Intelligent evaluation method for logical thinking training and related device

Abstract

The invention relates to the field of artificial intelligence, and discloses an intelligent evaluation method for logical thinking training and a related device, which are used for carrying out intelligent evaluation on a target test user and improving the accuracy of evaluation decision. The method comprises the following steps: performing motion state tracking and motion state feature extraction on a target evaluation user through a state space tracking model to obtain a plurality of motion state features; performing question-answer state feature extraction to obtain a plurality of question-answer state features; performing decision analysis through the plurality of initial intelligent evaluation decision models to obtain prediction data and initial model parameters; performing parameter optimization analysis through a particle swarm algorithm to obtain target model parameters; performing parameter execution strategy analysis and parameter updating to obtain a plurality of target intelligent evaluation decision models; and performing model weight analysis to obtain model weight data of each target intelligent evaluation decision model, and performing model integration through the model weight data to obtain an intelligent evaluation decision model set.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to an intelligent evaluation method and related device for logical thinking training. Background Art

[0002] With the development of artificial intelligence and machine learning technologies, researchers are seeking an intelligent assessment method to more accurately and efficiently evaluate and improve individuals' logical thinking abilities.

[0003] However, how to comprehensively establish an accurate evaluation system is a complex issue. Intelligent evaluation methods require large-scale data sets and complex algorithm models to support them, and the technology and resources for data collection and processing also face challenges. How to effectively integrate multiple intelligent evaluation models to improve the accuracy and robustness of the evaluation is also a problem that needs to be solved. Summary of the invention

[0004] The present invention provides an intelligent evaluation method and related devices for logical thinking training, which are used to perform intelligent evaluation on target test users and improve the accuracy of evaluation decisions.

[0005] The first aspect of the present invention provides an intelligent evaluation method for logical thinking training, which includes: tracking the motion state of a target evaluation user through a state space tracking model to obtain a target motion data set, and extracting motion state features from the target motion data set to obtain multiple motion state features; obtaining a target question and answer data set for the target evaluation user, and extracting question and answer state features from the target question and answer data set through a natural language processing model to obtain multiple question and answer state features; inputting the multiple motion state features and the multiple question and answer state features into multiple initial intelligent evaluation decision models for decision analysis, and obtaining prediction data and initial Model parameters; obtaining the model parameter range of each initial intelligent evaluation decision model, and performing parameter optimization analysis on the initial model parameters according to the model parameter range through the particle swarm algorithm to obtain the target model parameters of each initial intelligent evaluation decision model; performing parameter execution strategy analysis on the multiple initial intelligent evaluation decision models through the target model parameters to obtain multiple parameter execution strategies, and performing parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models; performing model weight analysis on the multiple target intelligent evaluation decision models to obtain model weight data for each target intelligent evaluation decision model, and performing model integration through the model weight data to obtain a set of intelligent evaluation decision models.

[0006] In combination with the first aspect, in a first implementation method of the first aspect of the present invention, the motion state of the target evaluation user is tracked through a state-space tracking model to obtain a target motion data set, and motion state feature extraction is performed on the target motion data set to obtain multiple motion state features, including: motion and speed detection is performed on the target evaluation user to obtain motion data and speed data, and vector initialization is performed on the motion data and speed data to obtain an initialized state vector, and the covariance matrix of process noise and measurement noise is initialized; the state transfer equation and measurement equation of the state-space tracking model are constructed, and time-varying measurement noise is estimated through variational Bayesian estimation to obtain state prediction data; according to the state prediction data, the initialized state vector is updated to obtain an updated state vector; motion state data analysis is performed on the updated state vector to generate a target motion data set, and motion state feature matching is performed on the target motion data set according to a preset evaluation motion feature knowledge base to obtain multiple motion state features.

[0007] In combination with the first aspect, in a second implementation method of the first aspect of the present invention, the target question and answer dataset of the target evaluation user is obtained, and question and answer state feature extraction is performed on the target question and answer dataset through a natural language processing model to obtain multiple question and answer state features, including: obtaining the target question and answer dataset of the target evaluation user; performing question and answer text conversion on the target question and answer dataset through a natural language processing model to obtain question and answer text data, and performing Mel spectrum conversion on the target question and answer dataset to obtain audio Mel spectrum; performing feature vectorization processing on the question and answer text data to obtain a high-dimensional space word vector, and performing semantic relationship analysis on the high-dimensional space word vector according to a preset evaluation feature word frequency knowledge base to obtain multiple semantic relationship features; performing speech feature matching on the audio Mel spectrum according to a preset evaluation feature speech segment knowledge base to obtain multiple speech change features; performing feature association processing on the multiple semantic relationship features and the multiple speech change features to obtain feature association relationships; performing feature enhancement on the multiple semantic relationship features and the multiple speech change features according to the feature association relationships to obtain multiple question and answer state features.

