Large-health customized talent training system based on artificial intelligence

Abstract

The invention relates to the technical field of talent training, and discloses an artificial intelligence-based large-health customized talent training system, which comprises a multi-modal data acquisition module, which is used for acquiring user biological index data, behavior log data, psychological assessment data and industry demand data; and the data processing and knowledge fusion module is used for performing time sequence alignment and missing value filling on the acquired data, and constructing a skill association knowledge graph based on industry demand data. According to the invention, through the personalized ability modeling and reinforcement learning recommendation module, the personalized training path of each user is dynamically generated, it is ensured that each user can obtain training matched with health requirements, ability levels and industry requirements, and compared with a universal training scheme in the prior art, the training efficiency is greatly improved. An accurate and customized training scheme can be provided according to real-time biological indexes, psychological evaluation and behavior data of the user, and the training effect is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of talent training technology, and specifically to an artificial intelligence-based customized talent training system for big health. Background Art

[0002] Currently, talent training in the healthcare sector generally adopts a standard curriculum system that cannot adapt to individual physiological characteristics and dynamic industry needs. Traditional solutions rely on manual experience to design training content and lack continuous tracking of users' real-time biological indicators and psychological states, resulting in a rigid training path. The existing system is single in data collection dimensions, unable to integrate multimodal information, and unable to support personalized capability modeling. In addition, virtual training scenarios are disconnected from real health management tasks, the feedback mechanism lags, and training strategies cannot be dynamically optimized. To address the above problems, there is an urgent need to build a customized talent training system based on artificial intelligence, which can achieve closed-loop control from data collection to path optimization through multi-source data fusion and adaptive algorithms.

[0003] Traditional technologies rely on fixed training plans, ignoring the changing physiological state and abilities of users. Existing systems only analyze biological indicators or behavioral data, failing to dynamically adjust training paths by integrating multiple sources of information. Training content is often manually preset, resulting in a slow response to individual differences and failing to meet users' actual needs for personalized and customized training.

[0004] Existing technologies are insufficiently capable of simulating health management scenarios. Virtual training often relies on static models, which provide no real-time feedback on training effectiveness and optimization suggestions. Model updates lack adaptive mechanisms, and the feedback process is simplistic. Due to the lack of multi-objective balancing and closed-loop control, existing solutions struggle to dynamically correct deviations, easily becoming stuck in fixed patterns and unable to respond promptly to environmental changes, impacting overall training effectiveness.

[0005] To this end, the present invention proposes a big health customized talent training system based on artificial intelligence to solve the above-mentioned problems. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the present invention provides a customized talent training system for big health based on artificial intelligence to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a large-scale health customized talent training system based on artificial intelligence, the talent training system comprising:

[0008] Multimodal data collection module, used to obtain user biometric data, behavioral log data, psychological assessment data and industry demand data;

[0009] The data processing and knowledge fusion module is used to align the acquired data in time series and fill in missing values, and to build a skill-related knowledge graph based on industry demand data;

[0010] The personalized capability modeling module receives the knowledge graph and aligned multimodal data to generate a user capability vector;

[0011] The reinforcement learning recommendation module dynamically generates personalized training paths based on the relationship between user ability vectors and skills in the knowledge graph;

[0012] The virtual training module receives personalized training paths, combines physiological dynamics models with user ability vectors to simulate health management scenarios, and outputs virtual training results;

[0013] The evaluation and optimization module receives the training results and generates optimization instructions by comparing the matching degree between the user capability vector and the knowledge graph.

[0014] Preferably, the data processing and knowledge fusion module further includes:

[0015] The time series alignment unit is used to align the biological indicator data, behavioral log data, and psychological assessment data obtained by the multimodal data acquisition module across devices. The specific algorithm is the dynamic time warping algorithm, and the formula is defined as:

[0016]

[0017] Where X={x1,x2,...,x n} is the timing data from device A,

[0018] Y={y1,y2,...,y m} is the timing data from device B,

[0019] w is the regular path set, λ is the time misalignment penalty coefficient, x i 、y j is the observation value at time i and j in the sequence;

[0020] The missing value filling unit receives the aligned data output by the time series alignment unit and fills the missing biological indicator data through the conditional generative adversarial network. The optimization objectives of the generator G and the discriminator D are:

[0021]

[0022] Among them, x partial is part of the observed data, z is the noise vector, G(z|x partial ) is the generator network D(x) is the discriminator network, p data is the real data distribution, p zis a Gaussian distribution, L G is the generator, L D is the discriminator, E is the expected operation;

[0023] The knowledge graph construction unit receives the complete data processed by the missing value filling unit and constructs the skill association knowledge graph based on the TransR model. The entity relationship scoring function is:

[0024]

[0025] Among them, h is the head entity embedding vector, t is the tail entity embedding vector, r is the relationship embedding vector, h r is the projection vector of the entity in the relational space r, t r is the projection form of the tail entity vector.

