A sports prescription generation system and method based on a sports intelligent large model

Abstract

The application relates to the technical field of exercise prescription generation, and discloses an exercise prescription generation system and method based on an exercise intelligent large model. The method constructs a mixed training system of general corpus and domain corpus, so that the model can obtain in-depth understanding of professional knowledge while keeping original language ability, the proportion of different types of data is dynamically adjusted through a data mixing optimization strategy, the catastrophic forgetting problem that may occur in the domain training process is effectively alleviated, the model is balanced between general ability and domain ability, the traditional recommendation problem is converted into a generation problem, an exercise prescription generation model based on a user portrait is proposed, and direct mapping from user information to an exercise scheme is realized.

Description

Technical Field

[0001] This invention relates to the field of exercise prescription generation technology, specifically to an exercise prescription generation system and method based on a large-scale intelligent exercise model. Background Technology

[0002] With the increasing demand for health management among residents, exercise intervention has gradually become an important means of chronic disease prevention, weight control, and physical fitness improvement. When developing an exercise plan, users usually need to comprehensively consider multiple factors such as age, gender, medical history, physical fitness level, exercise goals, and time conditions. Therefore, exercise prescription generation has a strong individualized characteristic.

[0003] Existing sports guidance methods mainly include three categories: human coaching, fixed template recommendations, and general intelligent question-and-answer systems.

[0004] in: (1) Human coaching is costly and difficult to provide large-scale, continuous service; (2) Recommendation methods based on rule bases or template bases have fixed output results, making it difficult to dynamically adjust the training intensity, frequency and cycle arrangement according to the differences in user physical characteristics; (3) Although the general large language model has natural language interaction capabilities, it lacks knowledge in fields such as sports science, sports risk control and training cycle design, and the generated results may have insufficient professionalism. (4) Existing methods generally lack continuous modeling of user performance feedback, movement preferences and historical compliance behavior, resulting in poor long-term performance of recommendation schemes.

[0005] Furthermore, under health constraints such as knee injuries, hypertension, and obesity, existing systems struggle to simultaneously meet the requirements of safety, scientific rigor, and feasibility.

[0006] Therefore, there is an urgent need for a method to generate exercise prescriptions that integrates sports expertise, individual user characteristics, and behavioral feedback information in order to generate and dynamically optimize personalized exercise plans. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a system and method for generating exercise prescriptions based on a large-scale intelligent motion model. This system effectively mitigates the catastrophic forgetting problem that may occur during domain training, enables the model to dynamically generate exercise plans that meet individual differences, and significantly improves personalization and adaptability, thus solving the aforementioned technical problems.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a motion prescription generation system based on a motion intelligence large model, comprising: The continuous pre-training module is used to acquire pre-training corpus and corpus in the field of scientific fitness guidance. After data processing, the initial model Qwen3-32B / 8B is pre-trained to obtain the base model. The supervised fine-tuning module is used to build user profiles and generate candidate exercise prescriptions. At the same time, it performs integrity filtering on the candidate exercise prescriptions. If the integrity filtering constraints are not met, the prescriptions are deemed unqualified. After the qualified exercise prescriptions are scored by multiple judges and ranked by consensus, LoRA fine-tuning is used to fine-tune the base model. The preference optimization module is used to read the base model fine-tuned by the supervised fine-tuning module and optimize it to obtain a large motion intelligence model. The model output module is used to receive user profiles input in real time and output exercise prescriptions through a large-scale intelligent motion model.

[0009] As a preferred technical solution of the present invention, the integrity filtering in the supervision and fine-tuning module includes performing integrity filtering on candidate exercise prescriptions through LLM-as-a-judge. If the prescription contains preset basic elements, it is judged as qualified; otherwise, it is judged as unqualified.