[0008] In combination with the first aspect, in a third implementation method of the first aspect of the present invention, the multiple motion state features and the multiple question-and-answer state features are respectively input into multiple initial intelligent evaluation decision models for decision analysis to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model, including: performing Pearson correlation coefficient analysis on the multiple motion state features and the multiple question-and-answer state features to obtain a target Pearson correlation coefficient; determining a first feature weight of the multiple motion state features according to the target Pearson correlation coefficient, and determining a second feature weight of the multiple question-and-answer state features according to the target Pearson correlation coefficient; performing Pearson correlation analysis on the multiple motion state features according to the first feature weight. Perform feature weighting and vector conversion to obtain a motion state feature vector, and perform feature weighting and vector conversion on the multiple question-answer state features according to the second feature weight to obtain a question-answer state feature vector; perform vector fusion on the motion state feature vector and the question-answer state feature vector to obtain a target state feature vector; input the target state feature vector into multiple initial intelligent evaluation decision models, and the multiple initial intelligent evaluation decision models include: a support vector machine model, a random forest model, and a gradient boosting decision tree model; perform decision analysis on the target state feature vector through the multiple initial intelligent evaluation decision models to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model.

[0009] In combination with the first aspect, in a fourth implementation method of the first aspect of the present invention, the model parameter range of each initial intelligent evaluation decision model is obtained, and the initial model parameters are optimized and analyzed according to the model parameter range by a particle swarm algorithm to obtain the target model parameters of each initial intelligent evaluation decision model, including: model parameter screening of the initial model parameters to obtain a corresponding model parameter set, wherein the model parameter set includes: a regularization parameter and an order of a polynomial kernel function in a support vector machine model, the number of features and a maximum depth of a random forest model, and the number of boosting stages and the degree of contribution of a gradient boosting decision tree model; the model parameter set Calibrate the parameter value range to obtain a parameter value range set; generate random initial values ​​for the model parameter set through the parameter value range set to obtain a corresponding random initial value set, and construct a particle population for the random initial value set through a particle swarm algorithm to obtain a particle population; calculate the particle fitness of the particle population to obtain a particle fitness set corresponding to the particle population, and at the same time, iterate the particle fitness set until a preset condition is met, and generate an optimal solution corresponding to the particle population; classify the initial model parameters through the optimal solution to obtain the target model parameters of each initial intelligent evaluation decision model.

[0010] In combination with the first aspect, in a fifth implementation method of the first aspect of the present invention, the parameter execution strategy analysis is performed on the multiple initial intelligent evaluation decision models through the target model parameters to obtain multiple parameter execution strategies, and the parameters are updated through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models, including: performing policy gradient analysis on the multiple initial intelligent evaluation decision models through the target model parameters to determine the policy gradient corresponding to each initial intelligent evaluation decision model; performing policy execution parameter analysis on the reward feedback parameters corresponding to each initial intelligent evaluation decision model through the policy gradient to obtain the policy execution parameters corresponding to each initial intelligent evaluation decision model; performing policy analysis on the initial action execution strategy corresponding to each initial intelligent evaluation decision model based on the policy update parameters corresponding to each initial intelligent evaluation decision model to generate multiple parameter execution strategies; and updating parameters through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models.

[0011] In combination with the first aspect, in a sixth implementation method of the first aspect of the present invention, the model weight analysis is performed on the multiple target intelligent evaluation decision models to obtain model weight data of each target intelligent evaluation decision model, and the models are integrated through the model weight data to obtain an intelligent evaluation decision model set, including: performing performance indicator analysis on the multiple target intelligent evaluation decision models respectively to obtain performance indicator data of each target intelligent evaluation decision model; performing model weight analysis on the multiple target intelligent evaluation decision models according to the performance indicator data to obtain model weight data of each target intelligent evaluation decision model; and using a stacking algorithm to integrate the multiple target intelligent evaluation decision models according to the model weight data to obtain an intelligent evaluation decision model set.

