Method and system for generating a large model exercise regimen with trusted image dynamic masking

Abstract

The present application relates to a method and system for generating a large model exercise health plan with trusted image dynamic masking, comprising: collecting multi-modal exercise health data and calculating original modal confidence, using subjective-objective conflict correction vectors to modify the confidence and gate fusion to generate a trusted multi-modal exercise health image; determining the current load state by a deterministic load state machine independent of the sports big model to generate a dynamic constraint mask matrix; performing mask access judgment on the target action, and when prohibited, searching for a replacement action based on a knowledge graph and constructing an allowed combination space; inputting the allowed combination space, the image and the load state into the sports big model to generate a structured plan, and using the same mask matrix to perform reverse verification on the plan, intercepting out-of-bound outputs and triggering reconstruction. The present application also relates to a system for implementing the method. The present application solves the problems of insufficient confidence of multi-modal data and uncontrollable safety boundary of large model generation, significantly improving the safety, stability and traceability of exercise health plans.

Description

Technical Field

[0001] This invention relates to the fields of intelligent sports and health management, personalized training program generation, multimodal data fusion, large model security control, and sports training risk constraint technology, specifically to a method and system for generating large model sports and health programs using a trusted profile dynamic mask. Background Technology

[0002] With the development of wearable devices, fitness testing terminals, motion video recognition, and large language models, exercise and health plans can be automatically generated based on user vital signs, exercise goals, historical exercise records, and multi-source sensor data. Existing solutions typically encode user static attributes, wearable time-series signals, video motion features, or subjective questionnaire data into a unified state vector, and then combine this with knowledge graphs, rule bases, reinforcement learning, or large language models to generate training plans. The output of this invention is used for exercise and health management and training suggestion generation, and is not used for disease diagnosis or disease treatment decisions.

[0003] The shortcomings of existing technologies are as follows: 1. Sports and health data exhibit significant heterogeneity from multiple sources and varying degrees of reliability. Physical fitness test data is typically infrequent but from stable sources; wearable data is frequent but easily affected by wearing position and device drift; sports video data reflects movement quality but is influenced by shooting angle; subjective fatigue and pain feedback exhibit individual differences and lag. Existing technologies that directly merge these data without distinguishing their reliability can easily lead to user profile contamination, resulting in subsequent solutions being based on distorted user states.

[0004] 2. Existing motion scheme generation methods based on large models typically rely on cue word constraints or post-generation text review. Due to the inherent uncertainty of the output of large models, issues such as assigning forbidden actions, exceeding intensity limits, insufficient recovery intervals, excessive continuous high-intensity training, or unreasonable action replacements may still occur. Relying solely on natural language cue words cannot guarantee that the generated results will always meet safety boundaries, and post-generation review can easily lead to problems such as scattered constraint sources, inconsistencies between review rules and generation rules, and untraceable error correction paths.

[0005] 3. There is a lack of methods to convert the safety boundary into a machine-executable combinatorial space before generation, which makes it impossible to generate large models based solely on a combination of allowed actions, allowed strengths, allowed training times, and allowed recovery intervals. Furthermore, there is a lack of mechanisms to perform reverse verification and intercept out-of-bounds outputs using structured output constraints and the same dynamic constraint mask. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method and system for generating large-scale motion and health schemes using a dynamic mask for a reliable profile. The aim is to solve the problems of insufficient reliability of multimodal motion and health profiles, uncontrollable large-scale model generation space, unstable load state switching, and lack of interpretable paths for action replacement.

[0007] To solve the above problems, the technical solution provided by the present invention is as follows: A method for generating a large-scale sports and health plan using a dynamic mask for a trusted profile includes the following steps: S100. Collect multimodal exercise and health data of the target user, including body measurement data, wearable device data, exercise video data, subjective fatigue data, pain feedback data, past injury data, exercise goal data, and historical execution feedback data; S200. Perform timestamp alignment, unit standardization, outlier filtering, missing value marking, and modal feature extraction on the multimodal exercise and health data, and calculate the original modal confidence of each modal data respectively; S300. Generate a subjective-objective conflict correction vector based on a preset conflict relationship between subjective and objective data; wherein, the subjective data includes subjective fatigue score, training intention, pain feedback set, and subjective recovery insufficiency score; the objective data includes motion deviation degree obtained from exercise video, heart rate zone deviation degree obtained from wearable device, sleep recovery index, heart rate variability index, resting heart rate deviation, and risk joint set obtained from video motion risk identification. S400. The original modal confidence of each modal data is corrected using the subjective-objective conflict correction vector to obtain the corrected modal confidence; S500. Map each modal feature to a shared feature space of a unified dimension, and perform gated fusion on each mapped modal feature based on the modified modal confidence to generate a credible multimodal sports and health profile. The credible multimodal sports and health profile includes a physical fitness level profile, a movement stability profile, a sports load profile, an injury risk profile, a recovery status profile, and a sports preference profile. S600. A deterministic load state machine, independent of the sports big model, calculates the load risk value based on the credible multimodal exercise health profile, and determines the current load state based on the entry threshold, exit threshold, and state lock-in period; the sports big model uses structured prompt templates for context input; S700. Generate a dynamic constraint mask matrix based on the current load status, forbidden action set, candidate action whitelist, training intensity upper limit, minimum recovery interval, continuous high-intensity training limit, and target matching conditions. , where a represents the candidate action, i represents the training intensity level, t represents the training time, and r represents the recovery interval; S800. Perform mask admission determination on the target action or initial candidate action; when the target action is prohibited by the dynamic constraint mask matrix, search for replacement actions that satisfy the dynamic constraint mask matrix based on the sports knowledge graph, and add the replacement actions that pass the admission determination to the candidate action set; S900. Construct an allowed combination space based on the dynamic constraint mask matrix and the candidate action set updated by the replacement action, and input the allowed combination space, the credible multimodal motion health profile, the current load state, the motion target data, and the output format requirements into the sports big model in a structured constraint manner, so that the sports big model generates a structured motion health plan based on the allowed combination space; the sports big model generates the structured motion health plan only based on the allowed combination space and enumeration fields written in the prompt context, and does not undertake the state transition determination of the deterministic load state machine; S1000. The structured sports and health program generated by the sports big model is parsed, and the dynamic constraint mask matrix is ​​used to perform reverse verification on the actions, intensity, training time and recovery interval in the structured sports and health program; when the generated result contains combinations outside the allowed combination space or lacks necessary identifiers, the corresponding program item is intercepted, and the candidate space is reconstructed or regenerated. S1100. Collect execution feedback data after the target user executes the structured exercise and health program, and update the trusted multimodal exercise and health profile, the corrected modal confidence, the parameters of the deterministic load state machine, and the dynamic constraint mask matrix based on the execution feedback data.

[0008] Preferably, the original modal confidence level in steps S200 to S400 is expressed by the following formula: Where k is used to characterize the mode number; Used to characterize the credibility of data sources; Used to characterize the freshness of the collection time; Used to characterize data integrity; To characterize data stability, all four scores are normalized to [0,1]. Determined by mapping data source authentication level, import verification status, manual entry status, and historical consistency level; Determined by the exponential decay value of the acquisition time difference; Determined by the ratio of the number of valid fields to the number of fields to be collected; Determined by the inverse normalized value of the variance or drift amplitude within the preset window; , , , Used to characterize the weighting coefficients, and ; Used to characterize minimum confidence level.

[0009] Preferably, the subjective-objective conflict correction vector in step S300 is expressed by the following formula: in, Used to characterize the conflict component between motion deviation and subjective fatigue; Used to characterize the conflicting component between heart rate deviation and subjective fatigue; Used to characterize the conflict component between the degree of inadequate recovery and the willingness to train. Used to characterize the conflict component between the pain site and the joint at risk of motion in video; The method for determining each conflicting component is expressed by the following formula: Where 1(·) is used to characterize the indicator function; PoseDev is used to characterize the degree of movement deviation; HrDev is used to characterize the degree of heart rate zone deviation; RecoveryDeficit is used to characterize the degree of recovery deficiency, and is obtained by weighting sleep gap, decreased heart rate variability, resting heart rate deviation and subjective recovery deficiency score; , , , , These are thresholds used to characterize motion deviation, subjective fatigue, heart rate deviation, degree of inadequate recovery, and training intention, respectively; PainSet represents the set of body parts corresponding to pain feedback; RiskJointSet represents the set of risk joints obtained from video motion risk identification; when PainSet∪RiskJointSet is an empty set... The value of is 0.

[0010] Preferably, the corrected modal confidence in step S400 is expressed by the following formula: in, The conflict suppression coefficient is used to characterize the k-th mode; The weighted value used to characterize one or more conflicting components corresponding to the k-th mode, and when there are multiple conflicting components , Used to characterize the weights of conflicting components.

