it with an alternative movement that puts less pressure on the knee joint, until the program simultaneously meets both causal benefit expectations and physical safety standards. This dual constraint mechanism ensures that the output personalized strength training program is both scientifically effective and safe and reliable.

[0029] Step S600: After the user executes the personalized strength training program, collect actual training data and recovery data, update the model parameters of the musculoskeletal digital twin based on the actual training data and recovery data, and adjust the causal edge weights in the individual training causal model based on the actual strength growth and fatigue recovery results to form a closed-loop adaptive mechanism.

[0030] Specifically, the closed-loop adaptive mechanism includes two levels of online updates: first, calibration at the physical model level, which uses actually collected kinematic, dynamic, and recovery data to correct physical parameters such as muscle stiffness, joint friction coefficient, and maximum voluntary contraction force in the musculoskeletal digital twin, allowing the virtual model to evolve synchronously with changes in the user's physical condition; second, iteration at the cognitive model level, which dynamically adjusts the weight coefficients of each causal edge in the individual training causal model based on the deviation between the user's actual strength increase and fatigue recovery and the expected values. For example, if a user's response to high-volume training is far lower than the model's prediction, the system will reduce the weight of the causal edge from "training volume" to "strength increase" and correspondingly increase the weight of the path related to "recovery strategy." This dual-track closed-loop update mechanism enables the system to continuously track and adapt to the heterogeneous changes of individual users, avoiding the performance degradation of general models after long-term application, and ensuring the accuracy and applicability of personalized strength training programs over long periods.

[0031] Embodiment 2 of this application is as follows: Based on Example 1, this example further details the specific implementation of adaptive fusion of multimodal training data based on dynamic correlation between modalities in step S100. As one implementation method, the temporal feature vectors corresponding to each modality in the multimodal training data are mapped to a common semantic space. Specifically, since the original signals such as inertial measurement unit data, surface electromyography data, and heart rate variability data have significant differences in sampling rate, dimensions, and physical meaning, direct concatenation would make it difficult for the model to capture cross-modal interaction information. Therefore, the system first converts the data of each modality into temporal feature vectors of a unified dimension through a feature extraction network, and then maps these heterogeneous feature vectors to the same high-dimensional semantic space through shared or independent projection layers. In this common semantic space, the features of different modalities are comparable and interactive, laying a geometric foundation for subsequent attention calculation. It should be understood that the mapping process can employ various feature extraction and transformation methods such as multilayer perceptrons, convolutional neural networks, or recurrent neural networks, as long as the function of aligning heterogeneous data to a unified representation space can be achieved; this application does not limit this.

[0032] Building upon this, an attention weight matrix is ​​calculated based on temporal feature vectors to determine the intermodal attention weight matrix. This matrix is ​​then scaled using a dot product based on the dynamic correlation between modalities. Specifically, the scaled dot product is essentially a similarity metric used to quantify the importance of information from a particular modality at a given moment for understanding the overall training state. After normalization by a scaling factor, the calculation results stabilize the gradients and reflect the real-time dependence strength between modalities. This content-based dynamic calculation method ensures that the fusion weights are no longer preset fixed values ​​but fluctuate in real time with factors such as user movement, fatigue levels, and environmental interference, thereby accurately capturing the nonlinear coupling relationships within the complex physiological system of the human body.

[0033] To further enhance the robustness and expressiveness of the fusion, as a preferred implementation, the temporal feature vectors corresponding to each modality are mapped to query vectors, key vectors, and value vectors through linear transformations. In the strength training data analysis scenario, the query vector represents the state intent that the system is attempting to parse at the current moment, the key vector represents the retrievalable feature indexes provided by each modality, and the value vector carries the actual physiological or kinematic information content. A multi-head attention mechanism is used to perform multiple parallel scaling dot product calculations on the query vector and key vector, resulting in multiple attention weight matrices. Multi-head parallel computation allows the model to simultaneously focus on different types of cross-modal dependencies in different representation subspaces. For example, one attention head might focus on capturing the temporal synchronicity of surface electromyography signals and inertial measurement unit acceleration signals to verify the standardization of action execution; while another attention head might focus on the long-term cumulative correlation between heart rate variability and subjective fatigue level to assess recovery status. This multi-view parallel perception mechanism effectively avoids the risk of a single attention head getting trapped in local optima or missing key interaction modes.

[0034] Subsequently, the value vector is weighted and summed based on multiple attention weight matrices to obtain the fused feature. This fused feature gathers high-value information from various modalities and has been dynamically filtered. Based on this, the fused feature is adaptively weighted using a gating mechanism and then input into a fully connected layer for dimensionality reduction to obtain an individual training state representation. The introduction of a gating mechanism is a key improvement in this embodiment that distinguishes it from conventional attention fusion. The gating mechanism can generate adjustment coefficients based on the instantaneous quality or confidence level of each modality signal, performing secondary calibration on the fused feature. For example, when a loose inertial measurement unit is detected, causing a sudden drop in the signal-to-noise ratio, the gating mechanism automatically outputs a small gain coefficient to suppress the contribution of that modality in the final representation and prevent noise contamination of the overall state estimation. As one implementation method, the gating mechanism can employ a gating unit based on the Sigmoid activation function, generating a gain coefficient between 0 and 1 by calculating the estimated signal-to-noise ratio of each modality signal; alternatively, a gated linear unit (GLU) structure can be used, implementing feature selection through element-wise multiplication. The dimensionality reduction operation of the fully connected layer is used to compress redundant information and extract compact and high-density individual training state representations, which are convenient for downstream causal models to process efficiently.