[0012] The second aspect of the present invention provides an intelligent evaluation device for logical thinking training, which includes: a motion tracking module, which is used to track the motion state of a target evaluation user through a state space tracking model to obtain a target motion data set, and extract motion state features from the target motion data set to obtain multiple motion state features; a feature extraction module, which is used to obtain a target question and answer data set for the target evaluation user, and extract question and answer state features from the target question and answer data set through a natural language processing model to obtain multiple question and answer state features; a decision analysis module, which is used to input the multiple motion state features and the multiple question and answer state features into multiple initial intelligent evaluation decision models for decision analysis, to obtain prediction data and initial Model parameters; a parameter optimization module, used to obtain the model parameter range of each initial intelligent evaluation decision model, and perform parameter optimization analysis on the initial model parameters according to the model parameter range through the particle swarm algorithm to obtain the target model parameters of each initial intelligent evaluation decision model; a parameter updating module, used to perform parameter execution strategy analysis on the multiple initial intelligent evaluation decision models through the target model parameters to obtain multiple parameter execution strategies, and perform parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models; a model integration module, used to perform model weight analysis on the multiple target intelligent evaluation decision models to obtain model weight data of each target intelligent evaluation decision model, and perform model integration through the model weight data to obtain a set of intelligent evaluation decision models.

[0013] In combination with the second aspect, in a first implementation method of the second aspect of the present invention, the motion tracking module is specifically used to: perform motion and speed detection on the target evaluation user to obtain motion data and speed data, and perform vector initialization on the motion data and speed data to obtain an initialized state vector, and initialize the covariance matrix of process noise and measurement noise; construct the state transfer equation and measurement equation of the state space tracking model, and perform time-varying measurement noise estimation through variational Bayesian estimation to obtain state prediction data; perform state update on the initialized state vector according to the state prediction data to obtain an updated state vector; perform motion state data analysis on the updated state vector to generate a target motion data set, and perform motion state feature matching on the target motion data set according to a preset evaluation motion feature knowledge base to obtain multiple motion state features.

[0014] In combination with the second aspect, in a second implementation method of the second aspect of the present invention, the feature extraction module is specifically used to: obtain a target question and answer data set of the target evaluation user; perform question and answer text conversion on the target question and answer data set through a natural language processing model to obtain question and answer text data, and perform Mel-spectrogram conversion on the target question and answer data set to obtain an audio Mel-spectrogram; perform feature vectorization processing on the question and answer text data to obtain a high-dimensional space word vector, and perform semantic relationship analysis on the high-dimensional space word vector according to a preset evaluation feature word frequency knowledge base to obtain multiple semantic relationship features; perform speech feature matching on the audio Mel-spectrogram according to a preset evaluation feature speech segment knowledge base to obtain multiple speech change features; perform feature association processing on the multiple semantic relationship features and the multiple speech change features to obtain feature association relationships; perform feature enhancement on the multiple semantic relationship features and the multiple speech change features according to the feature association relationships to obtain multiple question and answer status features.

[0015] In combination with the second aspect, in a third implementation method of the second aspect of the present invention, the decision analysis module is specifically used to: perform Pearson correlation coefficient analysis on the multiple motion state features and the multiple question-and-answer state features to obtain a target Pearson correlation coefficient; determine the first feature weights of the multiple motion state features according to the target Pearson correlation coefficient, and determine the second feature weights of the multiple question-and-answer state features according to the target Pearson correlation coefficient; perform feature weighting and vector conversion on the multiple motion state features according to the first feature weight to obtain a motion state feature vector, and perform feature weighting and vector conversion on the multiple question-and-answer state features according to the second feature weight to obtain a question-and-answer state feature vector; perform vector fusion on the motion state feature vector and the question-and-answer state feature vector to obtain a target state feature vector; input the target state feature vector into multiple initial intelligent evaluation decision models, and the multiple initial intelligent evaluation decision models include: a support vector machine model, a random forest model, and a gradient boosting decision tree model; perform decision analysis on the target state feature vector through the multiple initial intelligent evaluation decision models to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model.

[0016] In combination with the second aspect, in a fourth implementation of the second aspect of the present invention, the parameter optimization module is specifically used to: perform model parameter screening on the initial model parameters to obtain a corresponding model parameter set, wherein the model parameter set includes: the regularization parameter and the order of the polynomial kernel function in the support vector machine model, the number of features and the maximum depth of the random forest model, and the number of boosting stages and the degree of contribution of the gradient boosting decision tree model; perform parameter value range calibration on the model parameter set to obtain a parameter value range set; generate random initial values ​​for the model parameter set through the parameter value range set to obtain a corresponding random initial value set, and construct a particle population for the random initial value set through a particle swarm algorithm to obtain a particle population; calculate the particle fitness of the particle population to obtain a particle fitness set corresponding to the particle population, and at the same time, iteratively calculate the particle fitness set until the preset conditions are met, and generate the optimal solution corresponding to the particle population; perform parameter classification on the initial model parameters through the optimal solution to obtain the target model parameters of each initial intelligent evaluation decision model.