[0026] Preferably, the personalized capability modeling module further includes:

[0027] The data fusion unit is used to fuse multimodal data with skill information in the knowledge graph to construct a comprehensive user capability representation vector. The algorithm is a multi-head attention mechanism, and the formula is defined as:

[0028]

[0029] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k A is the key and query dimension, i is the attention output matrix;

[0030] The user capability vector generation unit generates the final user capability vector through the fused data. The algorithm is a weighted average calculation method, and the formula is defined as:

[0031]

[0032] Among them, α i is the weight of the i-th data source, A i is the attention output matrix, U is the final generated user capability vector, and N is the number of data sources;

[0033] The capability vector optimization unit optimizes the user capability vector and adjusts the vector to match the skill requirements of the industry. The optimization goal is to minimize the following loss function:

[0034]

[0035] Among them, S is the skill vector required by the industry, L(U) is the optimization loss, and U is the user ability vector.

[0036] Preferably, the reinforcement learning recommendation module further includes:

[0037] The state modeling unit is used to encode the relationship between the user capability vector and the skill in the knowledge graph into a state representation of reinforcement learning. The formula is defined as:

[0038]

[0039] Among them, U is the user capability vector, S i is the embedding vector of the i-th skill in the knowledge graph, d is the vector dimension, s t is the state vector at time step t, softmax is the normalization function, and M is the total number of relevant skills in the knowledge graph;

[0040] The policy optimization unit generates the action probability distribution based on the proximal policy optimization algorithm. The objective function is defined as:

[0041]

[0042] Among them, L CLIP (θ) is the objective function, π θ is the policy network parameterization function, is the advantage function estimate, ∈ is the policy update clipping threshold, is the probability distribution of the old policy network, π θ (a t |s t ) is the current policy network in state s t Next select action a t The probability of E t is the expected value at time step t;

[0043] The reward calculation unit generates an instant reward signal based on the user's ability improvement and path feasibility. The formula is defined as:

[0044] r t =λ1·ΔU t +λ2·Sim(U t ,S t )-λ3·Fatigue t ,

[0045] Among them, ΔU t is the user capability vector change, Sim(U t ,S t ) is the cosine similarity between user ability and target skill, Fatigue t is the learning fatigue, λ1, λ2, λ3 are weight coefficients;

[0046] The path generation unit generates a personalized training path based on the action probability distribution output by the policy network. The action selection formula is:

[0047]

[0048] Among them, Q(s t ,a) is the action value function, η is the exploration coefficient, A is the action space,

[0049] Entropy(π(·|s t )) is the policy entropy.

[0050] Preferably, the virtual training module further includes:

[0051] The physiological dynamics modeling unit generates personalized physiological parameters based on the user's ability vector and constructs a differential equation model of the cardiovascular system. The formula is defined as:

[0052]

[0053] Where P(t) is the blood pressure value at time t, Q in (t) is the blood flow into the heart, Q out (t) is the peripheral blood flow, C is the vascular compliance, and R is the peripheral vascular resistance.

[0054] Preferably, the virtual training module further includes:

[0055] The scenario simulation unit receives the skill training instructions in the personalized training path and outputs the virtual patient state in combination with the physiological dynamics model. The state transition formula is:

[0056] s t+1 =f NN (e t ,g t )+τ·N(0,1),

[0057] Among them, e t is the virtual patient state vector at time t, g t is the training action of step t in the training path, f NN is the neural network mapping function, τ is the noise intensity coefficient;

[0058] The real-time feedback unit compares the virtual patient state with the target physiological indicators to generate action suggestions. The loss function is defined as:

[0059]

[0060] Among them, e target is the target physiological index vector, e opt is the set of standard operating actions recommended by medical guidelines, and λ is the action deviation penalty coefficient.