[0035] To illustrate the technical effectiveness of the adaptive fusion mechanism more intuitively, let's take the scenario of fatigue accumulation during squat training as an example. In the early stages of training, the user has ample energy and high precision in their movements. At this time, the kinematic data collected by the inertial measurement unit (IMU) and the surface electromyography (SEMG) data are highly synchronized with a good signal-to-noise ratio, and the attention weight matrix shows a relatively balanced contribution of both to the state representation. As the number of training sets increases, the user enters a state of deep fatigue, the muscle recruitment pattern changes, the spectral characteristics of the SEMG signal show significant drift, and the correlation with the subjective fatigue level increases sharply. Simultaneously, due to movement distortion, the IMU data may exhibit atypical jitter noise. At this stage, the adaptive fusion mechanism, through scaling dot product calculations, keenly captures the increased dynamic correlation between SEMG and subjective fatigue, automatically increasing the attention weights corresponding to these two modalities. Simultaneously, the gating mechanism identifies abnormal fluctuations in the IMU data and reduces its fusion gain. The resulting individual training state representation accurately reflects the user's true fatigue level, rather than being misled by distorted kinematic data. This process fully demonstrates that this embodiment achieves a more accurate and robust state perception capability than fixed-weight fusion through dynamic weight allocation and gating adjustment, providing a reliable data foundation for subsequent causal inference and scheme generation.

[0036] Embodiment 3 of this application is as follows: Building upon Example 1, this example further details the specific implementation of step S200, which involves identifying the causal relationship between training intervention parameters and training outcome variables using an individual training causal model. As one implementation method, individual training state representations and historical training data are combined to form an individual longitudinal dataset. Specifically, unlike cross-sectional data, the individual longitudinal dataset records the state evolution trajectory of the same user within a continuous time window. This dataset not only contains structured records of each training session but also includes recovery monitoring data during training intervals, thus enabling the capture of the time-lag effect between training stimuli and physiological adaptation. For example, delayed onset muscle soreness caused by a high-intensity eccentric training session may not peak until 48 hours later, thus affecting subsequent training performance. Only a longitudinal dataset with sufficient time span and sampling density can support modeling such long-range causal dependencies. It should be understood that the construction of the longitudinal dataset can also integrate the user's static baseline information, such as age, gender, injury history, and genotype data, as time-invariant covariates to control for interference from individual heterogeneity on causal inference.

[0037] Building upon this foundation, an initial causal graph is constructed based on individual longitudinal datasets and prior knowledge from the field of sports science. This prior knowledge constrains the direction of causal edges in the initial causal graph. Specifically, purely data-driven causal discovery algorithms are highly susceptible to generating false causal edges that violate biological common sense when faced with high-dimensional, sparse physiological data containing a large number of unobserved confounding variables. For example, statistical data may show a high correlation between "joint pain" and "low training load." Without prior constraints, the algorithm might incorrectly infer reverse paths other than "pain leads to low load" or ignore the common cause of "old injury." Therefore, this embodiment introduces prior knowledge from the field of sports science as structural constraints, including but not limited to metabolic pathway mechanisms from physiology textbooks, force transmission principles from biomechanics, and peer-reviewed empirical conclusions from sports training. This prior knowledge is encoded as a mask or directional penalty term in the adjacency matrix, forcibly prohibiting edges that do not conform to physiological logic (such as "strength increase leads to training capacity"), and guiding the algorithm to search for the optimal graph structure within a reasonable hypothesis space. This hybrid modeling strategy of "knowledge-guided + data-driven" significantly improves the interpretability and robustness of causal graphs, ensuring that subsequent effect estimations are based on a structural foundation that conforms to the laws of human physiological function.

[0038] Subsequently, path coefficients are estimated based on the initial causal graph to identify direct and indirect causal effects between training intervention parameters and training outcome variables. To more accurately quantify the causal strength among variables, a constraint-based causal discovery algorithm is used to construct the initial causal graph as a preferred implementation. The training intervention parameters include training action selection, load intensity, training volume, rest time between sets, recovery strategy, and training frequency; the training outcome variables include strength gain, fatigue state, and injury risk. Constraint-based algorithms (such as the PC algorithm or its variants for time-series data) determine the skeleton and separating sets among variables by performing a series of conditional independence tests and combine this with the aforementioned prior knowledge to determine the direction of the edges. Compared to rating-based algorithms, this method has better computational efficiency when dealing with large-scale variables and can explicitly utilize domain knowledge to eliminate spurious correlations. It should be understood that although this embodiment uses a constraint-based algorithm as an example, rating-based or hybrid causal discovery methods can also be used in other implementations, as long as their construction process is effectively constrained by prior knowledge of sports science.

[0039] After determining the causal graph structure, structural equation modeling (SEM) is used to estimate the path coefficients in the initial causal graph, calculating the direct causal effect of the training intervention parameter on the training outcome variable. Specifically, SEM transforms the causal graph into a system of simultaneous equations, and obtains standardized path coefficients by fitting the covariance matrix of the observed data through maximum likelihood estimation or Bayesian inference. These path coefficients reflect the change in the standard deviation of the training outcome variable caused by a one-standard-deviation change in a particular training intervention parameter, while keeping other parent node variables constant. For example, after controlling for nutrient intake and sleep quality, the path coefficient from "load intensity" to "strength growth" is 0.45, indicating a direct positive promoting effect of load intensity on strength growth. Simultaneously, based on the mediating variables in the initial causal graph, the indirect causal effect of the training intervention parameter on the training outcome variable through these mediating variables is calculated. The indirect effect reveals the underlying mechanism by which the intervention works. For example, "load intensity" may not always directly act on "strength growth," but rather first promote "strength growth" by increasing the mediating variable of "muscle activation." At this point, the indirect causal effect equals the product of the path coefficient for "load intensity → muscle activation" and the path coefficient for "muscle activation → strength increase". If multiple parallel mediating paths exist, the total indirect effect is obtained by adding the products of all paths.

