Endocrine nursing teaching decision-making system based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence, and provides an endocrine nursing teaching decision-making system based on deep learning, and the system comprises a data collection module, a data preprocessing module, a feature extraction module, a teaching evaluation module, and a teaching decision-making module. The teaching evaluation module introduces a fairness constraint dual-BiGRU model, and improves the consistency and stability of teaching evaluation among different student groups through sub-group division and a fairness constraint mechanism; the teaching decision module introduces a risk awareness D3QN model, constructs a risk sensitive Q value function, performs segmented weighting on teaching action return in combination with a static spectrum risk measurement mechanism, and optimizes the decision ability and risk regulation ability of a strategy network; the system can realize deep fusion and personalized strategy recommendation of multi-source heterogeneous teaching data, is particularly suitable for complex medical teaching scenes such as endocrine nursing and the like, and has good fairness, adaptability and practicability.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an endocrine nursing teaching decision-making system based on deep learning. Background Art

[0002] With the rapid development of artificial intelligence technology, deep learning models are increasingly used in medical health, smart education and other fields. At present, intelligent teaching systems have been preliminarily applied in medical nursing teaching. Common methods include knowledge push systems based on knowledge graphs, student behavior prediction models based on convolutional neural networks (CNN) or long short-term memory networks (LSTM), and teaching strategy generation methods based on reinforcement learning. Such systems can, to a certain extent, realize the association modeling between teaching content and student behavior data, and assist teachers in teaching evaluation and strategy recommendation. However, existing intelligent nursing teaching systems generally have the following limitations: First, most systems use the average characteristics of the entire student group as the modeling basis, lack deep modeling of sub-group characteristic differences, and find it difficult to identify specific behavior patterns (such as "fast-paced skipping learning type"). ” or “video-dependent” students), which can easily cause the evaluation results to be distorted for some groups, thereby affecting the adaptability and fairness of teaching; secondly, the current application of reinforcement learning in teaching strategy optimization generally takes the traditional Q-value function as the core goal, ignoring the potential risk differences brought by teaching actions to different learning paths; this makes the generated teaching strategies lack stability and interpretability in complex teaching situations, and it is difficult to meet the comprehensive requirements of nursing education for strategy safety and expected benefits; finally, some systems lack the ability to integrate and model multi-source heterogeneous teaching data (such as case texts, answer records, video behaviors, etc.), and fail to fully extract the deep correlation characteristics between teaching content and student behavior, affecting the accuracy and responsiveness of subsequent evaluation and decision-making modules, resulting in a lack of stability and pertinence in teaching paths. Summary of the Invention

[0003] The present invention aims to overcome the problems of existing intelligent nursing teaching systems, such as insufficient personalized evaluation, significant group fairness deviation, and limited risk control ability of teaching decision-making. It proposes an endocrine nursing teaching decision-making system based on deep learning, which integrates multidimensional data modeling, fairness constraints and risk perception optimization strategies to improve the fairness of teaching evaluation and the scientific nature of teaching decision-making. The system introduces the fairness constraint dual-BiGRU model to cluster model student groups with different learning behavior characteristics, and imposes sub-group fairness constraints during the training process, effectively suppressing the systematic bias of the evaluation model for specific groups. At the same time, in order to improve the risk control ability of the teaching strategy, a static spectral risk measurement mechanism is introduced to perform piecewise weighted regulation on the reward distribution of teaching actions, and realize the construction of risk-sensitive Q-value function in reinforcement learning, thereby optimizing the stability and decision-making adaptability of the strategy network. The system can output personalized, risk-controllable and fair-optimized teaching decision paths based on multi-dimensional heterogeneous teaching data, which is particularly suitable for complex teaching scenarios such as endocrine nursing.

[0004] The present invention provides an endocrine nursing teaching decision-making system based on deep learning, which includes a data acquisition module, a data preprocessing module, a feature extraction module, a teaching evaluation module and a teaching decision-making module;

[0005] Data collection module, which collects multi-dimensional endocrine nursing teaching data and constructs the original teaching data set;

[0006] The data preprocessing module cleans, encodes and standardizes the original teaching data set to generate a cleaned teaching data set;

[0007] The feature extraction module extracts teaching content features and student behavior features from the cleaned teaching dataset to obtain a comprehensive feature set. Teaching content features include text semantic features, case difficulty labels, and knowledge point associations; student behavior features include time series learning trajectories, question mastery assessments, and interaction preference distributions.