[0061] Preferably, the evaluation and optimization module further includes:

[0062] The causal effect evaluation unit quantifies the effect of personalized training paths on user capabilities based on the double difference model. The formula is defined as:

[0063] Y it =β0+β1·Treat i +β2·Post t +δ·(Treat i Post t )+∈ it ,

[0064] Among them, Y it is the observation result of user i at time t, β0 is the intercept term of the model, and Treat i is a binary variable, β1 is the inherent difference between the treatment group and the control group, Post t is a binary variable, β2 is the time effect, δ is the double difference estimator, ∈ it is the error term.

[0065] Preferably, the evaluation and optimization module further includes:

[0066] The online A / B testing unit dynamically assigns users to different recommendation strategy groups through Thompson sampling. The action selection probability update formula is:

[0067]

[0068] Among them, P(a) is the probability of selecting strategy a in the next time step.

[0069] Beta(α a ,β a ) is the Beta prior distribution of strategy a, α a is the prior success number of strategy a, β a is the prior failure count of strategy a, n succ is the number of successful times of strategy a in the current experimental cycle, n fail is the number of failures of strategy a in the current experimental cycle;

[0070] The feedback control unit generates model optimization instructions based on the causal effect and A / B test results. The weight update formula is:

[0071]

[0072] Among them, w old is the current model parameter vector, w new is the updated model parameter vector, η is the learning rate, is the gradient of the loss function to the model parameters, λ1 and λ2 are weight coefficients, and Lcausal is the causal effect loss, L AB is the strategy diversity loss.

[0073] A terminal device includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the artificial intelligence-based big health customized talent training system.

[0074] A storage medium stores a computer program, which, when executed by a processor, implements the artificial intelligence-based customized health talent training system.

[0075] The present invention provides a customized talent training system for big health based on artificial intelligence.

[0076] Beneficial effects:

[0077] 1. The present invention dynamically generates personalized training paths for each user through personalized ability modeling and reinforcement learning recommendation modules, ensuring that each user can receive training that matches their health needs, ability level and industry requirements. Compared with the general training programs in the existing technology, it can provide accurate and tailored training programs based on the user's real-time biological indicators, psychological assessment and behavioral data, significantly improving the training effect.

[0078] 2. The present invention adopts virtual training modules and physiological dynamics modeling technology, combined with the user's personalized data, to simulate health management scenarios in real time and feedback training results, and further adjust the personalized training path. Compared with traditional training methods that rely on fixed content and steps, the system of the present invention has adaptive adjustment capabilities and can dynamically optimize the training process according to the user's ability changes and real-time feedback, solving the limitations of the traditional training model that is fixed and lacks flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a system diagram of the present invention. DETAILED DESCRIPTION

[0080] To help those skilled in the art understand the present invention, the following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0081] The present invention is described in detail below with reference to the accompanying drawings:

[0082] Example:

[0083] Please see the attached Figure 1 The embodiment of the present invention provides a customized talent training system for big health based on artificial intelligence, and the talent training system includes:

[0084] Multimodal data collection module, used to obtain user biometric data, behavioral log data, psychological assessment data and industry demand data;

[0085] The data processing and knowledge fusion module is used to align the acquired data in time series and fill in missing values, and to build a skill-related knowledge graph based on industry demand data;

[0086] The time series alignment unit is used to align the biological indicator data, behavioral log data, and psychological assessment data obtained by the multimodal data acquisition module across devices. The specific algorithm is the dynamic time warping algorithm, and the formula is defined as:

[0087]

[0088] Where X={x1,x2,...,x n} is the timing data from device A,

[0089] Y={y1,y2,...,y m} is the timing data from device B,

[0090] w is the regular path set, λ is the time misalignment penalty coefficient, x i 、y j is the observation value at time i and j in the sequence;

[0091] The missing value filling unit receives the aligned data output by the time series alignment unit and fills the missing biological indicator data through the conditional generative adversarial network. The optimization objectives of the generator G and the discriminator D are:

[0092]

[0093] Among them, x partial is part of the observed data, z is the noise vector, G(z|x partial ) is the generator network D(x) is the discriminator network, p data is the real data distribution, p z is a Gaussian distribution, L G is the generator, L D is the discriminator, E is the expected operation;

[0094] The knowledge graph construction unit receives the complete data processed by the missing value filling unit and constructs the skill association knowledge graph based on the TransR model. The entity relationship scoring function is:

[0095]

[0096] Among them, h is the head entity embedding vector, t is the tail entity embedding vector, r is the relationship embedding vector, h r is the projection vector of the entity in the relational space r, t r is the projection form of the tail entity vector;