[0011] Preferably, the gating fusion in step S500 includes the following steps: S510. Project the modal features onto the shared feature space of the same dimension using a modality-specific mapping function, and then calculate the fusion weights, expressed as follows: S520. Generate the portrait feature vector, expressed as follows: Wherein, Z is used to characterize the portrait feature vector; Used to characterize the mode-specific mapping function; Used to characterize the features of the k-th mode; The attention score, used to characterize the output of the cross-modal attention layer, is expressed as follows: Wherein, v, W, and b are used to characterize the attention layer parameters; the extraction of modal features from the motion video data includes: extracting the human skeleton key point sequence from the motion video frames, and calculating joint angles, trunk tilt, left and right limb symmetry, landing impact posture, and motion trajectory stability based on the human skeleton key point sequence to form video modal features; performing dynamic time warping matching between the human skeleton key point sequence and a standard action template, and combining the video modal features to determine the action deviation, action risk level, and video modal stability.

[0012] Preferably, the load risk value in step S600 is expressed by the following formula: Among them, RPEnorm is used to characterize the normalized value of subjective fatigue; RecoveryDeficit is used to characterize the degree of inadequate recovery; Injury is used to characterize the risk weight of past injuries; CompDrop is used to characterize the decline in training completion rate; and Conflict is used to characterize the combined value of subjective and objective conflict. , , , , , , Used to characterize the weighting coefficients, and ; The deterministic load state machine includes an initial assessment state, a low load adaptation state, a routine training state, a high load early warning state, a recovery intervention state, a reassessment state, and a locked observation state. when When entering a more conservative load state than the current load state; when Furthermore, when the exit condition is met for a consecutive preset number of detection cycles, the system exits the more conservative load state; switching to a higher load state is prohibited during the state lock-in period; where s is used to characterize the current load state; Used to characterize the entry threshold function, The exit threshold function is used to characterize the following formulas: in, The baseline risk threshold used to characterize the load state s Used to characterize the hysteresis margin. Used to characterize the exit hysteresis margin, and .

[0013] Preferably, the dynamic constraint mask matrix in step S700 is expressed by the following formula: Wherein, s represents the current load state; g represents the moving target; , , , , , These are used to characterize the gate function, each taking a value of 0 or 1, and are expressed as follows: in, W B is used to represent the whitelist of candidate actions, and A is used to represent the set of forbidden actions. Used to characterize the state-related time window containing training time t The number of consecutive high-intensity training days within a given period; The specific method for searching for the replacement action based on the sports knowledge graph in step S800 is as follows: When the candidate action is prohibited, candidate replacement actions are searched in the sports knowledge graph along the relationships of the same target muscle group, the same training goal, lower impact level, lower intensity level, suitable venue and equipment, and no triggering of previous injury tags; the replacement score of the candidate replacement action is expressed by the following formula: in, The values ​​µ1, µ2, µ3, µ4, and µ5 are used to characterize the candidate replacement action; µ1, µ2, µ3, µ4, and µ5 are used to characterize the weighting coefficients; the replacement score is only used for candidates that have passed the security screening and meet the requirements. The candidate replacement actions are sorted, and the candidate replacement action with the highest replacement score is added to the allowed combination space; when at least two candidate replacement actions have the same replacement score, the final replacement action is determined in sequence according to the rules of lower damage risk label triggering degree, lower impact level, higher site and equipment availability and higher action label ranking. The safety screening includes: the candidate replacement action does not belong to the prohibited action set; the candidate replacement action does not trigger the risk joints corresponding to the target user's past injury tags and current pain location; the impact level of the candidate replacement action is not higher than the impact level of the original target action; the training intensity level of the candidate replacement action does not exceed the upper limit of training intensity corresponding to the current load state; the venue and equipment conditions for the candidate replacement action meet the available resources of the target user; and the recovery interval and continuous high-intensity training limit corresponding to the candidate replacement action both satisfy the dynamic constraint mask matrix. The security screening is determined using a Boolean function, which is expressed by the following formula: The candidate replacement action enters the replacement score sorting process only if and only if the value of the Boolean function is equal to 1; where, Used to characterize the damage risk threshold Thresholds used to characterize the availability of site equipment.

[0014] The allowed combination space is obtained by filtering the Cartesian product of the candidate action set, intensity level set, training time set, and recovery interval set through a dynamic constraint mask, and is expressed as follows: Among them, A c Used to characterize the candidate action set; I c Used to characterize a set of intensity levels; T c Used to characterize the training time set; R c Used to characterize the set of recovery intervals; the action candidates, intensity candidates, training day candidates, and recovery interval candidates input into the sports model are all from the allowed combination space, and The combination of actions is not written into the prompt context of the sports model; the output field of the sports model is constrained by the enumerated values ​​of the candidate actions, the training intensity level, the training time and the recovery interval. When the output combination does not belong to the allowed combination space, it is determined to be an out-of-bounds output.

[0015] Preferably, the structured prompt template input to the sports big model includes a user profile summary, profile version number, profile confidence level, current load status, allowed combination space, dynamic constraint mask matrix identifier, candidate action whitelist, forbidden action set, training intensity upper limit, minimum recovery interval, available time and venue, output field requirements, enumeration constraint requirements, and risk warning requirements; the underlying model carrier of the sports big model is a text generation model based on the Transformer structure, a multimodal generation model, or a domain generation model fine-tuned with sports and health corpus, and receives the structured prompt template through an application programming interface, local inference service, or edge inference service; Each item in the structured exercise and health program includes an action identifier, the training intensity level, the training time, the recovery interval, the state machine state identifier, the dynamic constraint mask matrix identifier, the trigger rule identifier, and the confidence reference identifier.

[0016] Preferably, the reverse verification in step S1000 includes parsing the combination of the candidate action, the training intensity level, the training time, and the recovery interval corresponding to each scheme item, and querying the corresponding value of the dynamic constraint mask matrix. When any scheme item corresponds to M(a,i,t,r)=0 or the above identifier is missing, the verification is determined to fail. The determination of verification failure is used to trigger candidate space reconstruction, action replacement and reselection, or regeneration, and is not directly used as the basis for updating user profile and modal confidence. In step S1100, the corrected modal confidence is updated exponentially, as expressed by the following formula: in, The quality score for the k-th modality obtained from the (n+1)-th execution feedback is weighted by at least two of the following: execution completion rate, target heart rate zone achievement, change in movement deviation, subjective fatigue feedback, and pain feedback; ρ is used to characterize the smoothness coefficient, and 0 < ρ < 1. in, x Used to characterize the input value to be restricted. l and u These are used to characterize the lower and upper limits, respectively; in the formula for exponentially smoothing the updated modified modal confidence, l=T min , u=1.

[0017] A system for implementing the aforementioned method for generating large-model motion health schemes using a dynamic mask for a trusted profile, characterized in that it comprises: The data acquisition module is used to collect multimodal exercise and health data from target users. The confidence calculation module is used to preprocess and extract modal features from the multimodal exercise and health data, and calculate the original modal confidence of each modal data based on the credibility of the data source, the freshness of the collection time, the completeness of the data, and the stability of the data. The conflict correction module is used to generate a subjective-objective conflict correction vector based on a preset conflict relationship between subjective and objective data, and to use the subjective-objective conflict correction vector to correct the original modality confidence to obtain a corrected modality confidence. The subjective-objective conflict correction vector includes a recovery insufficiency degree and a training intention conflict component. The recovery insufficiency degree is obtained by weighting at least two of the following: sleep deficit, decreased heart rate variability, resting heart rate deviation, and subjective recovery insufficiency score. The trusted profile construction module is used to map the features of each modality to a shared feature space of a unified dimension, and to perform gated fusion on the mapped features based on the modified modality confidence to generate a trusted multimodal motion and health profile carrying version number, source identifier, timestamp and profile confidence. The deterministic load state machine module operates independently of the sports big model and is used to calculate the load risk value based on the credible multimodal sports health profile, and determine the current load state based on the entry threshold, exit threshold and state lock-in period. The dynamic constraint mask generation module is used to generate a dynamic constraint mask matrix based on the current load state, the set of forbidden actions, the whitelist of candidate actions, the upper limit of training intensity, the minimum recovery interval, the continuous high-intensity training limit and the target matching condition, and to construct an allowed combination space based on the combinations that satisfy M(a,i,t,r)=1 in the set of candidate actions, the set of intensity levels, the set of training time and the set of recovery interval. The sports knowledge graph module is used to store the relationships between movements, including target muscle groups, training goals, impact levels, venues and equipment, and injury tags. The alternative action search module is used to search for alternative actions that satisfy the dynamic constraint mask matrix based on the sports knowledge graph when the target action is prohibited by the dynamic constraint mask matrix, and to calculate a replacement score according to the similarity of the target muscle group, the similarity of the training target, the availability of venue and equipment, the impact level, and the triggering degree of the injury risk label; the alternative action search module is also used to calculate the Boolean function used for safety screening. When the Boolean function is 1, the replacement action search module inputs the corresponding candidate replacement action into the replacement score sorting process; when the Boolean function is 0, the corresponding candidate replacement action is discarded. The sports big data model generation module is used to input the allowed combination space, profile version number, current load status, dynamic constraint mask matrix identifier, and enumeration field constraints into the sports big data model, enabling the sports big data model to generate a structured sports and health solution based on the allowed combination space and structured output constraints. The sports big data model generation module includes a model carrier interface, which is used to call a Transformer-based text generation model, multimodal generation model, or domain generation model fine-tuned with sports and health corpus deployed on a server, edge computing device, or local computing device, and encapsulate the allowed combination space, enumeration fields, and output format requirements into a machine-parseable structured input. The mask reverse verification module is used to perform reverse verification on the structured sports and health scheme using the same dynamic constraint mask matrix. When any scheme item corresponds to M(a,i,t,r)=0 or lacks necessary identifiers, the verification is determined to be unsuccessful, triggering candidate space reconstruction, replacement action reselection or regeneration. The verification failure itself is not used as the basis for updating user profile or modal confidence. The feedback update module is used to update the trusted multimodal motion health profile, the corrected modal confidence, the parameters of the deterministic load state machine, and the dynamic constraint mask matrix based on the actual execution feedback data of the target user.