[0040] Finally, the direct and indirect causal effects are summed to obtain the total causal effect, and the training intervention parameters are ranked based on the magnitude of the total causal effect. This ranking logic is the key difference between this application and traditional correlation analysis. In traditional methods, a parameter may show a high correlation with the outcome variable, but this may be a spurious association caused by a third confounding factor; or a parameter may have a weak direct effect, but it generates a large cumulative indirect effect through multiple mediation paths, which is easily missed by simple regression analysis. By calculating the total causal effect, the system can comprehensively evaluate the true marginal contribution of each training intervention parameter. For example, in a user's current state, the direct effect of "inter-set rest time" may only be 0.1, but its indirect effect of improving "metabolic clearance rate" and reducing "nervous fatigue" is as high as 0.5, and the total causal effect reaches 0.6, which is higher than the total effect of "training capacity" (0.4). Accordingly, the system ranks "inter-set rest time" before "training capacity" as the preferred lever for adjusting the program in the next stage. This ranking mechanism based on total causal effect provides a scientific priority basis for the generation of candidate scheme combinations in the subsequent step S300, ensuring that personalized training schemes can accurately target the key bottleneck factors that restrict the user's current performance, rather than blindly adjusting those statistically relevant but actually ineffective variables.

[0041] Embodiment 4 of this application is as follows: Based on Example 1, this example further details the specific implementation of generating a personalized strength training plan in step S500. As mentioned earlier, generating a personalized strength training plan includes at least one of the following methods: knowledge graph retrieval and large language model generation, or reinforcement learning model generation. This example uses the knowledge graph retrieval and large language model generation method as an example. As one implementation method, user multi-source data is converted into semantic tags, and relevant subgraphs are retrieved in the pre-constructed strength training knowledge graph based on the semantic tags. Specifically, this process is performed in real time during the execution phase of plan generation, rather than an offline knowledge base construction process. User multi-source data includes, but is not limited to, the multimodal training data obtained in Example 1, the user's basic physiological information, and historical injury records. The system uses a preset entity linking and relation extraction model to map these unstructured or semi-structured raw data into standardized semantic tags in the knowledge graph, for example, converting "6 months after left anterior cruciate ligament reconstruction" into tags such as "knee joint stability risk - medium" and "contraindicated movement - squat > 90 degrees". Subsequently, using these semantic tags as anchors, subgraph retrieval is performed in the pre-constructed strength training knowledge graph. This knowledge graph contains structured domain knowledge such as action libraries, target muscles, load-volume-intensity relationships, recovery rules, injury contraindications, and individual difference labels. The retrieved relevant subgraphs not only include nodes directly related to the current user state but also their multi-hop neighbors, thus providing a rich and accurate structured context for subsequent generation. This Retrieval-Enhanced Generation (RAG) strategy effectively addresses the lack of vertical domain knowledge in general-purpose large language models, ensuring that the knowledge base generated by the solution is accurate and conforms to the principles of sports science.

[0042] Based on this, the retrieved relevant subgraphs are converted into natural language prompts. Specifically, the conversion process is not a simple text concatenation, but rather uses template filling or serialization algorithms to transform the graph structure information into a linear sequence that the large language model can understand. For example, the triple "Action A-Taboo-Condition B" is converted into the natural language instruction "Note: Due to the user's condition B, action A is strictly prohibited." Subsequently, the natural language prompts, semantic tags, physiological dynamic prediction results, and individual training state representations are input into the pre-trained large language model. The pre-trained large language model generates a draft training scheme, which conforms to the physiological dynamic prediction results and incorporates prior knowledge from the strength training knowledge graph. It is important to note that the physiological dynamic prediction results come from the simulation output of the digital twin in Example 1, providing a physical safety boundary for scheme generation; while the individual training state representation provides the user's current real-time functional state. The large language model acts as a comprehensive reasoning engine for multi-source heterogeneous information. The draft training scheme it generates not only includes expert rules from the knowledge graph but also adapts to the safety parameters verified by the digital twin, while also responding to the user's real-time state requirements.

[0043] To ensure the reliability of the generated content, the draft training plan is matched and screened against causal effect constraints and safety constraints to obtain a personalized strength training plan. Specifically, the causal effect constraints are derived from the ranking results of step S200 in Example 1, ensuring that the plan prioritizes the intervention parameters most effective for the user's current stage; the safety constraints include the physical boundaries determined by digital twin simulation and medical contraindications in the knowledge graph. The matching and screening process is a multi-dimensional verification step; only drafts that simultaneously meet causal validity, physical safety, and knowledge compliance can be adopted as the final plan.

[0044] To further suppress the illusion problem of large language models from the perspective of generation mechanism, as a preferred implementation, an embedding vector constraint of knowledge graph entities is introduced during the decoding process of the pre-trained large language model. This embedding vector constraint is achieved by constraining the vector representation of the knowledge graph entity with the decoded vector using cosine similarity. Specifically, this is a soft constraint mechanism that takes effect at execution time. At the moment the model generates each token, the system calculates the cosine similarity between the hidden state vector output by the current decoder and the embedding vector of the legitimate entity in the knowledge graph. If the similarity is lower than a preset threshold, it indicates that the model is generating irrelevant content or fictitious concepts that deviate from domain knowledge. In this case, the system will penalize the generation probability distribution at that position or directly block low-similarity candidate words. This mechanism forcibly pulls the free generation space of the large language model back to the semantic manifold defined by the knowledge graph, so that the generation process is guided by structured knowledge in real time.