[0010] As a preferred technical solution of the present invention, the multi-judge scoring and consensus ranking specifically includes using several inference models as independent evaluators to score candidate exercise prescriptions within the same score range and rank the candidate exercise prescriptions. When there is a discrepancy in the rankings of different evaluators, a consensus ranking algorithm based on dynamic programming is adopted. Starting from the initial ranking of one evaluator, the longest valid subsequence is found, such that the relative order of any two items is supported by at least K evaluators. The remaining candidate exercise prescriptions are added to the tail of the skeleton to form the final consensus ranking truth value. Finally, the screening results are verified by human experts.

[0011] As a preferred technical solution of the present invention, the optimization in the preference optimization module includes optimizing user profiles. Construct a pair of comparable samples of human preferences. and minimize sorting loss : in, This represents the sigmoid function. Representation of generative reward model User profile and normal prescription The output score; Representation of generative reward model User profile and negative prescription The output score, This indicates a positive example prescription; This indicates a negative prescription.

[0012] As a preferred technical solution of the present invention, the generative reward model Specifically, this involves generating natural language comments and then extracting reward points from the comments, as shown in the following expression: in, Represents a generative reward model The overall score field output at the k-th sampling time. The reward score is extracted from the comments. These are text comments generated by a generative reward model. This represents the total number of samples. This is the sampling index.

[0013] As a preferred embodiment of the present invention, the preference optimization module further includes minimizing the ranking loss. After completion, a second optimization is performed using the direct preference optimization method.

[0014] The present invention also provides a method for generating exercise prescriptions based on a large-scale intelligent motion model, wherein the exercise prescription generation process is performed based on the above-mentioned exercise prescription generation system based on a large-scale intelligent motion model.

[0015] Compared with existing technologies, this invention provides a system and method for generating exercise prescriptions based on a large-scale intelligent motion model, which has the following beneficial effects: This invention constructs a hybrid training system using general and domain-specific corpora, enabling the model to gain a deep understanding of professional knowledge while maintaining its original language capabilities. By dynamically adjusting the proportion of different data types through a data hybridization optimization strategy, it effectively mitigates the catastrophic forgetting problem that may occur during domain training, achieving a balance between general and domain capabilities. It transforms the traditional recommendation problem into a generation problem, proposing a user-profile-based exercise prescription generation model that achieves a direct mapping from user information to exercise plans. The user profile not only includes basic information but also integrates multi-dimensional information such as health status, exercise experience, behavioral characteristics, and real-world constraints, enabling the model to generate reasonable plans under complex conditions. This technology breaks through the limitations of traditional methods that rely on a fixed prescription database, allowing the model to dynamically generate exercise plans that meet individual differences, significantly improving personalization and adaptability. First, multiple large models are used to generate candidate user profiles and corresponding exercise prescriptions, and an inference process is introduced to enhance the logic of the results. Then, a constraint mechanism is used to screen samples that meet safety and completeness requirements. Furthermore, a multi-judge evaluation mechanism is introduced to independently score candidate results and use a consensus ranking method to screen high-quality samples. Finally, human expert review ensures data reliability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 A system for generating exercise prescriptions based on a large-scale intelligent motion model, comprising: The continuous pre-training module specifically includes the following steps: Step A1: Data construction includes: High-quality pre-training corpus: This invention selects multiple types of high-quality open corpora as general corpus sources, including WanJuan-CC, RedPajama series, and FineWeb. To improve training efficiency, the general corpus is then uniformly converted into single-sentence or short paragraph text format, and segmented and concatenated according to a predefined length threshold. Scientific Fitness Guidance Corpus: This invention constructs a large-scale domain corpus around scientific fitness guidance. The corpus sources cover academic databases, official institutions, authoritative guidelines, and standardized knowledge pages, ultimately forming a sports domain corpus exceeding 2 billion tokens, used for domain pre-training and subsequent task development. This corpus system organizes knowledge from both scientific and practical dimensions, dividing scientific fitness content into 7 core domains: exercise methods and techniques, weight and body composition management, strength and endurance training, sports medicine, exercise biochemistry and metabolism, exercise physiology, and sports nutrition. Step A2: Data processing, including standardization, data filtering, structured knowledge extraction, and data blending; Data standardization: First, heterogeneous texts from multiple sources are uniformly converted into a processable format (such as Markdown or plain text), using UTF-8 encoding, and punctuation and space formatting noise are standardized through rule processing. For long texts, a dynamic window segmentation strategy is adopted based on semantic integrity to ensure that each slice retains as complete semantic units as possible, avoiding knowledge fragmentation caused by coarse-grained truncation.