[0018] Compared with the prior art, the present invention has the following advantages: 1. This invention uses modal confidence, subjective-objective conflict correction, and gating fusion to prevent body measurements, wearable devices, video, subjective fatigue, pain feedback, and executive feedback from being simply averaged when conflicts occur, thereby reducing the impact of drift or distortion from a single data source.

[0019] 2. This invention separates the determination of safe state from the natural language generation process by using a deterministic load state machine independent of the large sports model, thereby making the load state transition reproducible, auditable, and controllable in engineering.

[0020] 3. This invention uses a dynamic constraint mask matrix M(a,i,t,r) to pre-contain forbidden actions, training intensity limits, recovery intervals, and continuous high-intensity training restrictions as an allowed combination space. It also uses structured output constraints and reverse verification with the same mask to intercept out-of-bounds outputs, thereby reducing the risk of large sports models generating combinations of forbidden actions, intensities, training times, and recovery intervals.

[0021] 4. This invention provides an action replacement search path through a sports knowledge graph and uses the same dynamic constraint mask to perform reverse verification on the replacement results and the final solution, thereby improving the interpretability of the replacement actions and the closed-loop update capability of the solution. Attached Figure Description

[0022] Figure 1This is a schematic diagram of the overall system architecture of a specific embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the dynamic constraint mask generation in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of a deterministic load state machine according to a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the action replacement and generation verification closed loop in a specific embodiment of the present invention.

[0023] The system comprises: 170. Overall System; 100. Data Access and Trusted Storage Unit; 110. Feature Preprocessing Unit; 120. Trusted Profile Feature Matrix Construction Unit; 130. Sports Knowledge Graph Unit; 140. Candidate Action and Load Assessment Unit; 150. Sports Large Model Generation Unit; 160. Feedback and Verification Update Unit; 200. Multimodal Profile and Profile Data Input Unit; 210. Quality Evaluation and State Parameter Unit; 220. Candidate Action Whitelist Gating Unit; 230. Prohibited Action Gating Unit; 240. Training Intensity Gating Unit; 250. Recovery Interval Gating Unit; 260. Continuous High-Intensity Training Gating Unit; 270. Sports Target Matching Gating Unit; 280. Dynamic... State constraint mask matrix; 300. Initial evaluation state, 310. Low load adaptation state, 320. Regular training state, 330. High load warning state, 340. Recovery intervention state, 350. Re-evaluation state, 360. Locked observation state, 370. Conservative transfer path, 380. Recovery transfer path; 400. Injury or restriction label, 410. Sports knowledge graph unit, 420. Target muscle group equivalent mapping unit, 430. Safety screening unit, 440. Action admission judgment unit, 450. Replacement score ranking unit, 460. Allowable combination space construction unit, 470. Sports large model generation unit, 480. Mask reverse verification unit, 490. Final scheme output unit. Detailed Implementation

[0024] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0025] This invention application claims a method for generating a large-model motion health scheme using a trusted dynamic image mask, comprising the following steps: S100. Collect multimodal exercise and health data of the target user. The multimodal exercise and health data includes physical test data, wearable device data, exercise video data, subjective fatigue data, pain feedback data, past injury data, exercise goal data, and historical execution feedback data.

[0026] S200. Perform timestamp alignment, unit standardization, outlier filtering, missing value labeling, and modal feature extraction on multimodal exercise and health data, and calculate the original modal confidence of each modal data separately.

[0027] S300. Generate a subjective-objective conflict correction vector based on the preset conflict relationship between subjective and objective data; wherein, subjective data includes subjective fatigue score, training intention, pain feedback set and subjective recovery insufficiency score; objective data includes movement deviation degree obtained from exercise video, heart rate zone deviation degree obtained from wearable device, sleep recovery index, heart rate variability index, resting heart rate deviation and risk joint set obtained from video movement risk identification.

[0028] S400. The original modal confidence of each modal data is corrected using the subjective-objective conflict correction vector to obtain the corrected modal confidence.

[0029] S500. Map each modal feature to a shared feature space of a unified dimension, and perform gated fusion on each mapped modal feature based on the modified modal confidence to generate a credible multimodal sports and health profile. The credible multimodal sports and health profile includes a physical fitness profile, a movement stability profile, a sports load profile, an injury risk profile, a recovery status profile, and a sports preference profile.

[0030] S600. A deterministic load state machine, independent of the sports big model, calculates the load risk value based on a reliable multimodal motion health profile, and determines the current load state based on the entry threshold, exit threshold, and state lock-in period; the sports big model uses structured prompt templates for context input.

[0031] S700. Based on the current load status, forbidden action set, candidate action whitelist, training intensity limit, minimum recovery interval, continuous high-intensity training limit, and target matching conditions, generate a dynamic constraint mask matrix 280. , where a represents the candidate action, i represents the training intensity level, t represents the training time, and r represents the recovery interval.

[0032] S800. Perform mask admission determination on the target action or initial candidate action; when the target action is prohibited by the dynamic constraint mask matrix 280, search for replacement actions that satisfy the dynamic constraint mask matrix 280 based on the sports knowledge graph, and add the replacement actions that pass the admission determination to the candidate action set.

[0033] S900. Construct an allowed combination space based on the dynamic constraint mask matrix 280 and the candidate action set updated by the replacement action. Input the allowed combination space, the credible multimodal motion health profile, the current load state, the motion target data, and the output format requirements into the sports big model in a structured constraint manner, so that the sports big model generates a structured motion health solution based on the allowed combination space. The sports big model only generates a structured motion health solution based on the allowed combination space and enumeration fields written in the prompt context, and does not undertake the state transition determination of the deterministic load state machine.

[0034] S1000. Analyze the structured sports and health programs generated by the sports big model, and use the dynamic constraint mask matrix 280 to perform reverse verification on the actions, intensity, training time and recovery interval in the structured sports and health programs; when the generated results contain combinations outside the allowed combination space or lack necessary identifiers, the corresponding program items are intercepted, and the candidate space is reconstructed or regenerated.

[0035] S1100. Collect execution feedback data after the target user executes the structured exercise and health program, and update the trusted multimodal exercise and health profile, correct the modal confidence, parameters of the deterministic load state machine, and dynamic constraint mask matrix 280 based on the execution feedback data.

[0036] It should be noted that, as Figure 1 As shown, the overall system 170 includes a data access and trusted storage unit 100, a feature preprocessing unit 110, a trusted profile feature matrix construction unit 120, a sports knowledge graph unit 130, a candidate action and load assessment unit 140, a sports large model generation unit 150, and a feedback and verification update unit 160. The units are linked through profile version numbers, state machine state identifiers, mask matrix identifiers, and trigger rule identifiers.

[0037] The data access and trusted storage unit 100 collects and stores the target user's physical test data, wearable device data, sports video data, subjective fatigue data, pain feedback data, past injury data, sports goal data, and execution feedback data, and performs compliance verification and secure storage on the data sources. The physical test data includes height, weight, body fat percentage, lung capacity, grip strength, flexibility, endurance test scores, and strength test scores; the wearable device data includes heart rate, heart rate variability, steps, cadence, sleep duration, sleep quality, blood oxygen saturation, and energy consumption; and the sports video data includes training movement videos, gait videos, and videos of specific technical movements.