[0045] Simultaneously, constraint decoding restricts the token sequences generated by the pre-trained large language model. This constraint decoding is achieved by setting allowed and prohibited token sets, ensuring that the output training scheme draft meets safety rules and individual injury taboos. Specifically, this is a more stringent hard constraint mechanism than embedding vector constraints. The allowed and prohibited token sets are dynamically generated based on the current user's semantic tags and safety rules. For example, for a user with a history of lumbar disc herniation, the prohibited token set will include all tokens and their synonym variations corresponding to high-risk movements such as "barbell bent-over rows" and "good morning forward bends"; while for a user in an acute fatigue phase, the allowed token set may only retain tokens related to low-intensity recovery movements. At each step of decoding, the model can only sample from the allowed set or is forced to avoid words in the prohibited set. This token-level precise control ensures that regardless of the model's internal reasoning, its output text sequence cannot violate key safety bottom lines at the grammatical level.

[0046] In addition, a rule validator is set up to verify the safety rules of the draft training program. These safety rules include load safety range, movement contraindications, fatigue state adaptability, and movement combination safety. Specifically, the rule validator is a deterministic logic module independent of the large language model, used to compensate for the uncertainty defects of neural network probabilistic generation. Load safety range verification ensures that the weight, number of sets, and repetitions in the program do not exceed the individual thresholds determined by the digital twin simulation; movement contraindication verification re-checks whether any medical restrictions have been omitted from the program; fatigue state adaptability verification checks whether reasonable load reduction or rest is arranged after high-intensity training days; and movement combination safety verification focuses on the upper limit of the number of training movements targeting the same muscle group within the same training day and the joint coordination requirements between adjacent movements to prevent cumulative injuries caused by local overuse. If the draft training program fails the safety rule verification, it is returned to be regenerated until it passes verification. This feedback loop mechanism constitutes the last line of defense in program generation. When verification fails, the system will re-input the specific reason for the violation (such as "squat load exceeds the knee joint shear force safety threshold of 15%) as negative feedback to the large language model, guiding it to make targeted corrections in the next round of generation. This closed-loop iterative process of "generation-verification-correction" ensures that the final output of personalized strength training programs is logically consistent, knowledge-accurate, and physically safe, achieving a deep integration of large language model generation capabilities and the rigor of sports science.

[0047] Embodiment 5 of this application is as follows: Building upon Example 1, this example provides another implementation method for generating personalized strength training programs in step S500. Unlike the knowledge-based reasoning generation path based on knowledge graphs and large language models in Example 4, this example employs a numerical optimization generation path based on reinforcement learning models. This aims to seek the optimal solution for long-term training benefits under complex constraints through mathematical modeling and policy search. As one implementation method, the individual training state representation and digital twin simulation state are used as the state space of the reinforcement learning model, while training action selection, load intensity, training volume, rest time between sets, recovery strategy, and training frequency are used as the action space of the reinforcement learning model.

[0048] Specifically, the state space is constructed using a dual-channel design that combines virtual and real-world elements. The individual training state representation comes from the adaptive fusion output in Example 1, reflecting the user's current external functional level that can be directly observed by sensors. The digital twin simulation state, on the other hand, comes from a pre-constructed musculoskeletal digital twin, containing internal physiological variables crucial to safety, such as muscle tension distribution, articular cartilage contact pressure, and cumulative concentration of metabolic products—variables that cannot be directly measured by wearable devices. As one implementation method, the digital twin simulation state can be structured into a state vector containing the following dimensions: muscle force vector (the ratio of real-time activation level to maximum voluntary contractile force of each target muscle group), joint torque vector (the ratio of net torque to ultimate bearing capacity of each joint), and metabolic index vector (lactate concentration, blood oxygen saturation, energy consumption rate, etc.). This combination allows the reinforcement learning agent to perceive both real-time user feedback and potential physiological risks, solving the problem of decision-making blind spots caused by relying solely on external data in traditional methods. The movement space encompasses the core controllable variables of a strength training program, including both discrete training movement selection and continuous variables such as load intensity, training volume, rest time between sets, recovery strategies, and training frequency. It should be understood that the specific dimensions of the movement space can be adjusted according to the actual application scenario. For example, in rehabilitation training scenarios, "range of motion limitation" can be added as an additional movement dimension, while in competitive sports scenarios, dimensions such as "proportion of explosive power training" can be added, as long as it can fully parameterize and describe an executable training unit.

[0049] Building upon this foundation, strength gain, fatigue recovery, and injury risk are used as multi-objective reward functions. Proximal policy optimization algorithms or soft actor critic algorithms are employed to optimize the reinforcement learning model, outputting a dynamic training strategy. Specifically, the design of the multi-objective reward function embodies the "benefit-cost-risk" triangular balance principle in training science. Strength gain, as a positive incentive, encourages the agent to explore training loads that effectively improve performance; fatigue recovery, as a moderating term, penalizes schemes that lead to excessive fatigue or insufficient recovery, guiding the agent to rationally arrange the training rhythm; injury risk, as a negative penalty, severely punishes any high-risk parameter combination that may cause tissue damage, ensuring that the strategy search always stays within safe boundaries. These three objectives often have inherent conflicts. For example, pursuing maximum strength gain usually accompanies higher injury risk and longer fatigue recovery time. Therefore, the reward function uses dynamic weight coefficients to adapt to the preferences of different users at different stages. For example, for athletes in the peak of their training season, the weight of strength gain may be increased; while for users in post-operative recovery, the penalty weight of injury risk dominates.