[0008] The teaching evaluation module introduces the Fair Constraint Dual-BiGRU model. The Fair Constraint Dual-BiGRU model is constructed by establishing a BiGRU model, introducing the K-Means subgroup division mechanism and the subgroup fairness constraint mechanism during the BiGRU model training process, and optimizing the evaluation performance of the BiGRU model. The Fair Constraint Dual-BiGRU model processes the comprehensive feature set to generate teaching evaluation results. The teaching evaluation results include a knowledge point matrix, a weakness diagnosis report, and a teaching adaptability curve. The Fair Constraint Dual-BiGRU model includes the BiGRU model.

[0009] The teaching decision module defines the teaching action space and introduces the risk-aware D3QN model. The risk-aware D3QN model is constructed by establishing a D3QN model and introducing a static spectral risk measurement mechanism during the D3QN model training process to optimize the traditional Q-value function of the D3QN model. The risk-aware D3QN model is used to evaluate the risk sensitivity and optimize the strategy of the teaching actions in the teaching action space to obtain the teaching decision results. The risk-aware D3QN model includes the D3QN model.

[0010] Furthermore, the process of processing the comprehensive feature set through the fairness constraint dual-BiGRU model to generate the teaching evaluation results specifically includes the following steps:

[0011] Step S1: Input the comprehensive feature set into the BiGRU model for initial training to generate initial teaching evaluation results;

[0012] Step S2: Set the error tolerance threshold and reference evaluation benchmark value, perform multi-dimensional feature partitioning on the comprehensive feature set through K-Means, obtain typical sub-groups, input the feature data of each individual in the typical sub-group into the BiGRU model, and generate sub-group evaluation results; introduce a sub-group fairness constraint mechanism during the model training process to constrain the deviation between the mean of the sub-group evaluation results and the reference evaluation benchmark value, and generate sub-group error constraints;

[0013] Step S3: Taking the subpopulation error constraint as a training constraint, the BiGRU model training task is formalized as a multiple fairness constraint-nonconvex optimization problem. The primal-dual optimization algorithm is introduced to transform the multiple fairness constraint-nonconvex optimization problem into a dual objective function for joint training. The dual objective function is used during the training process to perform alternating iterative optimization of the BiGRU model's primary parameters and dual variables.

[0014] Furthermore, the risk-aware D3QN model is used to obtain the teaching decision results, which specifically includes the following steps:

[0015] Step B1: Construct a teaching state representation based on the teaching evaluation results. The teaching action space includes basic content review, targeted training push, case video guidance, and assessment task arrangement. Execute the teaching actions in the teaching action space, monitor student state changes, and construct state transition pairs.

[0016] Step B2: Design a reward function and combine it with state transition pairs to generate an immediate reward value. Input the state transition pairs and immediate reward values ​​into the D3QN model for training. Use the traditional Q-value function to evaluate the long-term benefits of the teaching action, obtain the state-action reward distribution, and obtain the initial policy network.

[0017] Step B3: During the training process, a static spectral risk measurement mechanism is introduced to replace the traditional Q-value function-based training method in the D3QN model. Specifically, the state-action reward distribution is modeled as a complete reward distribution. Based on the complete reward distribution, a risk spectrum function is introduced to construct a piecewise linear approximation function as a risk-weighted mapping. The piecewise linear approximation function is embedded in the traditional Q-value function to construct a risk-sensitive Q-value function. This realizes the perceptual control of different risk quantiles in the teaching action reward distribution, generates risk-sensitive Q values, and optimizes the policy network parameters in the D3QN model to obtain the current teaching state and the optimized policy network.

[0018] Step B4: Based on the current teaching state and the optimization strategy network, the optimal teaching action is selected from the teaching action space to generate the teaching decision result.

[0019] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0020] By introducing the fairness-constrained dual-BiGRU model, the present invention realizes the sub-group identification and structured modeling of multiple types of student learning behaviors in endocrine nursing teaching. It can dynamically calibrate the teaching evaluation path for typical learning groups such as high-frequency interactive, video-dependent, and low-scoring review groups. The model introduces a sub-group fairness constraint mechanism during the training process and dynamically compresses the evaluation deviations of different groups through the primal-dual optimization method, thereby significantly improving the consistency and fairness of teaching evaluation results among diverse student groups, solving the evaluation distortion problem caused by ignoring group differences in traditional models, and enhancing the system's ability to accurately feedback personalized learning status in complex nursing teaching environments.