[0038] The feature preprocessing unit 110 performs denoising, timestamp alignment, unit standardization, outlier filtering, and missing value labeling on the above data. For motion video data, it extracts the human skeleton key point sequence and calculates joint angles, trunk tilt, left and right limb symmetry, landing impact posture, and motion trajectory stability based on the human skeleton key point sequence to form video modal features; it then performs dynamic time warping matching between the human skeleton key point sequence and a standard motion template, and combines the video modal features to obtain motion deviation PoseDev, motion risk level, and video modal stability.

[0039] It should be noted that the raw modal confidence scores for each modality are obtained by weighting and truncating the data source reliability, acquisition time freshness, data completeness, and data stability. Data source reliability can be mapped according to device authentication status, import verification status, manual entry status, or historical consistency level. The values ​​within the interval; the freshness of the collection time can be determined by the exponential decay value of the collection time difference; the data completeness can be determined by the ratio of the number of valid fields to the number of fields to be collected; and the data stability can be determined by the inverse normalization value of the variance or drift amplitude within the preset window.

[0040] It should be further explained that the original modal confidence in steps S200 to S400 is expressed by the following formula (1): (1) Where k is used to characterize the mode number; Used to characterize the credibility of data sources; Used to characterize the freshness of the collection time; Used to characterize data integrity; To characterize data stability, all four scores are normalized to [0,1]. Determined by mapping data source authentication level, import verification status, manual entry status, and historical consistency level; Determined by the exponential decay value of the acquisition time difference; Determined by the ratio of the number of valid fields to the number of fields to be collected; Determined by the inverse normalized value of the variance or drift amplitude within the preset window; , , , Used to characterize the weighting coefficients, and ; Used to characterize minimum confidence level.

[0041] In this specific embodiment, the credibility of the data source The mapping is determined according to the data source level: 1.0 for data directly collected by certified devices, 0.8 for imported records that have been signed or verified, 0.6 for manually entered data that is consistent with historical records, 0.3 for data from unverified sources, and 0.1 for data that conflicts with historical consistency. The freshness of collection time, data integrity, and data stability are calculated as follows (2): (2) in, Let k be the time difference between the acquisition of the k-th mode and the current calculation time. Let be the time decay constant corresponding to the k-th mode. The number of valid fields. The number of fields to be sampled is σk, where σk is the fluctuation or drift amplitude within the preset window. This represents the maximum permissible fluctuation range.

[0042] Data source credibility It is obtained by mapping based on device authentication status, manual entry marks, historical consistency level, or data source whitelist.

[0043] It should be noted that the subjective-objective conflict correction vector includes conflict components related to movement deviation and subjective fatigue, heart rate deviation and subjective fatigue, inadequate recovery and training motivation, and pain location and video-based movement risk joints. The values ​​of each conflict component are determined by an indicator function and set similarity. The degree of inadequate recovery is obtained by weighting at least two of the following: sleep deficit, decreased heart rate variability, resting heart rate deviation, and subjective inadequate recovery score.

[0044] It should be further explained that the subjective-objective conflict correction vector in step S300 is expressed by the following formula (3): (3) in, Used to characterize the conflict component between motion deviation and subjective fatigue; Used to characterize the conflicting component between heart rate deviation and subjective fatigue; Used to characterize the conflict component between the degree of inadequate recovery and the willingness to train. Used to characterize the conflict component between the pain site and the joint at risk of motion in video.

[0045] The method for determining each conflicting component is expressed by the following formula (4): (4) Where 1(·) is used to characterize the indicator function; PoseDev is used to characterize the degree of movement deviation; HrDev is used to characterize the degree of heart rate zone deviation; RecoveryDeficit is used to characterize the degree of recovery deficiency, and is obtained by weighting sleep gap, decreased heart rate variability, resting heart rate deviation and subjective recovery deficiency score; , , , , These are thresholds used to characterize motion deviation, subjective fatigue, heart rate deviation, degree of inadequate recovery, and training intention, respectively; PainSet represents the set of body parts corresponding to pain feedback; RiskJointSet represents the set of risk joints obtained from video motion risk identification; when PainSet∪RiskJointSet is an empty set... The value of is 0.

[0046] RecoveryDeficit is calculated as follows (5): (5) Where SleepNorm is the normalized value for sleep recovery, HRVNorm is the normalized value for heart rate variability, RHRDev is the resting heart rate deviation, RecoverSubj is the normalized value for subjective inadequate recovery score, and λ1 to λ4 are weighting coefficients that sum to 1. When PainSet∪RiskJointSet is an empty set, Cpain_pose is 0. The above conflict relationship is used to identify situations where objective risk is high but subjective feedback is low, recovery is insufficient but training intention is high, or pain markers are inconsistent with video risk joints.

[0047] In the above formula, PoseDev is obtained by normalizing the deviation between the video action key points and the standard action template, HrDev is obtained by normalizing the deviation between the target heart rate zone and the measured heart rate zone, RPE uses a subjective fatigue scale of 0 to 10, and Will uses the normalized value of training intention in the range of 0 to 1.

[0048] It should be noted that the modified modal confidence in step S400 is expressed by the following formula (6): (6) in, The conflict suppression coefficient is used to characterize the k-th mode; The weighted value used to characterize one or more conflicting components corresponding to the k-th mode, and when there are multiple conflicting components , This is used to characterize the weights of conflict components. Using the above method, video, heart rate, sleep, subjective fatigue, and pain feedback are not simply averaged when conflicts occur, but rather their credibility is suppressed according to the source of the conflict.

[0049] It should be noted that the gating fusion in step S500 includes the following steps: S510. Project the modal features to a shared feature space of uniform dimension using a modality-specific mapping function, and then calculate the fusion weights as expressed in the following formula (7): (7) S520. Generate the portrait feature vector, expressed as follows (8): (8) Where Z is used to represent the portrait feature vector; Used to characterize mode-specific mapping functions; Used to characterize the features of the k-th mode; The attention score used to characterize the output of the cross-modal attention layer is expressed by the following equation (9): (9) Here, v, W, and b are used to characterize the attention layer parameters; modal feature extraction from motion video data includes: extracting human skeletal keypoint sequences from motion video frames, and calculating joint angles, trunk tilt, left and right limb symmetry, landing impact posture, and motion trajectory stability based on the human skeletal keypoint sequences to form video modal features; dynamically warping and matching the human skeletal keypoint sequences with standard action templates, and combining the video modal features to determine the action deviation degree, action risk level, and video modal stability. During the fusion process, the sports knowledge graph unit 130 provides the trustworthy profile feature matrix construction unit 120 with the relationship constraints between actions, muscle groups, equipment, and injury tags for the structured alignment of profile features.

[0050] The fused credible multimodal sports and health profile includes at least a physical fitness profile, a movement stability profile, a sports load profile, an injury risk profile, a recovery status profile, and a sports preference profile, and is stored together with the profile version number, data source identifier, collection timestamp, modal confidence level, and conflict trigger identifier.

[0051] The candidate action and load assessment unit 140, based on a reliable multimodal exercise health profile, performs load statistics and intensity distribution assessment on various exercises (such as running, cycling, swimming, strength training, etc.) in the candidate action library, providing action-level load parameters for subsequent state machine determination and mask generation.

[0052] It should be noted that the load risk value in step S600 is expressed by the following formula (10): (10) Among them, RPEnorm is used to characterize the normalized value of subjective fatigue; RecoveryDeficit is used to characterize the degree of inadequate recovery; Injury is used to characterize the risk weight of past injuries; CompDrop is used to characterize the decline in training completion rate; and Conflict is used to characterize the combined value of subjective and objective conflict. , , , , , , Used to characterize the weighting coefficients, and .

[0053] like Figure 3 As shown, the deterministic load state machine includes initial assessment state 300, low load adaptation state 310, routine training state 320, high load warning state 330, recovery intervention state 340, reassessment state 350, and locked observation state 360. State transitions employ a hysteresis mechanism: when... When the load state is changed to a more conservative one than the current load state, for example, from the normal training state 320 to the high load warning state 330; when Furthermore, when the exit condition is met for a consecutive preset number of detection cycles, the system exits the more conservative load state; switching to a higher load state is prohibited during the state lockout period; where s is used to characterize the current load state; Used to characterize the entry threshold function, The exit threshold function is used to characterize the function, and is expressed by the following formula (11): (11) in, The baseline risk threshold used to characterize the load state s Used to characterize the hysteresis margin. Used to characterize the exit hysteresis margin, by making This creates a state transition hysteresis interval, reducing repeated switching of load states near the threshold. When the high load warning state 330 is continuously triggered, the system enters the recovery intervention state 340. After the recovery intervention ends, it enters the reassessment state 350. After the reassessment state 350 or the low load adaptation state 310, the system enters the lockout observation state 360. During the lockout period L corresponding to the lockout observation state 360, the system is prohibited from switching to a higher load state. After the lockout period expires, the system is reassessed according to the transition rules.