[0050] At the algorithmic level, the Proximal Policy Optimization (PPO) algorithm is preferred for handling high-dimensional continuous action spaces due to its stable training process and low hyperparameter sensitivity; the Soft Actor Critics (SAC) algorithm, with its strong exploration capabilities and sample efficiency brought by the maximum entropy framework, is suitable for scenarios that require rapid adaptation to individual user differences. It is important to emphasize that the policy optimization process here is an online adaptation performed at runtime for a specific user, rather than general offline pre-training. The system utilizes the musculoskeletal digital twin constructed in Example 1 as a virtual interactive environment, allowing the reinforcement learning agent to conduct thousands of simulated trials and errors within it, thereby quickly converging a personalized strategy suitable for the user's current state without consuming the user's real physical resources. Finally, the dynamic training strategy is used as a personalized strength training scheme. This process transforms the abstract training objective into a concrete, numerically validated combination of parameters, complementing the semantic understanding-based generation method in Example 4: the former excels at refined numerical optimization within clear safety boundaries, while the latter excels at handling unstructured knowledge reasoning and natural language interaction. Together, they constitute the multi-layered, comprehensive scheme generation defense system of this application.

[0051] Embodiment 6 of this application is as follows: Building upon Example 5, this example further details the specific mathematical implementation and parameter adjustment logic of the closed-loop adaptive mechanism in step S600. The closed-loop adaptive mechanism includes: updating the model parameters of the musculoskeletal digital twin, adjusting the causal edge weights in the individual training causal model, and adjusting the weight coefficients of each objective in the multi-objective reward function. As one implementation method, an extended Kalman filter or Bayesian optimization method is used to update the model parameters of the musculoskeletal digital twin based on actual training and recovery data. Specifically, musculoskeletal digital twins are often initially constructed based on population averages or limited static test data, making it difficult to accurately reflect the dynamic physiological adaptation of individuals during long-term training. Actual training and recovery data contain the user's kinematic response, dynamic output, and physiological recovery trajectory in a real physical environment, serving as the most direct basis for calibrating the virtual model. The extended Kalman filter method is suitable for processing biomechanical systems with nonlinear state transition characteristics, enabling optimal fusion of sensor observations and model predictions to estimate and correct in real time internal state parameters that cannot be directly measured, such as muscle stiffness coefficients, tendon resting length, maximum voluntary contractile force, and joint friction coefficients. Bayesian optimization methods excel at finding globally optimal parameter combinations in scenarios with limited sample sizes and high evaluation costs, making them particularly suitable for calibrating anatomical parameters that are highly sensitive to model output but change slowly. Through continuous iterative updates of these algorithms, the musculoskeletal digital twin can gradually evolve from a generalized, standardized model into a highly personalized, user-specific digital image, thereby significantly improving the reliability and security of subsequent simulation predictions.

[0052] Based on this, the causal edge weights of the individual training causal model are updated according to actual strength gains and fatigue recovery results. Specifically, the individual training causal model is not a static structure; its path coefficients should dynamically evolve as user training data accumulates. Historical training data includes not only immediate feedback from the current training session but also long-term longitudinal records. The system uses this accumulated data to refit the causal graph structure or update path coefficients to capture the drift in the user's body's adaptation to training stimuli. For example, a user may show a significant strength gain response to high-volume training in the early stages of training, but as the training years increase, the marginal benefit diminishes. In this case, the system should automatically reduce the causal edge weight from "training volume" to "strength gain" and correspondingly increase the weight of paths related to "load intensity" or "movement variation." This online update mechanism ensures that the causal model always reflects the user's current physiological response characteristics, avoiding decision bias caused by model aging.

[0053] To further optimize the quality of personalized solutions, the reward function prediction error is calculated based on actual strength gain and fatigue recovery results, and the weight coefficients of each objective in the multi-objective reward function are adjusted accordingly. Specifically, the weight coefficients in the multi-objective reward function represent the system's trade-off preference between strength gain, fatigue recovery, and injury risk. However, different users may experience differences in subjective feelings and objective physiological responses at different stages, making it difficult for preset fixed weights to perfectly suit all individuals. The reward function prediction error refers to the deviation between the training results predicted by the system based on the current model and the actual feedback results from the user. For example, if the system predicts that a certain solution will lead to moderate fatigue accumulation, but the user's actual feedback shows that the fatigue recovery time is much longer than expected, it indicates that the current model underestimates the user's fatigue sensitivity. In this case, the system will calculate a positive fatigue prediction error and automatically increase the weight coefficient of the "fatigue recovery" term in the multi-objective reward function, while possibly moderately decreasing the weight of the "strength gain" term. Conversely, if the user's actual strength gain is significantly higher than expected and there is no obvious discomfort, the weight incentive for strength gain can be appropriately increased. This weight adaptive adjustment mechanism based on prediction error essentially encodes the user's implicit preferences and physiological thresholds into the optimization objective, enabling the reinforcement learning strategy to automatically converge in a direction that better matches the user's true physical and mental state without human intervention, thus achieving truly human-centered personalized training services.

[0054] It should be understood that although this embodiment uses extended Kalman filtering and Bayesian optimization as examples to illustrate the parameter update method, in other embodiments, particle filtering, gradient descent, or other parameter estimation algorithms suitable for nonlinear system identification can also be used, as long as they can use actual observation data to correct the internal parameters of the digital twin. Similarly, the adjustment logic of the reward function weights is not limited to linear adjustment based on prediction error; fuzzy logic, meta-learning, or gradient-based hyperparameter optimization strategies can also be used to achieve more refined preference alignment. This application does not limit this.

[0055] Embodiment 7 of this application is as follows: Based on Example 1, this example further details the preprocessing step after acquiring the user's multimodal training data in step S100. As one implementation, the acquired multimodal training data is preprocessed, including clock synchronization and time alignment using interpolation and resampling. Specifically, since inertial measurement units, surface electromyography sensors, force platforms, and heart rate monitoring devices typically have different sampling frequencies and data transmission delays, the original data is often misaligned on the time axis. The system first synchronizes the clocks of each acquisition device using a hardware trigger signal or a unified timestamp protocol to eliminate absolute time deviations between devices. Subsequently, to address the issue of inconsistent sampling rates, cubic spline interpolation or linear interpolation algorithms are used to upsample low-frequency data, or anti-aliasing filtering is used to downsample high-frequency data, unifying the data of all modalities to the same time reference. This step is a prerequisite for subsequent cross-modal attention calculation and causal inference. Without accurate time alignment, the calculation of dynamic correlations between different modalities will lose physical meaning, leading to distorted fusion results.