[0019] Data filtering: The filtering process includes two types of mechanisms: Rule-based filtering: Removes special characters, HTML remnants, template noise, repeated sentences, anomalously encoded text, and low-information-density fragments; Domain relevance filtering based on small models: Identify content that is irrelevant to scientific fitness, overly marketing-oriented, has low fact density, or contains obviously misleading information.

[0020] This dual filtering approach, combining rules and semantics, ensures that corpus quality control is no longer limited to surface cleaning, but also takes into account task relevance and professional credibility.

[0021] Structured Knowledge Extraction: To expose the model to higher-density professional relationships during the pre-training phase, this invention performs structured knowledge extraction on the original corpus, including named entity recognition, dependency parsing, part-of-speech tagging, and relation extraction. For the fitness scenario, this invention defines 8 types of entities and 11 types of relationships. Through these structured units, the "action—target muscle group—equipment—target—nutrition—constraints" relationship in the original text can be more systematically made explicit, which is beneficial for the model to learn more stable movement knowledge patterns in subsequent training.

[0022] In terms of structured knowledge extraction, this invention defines 8 types of entities (body parts, sports, fitness movements, equipment, exercise goals, anatomical structures, nutrients, and technical terms) and 11 types of relationships (position, shape, origin, end, inclusion, subordination, function, exercise, use, realization, and need) for fitness scenarios, making the "movement-target muscle group-equipment-goal-nutrient-constraint" relationship in the original text explicit, thereby enhancing the model's ability to model the structure of exercise knowledge.

[0023] Data Mixing: To minimize catastrophic forgetting while injecting domain knowledge, this invention does not employ a fixed ratio of general-purpose to domain-specific data. Instead, it uses the RegMix approach to find a better data ratio. RegMix models the selection of the data mixing ratio as a regression problem, predicting a better mixing distribution for large-scale training through small-scale agent training. Based on this idea, this invention determines a more suitable ratio of general / specialized data for injection into the motion domain. Through the above process, a high-quality corpus for continuous pre-training is finally obtained, with the domain-specific corpus size approximately 2 bytes. Step A3: Continuous pre-training of domain awareness: After completing the data construction, this invention continued training on the Qwen3-32B / 8B base, with the overall optimizer configuration and the main engineering configuration of the tokenizer scheme being implemented using the corresponding base model. After this stage of training was completed, this invention formed a domain base model with the ability to perceive scientific fitness language patterns: it retains the ability to understand general question-and-answer and instructions, and can better understand expressions related to kinematics, training, nutrition and rehabilitation. The core task of this module is to build a domain-specific foundation that is both general and specialized. This module does not aim for models to only answer motion-related questions, but rather for them to retain basic language skills, reasoning abilities, and common-sense transfer capabilities while incorporating specialized knowledge and expression patterns from the motion domain.

[0024] In its implementation, this module does not simply introduce domain data, but rather achieves effective injection of domain knowledge through the joint design of data engineering and training mechanisms.

[0025] The monitoring and fine-tuning module specifically includes the following steps: Step B1: Dataset Design and Candidate Data Generation Exercise prescription recommendation is one of the core tasks in scientific fitness guidance. Unlike traditional recommendation systems, this invention focuses on generative exercise prescription recommendation, that is, generating a new, personalized exercise prescription directly based on the user profile, rather than simply retrieving prescriptions from an existing prescription database. This invention is no longer limited by an existing prescription database, but can generate truly individualized new prescriptions for complex health conditions, specific equipment conditions, and specific lifestyle constraints.