[0054] It should be further explained that the deterministic load state machine adopts a hysteresis mechanism where the entry threshold is higher than the exit threshold, and sets a state lock-in period. During the lock-in period, switching to a higher load state is prohibited to avoid the load state from oscillating repeatedly around the threshold.

[0055] It should be noted that, as Figure 2 As shown, the generation process of the dynamic constraint mask is as follows: the multimodal portrait and portrait data input unit 200 outputs a reliable portrait and its derived parameters; the quality evaluation and state parameter unit 210 outputs the current load state s, training intensity upper limit, minimum recovery interval and other rule parameters; the candidate action whitelist gating unit 220, the forbidden action gating unit 230, the training intensity gating unit 240, the recovery interval gating unit 250, the continuous high intensity training gating unit 260 and the moving target matching gating unit 270 respectively output 0 / 1 judgment vectors for the candidate combinations, and the dynamic constraint mask matrix 280 is obtained by multiplying each judgment vector element by element.

[0056] The dynamic constraint mask matrix 280 in step S700 is expressed by the following formula (12): (12) Where s represents the current load state; g represents the moving target; , , , , , They are used to characterize the gate functions, each taking a value of 0 or 1, and are expressed as follows (13): (13) in, W B is used to represent the whitelist of candidate actions, and B is used to represent the set of forbidden actions. Used to characterize the state-related time window containing training time t The number of consecutive high-intensity training days within the period.

[0057] The specific method for searching for replacement actions based on the sports knowledge graph in step S800 is as follows: When a candidate action is prohibited, the system searches for candidate replacement actions along the relationships of the same target muscle group, the same training goal, lower impact level, lower intensity level, suitable venue and equipment, and no triggering of previous injury tags in the sports knowledge graph, and then constructs the final allowed combination space. When the number of candidate combinations exceeds the preset upper limit, the system generates and filters the final allowed combination space layer by layer according to the action category, current load state, or sports target tag. Combinations in the allowed combination space are written into the structured prompt template of the sports big model, and combinations with M(a,i,t,r)=0 do not enter the big model context. Thus, the safety rule is not the generated text suggestion, but the candidate space boundary before generation; at the same time, the system intercepts the out-of-bounds output of the big model through structured fields, enumeration identifiers, and reverse verification mechanisms.

[0058] The sports large-scale model generation unit 470 receives user profile summaries, profile version numbers, profile confidence levels, current load status, allowed combination space, dynamic constraint mask matrix 280 identifiers, candidate action whitelist, prohibited action set, training intensity upper limit, minimum recovery interval, available time and venue, output field requirements, and enumeration constraint requirements. Based on the allowed combination space, it generates structured sports and health plans. Each plan item includes an action identifier, intensity level, training time, recovery interval, state machine state identifier, dynamic constraint mask matrix 280 identifier, trigger rule identifier, and confidence reference identifier.

[0059] The feedback and verification update unit 160 performs reverse verification after the solution output: it parses each solution item in the structured sports and health solution, extracts the corresponding (a,i,t,r) combination, and queries M(a,i,t,r). When any solution item corresponds to M(a,i,t,r)=0, or is missing a state machine state identifier, dynamic constraint mask matrix 280 identifier, trigger rule identifier, or confidence reference identifier, the verification is deemed to have failed, and a correction instruction containing anomalies, trigger rules, and correction directions is generated. The correction instruction is fed back to the allowable combination space construction unit 460 or the sports large model generation unit 470 to trigger candidate space reconstruction, replacement action reselection, or regeneration. Verification failure indicates that the generated result does not meet the current mask constraints and is not directly used as the basis for updating the user profile and modal confidence; the update basis for the user profile and modal confidence comes from the actual execution feedback data of the target user.

[0060] like Figure 4As shown, when the target user has an injury or restriction label 400 (e.g., a previous knee injury), the action replacement and generation verification closed loop operates as follows: The sports knowledge graph unit 410 stores the relationship edges between actions and target muscle groups, training objectives, impact levels, venues and equipment, and injury labels; the target muscle group equivalence mapping unit 420 determines candidate actions equivalent to the restricted action target muscle group based on the graph; the safety screening unit 430 screens candidate actions based on constraints such as reduced impact level, reduced intensity level, and not triggering injury labels; the action admission judgment unit 440 outputs the admission judgment result for each candidate action; the replacement score sorting unit 450 calculates and sorts only the replacement scores of candidate replacement actions that pass the safety screening and satisfy M(a',i,t,r)=1.

[0061] The substitution score of the candidate substitution action is expressed by the following formula (14): (14) in, Used to characterize candidate replacement actions; µ1, µ2, µ3, µ4, and µ5 are used to characterize weighting coefficients; the replacement score is only used for those that have passed the security screening and meet the requirements. The candidate replacement actions are sorted, and the candidate replacement action with the highest replacement score is added to the allowed combination space; when at least two candidate replacement actions have the same replacement score, the final replacement action is determined in order of lower damage risk label triggering degree, lower impact level, higher site and equipment availability, and higher action label ranking.

[0062] Safety screening includes: the candidate replacement action does not belong to the prohibited action set; the candidate replacement action does not trigger the target user's past injury tags and the risk joints corresponding to the current pain location; the impact level of the candidate replacement action is not higher than the impact level of the original target action; the training intensity level of the candidate replacement action does not exceed the upper limit of training intensity corresponding to the current load state; the venue and equipment conditions for the candidate replacement action meet the target user's available resources; and the recovery interval and continuous high-intensity training restrictions corresponding to the candidate replacement action both meet the dynamic constraint mask matrix 280. Candidate replacement actions that do not simultaneously meet the above rules will not be included in the replacement score ranking.

[0063] Security screening is performed using a Boolean function, which is expressed by the following formula (15): (15) Candidate replacement actions are included in the replacement score sorting only if and only if the value of the Boolean function is equal to 1; where, Used to characterize the damage risk threshold This is used to characterize the availability threshold of venue equipment. When the value of the Boolean function is equal to 0, the replacement action search module discards the candidate action.

[0064] It should be further explained that the dynamic constraint mask matrix 280 is obtained by element-wise multiplication of the candidate action whitelist gating, forbidden action gating, training intensity gating, recovery interval gating, continuous high-intensity training gating, and motion target matching gating, and the value of each gating function is 0 or 1.

[0065] It should be further explained that when the target action is prohibited, candidate replacement actions are searched along the relationship edges in the sports knowledge graph that are the same target muscle group, the same training goal, lower impact level, lower intensity level, satisfactory venue and equipment, and do not trigger previous injury tags. The optimal replacement action that meets the mask constraints is then selected after sorting by replacement score.

[0066] It should be further explained that the allowed combination space is obtained by filtering the Cartesian product of the candidate action set, intensity level set, training time set, and recovery interval set through a dynamic constraint mask, as expressed by the following equation (16): (16) Among them, A c Used to characterize the candidate action set; I c Used to characterize a set of intensity levels; T c Used to characterize the training time set; R c Used to characterize the set of recovery intervals; the action candidates, intensity candidates, training day candidates, and recovery interval candidates input into the sports model all come from the allowed combination space, and The combinations are not written into the prompt context of the sports big model; the output field of the sports big model is constrained by enumerated values ​​of candidate actions, training intensity level, training time and recovery interval. When the output combination does not belong to the allowed combination space, it is judged as an out-of-bounds output.

[0067] The permissible combination space construction unit 460 constructs a tensor-form permissible combination space by combining actions, intensity, training time, and recovery intervals through admission judgment and ranking; the sports large model generation unit 470 generates a structured sports and health program based on the permissible combination space and structured output constraints; the mask reverse verification unit 480 verifies the program item by item using the same dynamic constraint mask matrix 280, and generates a correction instruction when the verification fails and feeds it back to the permissible combination space construction unit 460 or the sports large model generation unit 470 to reconstruct the candidate space or regenerate; the final program output unit 490 outputs the structured sports and health program that has passed the verification, and collects data on heart rate changes, training completion rate, movement deviation, subjective fatigue score, pain feedback, and recovery status after execution.

[0068] It should be noted that the structured prompt templates input to the sports big data model include user profile summary, profile version number, profile confidence level, current load status, allowed combination space, dynamic constraint mask matrix 280 identifier, candidate action whitelist, forbidden action set, training intensity upper limit, minimum recovery interval, available time and venue, output field requirements, enumeration constraint requirements, and risk warning requirements. The underlying model carrier of the sports big data model is a text generation model based on the Transformer structure, a multimodal generation model, or a domain generation model fine-tuned with sports and health corpus, and receives structured prompt templates through application programming interfaces, local inference services, or edge inference services.