[0017] In combination with the second aspect, in the fifth implementation method of the second aspect of the present invention, the parameter updating module is specifically used to: perform policy gradient analysis on the multiple initial intelligent evaluation decision models through the target model parameters to determine the policy gradient corresponding to each initial intelligent evaluation decision model; perform policy execution parameters on the reward feedback parameters corresponding to each initial intelligent evaluation decision model through the policy gradient to obtain the policy execution parameters corresponding to each initial intelligent evaluation decision model; perform policy analysis on the initial action execution strategy corresponding to each initial intelligent evaluation decision model based on the policy update parameters corresponding to each initial intelligent evaluation decision model to generate multiple parameter execution strategies; perform parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models.

[0018] In combination with the second aspect, in a sixth implementation method of the second aspect of the present invention, the model integration module is specifically used to: perform performance indicator analysis on the multiple target intelligent evaluation decision models respectively to obtain performance indicator data of each target intelligent evaluation decision model; perform model weight analysis on the multiple target intelligent evaluation decision models according to the performance indicator data to obtain model weight data of each target intelligent evaluation decision model; and use a stacking algorithm to integrate the multiple target intelligent evaluation decision models according to the model weight data to obtain an intelligent evaluation decision model set.

[0019] In the technical solution provided by the present invention, the multimodal information of motion state data and question-and-answer data is combined, and these information are integrated through natural language processing and state space tracking models to provide more comprehensive user evaluation data. Through Pearson correlation coefficient analysis and feature association processing, the method can effectively extract and select features with key influences, thereby improving the accuracy and interpretability of the evaluation model. The particle swarm algorithm is used to optimize the model parameters, and the model can be better adjusted to adapt to different types of evaluation tasks, thereby improving the generalization ability and performance of the model. Through policy gradient analysis and parameter execution strategy, the method can optimize the strategy of the evaluation decision model according to the target model parameters, so that it can better adapt to different evaluation conditions and requirements. Using the stacking algorithm for model integration, combining multiple intelligent evaluation decision models together can improve the overall evaluation performance and reduce the uncertainty of model selection. Since the model parameters and strategies can be automatically adjusted according to the user's motion and question-and-answer data, a more personalized intelligent evaluation can be achieved to better meet the needs of users. Through the state space tracking model and strategy analysis, the method can monitor and adjust the user's evaluation situation in real time to adapt to the user's changing needs and performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of an embodiment of an intelligent evaluation method for logical thinking training in an embodiment of the present invention;

[0021] Figure 2 This is a flow chart of question and answer status feature extraction in an embodiment of the present invention;

[0022] Figure 3 A flowchart of decision analysis in an embodiment of the present invention;

[0023] Figure 4 A flow chart of parameter optimization analysis in an embodiment of the present invention;

[0024] Figure 5 Schematic diagram of an embodiment of an intelligent evaluation device for logical thinking training in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] Embodiments of the present invention provide an intelligent evaluation method and related devices for logical thinking training, which are used to perform intelligent evaluation on target test users and improve the accuracy of evaluation decisions. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the intelligent evaluation method of logical thinking training in the embodiment of the present invention includes:

[0027] S101, tracking the motion state of the target evaluation user through the state space tracking model to obtain a target motion data set, and extracting motion state features from the target motion data set to obtain multiple motion state features;

[0028] It is understandable that the execution subject of the present invention may be an intelligent evaluation device for logical thinking training, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0029] Specifically, the target user is subjected to detailed motion and speed detection, which can be achieved through various sensor technologies, such as accelerometers, gyroscopes, or advanced motion capture systems. These technologies can capture the subtle differences in human movement, such as small hand tremors or body coordination problems, which are characteristics exhibited by the target test user. Vector initialization of these data is performed. The collected data is converted into a format that can be processed by the state space model. This usually involves standardization and normalization of the data to ensure that the data can be effectively used in subsequent analysis. At the same time, the initialization process also includes the setting of the covariance matrix of process noise and measurement noise. This determines the sensitivity of the model to random fluctuations in the data. The state transition equation and measurement equation of the state space tracking model are constructed. The state transition equation describes how the system state changes over time, while the measurement equation defines the relationship between the observed data and the system state. In this process, the variational Bayesian estimation method is used to estimate the time-varying measurement noise. This estimation method can improve the accuracy of the model prediction by considering the uncertainty of the data. The obtained state prediction data is a prediction of the future motion state of the target user. Update the initialization state vector according to the state prediction data. Correct and adjust the original data to more accurately reflect the user's actual motion state. The updated state vector provides more accurate and detailed motion state information. Perform motion state data analysis and match motion state features according to the preset evaluation motion feature knowledge base. The evaluation motion feature knowledge base contains various motion features displayed by the target test user. By comparing with this knowledge base, specific motion patterns related to logical thinking training assessment can be identified.