[0097] The personalized capability modeling module receives the knowledge graph and aligned multimodal data to generate a user capability vector;

[0098] The data fusion unit is used to fuse multimodal data with skill information in the knowledge graph to construct a comprehensive user capability representation vector. The algorithm is a multi-head attention mechanism, and the formula is defined as:

[0099]

[0100] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k A is the key and query dimension, i is the attention output matrix;

[0101] The user capability vector generation unit generates the final user capability vector through the fused data. The algorithm is a weighted average calculation method, and the formula is defined as:

[0102]

[0103] Among them, α i is the weight of the i-th data source, A i is the attention output matrix, U is the final generated user capability vector, and N is the number of data sources;

[0104] The capability vector optimization unit optimizes the user capability vector and adjusts the vector to match the skill requirements of the industry. The optimization goal is to minimize the following loss function:

[0105]

[0106] Among them, S is the skill vector required by the industry, L(U) is the optimization loss, and U is the user ability vector;

[0107] The reinforcement learning recommendation module dynamically generates personalized training paths based on the relationship between user ability vectors and skills in the knowledge graph;

[0108] The state modeling unit is used to encode the relationship between the user capability vector and the skills in the knowledge graph into a state representation of reinforcement learning. The formula is defined as:

[0109]

[0110] Among them, U is the user capability vector, S i is the embedding vector of the i-th skill in the knowledge graph, d is the vector dimension, s t is the state vector at time step t, softmax is the normalization function, and M is the total number of relevant skills in the knowledge graph;

[0111] The policy optimization unit generates the action probability distribution based on the proximal policy optimization algorithm. The objective function is defined as:

[0112]

[0113] Among them, L CLIP (θ) is the objective function, π θ is the policy network parameterization function, is the advantage function estimate, ∈ is the policy update clipping threshold, is the probability distribution of the old policy network, π θ (a t |s t ) is the current policy network in state s t Next select action a t The probability of E t is the expected value at time step t;

[0114] The reward calculation unit generates an instant reward signal based on the user's ability improvement and path feasibility. The formula is defined as:

[0115] r t =λ1·ΔU t +λ2·Sim(U t ,S t )-λ3·Fatigue t ,

[0116] Among them, ΔU t is the user capability vector change, Sim(U t ,S t ) is the cosine similarity between user ability and target skill, Fatigue t is the learning fatigue, λ1, λ2, λ3 are weight coefficients;

[0117] The path generation unit generates a personalized training path based on the action probability distribution output by the policy network. The action selection formula is:

[0118]

[0119] Among them, Q(s t ,a) is the action value function, η is the exploration coefficient, A is the action space,

[0120] Entropy(π(·|s t )) is the strategy entropy;

[0121] The virtual training module receives personalized training paths, combines physiological dynamics models with user ability vectors to simulate health management scenarios, and outputs virtual training results;

[0122] The physiological dynamics modeling unit generates personalized physiological parameters based on the user's ability vector and constructs a differential equation model of the cardiovascular system. The formula is defined as:

[0123]

[0124] Where P(t) is the blood pressure value at time t, Q in (t) is the blood flow into the heart, Q out (t) is the peripheral blood flow, C is the vascular compliance, and R is the peripheral vascular resistance;

[0125] The scenario simulation unit receives the skill training instructions in the personalized training path and outputs the virtual patient state in combination with the physiological dynamics model. The state transition formula is:

[0126] s t+1 =f NN (e t ,g t )+τ·N(0,1),

[0127] Among them, e t is the virtual patient state vector at time t, g t is the training action of step t in the training path, f NN is the neural network mapping function, τ is the noise intensity coefficient;

[0128] The real-time feedback unit compares the virtual patient state with the target physiological indicators to generate action suggestions. The loss function is defined as:

[0129]

[0130] Among them, e target is the target physiological index vector, e opt is the set of standard operating actions recommended by medical guidelines, and λ is the action deviation penalty coefficient;

[0131] The evaluation and optimization module receives training results and generates optimization instructions by comparing the matching degree between the user capability vector and the knowledge graph;

[0132] The causal effect evaluation unit quantifies the effect of personalized training paths on user capabilities based on the double difference model. The formula is defined as:

[0133] Y it=β0+β1·Treat i +β2·Post t +δ·(Treat i Post t )+∈ it ,

[0134] Among them, Y it is the observation result of user i at time t, β0 is the intercept term of the model, and Treat i is a binary variable, β1 is the inherent difference between the treatment group and the control group, Post t is a binary variable, β2 is the time effect, δ is the double difference estimator, ∈ it is the error term;

[0135] The online A / B testing unit dynamically assigns users to different recommendation strategy groups through Thompson sampling. The action selection probability update formula is:

[0136]

[0137] Among them, P(a) is the probability of selecting strategy a in the next time step.