[0069] The underlying model carrier does not directly rewrite the load state machine parameters and dynamic constraint mask matrix 280; the load state machine and mask matrix are calculated by an independent rule module.

[0070] Each item in the structured sports and health program includes an action identifier, training intensity level, training time, recovery interval, state machine state identifier, dynamic constraint mask matrix 280 identifier, trigger rule identifier, and confidence reference identifier.

[0071] It should be noted that the reverse verification in step S1000 includes parsing the combination of candidate actions, training intensity level, training time, and recovery interval corresponding to each scheme item, and querying the corresponding dynamic constraint mask matrix 280 value. When any scheme item corresponds to M(a,i,t,r)=0 or the above identifier is missing, the verification is determined to fail. The determination of verification failure is used to trigger candidate space reconstruction, action replacement and reselection, or regeneration, and is not directly used as the basis for updating user profile and modal confidence.

[0072] In step S1100, the modified mode confidence is updated exponentially, as expressed by the following formula (17): (17) in, The quality score for the k-th modality obtained from the (n+1)-th execution feedback is weighted by at least two of the following: execution completion rate, target heart rate zone achievement, change in movement deviation, subjective fatigue feedback, and pain feedback; ρ is used to characterize the smoothness coefficient, and 0 < ρ < 1. in, x Used to characterize the input value to be restricted. l and u These are used to characterize the lower and upper limits, respectively; in the aforementioned update formula, l = Tmin , u=1. Through this update method, the system can both absorb new execution feedback and avoid drastic fluctuations in the profile caused by a single abnormal feedback.

[0073] in, Calculate as follows (18): (18) in, To achieve the completion rate, For the achievement of the target heart rate zone, This represents the change in motion deviation. The consistency between subjective fatigue and planned intensity. For pain feedback safety, to The feedback quality weights are summed to 1.

[0074] It should be further explained that the execution feedback data is used to update and correct the modal confidence in an exponentially smooth manner, so that the profile can both absorb new execution feedback and avoid drastic fluctuations in the profile caused by a single abnormal feedback; the failure of reverse verification itself is not directly used as the basis for updating the profile or modal confidence.

[0075] In one specific embodiment, the parameters can be valued as follows: confidence weights are α1=0.35, α2=0.25, α3=0.20, α4=0.20, and the minimum confidence level T is... min =0.2; Time decay constant τ k Physical fitness test data was collected over 30 days, wearable time-series data over 24 hours, and motion video data over 7 days; the inadequate recovery weights were set to λ1=0.30, λ2=0.30, λ3=0.20, and λ4=0.20; the conflict threshold was set to θ. pose =0.30、θ rpe =4 (RPE uses a 0-10 scale), θ hr =0.15, θ rec =0.60, θ will =0.70; Load risk weights are η1=0.20, η2=0.15, η3=0.15, η4=0.15, η5=0.15, η6=0.10, η7=0.10; Entry threshold T for transitioning from normal training state 320 to high load warning state 330 enter =0.65, Exit Threshold T exit =0.45, the number of consecutive detection periods for the exit condition N=3, the lockout period L=7 days; the replacement score weights are μ1=0.30, μ2=0.25, μ3=0.15, μ4=0.15, μ5=0.15; the target matching threshold θ goal =0.50; Smoothing coefficient ρ=0.80.

[0076] The baseline risk threshold can be T. normal0 =0.55, T warning0 =0.65, T recovery0 =0.75, entering the hysteresis margin Δ senter A value of 0.05 to 0.10 can be used to exit the hysteresis margin Δ. sexit The threshold can be taken as 0.10 to 0.20; the injury risk threshold θrisk in safety screening can be taken as 0.40, and the site and equipment sufficiency threshold θ equip A value of 0.60 can be used; the feedback quality weights can be ξ1=0.25, ξ2=0.20, ξ3=0.20, ξ4=0.15, and ξ5=0.20.

[0077] The values ​​mentioned above are merely examples, and those skilled in the art can adjust them according to the user group, sport, and security strategy.

[0078] In one application example, the target users were individuals with a history of knee injury who were trying to lose weight. After collecting their body composition analysis, wearable devices, running training videos, and subjective fatigue data, the system found that the video motion deviation PoseDev = 0.42 was higher than θ. pose Subjective fatigue RPE=3 is lower than θ rpe Triggering conflict component C pose_rpe =1, the confidence levels of video and subjective modalities are suppressed accordingly; simultaneously, the system calculates the recovery deficiency degree (RecoveryDeficit) and incorporates it into the load risk value. The deterministic load state machine calculation yields R=0.71>T. enter =0.65, the state transitions from the normal training state 320 along the conservative transition path 370 to the high-load warning state 330 and enters the lockout period. Based on this, the dynamic constraint mask sets high-impact running movements to M(a,i,t,r)=0; the replacement search finds swimming and elliptical training as candidate replacement movements along the knowledge graph that are equivalent to the target muscle group, have lower impact levels, and are supported by suitable equipment. Among these, swimming has the highest SubScore and satisfies M(a',i,t,r)=1, and is added to the allowed combination space. The sports big model generates a weekly plan containing swimming, elliptical training, and core strength training based on this allowed combination space and structured output constraints; after back-checking to confirm that the mask value of all plan items is 1, the plan is output. After one week of user execution, the execution feedback data updates the profile and confidence level in an exponentially smoothed manner. The state machine, upon the expiration of the lockout period and R... <T exit After satisfying the exit condition for N consecutive cycles, the system returns to the normal training state 320 along the recovery transfer path 380.

[0079] This application also claims a system for generating large-model motion health schemes using a dynamic mask for a trusted profile, characterized by comprising: The data acquisition module is used to collect multimodal exercise and health data from the target user.

[0080] The confidence calculation module is used to preprocess multimodal exercise and health data and extract modal features, and calculate the original modal confidence of each modality based on the credibility of the data source, the freshness of the collection time, the completeness of the data, and the stability of the data.

[0081] The conflict correction module is used to generate a subjective-objective conflict correction vector based on a preset conflict relationship between subjective and objective data, and to use the subjective-objective conflict correction vector to correct the original modality confidence to obtain the corrected modality confidence. The subjective-objective conflict correction vector includes a recovery insufficiency degree and a training intention conflict component. The recovery insufficiency degree is obtained by weighting at least two of the following: sleep gap, decreased heart rate variability, resting heart rate deviation, and subjective recovery insufficiency score.

[0082] The trusted profile building module is used to map the features of each modality to a shared feature space of a unified dimension. Based on the modified modality confidence, the mapped features of each modality are gated and fused to generate a trusted multimodal sports and health profile carrying version number, source identifier, timestamp and profile confidence.

[0083] The deterministic load state machine module operates independently of the sports big model. It is used to calculate the load risk value based on the credible multimodal sports health profile and determine the current load state based on the entry threshold, exit threshold and state lock-in period.

[0084] The dynamic constraint mask generation module is used to generate a dynamic constraint mask matrix 280 based on the current load state, the set of forbidden actions, the whitelist of candidate actions, the upper limit of training intensity, the minimum recovery interval, the limit of continuous high-intensity training, and the target matching condition. It also constructs an allowed combination space based on the combinations that satisfy M(a,i,t,r)=1 in the set of candidate actions, the set of intensity levels, the set of training time, and the set of recovery intervals.

[0085] The sports knowledge graph module is used to store the relationships between movements, including target muscle groups, training goals, impact levels, venues and equipment, and injury tags.

[0086] The alternative action search module is used to search for alternative actions that satisfy the dynamic constraint mask matrix 280 when the target action is prohibited by the dynamic constraint mask matrix 280. It calculates the replacement score based on the similarity of the target muscle group, the similarity of the training target, the availability of venue and equipment, the impact level, and the triggering degree of the injury risk label. The alternative action search module is also used to calculate a Boolean function for safety screening. When the Boolean function is 1, the alternative action search module inputs the corresponding candidate alternative action into the replacement score sorting process. When the Boolean function is 0, the corresponding candidate alternative action is discarded.

[0087] The sports big data model generation module is used to input the allowed combination space, profile version number, current load status, dynamic constraint mask matrix 280 identifier, and enumeration field constraints into the sports big data model, enabling the sports big data model to generate a structured sports and health solution based on the allowed combination space and structured output constraints. The sports big data model generation module includes a model carrier interface, which is used to call a Transformer-based text generation model, multimodal generation model, or domain generation model fine-tuned with sports and health corpus deployed on a server, edge computing device, or local computing device, and encapsulate the allowed combination space, enumeration fields, and output format requirements into a machine-parseable structured input.