[0056] Based on this, wavelet threshold denoising is used to reduce noise in inertial measurement unit data and surface electromyography (SEMG) data, Z-score normalization is used to normalize heart rate variability data and subjective fatigue level data, and multiple imputation is used to fill in missing data. Specifically, physiological signals in strength training scenarios are often accompanied by motion artifacts and environmental noise. Taking SEMG data as an example, although traditional low-pass filtering can remove high-frequency noise, it easily smooths out key transient features generated by rapid muscle contraction. This embodiment uses wavelet threshold denoising, which decomposes the signal into different scales and adaptively sets the threshold according to the noise distribution characteristics at each scale. It only suppresses the noise coefficient while retaining the effective high-frequency components reflecting muscle activation patterns, thereby maintaining the time-frequency localization characteristics of the signal to the greatest extent while denoising. For heart rate variability data and subjective fatigue level data, since the former is in milliseconds and the latter is a dimensionless subjective rating, direct fusion will lead to features with large numerical magnitudes dominating the model gradient. By standardizing using Z-scores, the data from each modality are transformed into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating dimensional differences and enabling the subsequent adaptive fusion mechanism to fairly evaluate the information contribution of each modality. Furthermore, for missing data caused by sensor detachment or wireless transmission packet loss, a multiple imputation method is used to generate multiple possible complete datasets and perform a comprehensive estimation. Compared to single mean imputation, this method better reflects the true variance structure of the data and avoids statistical biases introduced by humans.

[0057] Finally, the preprocessed multimodal training data was segmented into three time-window segments: pre-training, during-training, and post-training, for subsequent feature extraction. Specifically, this segmentation strategy is not a simple fixed-duration segmentation, but rather a semantic segmentation based on training event labels. The pre-training segment captures the user's resting baseline state and warm-up level, providing a reference for assessing initial functional reserves; the during-training segment records the dynamic physiological response under load stimulation, serving as the core region for identifying fatigue accumulation and changes in movement quality; and the post-training segment reflects the body's immediate recovery ability and homeostasis reconstruction process. For example, in analyzing squat training, the surface electromyography (EMG) baseline of the pre-training segment can be used to calibrate individual differences, the coupling relationship between joint angles and EMG signals in the during-training segment is used to assess technical stability, and the heart rate recovery rate in the post-training segment is a key indicator for quantifying cardiopulmonary adaptability. By decoupling continuous data streams into phase sequences with clear physiological meanings, the system can provide structured input priors for subsequent temporal feature extraction networks. This allows the network to learn specific representations for different training phases, rather than blindly searching for features in a long sequence, significantly improving the efficiency and interpretability of individual training state representation generation. It should be understood that although this embodiment uses a three-phase segmentation as an example, in practical applications, the segmentation granularity can be flexibly adjusted according to the specific training type (such as cyclic training or interval training), as long as different physiological load states can be distinguished.

[0058] Embodiment 8 of this application is as follows: To more clearly illustrate the technical effects and data flow logic of the strength training data fusion analysis and personalized training program generation method provided in this application in practical applications, a specific comprehensive application scenario is used as an example for detailed explanation below. In this scenario, the user is set as a strength trainee who has undergone left anterior cruciate ligament reconstruction surgery 8 months ago. Their current training goal is to restore maximum lower limb strength while strictly controlling knee joint shear force to avoid secondary injury. The user wears intelligent training equipment including an inertial measurement unit, surface electromyography sensor, and heart rate monitoring device, and performs squat training on the strength equipment. This embodiment will fully demonstrate how the system coordinates various functional modules to provide the user with safe and efficient personalized training services.

[0059] Before and during training, the system first performs a multimodal data acquisition and preprocessing process. Sensors on the user collect inertial measurement unit data, surface electromyography data, heart rate variability data, and subjective fatigue level data in real time. After preprocessing operations such as clock synchronization, wavelet threshold denoising, and Z-score normalization, this data is segmented into sequence segments in three stages: pre-training baseline, dynamic response during training, and post-training recovery. This preprocessing process ensures the spatiotemporal alignment and quality reliability of the data used in subsequent analysis, laying the foundation for accurately sensing the user's state.

[0060] Subsequently, the system adaptively fuses the preprocessed multimodal training data based on the dynamic correlation between modalities to generate an individual training state representation. In this scenario, as the number of squat sets increases, early signs of decreased neuromuscular control gradually appear in the muscles around the user's left knee. At this point, a significant increase in dynamic correlation is observed between the spectral characteristics of the surface electromyography (EMG) signal and the subtle knee joint swaying patterns captured by the inertial measurement unit. The adaptive fusion mechanism keenly captures this cross-modal interaction change, automatically increasing the attention weights of these two modalities in the fusion calculation, while suppressing kinematic data interference caused by reduced signal-to-noise ratio due to motion deformation through a gating mechanism. The resulting individual training state representation accurately quantifies the user's current level of "compensatory fatigue," rather than merely reflecting overall metabolic fatigue or subjective exertion. This precise state perception capability allows the system to identify potential biomechanical risks before the user subjectively perceives any danger.