[0021] At the same time, the present invention introduces a static spectral risk measurement mechanism into the teaching strategy optimization module and constructs a risk-sensitive Q-value function for piecewise weighted control of the distribution of teaching action rewards. Unlike traditional reinforcement learning models that only target expected returns, the present invention can perceive the risk quantile differences brought about by different actions in the teaching path and dynamically adjust the selection priority of teaching behaviors. Through the synergistic effect of the risk spectrum function and the piecewise linear approximation function, the system can take into account both teaching benefits and learning risks when generating teaching strategies, effectively improving the stability and controllability of the teaching path output, and providing reliable guarantees for scenarios such as pushing high-difficulty content and recommending key training tasks in nursing education.

[0022] In summary, the present invention integrates core technologies such as multidimensional data-driven modeling, fairness constraint optimization and risk-aware strategy control to construct an intelligent teaching decision-making system with high adaptability and regulation capabilities; the system not only improves the comprehensive perception of students' learning status, but also enhances the intervention value of the system in complex teaching decisions through a refined strategy output mechanism; especially in professional fields such as endocrine nursing, where the knowledge system is complex and teaching feedback requirements are high, the present invention can greatly improve teaching accuracy and adaptability, significantly improve the problems of uneven distribution of traditional teaching resources and lagging strategy regulation, and promote nursing education to develop in a more intelligent, fair and safe direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A comparison chart of the training effects of the teaching decision-making models proposed in Examples 4 and 5;

[0024] Figure 2 This is the teaching adaptability curve diagram proposed in Example 6;

[0025] Figure 1 In the figure, the horizontal axis is the number of training rounds (1-20), and the vertical axis is the average return value. The comparison results show that the risk-aware D3QN model (blue curve) converges faster in the early stages and achieves a higher average return value in the later stages, outperforming the traditional D3QN model (orange curve) overall.

[0026] Figure 2 In the figure, the left side shows the teaching adaptability score, and the bottom side shows the week number. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] In embodiment 1, the present invention provides an endocrine nursing teaching decision system based on deep learning, which includes a data acquisition module, a data preprocessing module, a feature extraction module, a teaching evaluation module and a teaching decision module;

[0029] Data collection module, which collects multi-dimensional endocrine nursing teaching data and constructs the original teaching data set;

[0030] Multi-dimensional endocrine nursing teaching data includes:

[0031] Teaching content data: courseware, cases, test questions, video explanations;

[0032] Nursing knowledge graph data: endocrine disease classification, nursing standards;

[0033] Student learning behavior data: learning time, answer records, and interaction frequency;

[0034] Learning assessment data: regular grades, test scores, and subjective evaluations;

[0035] The data preprocessing module cleans, encodes and standardizes the original teaching data set to generate a cleaned teaching data set;

[0036] The feature extraction module extracts teaching content features and student behavior features from the cleaned teaching dataset to obtain a comprehensive feature set. Teaching content features include text semantic features, case difficulty labels, and knowledge point associations; student behavior features include time series learning trajectories, question mastery assessments, and interaction preference distributions.

[0037] The teaching evaluation module introduces the Fair Constraint Dual-BiGRU model. The Fair Constraint Dual-BiGRU model is constructed by establishing a BiGRU model, introducing the K-Means subgroup division mechanism and the subgroup fairness constraint mechanism during the BiGRU model training process, and optimizing the evaluation performance of the BiGRU model. The Fair Constraint Dual-BiGRU model processes the comprehensive feature set to generate teaching evaluation results. The teaching evaluation results include a knowledge point matrix, a weakness diagnosis report, and a teaching adaptability curve. The Fair Constraint Dual-BiGRU model includes the BiGRU model.

[0038] The teaching decision module defines the teaching action space and introduces the risk-aware D3QN model. The risk-aware D3QN model is constructed by establishing a D3QN model and introducing a static spectral risk measurement mechanism during the D3QN model training process to optimize the traditional Q-value function of the D3QN model. The risk-aware D3QN model is used to evaluate the risk sensitivity and optimize the strategy of the teaching actions in the teaching action space to obtain the teaching decision results. The risk-aware D3QN model includes the D3QN model.