[0026] The supervised fine-tuning module also includes a user profile construction submodule, a candidate exercise prescription generation submodule, and a prescription constraint and screening submodule. These three submodules work together to generate and screen high-quality training data. The first module is the user profile construction module. Instead of using simple labels, this module constructs a structured profile containing multi-dimensional information such as basic information, health status, exercise background, psychological characteristics, and life constraints. This allows the model to learn the complex decision-making environment of a "real user." The structured profile is shown in Table 1 below: Table 1 The candidate exercise prescription generation submodule includes two main categories of patients: the general population and those with chronic diseases. The general population is organized according to demographic characteristics such as age and sex, covering: children aged 6–12 years; adolescents aged 12–17 years; university students aged 18–25 years; adults aged 18–59 years; and the elderly aged 60–79 years. For those with chronic diseases, priority is given to five categories with high incidence rates and relatively clear evidence of exercise intervention: obesity combined with hyperlipidemia, hypertension, coronary heart disease, malignant tumors, and diabetes. Prescription Constraint and Screening Submodule: This submodule constrains the candidate results. Candidate exercise prescriptions must contain all eight basic elements, as shown in Table 2 below. Before formal scoring, the candidate exercise prescriptions are first filtered for completeness using LLM-as-a-judge: if all eight elements are incomplete, or if there are obvious errors in the reasoning chain, they are directly judged as unqualified samples and will not proceed to the subsequent ranking stage; Table 2 Step B2: Multi-judge scoring and consensus ranking. Specifically, this invention introduces five strong inference models (Claude-4Sonnet, GPT-o3, DeepSeek-R1-0528, Doubao-seed-1.6, and Gemini2.5Pro) as independent evaluators. Each evaluator scores candidates according to six basic principles: safety, effectiveness, specialization, whole-body application, gradual progression, and periodicity, from 1 to 5. The total score is calculated, and the candidate exercise prescriptions are ranked. When there is disagreement among the rankings of different evaluators, a consensus ranking algorithm based on dynamic programming is used. Starting from the initial ranking of one evaluator, the longest valid subsequence is found, such that the relative order of any two items is supported by at least K evaluators (e.g., at least 3 out of 5 judges). This extracts a stable ranking skeleton that is "accepted by the majority of models." The remaining candidate exercise prescriptions are then appended to the end of the skeleton to form the final consensus ranking truth value. in This represents the rating across six principle dimensions. This indicates the candidate ranking from best to worst; Finally, the screening results were verified by human experts to ensure the reliability of the data in terms of professionalism and security. This invention involved independent evaluation by three experts with a Master's degree in Sports Science or intermediate / senior national fitness coach qualifications, using a 1-5 Likert scale. Only samples deemed valid by at least two experts and with an average score greater than 3 were retained. The final consistency rate was 86%, indicating good consensus among the experts; prescriptions lacking sufficient consistency were discarded. Step B3: Two-step hybrid supervision fine-tuning: The LoRA (Low-Rank Adaptation) fine-tuning method is adopted to update the low-rank adapter parameters while keeping the main parameters of the base model frozen, in order to reduce the risk of catastrophic forgetting and improve training efficiency. The overall approach adopts a two-step hybrid fine-tuning strategy, which includes wide-coverage hybrid fine-tuning and domain-enhanced fine-tuning. The first phase, broad-coverage hybrid fine-tuning, uses a 1:2 ratio of exercise prescription datasets to general instruction datasets (i.e., one exercise prescription dataset paired with two general instruction datasets). The general instruction dataset contains approximately 100,000 entries, covering instruction compliance, multi-turn dialogue, and logical reasoning tasks. This phase involves training for three epochs, allowing the model to initially learn the output format and basic patterns of exercise prescriptions while maintaining its general instruction responsiveness.