[0088] The mask reverse verification module is used to perform reverse verification of the structured sports and health scheme using the same dynamic constraint mask matrix 280. When any scheme item corresponds to M(a,i,t,r)=0 or lacks necessary identifiers, the verification is judged to be unsuccessful, triggering candidate space reconstruction, replacement action reselection or regeneration. The verification failure itself is not used as the basis for updating user profile or modal confidence.

[0089] The feedback update module is used to update the trusted multimodal motion health profile, correct modal confidence, parameters of the deterministic load state machine, and dynamic constraint mask matrix 280 based on the actual execution feedback data of the target user.

[0090] It should be noted that the present invention is based on a computer device, including a processor, a memory, and an input / output unit. The memory stores a computer program, and the processor calls the computer program to execute the method of the present invention. Simultaneously, the present invention is based on a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method of the present invention.

[0091] In the above detailed description, various features are combined together in a single embodiment to simplify this disclosure. This approach to disclosure should not be construed as reflecting an intention that embodiments of the claimed subject matter require more features than are explicitly stated in each claim. Rather, as reflected in the appended claims, the invention is presented with fewer features than all of the features of the single disclosed embodiment. Therefore, the appended claims are hereby explicitly incorporated into the detailed description, wherein each claim stands alone as a preferred embodiment of the invention.

[0092] The disclosed embodiments have been described above to enable any person skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments without departing from the spirit and scope of this disclosure. Therefore, this disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0093] The foregoing description includes examples of one or more embodiments. It is certainly impossible to describe all possible combinations of components or methods in order to describe the above embodiments, but those skilled in the art will recognize that further combinations and arrangements of the various embodiments are possible. Therefore, the embodiments described herein are intended to cover all such changes, modifications, and variations that fall within the scope of the appended claims. Furthermore, the term "comprising" as used in the specification or claims is interpreted in a manner similar to the term "including," as interpreted when used as a conjunction in the claims. Additionally, the use of any term "or" in the specification of the claims is intended to mean "non-exclusive or."

[0094] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating a large-scale model of exercise and health scheme using a reliable dynamic image mask, characterized in that, Includes the following steps: S100. Collect multimodal exercise and health data of the target user, including body measurement data, wearable device data, exercise video data, subjective fatigue data, pain feedback data, past injury data, exercise goal data, and historical execution feedback data; S200. Perform timestamp alignment, unit standardization, outlier filtering, missing value marking, and modal feature extraction on the multimodal exercise and health data, and calculate the original modal confidence of each modal data respectively; S300. Generate a subjective-objective conflict correction vector based on a preset conflict relationship between subjective and objective data; wherein, the subjective data includes subjective fatigue score, training intention, pain feedback set, and subjective recovery insufficiency score; the objective data includes motion deviation degree obtained from exercise video, heart rate zone deviation degree obtained from wearable device, sleep recovery index, heart rate variability index, resting heart rate deviation, and risk joint set obtained from video motion risk identification. S400. The original modal confidence of each modal data is corrected using the subjective-objective conflict correction vector to obtain the corrected modal confidence; S500. Map each modal feature to a shared feature space of a unified dimension, and perform gated fusion on each mapped modal feature based on the modified modal confidence to generate a credible multimodal sports and health profile. The credible multimodal sports and health profile includes a physical fitness level profile, a movement stability profile, a sports load profile, an injury risk profile, a recovery status profile, and a sports preference profile. S600. A deterministic load state machine, independent of the sports big model, calculates the load risk value based on the credible multimodal exercise health profile, and determines the current load state based on the entry threshold, exit threshold, and state lock-in period; the sports big model uses structured prompt templates for context input; S700. Generate a dynamic constraint mask matrix based on the current load status, forbidden action set, candidate action whitelist, training intensity upper limit, minimum recovery interval, continuous high-intensity training limit, and target matching conditions. , where a represents the candidate action, i represents the training intensity level, t represents the training time, and r represents the recovery interval; S800. Perform mask admission determination on the target action or initial candidate action; when the target action is prohibited by the dynamic constraint mask matrix, search for replacement actions that satisfy the dynamic constraint mask matrix based on the sports knowledge graph, and add the replacement actions that pass the admission determination to the candidate action set; S900. Construct an allowed combination space based on the dynamic constraint mask matrix and the candidate action set updated by the replacement action, and input the allowed combination space, the credible multimodal motion health profile, the current load state, the motion target data, and the output format requirements into the sports big model in a structured constraint manner, so that the sports big model generates a structured motion health plan based on the allowed combination space; the sports big model generates the structured motion health plan only based on the allowed combination space and enumeration fields written in the prompt context, and does not undertake the state transition determination of the deterministic load state machine; S1000. The structured sports and health program generated by the sports big model is parsed, and the dynamic constraint mask matrix is ​​used to perform reverse verification on the actions, intensity, training time and recovery interval in the structured sports and health program; when the generated result contains combinations outside the allowed combination space or lacks necessary identifiers, the corresponding program item is intercepted, and the candidate space is reconstructed or regenerated. S1100. Collect execution feedback data after the target user executes the structured exercise and health program, and update the trusted multimodal exercise and health profile, the corrected modal confidence, the parameters of the deterministic load state machine, and the dynamic constraint mask matrix based on the execution feedback data.

2. The method for generating a large-model motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The original modal confidence in steps S200 to S400 is expressed by the following formula: Where k is used to characterize the mode number; Used to characterize the credibility of data sources; Used to characterize the freshness of the collection time; Used to characterize data integrity; To characterize data stability, all four scores are normalized to [0,1]. Determined by mapping data source authentication level, import verification status, manual entry status, and historical consistency level; Determined by the exponential decay value of the acquisition time difference; Determined by the ratio of the number of valid fields to the number of fields to be collected; Determined by the inverse normalized value of the variance or drift amplitude within the preset window; , , , Used to characterize the weighting coefficients, and ; Used to characterize minimum confidence level.

3. The method for generating a large-model motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The subjective-objective conflict correction vector in step S300 is expressed by the following formula: in, Used to characterize the conflict component between motion deviation and subjective fatigue; Used to characterize the conflicting component between heart rate deviation and subjective fatigue; Used to characterize the conflict component between the degree of inadequate recovery and the willingness to train. Used to characterize the conflict component between the pain site and the joint at risk of motion in video; The method for determining each conflicting component is expressed by the following formula: Where 1(·) is used to characterize the indicator function; PoseDev is used to characterize the degree of movement deviation; HrDev is used to characterize the degree of heart rate zone deviation; RecoveryDeficit is used to characterize the degree of recovery deficiency, and is obtained by weighting sleep gap, decreased heart rate variability, resting heart rate deviation and subjective recovery deficiency score; , , , , These are thresholds used to characterize motion deviation, subjective fatigue, heart rate deviation, degree of inadequate recovery, and training intention, respectively; PainSet represents the set of body parts corresponding to pain feedback; RiskJointSet represents the set of risk joints obtained from video motion risk identification; when PainSet∪RiskJointSet is an empty set... The value of is 0.

4. The method for generating a large-model motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The modified modal confidence in step S400 is expressed by the following formula: in, The conflict suppression coefficient is used to characterize the k-th mode; The weighted value used to characterize one or more conflicting components corresponding to the k-th mode, and when there are multiple conflicting components , Used to characterize the weights of conflicting components.

5. The method for generating a large-model motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The gated fusion in step S500 includes the following steps: S510. Project the modal features onto the shared feature space of the same dimension using a modality-specific mapping function, and then calculate the fusion weights, expressed as follows: S520. Generate the portrait feature vector, expressed as follows: Wherein, Z is used to characterize the portrait feature vector; Used to characterize the mode-specific mapping function; Used to characterize the features of the k-th mode; The attention score, used to characterize the output of the cross-modal attention layer, is expressed as follows: Wherein, v, W, and b are used to characterize the attention layer parameters; the extraction of modal features from the motion video data includes: extracting the human skeleton key point sequence from the motion video frames, and calculating joint angles, trunk tilt, left and right limb symmetry, landing impact posture, and motion trajectory stability based on the human skeleton key point sequence to form video modal features; performing dynamic time warping matching between the human skeleton key point sequence and a standard action template, and combining the video modal features to determine the action deviation, action risk level, and video modal stability.

6. The method for generating a large-model motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The load risk value in step S600 is expressed by the following formula: Among them, RPEnorm is used to characterize the normalized value of subjective fatigue; RecoveryDeficit is used to characterize the degree of inadequate recovery; Injury is used to characterize the risk weight of past injuries; CompDrop is used to characterize the decline in training completion rate; and Conflict is used to characterize the combined value of subjective and objective conflict. , , , , , , Used to characterize the weighting coefficients, and ; The deterministic load state machine includes an initial assessment state, a low load adaptation state, a routine training state, a high load warning state, a recovery intervention state, a reassessment state, and a locked observation state; when When entering a more conservative load state than the current load state; when Furthermore, when the exit condition is met for a consecutive preset number of detection cycles, the system exits the more conservative load state; switching to a higher load state is prohibited during the state lock-in period; where s is used to characterize the current load state; Used to characterize the entry threshold function, The exit threshold function is used to characterize the following formulas: in, The baseline risk threshold used to characterize the load state s Used to characterize the hysteresis margin. Used to characterize the exit hysteresis margin, and .