[0039] Example 2: This example is based on Example 1. In this example, the process of processing the comprehensive feature set through the fairness constraint dual-BiGRU model to generate the teaching evaluation results specifically includes the following steps:

[0040] Step S1: Input the comprehensive feature set into the BiGRU model for initial training to generate initial teaching evaluation results;

[0041] Step S2: Set an error tolerance threshold and a reference evaluation benchmark value, perform multi-dimensional feature partitioning on the comprehensive feature set using K-Means, obtain typical subgroups, and input the characteristic data of each individual in the typical subgroup into the BiGRU model to generate subgroup evaluation results. In the model training process, a subgroup fairness constraint mechanism is introduced to constrain the deviation between the mean of the subgroup evaluation results and the reference evaluation benchmark value, generating a subgroup error constraint to ensure that the deviation does not exceed the error tolerance threshold, thereby suppressing the model from generating systematic evaluation bias between different types of students.

[0042] The formula of the subgroup fairness constraint mechanism is as follows:

[0043] ;

[0044] in, represents the typical subgroup number, represents the sample number within the typical subpopulation, Indicates the The number of samples of a typical subpopulation, Indicates the Among the typical subgroups Comprehensive feature input data of samples; Represents the main parameters of the Bi-GRU model, Indicates that the Bi-GRU model is samples The predicted results of teaching evaluation, that is, the subgroup evaluation results; Indicates the The average of all sample prediction evaluation results in a typical subgroup; Indicates the Reference assessment benchmark values ​​for typical subgroups; represents the distance function; Indicates the The error tolerance threshold of a typical subgroup; means that this calibration constraint is imposed on all representative subpopulations;

[0045] Typical subgroups include: high-frequency interactive students, low-scoring review students, video-dependent students, fast-paced skipping students, and exam-oriented students.

[0046] Step S3: Taking the subgroup error constraint as a training constraint, the training task of the BiGRU model is formalized as a multiple fairness constraint-nonconvex optimization problem; introducing the primal-dual optimization algorithm, the multiple fairness constraint-nonconvex optimization problem is converted into a dual objective function for joint training; using the dual objective function during the training process, the main parameters and dual variables of the BiGRU model are alternately iteratively optimized; the main parameters are used to minimize the overall teaching error, and the dual variables dynamically adjust the fairness constraint strength of each subgroup; through continuous iterative optimization, the model can automatically identify and suppress the path with the largest subgroup evaluation deviation, effectively improving the fairness, stability and generalization ability of the teaching evaluation model, optimizing the initial teaching evaluation results, and generating teaching evaluation results;

[0047] Optimization objectives of multiple fairness constraints-non-convex optimization problems:

[0048] ;

[0049] in, represents the mean square error function, which is used as the teaching evaluation loss function, represents the total number of training samples, Indicates the The true teaching evaluation value corresponding to the samples, By adjusting the main parameters , minimize the overall teaching evaluation error; Indicates the The average difference measurement function between the model prediction value and the reference evaluation benchmark value in a typical subgroup;

[0050] Transforming the multi-fairness-constrained non-convex optimization problem into a dual objective function helps restructure the multi-constrained non-convex training task, which is difficult to solve directly, into a scalable and controllable joint optimization problem. By introducing dual variables, the model can automatically adjust the constraint strength of each sub-group, effectively suppressing sub-group prediction bias while optimizing overall teaching performance, thereby improving the fairness, robustness, and generalization ability of the model.

[0051] Dual objective function formula:

[0052] ;

[0053] in, represents the dual objective function, represents the total number of typical subpopulations, Indicates the The dual variables of the typical subgroups.

[0054] Example 3: This example is based on Example 1. In this example, the process of generating teaching evaluation results specifically includes the following steps: inputting the comprehensive feature set into the BiGRU model for initial training to generate teaching evaluation results.

[0055] Example 4, according to Figure 1 This embodiment is based on the second embodiment. In this embodiment, the risk-aware D3QN model is used to obtain the teaching decision results. The process specifically includes the following steps:

[0056] Step B1: Construct a teaching state representation based on the teaching evaluation results. The teaching action space includes basic content review, targeted training push, case video guidance, and assessment task arrangement. Execute the teaching actions in the teaching action space, monitor student state changes, and construct state transition pairs.