[0027] The second stage involves domain-enhanced fine-tuning: once the validation loss from the first stage has plateaued, the model is switched to the second stage. In this stage, the data mixing ratio is adjusted to 3:1 (exercise prescription data: general instruction data), with approximately 50,000 exercise prescription data entries, and training lasts for 2 epochs. This stage focuses the model more intently on the exercise prescription generation task, strengthening individualized condition modeling, constraint compliance, and structural integrity. This progressive training method can effectively avoid the problem of model capability degradation during training for specific tasks; The preference optimization module, after supervised fine-tuning, further aligns the model output, specifically including the following steps: Step C1: Preference Data Construction Strategy First, in terms of constructing preference data, this invention divides preference information into two categories, forming binary pairs: (1) Tasks with strong subjectivity: For subjective tasks such as emotional encouragement, companionship expression, and style preference suggestions, ordinary users or domain experts will label multiple candidate answers with preferences to more directly depict real human preferences.

[0028] (2) Objective tasks with obvious advantages and disadvantages: For tasks that are closer to "having clear good and bad standards", such as exercise prescription generation, risk warning, and movement error correction knowledge Q&A, preference pairs are constructed based on expert judgment of structural integrity and principle compliance, and higher quality samples are set as positive examples and lower quality samples are set as negative examples. Step C2 expands the preference data using a generative reward model. Specifically, the reward model in this invention is a GRM model. To expand the scale of preference data, a generative reward model is trained based on manually generated preference data. The core idea of ​​the GRM model is to first generate natural language comments, and then extract reward points from the comments. For a given user profile... Candidate exercise prescription Its formal definition is: in, These are text comments generated by a generative reward model. The reward score is extracted from the comments; the extraction function... The specific calculation method is as follows: Generative reward model required during training Along with the comments, output an overall rating on a scale of 1 to 5. Extraction function This score is directly analyzed as a reward score. : Where Score( ) represents the overall score field carried in the output of the generative reward model U.

[0029] To enhance the stability of the scoring, multiple sampling and averaging can be used: in, Represents a generative reward model The overall score field (1-5 points) output at the kth sampling time; This represents the total number of samples. For sampling index, for training For the same user profile Construct a pair of comparable samples of human preferences. and minimize sorting loss : in, This represents the sigmoid function, used to map differences to the interval (0,1); Representation of generative reward model User profile and normal prescription The output score; Representation of generative reward model User profile and negative prescription The output score; This indicates a positive example prescription (a prescription for exercise that humans prefer). The negative example prescription (an exercise prescription that humans are less likely to prefer) is represented by the generative reward model, which aims to learn that "under the same user conditions, exercise prescriptions that humans prefer should receive a higher score." Compared to directly outputting a single scalar score, generative reward models can retain more interpretable commentary information, facilitating subsequent analysis of "why this prescription is more preferred."

[0030] This invention addresses preference pairs The construction method is as follows: For the same user profile Using the generative reward model obtained through training Scoring of all candidate exercise prescriptions The prescription with the highest score will be used as... (Best Prescription), the prescription with the lowest score is considered as... (Poor prescription); In terms of model optimization, this invention employs the direct preference optimization method, and its objective function is as follows: This formula can be interpreted as follows: This represents the sigmoid function. This indicates the input (user profile). This indicates a better prescription (higher user preference). Indicates a poor prescription. Indicates the current model, Reference model This indicates a preference for optimizing the temperature coefficient. Indicates in the current model Given a user profile At that time, the model generates positive prescriptions. The probability of minimizing the sorting loss. After completion, a second optimization is performed using the direct preference optimization method.

[0031] By combining these two types of preference construction methods, a preference dataset that can both reflect human subjective preferences and retain objective structural information of the task is obtained; The essence of the objective function is to increase the probability of the model generating the "optimal solution" while decreasing the probability of the "poor solution".

[0032] Where: the fractional term represents the degree of preference of the current model relative to the reference model, the parameter β controls the optimization intensity, and the sigmoid function is used to convert the difference into a probabilistic form; Model output module: After a series of steps involving continuous pre-training, supervised fine-tuning, and preference optimization, this invention yields a directly deployable large-scale intelligent motion model πθ. During the inference phase, the user inputs their motion profile (including basic information, health status, exercise background, lifestyle constraints, etc., see 2.1 Structured Profile), and model πθ directly generates a structured exercise prescription based on the capabilities gained from the three-stage training.