[0061] Based on the generated individual training state representations and historical training data, the system evaluates the intervention parameters in the original training plan using an individual training causal model. Causal model analysis shows that, under the current state of "compensatory fatigue," while the intervention parameter "load intensity" has a positive direct causal effect on "strength increase," its indirect causal effect on "injury risk" is significantly amplified through the mediating variable of "decreased knee joint stability," leading to a severe imbalance in the risk-benefit ratio of the overall causal effect. In contrast, "movement variation" and "rest time between sets" rank higher in the overall causal effect under the current state, indicating that adjusting the movement pattern or increasing rest is more effective in balancing safety and training benefits than simply maintaining the load. Therefore, the system determines that the original high-load squat program is no longer applicable and identifies "movement variation" as the priority lever for adjustment.

[0062] After determining the direction of adjustment, the system inputs the original plan and adjustment suggestions into a pre-constructed musculoskeletal digital twin for simulation prediction. The digital twin is modeled based on the user's personalized anatomical structure and postoperative rehabilitation progress. Simulation results show that if the original load is continued, the peak shear force on the left anterior cruciate ligament will exceed the safety threshold by 15%. However, if a "box squat" is used as an alternative movement and the load is appropriately reduced, the joint shear force can be controlled within a safe range while maintaining quadriceps activation. The simulation prediction results not only rejected the original plan but also output clear physiological dynamic prediction results and safety constraints, defining insurmountable physical boundaries for subsequent plan generation.

[0063] Next, under the constraint of causal effects and based on physiological dynamic prediction results and safety constraints, the system generates a personalized strength training plan. In this embodiment, the system adopts a generation path based on knowledge graphs and large language models. First, the semantic tags of the user, such as "left knee ACL surgery," "compensatory fatigue," and "box squat," are retrieved from the strength training knowledge graph to obtain the movement specifications, contraindications, and recommended load range for postoperative squat variations. Subsequently, this structured knowledge is converted into natural language prompts and input into the pre-trained large language model along with the individual training state representation and safety constraints. During the decoding and generation process, the system ensures that the generated movement names are consistent with the knowledge graph entities through embedded vector constraints and prohibits the generation of any high-risk tokens involving "over the knee" or "explosive force" through constraint decoding. At the same time, the rule validator performs a secondary check on the generated draft to confirm that it meets the requirements of load safety range and movement combination safety. Finally, the system outputs a personalized strength training plan containing "4 sets of 8 repetitions of box squats, load reduced to 60% of 1RM, and rest between sets extended to 120 seconds." It should be understood that in other implementations, this step can also adopt a reinforcement learning path, by optimizing the policy on the digital twin and directly searching for the optimal parameter combination that satisfies the multi-objective reward function. Both can achieve scheme generation under security constraints.

[0064] After the user completes the newly generated personalized strength training program, the system enters a closed-loop adaptive update phase. By collecting the user's actual kinematic data, electromyographic response, and recovery data over 24 hours post-training, the system uses an extended Kalman filter to update the muscle stiffness coefficients and joint stability parameters in the musculoskeletal digital twin, enabling the virtual model to more accurately reflect the user's biomechanical characteristics at this specific stage eight months post-surgery. Simultaneously, based on the deviation between the actual strength gain and fatigue recovery and the expected values, the system adjusts the causal edge weights of "movement variation" pointing to "strength gain" in the individual training causal model. For example, if actual data shows that box squats have a better strength stimulation effect on the user's current condition than expected, the system will increase the weight of this path, giving it higher priority in future program generation. This continuous iterative mechanism based on actual feedback ensures that the personalized strength training program can dynamically evolve with the user's recovery process and changes in physical adaptation, truly achieving long-term, safe, and effective personalized training guidance.

[0065] As can be seen from the complete application scenarios described above, the method provided in this application is not a simple superposition of various algorithm modules, but an organically coordinated intelligent closed loop. Adaptive fusion solves the problem of perceptual distortion under complex human conditions, providing reliable input for causal inference; the causal model eliminates the interference of confounding factors, ensuring the scientific nature and interpretability of decision-making; digital twin simulation provides a physical safety net, making up for the blind spots of purely data-driven methods; the knowledge-guided scheme generation mechanism achieves accurate output of personalized content while ensuring safety; and the closed-loop update mechanism endows the system with the ability to continuously learn and adapt to individual heterogeneity. This entire set of technical solutions works together to effectively overcome the shortcomings of existing technologies, such as insufficient security of training schemes, poor individual adaptability, and lack of dynamic evolution capabilities, demonstrating significant technological progress and application value.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for fusion analysis of strength training data and generation of personalized training programs, characterized in that, include: Acquire the user's multimodal training data, and adaptively fuse the multimodal training data based on the dynamic correlation between modalities to generate an individual training state representation; Based on the individual training state representation and historical training data, the direct and indirect causal effects between training intervention parameters and training outcome variables are identified through an individual training causal model. The training intervention parameters are then ranked according to the direct and indirect causal effects to obtain the ranking results. Based on the ranking results, multiple candidate training schemes are generated within a preset safety constraint range. The simulation prediction is performed on the multiple candidate training schemes to obtain the physiological dynamic prediction results and safety constraints corresponding to each candidate training scheme. The simulation prediction is performed based on a pre-constructed musculoskeletal digital twin. Under the constraints of the causal effect, and based on the physiological dynamic prediction results and the safety constraints, a personalized strength training program is generated. After the user executes the personalized strength training program, actual training data and recovery data are collected. The model parameters of the musculoskeletal digital twin are updated based on the actual training data and recovery data. The causal edge weights in the individual training causal model are adjusted based on the actual strength growth and fatigue recovery results to form a closed-loop adaptive mechanism.

2. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 1, characterized in that, The adaptive fusion of the multimodal training data based on the dynamic correlation between modalities to generate individual training state representations includes: Map the temporal feature vectors corresponding to each modality in the multimodal training data to a common semantic space; Based on the temporal feature vector, an attention weight matrix is ​​calculated between different modalities, and the attention weight matrix is ​​calculated by scaling the dot product according to the dynamic correlation between modalities. The temporal feature vector is weighted and fused based on the attention weight matrix to generate the individual training state representation.

3. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 2, characterized in that, The calculation of the attention weight matrix between different modalities based on the temporal feature vector includes: The temporal feature vectors corresponding to each modality are mapped to query vectors, key vectors, and value vectors respectively through linear transformations. Multiple attention weight matrices are obtained by performing multiple parallel scaling dot product calculations on the query vector and the key vector using a multi-head attention mechanism. The value vector is weighted and summed based on the multiple attention weight matrices to obtain the fused features; The fused features are adaptively weighted using a gating mechanism, and then input into a fully connected layer for dimensionality reduction to obtain the individual training state representation.

4. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 1, characterized in that, The process of identifying direct and indirect causal effects between training intervention parameters and training outcome variables based on the individual training state representation and historical training data through an individual training causal model includes: The individual training state representation and the historical training data are combined to form an individual longitudinal dataset; An initial causal graph is constructed based on the individual longitudinal dataset and prior knowledge in the field of sports science, wherein the prior knowledge in the field of sports science is used to constrain the direction of the causal edges in the initial causal graph. Based on the path coefficients estimated from the initial causal graph, the direct and indirect causal effects between the training intervention parameters and the training outcome variables are identified.

5. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 4, characterized in that, The step of estimating path coefficients based on the initial causal graph and identifying the direct and indirect causal effects between the training intervention parameters and the training outcome variables includes: The initial causal graph is constructed using a constraint-based causal discovery algorithm. The training intervention parameters include training action selection, load intensity, training volume, rest time between sets, recovery strategy, and training frequency. The training outcome variables include strength gain, fatigue state, and injury risk. The path coefficients in the initial causal graph are estimated using structural equation modeling, and the direct causal effect of the training intervention parameters on the training outcome variables is calculated. Based on the mediating variables in the initial causal graph, calculate the indirect causal effect of the training intervention parameters on the training outcome variables through the mediating variables; The total causal effect is obtained by adding the direct causal effect and the indirect causal effect, and the training intervention parameters are ranked based on the magnitude of the total causal effect.

6. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 1, characterized in that, The generation of a personalized strength training program, under the constraint of the causal effect and based on the physiological dynamic prediction results and the safety constraints, includes at least one of the following methods: The personalized strength training program is generated based on knowledge graph retrieval and a large language model. The personalized strength training program is generated based on a reinforcement learning model.

7. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 6, characterized in that, The personalized strength training program generated based on knowledge graph retrieval and a large language model includes: User multi-source data is transformed into semantic tags, and relevant subgraphs are retrieved in a pre-constructed strength training knowledge graph based on the semantic tags; The retrieved relevant sub-graphs are converted into natural language prompt text; The natural language prompt text, the semantic label, the physiological dynamic prediction result, and the individual training state representation are jointly input into the pre-trained large language model. The pre-trained large language model generates a training plan draft, which conforms to the physiological dynamic prediction result and integrates the prior knowledge of the strength training knowledge graph. In the decoding process of the pre-trained large language model, the embedding vector constraint of knowledge graph entities is introduced. The embedding vector constraint is achieved by applying cosine similarity constraint between the vector representation of the knowledge graph entity and the decoding vector. The token sequences generated by the pre-trained large language model are restricted by constraint decoding, which is achieved by setting allowed token sets and prohibited token sets, so that the output training scheme draft satisfies safety rules and individual impairment taboos. A rule validator is set up to perform safety rule verification on the draft training scheme. The safety rules include load safety range, action prohibition, fatigue state adaptability and action combination safety. If the draft training plan fails the security rule verification, it will be regenerated until it passes the verification to obtain the personalized strength training plan.

8. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 1, characterized in that, After acquiring the user's multimodal training data, the process also includes: The acquired multimodal training data is preprocessed, including clock synchronization and time alignment using interpolation and resampling. Wavelet threshold denoising was used to reduce noise in inertial measurement unit data and surface electromyography data. Z-score normalization was used to normalize heart rate variability data and subjective fatigue level data. Missing data were filled using multiple imputation methods. The preprocessed multimodal training data is divided into sequence segments in three stages—pre-training, during-training, and post-training—according to a time window for subsequent feature extraction.

9. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 6, characterized in that, The generation of the personalized strength training program based on the reinforcement learning model includes: The individual training state representation and the digital twin simulation state are used as the state space of the reinforcement learning model, and the training action selection, load intensity, training capacity, rest time between sets, recovery strategy and training frequency are used as the action space of the reinforcement learning model. Using strength growth, fatigue recovery, and injury risk as multi-objective reward functions, the reinforcement learning model is optimized using a proximal policy optimization algorithm or a soft actor critic algorithm to output a dynamic training policy. The dynamic training strategy is used as the personalized strength training program.

10. The method for fusion analysis of strength training data and generation of personalized training programs according to claim 9, characterized in that, The process involves updating the model parameters of the musculoskeletal digital twin based on the actual training data and the recovery data, adjusting the causal edge weights in the individual training causal model based on actual strength growth and fatigue recovery results, and adjusting the weight coefficients of each objective in the multi-objective reward function, including: The model parameters of the musculoskeletal digital twin are updated based on the actual training data and the recovery data using either the extended Kalman filter method or the Bayesian optimization method. The causal edge weights of the individual training causal model are updated based on actual strength growth and fatigue recovery results; The prediction error of the reward function is calculated based on the actual strength growth and fatigue recovery results, and the weight coefficients of each objective in the multi-objective reward function are adjusted based on the prediction error of the reward function.