[0138] Beta(α a ,β a ) is the Beta prior distribution of strategy a, α a is the prior success number of strategy a, β a is the prior failure count of strategy a, n succ is the number of successful times of strategy a in the current experimental cycle, n fail is the number of failures of strategy a in the current experimental cycle;

[0139] The feedback control unit generates model optimization instructions based on the causal effect and A / B test results. The weight update formula is:

[0140]

[0141] Among them, w old is the current model parameter vector, w new is the updated model parameter vector, η is the learning rate, is the gradient of the loss function to the model parameters, λ1 and λ2 are weight coefficients, and L causal is the causal effect loss, L AB is the strategy diversity loss.

[0142] The multimodal data collection module overcomes the limitations of traditional single-dimensional data collection by integrating four types of data: biological indicators, behavioral logs, psychological assessments, and industry requirements. Biosensors capture physiological signals such as heart rate and blood pressure in real time, behavioral logs record users' daily health management activities, psychological assessment scales quantify emotional states, and industry requirements data dynamically updates job skill requirements. The collection of multi-source, heterogeneous data provides high-density, multi-perspective user portraits for subsequent analysis, avoiding training biases caused by one-sided data.

[0143] The data processing and knowledge fusion module addresses the time misalignment issue of multi-device data. The dynamic time warping algorithm aligns cross-device time series data, eliminating interference caused by differences in device sampling frequencies. A conditional generative adversarial network fills in missing values, generating synthetic data consistent with the user's true physiological state and addressing data sparsity. Complete, aligned, and structured data provides reliable input for personalized modeling, avoiding model distortion caused by data fragmentation in traditional solutions.

[0144] The personalized capability modeling module uses a multi-head attention mechanism to fuse multimodal data with the knowledge graph to extract key features of user capabilities. For example, the "anxiety index" in psychological assessment data is associated with the "emergency handling ability" in the knowledge graph to generate a vector representation of the user's emergency response capabilities. Weighted average calculations dynamically balance the contributions of various data sources. For example, when biological indicators have a higher weight than behavioral logs, the system focuses on the user's physiological state. The capability vector optimization unit ensures that the user's capability model is closely aligned with job requirements by minimizing industry skill gaps. Compared to static scoring models, this module supports real-time updates. The vector is automatically iterated after each user training session, accurately reflecting the changing trends in capabilities.

[0145] In the reinforcement learning recommendation module, the state modeling unit encodes the user's ability vector and knowledge graph into reinforcement learning states. For example, when a user has a "low cardiovascular health score" and "lack of first aid skills," the state vector triggers a recommendation for first aid training. The PPO algorithm introduces a clipping threshold in policy updates to balance exploration and utilization, avoiding the overfitting problem of traditional Q-learning. The reward function design integrates ability improvement, path feasibility, and fatigue. For example, if short-term high-intensity training improves ability but increases fatigue, the system automatically reduces the reward weight for such actions. The path generation unit dynamically expands skill combinations through an entropy-increasing exploration mechanism, for example, recommending a cross-domain skill package of "nutrition + sports rehabilitation," breaking through the rigidity of traditional linear training paths.

[0146] In the virtual training module, the physiological dynamics modeling unit generates personalized parameters based on the user's ability vector. For example, vascular resistance parameters are calculated based on the user's age and blood pressure history, simulating real cardiovascular responses. The scenario simulation unit uses a neural network to map training actions to changes in physiological state. For example, the action of "adjusting insulin dose" directly affects the blood glucose curve of the virtual patient. The real-time feedback unit compares operation results with medical guidelines. For example, when the blood pressure control deviation exceeds 10%, it generates a recommendation of "adjusting medication dosage" or "increasing monitoring frequency." This closed-loop simulation of virtual scenarios allows users to master complex health management skills in a zero-risk environment, addressing the pain points of high cost and high risk of traditional practical training.