7. The method for generating a large-model motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The dynamic constraint mask matrix in step S700 is expressed by the following formula: Wherein, s represents the current load state; g represents the moving target; , , , , , These are used to characterize the gate function, each taking a value of 0 or 1, and are expressed as follows: Wherein, W represents the whitelist of candidate actions, and B represents the set of forbidden actions. Used to characterize the state-related time window containing training time t The number of consecutive high-intensity training days within a given period; The specific method for searching for the replacement action based on the sports knowledge graph in step S800 is as follows: When the candidate action is prohibited, candidate replacement actions are searched in the sports knowledge graph along the relationships of the same target muscle group, the same training goal, lower impact level, lower intensity level, suitable venue and equipment, and no triggering of previous injury tags; the replacement score of the candidate replacement action is expressed by the following formula: in, The values ​​µ1, µ2, µ3, µ4, and µ5 are used to characterize the candidate replacement action; µ1, µ2, µ3, µ4, and µ5 are used to characterize the weighting coefficients; the replacement score is only used for candidates that have passed the security screening and meet the requirements. The candidate replacement actions are sorted, and the candidate replacement action with the highest replacement score is added to the allowed combination space; when at least two candidate replacement actions have the same replacement score, the final replacement action is determined in sequence according to the rules of lower damage risk label triggering degree, lower impact level, higher site and equipment availability and higher action label ranking. The safety screening includes: the candidate replacement action does not belong to the prohibited action set; the candidate replacement action does not trigger the risk joints corresponding to the target user's past injury tags and current pain location; the impact level of the candidate replacement action is not higher than the impact level of the original target action; the training intensity level of the candidate replacement action does not exceed the upper limit of training intensity corresponding to the current load state; the venue and equipment conditions for the candidate replacement action meet the available resources of the target user; and the recovery interval and continuous high-intensity training limit corresponding to the candidate replacement action both satisfy the dynamic constraint mask matrix. The security screening is determined using a Boolean function, which is expressed by the following formula: The candidate replacement action enters the replacement score sorting process only if and only if the value of the Boolean function is equal to 1; where, Used to characterize the damage risk threshold Thresholds used to characterize the availability of site equipment; The allowed combination space is obtained by filtering the Cartesian product of the candidate action set, intensity level set, training time set, and recovery interval set through a dynamic constraint mask, and is expressed as follows: Among them, A c Used to characterize the candidate action set; I c Used to characterize a set of intensity levels; T c Used to characterize the training time set; R c Used to characterize the set of recovery intervals; the action candidates, intensity candidates, training day candidates, and recovery interval candidates input into the sports model are all from the allowed combination space, and The combination of actions is not written into the prompt context of the sports model; the output field of the sports model is constrained by the enumerated values ​​of the candidate actions, the training intensity level, the training time and the recovery interval. When the output combination does not belong to the allowed combination space, it is determined to be an out-of-bounds output.

8. The method for generating a large-scale motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The structured prompt template input to the sports big data model includes a user profile summary, profile version number, profile confidence level, current load status, allowed combination space, dynamic constraint mask matrix identifier, candidate action whitelist, forbidden action set, training intensity upper limit, minimum recovery interval, available time and venue, output field requirements, enumeration constraint requirements, and risk warning requirements. The underlying model carrier of the sports big data model is a text generation model based on the Transformer structure, a multimodal generation model, or a domain generation model fine-tuned with sports and health corpus, and receives the structured prompt template through an application programming interface, local inference service, or edge inference service. Each item in the structured sports and health program includes an action identifier, the training intensity level, the training time, the recovery interval, the state machine state identifier, the dynamic constraint mask matrix identifier, the trigger rule identifier, and the confidence reference identifier.

9. The method for generating a large-scale motion health scheme using a trusted dynamic image mask according to claim 1, characterized in that: The reverse verification in step S1000 includes parsing the combination of the candidate action, training intensity level, training time, and recovery interval corresponding to each scheme item, and querying the corresponding dynamic constraint mask matrix value. When any scheme item corresponds to M(a,i,t,r)=0 or the above identifier is missing, the verification is determined to fail. The determination of verification failure is used to trigger candidate space reconstruction, action replacement and reselection, or regeneration, and is not directly used as the basis for updating user profile and modal confidence. In step S1100, the corrected modal confidence is updated exponentially, as expressed by the following formula: in, The quality score for the k-th modality obtained from the (n+1)-th execution feedback is weighted by at least two of the following: execution completion rate, target heart rate zone achievement, change in movement deviation, subjective fatigue feedback, and pain feedback; ρ is used to characterize the smoothness coefficient, and 0 < ρ < 1. Where x represents the input value to be restricted, and l and u represent the lower and upper limits, respectively; in the formula for exponential smoothing update of the modified modal confidence, l = T min , u=1.

10. A system for implementing the method for generating large-model motion health schemes using a trusted image dynamic mask as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect multimodal exercise and health data from target users. The confidence calculation module is used to preprocess and extract modal features from the multimodal exercise and health data, and calculate the original modal confidence of each modal data based on the credibility of the data source, the freshness of the collection time, the completeness of the data, and the stability of the data. The conflict correction module is used to generate a subjective-objective conflict correction vector based on a preset conflict relationship between subjective data and objective data, and to use the subjective-objective conflict correction vector to correct the original modality confidence to obtain a corrected modality confidence. The subjective-objective conflict correction vector includes a recovery insufficiency degree and a training intention conflict component. The recovery insufficiency degree is obtained by weighting at least two of the following: sleep deficit, decreased heart rate variability, resting heart rate deviation, and subjective recovery insufficiency score. The trusted profile construction module is used to map the features of each modality to a shared feature space of a unified dimension, and to perform gated fusion on the mapped features based on the modified modality confidence to generate a trusted multimodal motion and health profile carrying version number, source identifier, timestamp and profile confidence. The deterministic load state machine module operates independently of the sports big model and is used to calculate the load risk value based on the credible multimodal sports health profile, and determine the current load state based on the entry threshold, exit threshold and state lock-in period. The dynamic constraint mask generation module is used to generate a dynamic constraint mask matrix based on the current load state, the set of forbidden actions, the whitelist of candidate actions, the upper limit of training intensity, the minimum recovery interval, the continuous high-intensity training limit and the target matching condition, and to construct an allowed combination space based on the combinations that satisfy M(a,i,t,r)=1 in the set of candidate actions, the set of intensity levels, the set of training time and the set of recovery interval. The sports knowledge graph module is used to store the relationships between movements, including target muscle groups, training goals, impact levels, venues and equipment, and injury tags. The alternative action search module is used to search for alternative actions that satisfy the dynamic constraint mask matrix based on the sports knowledge graph when the target action is prohibited by the dynamic constraint mask matrix, and to calculate a replacement score according to the similarity of the target muscle group, the similarity of the training target, the availability of venue and equipment, the impact level, and the triggering degree of the injury risk label; the alternative action search module is also used to calculate the Boolean function used for safety screening. ; When the Boolean function takes the value 1, the replacement action search module inputs the corresponding candidate replacement action into the replacement score sorting process; When the Boolean function takes the value of 0, the corresponding candidate replacement action is discarded; The sports big data model generation module is used to input the allowed combination space, profile version number, current load status, dynamic constraint mask matrix identifier, and enumeration field constraints into the sports big data model, enabling the sports big data model to generate a structured sports and health solution based on the allowed combination space and structured output constraints. The sports big data model generation module includes a model carrier interface, which is used to call a Transformer-based text generation model, multimodal generation model, or domain generation model fine-tuned with sports and health corpus deployed on a server, edge computing device, or local computing device, and encapsulate the allowed combination space, enumeration fields, and output format requirements into a machine-parseable structured input. The mask reverse verification module is used to perform reverse verification on the structured sports and health scheme using the same dynamic constraint mask matrix. When any scheme item corresponds to M(a,i,t,r)=0 or lacks necessary identifiers, the verification is determined to be unsuccessful, triggering candidate space reconstruction, replacement action reselection or regeneration. The verification failure itself is not used as the basis for updating user profile or modal confidence. The feedback update module is used to update the trusted multimodal motion health profile, the corrected modal confidence, the parameters of the deterministic load state machine, and the dynamic constraint mask matrix based on the actual execution feedback data of the target user.

Images