[0033] The specific generation process is as follows: Input construction: The user's structured motion profile is converted into natural language prompts, which include key information such as the user's basic vital signs, health constraints, exercise goals, and time and location conditions; Model inference: Input the prompt words into the model πθ, and the model performs autoregressive generation based on the preference signal aligned with the prescription generation ability DPO module learned by the supervised fine-tuning module; Output Analysis: After structured parsing, the text output by the model yields the following motion prescription: The output, which includes exercise goals, training methods, intensity and time arrangements, cycle and progression strategies, and individualized considerations, is actionable, interpretable, and personalized, and can be directly applied to actual exercise guidance scenarios.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for generating exercise prescriptions based on a large-scale intelligent motion model, characterized in that: include: The continuous pre-training module is used to acquire pre-training corpus and corpus in the field of scientific fitness guidance. After data processing, the initial model Qwen3-32B / 8B is pre-trained to obtain the base model. The supervised fine-tuning module is used to build user profiles and generate candidate exercise prescriptions. At the same time, it performs integrity filtering on the candidate exercise prescriptions. If the integrity filtering constraints are not met, the prescriptions are deemed unqualified. After the qualified exercise prescriptions are scored by multiple judges and ranked by consensus, LoRA fine-tuning is used to fine-tune the base model. The preference optimization module is used to read the base model fine-tuned by the supervised fine-tuning module and optimize it to obtain a large motion intelligence model. The model output module is used to receive user profiles input in real time and output exercise prescriptions through a large-scale intelligent motion model.

2. The exercise prescription generation system based on a large-scale motion intelligence model according to claim 1, characterized in that: The integrity filtering in the supervision and fine-tuning module includes using LLM-as-a-judge to perform integrity filtering on candidate exercise prescriptions. If a prescription contains preset basic elements, it is deemed qualified; otherwise, it is deemed unqualified.

3. The exercise prescription generation system based on a large-scale motion intelligence model according to claim 2, characterized in that: The multi-judge scoring and consensus ranking specifically involves using several inference models as independent evaluators to score candidate exercise prescriptions within the same score range and rank the candidate exercise prescriptions. When there are discrepancies in the rankings of different evaluators, a consensus ranking algorithm based on dynamic programming is used. Starting from the initial ranking of one evaluator, the algorithm finds the longest valid subsequence such that the relative order of any two items is supported by at least K evaluators. The remaining candidate exercise prescriptions are then appended to the end of the skeleton to form the final consensus ranking truth value. Finally, human experts review the screening results for final verification.

4. The exercise prescription generation system based on a large-scale motion intelligence model according to claim 3, characterized in that: The optimization process in the preference optimization module includes optimizing user profiles. Construct a pair of comparable samples of human preferences. and minimize sorting loss : in, This represents the sigmoid function. Representation of generative reward model User profile and normal prescription The output score; Representation of generative reward model User profile and negative prescription The output score, This indicates a positive example prescription; This indicates a negative prescription.

5. The exercise prescription generation system based on a large-scale motion intelligence model according to claim 4, characterized in that: The generative reward model Specifically, this involves generating natural language comments and then extracting reward points from the comments, as shown in the following expression: in, Represents a generative reward model The overall score field output at the k-th sampling time. The reward score is extracted from the comments. These are text comments generated by a generative reward model. This represents the total number of samples. This is the sampling index.

6. The exercise prescription generation system based on a large-scale motion intelligence model according to claim 5, characterized in that: The preference optimization module also includes minimizing the ranking loss. After completion, a second optimization is performed using the direct preference optimization method.

7. A method for generating exercise prescriptions based on a large-scale motion intelligence model, characterized in that: The process of generating an exercise prescription is based on the exercise prescription generation system based on a motion intelligence big model as described in any one of claims 1-6.

Images