[0147] In the evaluation and optimization module, a difference-in-difference model quantifies the causal effects of the training pathway, eliminating the influence of temporal trends and individual differences. For example, of the 30% improvement in competence in the intervention group, 15% can be attributed to the pathway itself. Thompson sampling dynamically allocates resources. For example, new nurses prioritize the combination of "first aid skills + communication skills," while veteran employees focus on "chronic disease management + data analysis." The feedback control unit balances effectiveness and diversity through a multi-objective loss function. A continuous optimization mechanism enables the system to adapt to changes in industry standards. For example, after the introduction of new medical insurance policies, the knowledge graph and reward function are updated simultaneously to ensure the forward-looking nature of the training program.

[0148] A terminal device includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, a customized big health talent training system based on artificial intelligence is implemented.

[0149] A storage medium stores a computer program, which, when executed by a processor, implements a large-scale health customized talent training system based on artificial intelligence.

[0150] The terminal device uses the processor to efficiently run multimodal data collection, knowledge graph construction, and reinforcement learning recommendation algorithms, achieving real-time analysis of user biological indicators and behavioral data and dynamic path generation, ensuring millisecond-level response for personalized training programs;

[0151] The storage medium provides standardized program packaging and cross-platform deployment capabilities, supporting medical institutions and health management platforms to flexibly call system functions, such as quickly updating the skill knowledge base through local storage media in an offline environment.

[0152] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A big health customized talent training system based on artificial intelligence, characterized by: The talent training system includes: Multimodal data collection module, used to obtain user biometric data, behavioral log data, psychological assessment data and industry demand data; The data processing and knowledge fusion module is used to align the acquired data in time series and fill in missing values, and to build a skill-related knowledge graph based on industry demand data; The personalized capability modeling module receives the knowledge graph and aligned multimodal data to generate a user capability vector; The reinforcement learning recommendation module dynamically generates personalized training paths based on the relationship between user ability vectors and skills in the knowledge graph; The virtual training module receives personalized training paths, combines physiological dynamics models with user ability vectors to simulate health management scenarios, and outputs virtual training results; The evaluation and optimization module receives the training results and generates optimization instructions by comparing the matching degree between the user capability vector and the knowledge graph.

2. The artificial intelligence-based customized health talent training system according to claim 1 is characterized in that: The data processing and knowledge fusion module further includes: The time series alignment unit is used to align the biological indicator data, behavioral log data, and psychological assessment data obtained by the multimodal data acquisition module across devices. The specific algorithm is the dynamic time warping algorithm, and the formula is defined as: Where X={x1,x2,...,x n } is the timing data from device A, Y={y1,y2,...,y m } is the timing data from device B, w is the regular path set, λ is the time misalignment penalty coefficient, x i 、y j is the observation value at time i and j in the sequence; The missing value filling unit receives the aligned data output by the time series alignment unit and fills the missing biological indicator data through the conditional generative adversarial network. The optimization objectives of the generator G and the discriminator D are: Among them, x partial is part of the observed data, z is the noise vector, G(z|x partial ) is the generator network D(x) is the discriminator network, p data is the real data distribution, p z is a Gaussian distribution, L G is the generator, L D is the discriminator, E is the expected operation; The knowledge graph construction unit receives the complete data processed by the missing value filling unit and constructs the skill association knowledge graph based on the TransR model. The entity relationship scoring function is: Among them, h is the head entity embedding vector, t is the tail entity embedding vector, r is the relationship embedding vector, h r is the projection vector of the entity in the relational space r, t r is the projection form of the tail entity vector.

3. The AI-based customized talent training system for big health according to claim 1 is characterized in that: The personalized capability modeling module further includes: The data fusion unit is used to fuse multimodal data with skill information in the knowledge graph to construct a comprehensive user capability representation vector. The algorithm is a multi-head attention mechanism, and the formula is defined as: Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k A is the key and query dimension, i is the attention output matrix; The user capability vector generation unit generates the final user capability vector through the fused data. The algorithm is a weighted average calculation method, and the formula is defined as: Among them, α i is the weight of the i-th data source, A i is the attention output matrix, U is the final generated user capability vector, and N is the number of data sources; The capability vector optimization unit optimizes the user capability vector and adjusts the vector to match the skill requirements of the industry. The optimization goal is to minimize the following loss function: Among them, S is the skill vector required by the industry, L(U) is the optimization loss, and U is the user ability vector.