[0057] Step B2: Design a reward function and combine it with state transition pairs to generate an immediate reward value. Input the state transition pairs and immediate reward values ​​into the D3QN model for training. Use the traditional Q-value function to evaluate the long-term benefits of the teaching action, obtain the state-action reward distribution, and obtain the initial policy network.

[0058] Step B3: During the training process, a static spectral risk measurement mechanism is introduced to replace the traditional Q-value function-based training method in the D3QN model. Specifically, the state-action reward distribution is modeled as a complete reward distribution. Based on the complete reward distribution, a risk spectrum function is introduced to construct a piecewise linear approximation function as a risk-weighted mapping. The piecewise linear approximation function is embedded in the traditional Q-value function to construct a risk-sensitive Q-value function. This realizes the perceptual control of different risk quantiles in the teaching action reward distribution, generates risk-sensitive Q values, and optimizes the policy network parameters in the D3QN model to obtain the current teaching state and the optimized policy network.

[0059] The formula used in the static spectral risk measurement mechanism is:

[0060] ;

[0061] in, represents the risk adjustment function, Indicates that in the strategy Next, status and actions The risk-sensitive Q-value function of represents the scaling factor, represents the mathematical expectation; represents the risk spectrum function, represents the piecewise linear approximation function, represents the offset term, represents the stretch factor, represents the full return distribution;

[0062] Piecewise linear approximation function formula:

[0063] ;

[0064] in, represents the predicted return for a state-action combination, The number of segments representing the state-action reward distribution, Indicates the Return quantiles, Indicates the The weighting coefficient of each segment, Indicates the The quantile corresponding to the cumulative distribution function, Indicates the Segment intermediate points, used for numerical integration approximation; Represents the risk spectrum function value, at the quantile The risk preference weight of Represents the risk spectrum function value, at the quantile Risk preference weight; represents the negative partial function;

[0065] Step B4: Based on the current teaching state and the optimization strategy network, the optimal teaching action is selected from the teaching action space to generate the teaching decision result.

[0066] Example 5, according to Figure 1 This embodiment is based on the second embodiment. In this embodiment, the risk-aware D3QN model is used to obtain the teaching decision results. The process specifically includes the following steps:

[0067] Step R1: Construct a teaching state representation based on the teaching evaluation results. The teaching action space includes basic content review, targeted training push, case video guidance, and assessment task arrangement; execute the teaching actions in the teaching action space, monitor student state changes, and construct state transition pairs;

[0068] Step R2: Design a reward function and combine it with the state transition pair to generate an immediate reward value. Input the state transition pair and the immediate reward value into the D3QN model for training. Use the traditional Q-value function to evaluate the long-term benefits of the teaching action and obtain the current teaching state and policy network.

[0069] Step R3: Based on the current teaching state and strategy network, select the optimal teaching action from the teaching action space and generate the teaching decision result.

[0070] Example 6, according to Figure 2 This embodiment is based on the fourth embodiment. In this embodiment,

[0071] The teaching evaluation module introduces the Fair Constraint Dual-BiGRU model. The Fair Constraint Dual-BiGRU model is constructed by establishing a BiGRU model, introducing the K-Means subgroup division mechanism and the subgroup fairness constraint mechanism during the BiGRU model training process, and optimizing the evaluation performance of the BiGRU model. The Fair Constraint Dual-BiGRU model processes the comprehensive feature set to generate teaching evaluation results. The teaching evaluation results include a knowledge point matrix, a weakness diagnosis report, and a teaching adaptability curve. The Fair Constraint Dual-BiGRU model includes the BiGRU model.

[0072] The teaching evaluation results of student S_17 are shown in Table 1:

[0073] Table 1 Teaching evaluation results (student S_17)

[0074] ;

[0075] Weakness diagnosis report for student S_17:

[0076] High-risk knowledge points:

[0077] K4: Thyroid crisis management – ​​The correct answer rate was low (29%), the corresponding case video playback interruption rate was high, and students spent much less time studying this module than the class average;

[0078] K2: Hypoglycemia emergency procedures - there is a tendency to confuse options, and many errors are concentrated in the operation sequence judgment questions;

[0079] Student S_17, teaching adaptability curve for weeks 1-8, shown in the attached figure; the teaching adaptability score is on the left and the week number is below;

[0080] The teaching decision module defines the teaching action space and introduces the risk-aware D3QN model. The risk-aware D3QN model is constructed by establishing a D3QN model and introducing a static spectral risk measurement mechanism during the D3QN model training process to optimize the traditional Q-value function of the D3QN model. The risk-aware D3QN model is used to evaluate the risk sensitivity and optimize the strategy of the teaching actions in the teaching action space to obtain the teaching decision results.