4. The AI-based customized talent training system for big health according to claim 1 is characterized in that: The reinforcement learning recommendation module further includes: The state modeling unit is used to encode the relationship between the user capability vector and the skills in the knowledge graph into a state representation of reinforcement learning. The formula is defined as: Among them, U is the user capability vector, S i is the embedding vector of the i-th skill in the knowledge graph, d is the vector dimension, s t is the state vector at time step t, softmax is the normalization function, and M is the total number of relevant skills in the knowledge graph; The policy optimization unit generates the action probability distribution based on the proximal policy optimization algorithm. The objective function is defined as: Among them, L CLIP (θ) is the objective function, π θ is the policy network parameterization function, is the advantage function estimate, ∈ is the policy update clipping threshold, is the probability distribution of the old policy network, π θ (a t |s t ) is the current policy network in state s t Next select action a t The probability of E t is the expected value at time step t; The reward calculation unit generates an instant reward signal based on the user's ability improvement and path feasibility. The formula is defined as: r t =λ1·ΔU t +λ2·Sim(U t ,S t )-λ3·Fatigue t , Among them, ΔU t is the user capability vector change, Sim(U t ,S t ) is the cosine similarity between user ability and target skill, Fatigue t is the learning fatigue, λ1, λ2, λ3 are weight coefficients; The path generation unit generates a personalized training path based on the action probability distribution output by the policy network. The action selection formula is: Among them, Q(s t ,a) is the action value function, η is the exploration coefficient, A is the action space, Entropy(π(·|s t )) is the policy entropy.

5. The AI-based customized talent training system for big health according to claim 1 is characterized in that: The virtual training module further includes: The physiological dynamics modeling unit generates personalized physiological parameters based on the user's ability vector and constructs a differential equation model of the cardiovascular system. The formula is defined as: Where P(t) is the blood pressure value at time t, Q in (t) is the blood flow into the heart, Q out (t) is the peripheral blood flow, C is the vascular compliance, and R is the peripheral vascular resistance.

6. The AI-based customized talent training system for big health according to claim 1 is characterized in that: The virtual training module further includes: The scenario simulation unit receives the skill training instructions in the personalized training path and outputs the virtual patient state in combination with the physiological dynamics model. The state transition formula is: s t+1 =f NN (e t ,g t )+τ·N(0,1), Among them, e t is the virtual patient state vector at time t, g t is the training action of step t in the training path, f NN is the neural network mapping function, τ is the noise intensity coefficient; The real-time feedback unit compares the virtual patient state with the target physiological indicators to generate action suggestions. The loss function is defined as: Among them, e target is the target physiological index vector, e opt is the set of standard operating actions recommended by medical guidelines, and λ is the action deviation penalty coefficient.

7. The AI-based customized talent training system for big health according to claim 1 is characterized in that: The evaluation and optimization module further includes: The causal effect evaluation unit quantifies the effect of personalized training paths on user capabilities based on the double difference model. The formula is defined as: Y it =β0+β1·Treat i +β2·Post t +δ·(Treat i ·Post t )·+∈ it , Among them, Y it is the observation result of user i at time t, β0 is the intercept term of the model, and Treat i is a binary variable, β1 is the inherent difference between the treatment group and the control group, Post t is a binary variable, β2 is the time effect, δ is the double difference estimator, ∈ it is the error term.

8. The AI-based customized talent training system for big health according to claim 1 is characterized by: The evaluation and optimization module further includes: The online A / B testing unit dynamically assigns users to different recommendation strategy groups through Thompson sampling. The action selection probability update formula is: Among them, P(a) is the probability of selecting strategy a in the next time step. Beta(α a ,β a ) is the Beta prior distribution of strategy a, α a is the prior success number of strategy a, β a is the prior failure count of strategy a, n succ is the number of successful times of strategy a in the current experimental cycle, n fail is the number of failures of strategy a in the current experimental cycle; The feedback control unit generates model optimization instructions based on the causal effect and A / B test results. The weight update formula is: Among them, w old is the current model parameter vector, w new is the updated model parameter vector, η is the learning rate, is the gradient of the loss function to the model parameters, λ1 and λ2 are weight coefficients, and L causal is the causal effect loss, L AB is the strategy diversity loss.

9. A terminal device, characterized in that: It includes a processor and a memory, the memory stores a computer program, and when the processor executes the computer program, it implements an artificial intelligence-based customized health talent training system according to any one of claims 1 to 8.

10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, it implements an artificial intelligence-based customized health talent training system according to any one of claims 1 to 8.

Images