[0081] Table 2 shows the teaching decision results of S_17, as shown in Table 1:

[0082] Table 2 Teaching decision results (student S_17)

[0083] .

[0084] The present invention and its embodiments are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. An endocrine nursing teaching decision system based on deep learning, comprising a data preprocessing module that generates a cleaned teaching data set; characterized in that: The system also includes a feature extraction module, a teaching evaluation module and a teaching decision module; Feature extraction module, extracts teaching content features and student behavior features from the cleaned teaching data set to obtain a comprehensive feature set; The teaching evaluation module introduces the Fair Constraint Dual-BiGRU model, which processes the comprehensive feature set through the Fair Constraint Dual-BiGRU model to generate teaching evaluation results. The Fair Constraint Dual-BiGRU model includes the BiGRU model. The teaching decision module defines the teaching action space and introduces the risk-aware D3QN model. Based on the teaching evaluation results, the risk-aware D3QN model is used to perform risk sensitivity assessment and strategy optimization on the teaching actions in the teaching action space to obtain the teaching decision results. The risk-aware D3QN model includes a D3QN model.

2. The endocrine nursing teaching decision system based on deep learning according to claim 1 is characterized by: The teaching content features include text semantic features, case difficulty labels and knowledge point associations; the student behavior features include time series learning trajectories, question mastery assessment and interaction preference distribution.

3. The endocrine nursing teaching decision system based on deep learning according to claim 1 is characterized by: The process of generating teaching evaluation results through the fairness constraint dual-BiGRU model includes the following steps: Step S1: Input the comprehensive feature set into the BiGRU model for initial training to generate initial teaching evaluation results; Step S2: Set a reference evaluation benchmark value, perform multi-dimensional feature partitioning on the comprehensive feature set using K-Means, obtain typical subgroups, and generate subgroup evaluation results; introduce a subgroup fairness constraint mechanism to constrain the deviation between the mean of the subgroup evaluation results and the reference evaluation benchmark value, and generate a subgroup error constraint; Step S3: Taking the sub-population error constraint as the training constraint term, the training task of the BiGRU model is formalized as a multiple fairness constraint-non-convex optimization problem; introducing the primal-dual optimization algorithm, the multiple fairness constraint-non-convex optimization problem is converted into a dual objective function; using the dual objective function during the training process, the main parameters and dual variables of the Bi-GRU model are alternately iteratively optimized to optimize the initial teaching evaluation results and generate teaching evaluation results.

4. The endocrine nursing teaching decision system based on deep learning according to claim 3 is characterized by: Typical subgroups include: high-frequency interactive student group, low-scoring review student group, video-dependent student group, fast-paced skipping student group and exam-oriented student group.

5. The endocrine nursing teaching decision system based on deep learning according to claim 1 is characterized by: The process of obtaining teaching decision results using the risk-aware D3QN model includes the following steps: Step B1: Construct the teaching state representation based on the teaching evaluation results, execute the teaching action, and construct the state transition pair; Step B2: Design a reward function and combine it with state transition pairs to generate an immediate reward value. Input the state transition pairs and immediate reward values ​​into the D3QN model for training. Use the traditional Q-value function to evaluate the teaching action and obtain the state-action reward distribution. Step B3: During the training process, a static spectral risk measurement mechanism is introduced to replace the traditional Q-value function to obtain the optimized policy network; Step B4: Based on the optimization strategy network, select the optimal teaching action and generate the teaching decision result.

6. The deep learning-based endocrine nursing teaching decision system according to claim 5, characterized in that: The static spectral risk measurement mechanism in step B3 is specifically as follows: modeling the state-action reward distribution as a complete reward distribution, introducing the risk spectrum function based on the complete reward distribution, constructing a piecewise linear approximation function, embedding the piecewise linear approximation function into the traditional Q-value function, constructing a risk-sensitive Q-value function, generating risk-sensitive Q-values, optimizing the policy network parameters in the D3QN model, and obtaining the optimized policy network.

Images