[0030] S102, obtaining a target question and answer data set of a target evaluation user, and extracting question and answer state features from the target question and answer data set through a natural language processing model to obtain multiple question and answer state features;

[0031] Specifically, obtain the question-and-answer datasets of the target assessment users. These datasets contain the user's daily conversation records, interview answers, or other interactive communication records. These datasets are the basis for analysis and provide the original data of the user's language and communication mode. They are processed through the natural language processing (NLP) model. This process is divided into two main parts: question-and-answer text conversion and audio mel-spectrogram conversion. The purpose of question-and-answer text conversion is to convert the collected spoken or written language into text data that can be processed by a computer. This involves the conversion from speech to text, or from unstructured text to structured text. Audio mel-spectrogram conversion converts the user's speech data into mel-spectrogram, which is a method of representing audio signals and is particularly suitable for capturing and analyzing speech features. The question-and-answer text data is feature vectorized to convert the text into a series of numerical vectors so that it can be effectively processed by computer algorithms. This is usually achieved by converting each word in the text into a word vector in a high-dimensional space, for example, using word embedding technology such as Word2Vec or GloVe. Once the word vectorization is completed, semantic relationship analysis can be performed based on the preset assessment feature word frequency knowledge base. This knowledge base contains specific words and expressions related to assessment behaviors and communication patterns. By analyzing the relationship between high-dimensional space word vectors and these feature words, multiple semantic relationship features can be extracted. In terms of audio analysis, the audio Mel spectrogram is matched with speech features according to the preset assessment feature speech segment knowledge base. This knowledge base contains specific speech patterns related to the assessment, such as specific pitch changes, rhythm, or speech speed. By matching the audio Mel spectrogram with these features, multiple speech change features can be obtained. The extracted semantic relationship features and speech change features are subjected to feature association processing in order to understand the relationship between these features. For example, certain specific semantic expressions are accompanied by specific speech patterns. By understanding these associations, these features can be analyzed at a deeper level. The semantic relationship features and speech change features are enhanced according to the feature association relationships. The identified association relationships are used to strengthen individual features, thereby obtaining more accurate and comprehensive question and answer status features.

[0032] S103, inputting the multiple motion state features and the multiple question-answer state features into multiple initial intelligent evaluation decision models for decision analysis, and obtaining prediction data and initial model parameters of each initial intelligent evaluation decision model;

[0033] It should be noted that the Pearson correlation coefficient analysis is performed on multiple motion state features and multiple question-answer state features to obtain the target Pearson correlation coefficient. This analysis method can quantify the linear relationship between different features, that is, how one feature changes with the change of another feature. By calculating the Pearson correlation coefficient of the motion state features and the question-answer state features, the correlation between these different dimensions of data can be understood. The feature weights are determined based on the obtained target Pearson correlation coefficient. For the motion state features, the first feature weight is assigned according to the size of their correlation with the assessment recognition; for the question-answer state features, the second feature weight is assigned. This feature weighting process is based on the assumption that features that are more strongly related to the assessment recognition should have a greater influence in the model. After the feature weights are obtained, feature weighting and vector conversion are performed. According to their respective weights, the numerical values ​​of the motion state features and the question-answer state features are adjusted to generate feature vectors representing these weighted features. This conversion process is to enable these features to be more effectively processed and analyzed by the initial intelligent assessment decision model. The motion state feature vector and the question-answer state feature vector are vector fused to form a comprehensive target state feature vector. This fusion process is to create a unified representation of all relevant features, which ensures that all relevant dimensions are considered in the decision analysis. The fused target state feature vector is input into multiple initial intelligent assessment decision models. These models include support vector machines (SVM), random forests, and gradient boosted decision trees (GBDT). Each model has its own unique way to analyze data and make predictions. For example, support vector machines find the best separating hyperplane in high-dimensional space; random forests improve the accuracy and robustness of predictions by building multiple decision trees and combining their predictions; gradient boosted decision trees improve prediction performance by gradually optimizing decision trees. In these models, the target state feature vector is used for decision analysis, and each model generates prediction data and initial model parameters based on the input feature vector. These prediction data and model parameters are the basis for subsequent model optimization and evaluation. For example, suppose that during the analysis process, it is found that certain specific movement patterns in the motion state features (such as the frequency and amplitude of hand swings) are highly correlated with the judgment results of the assessment. Similarly, in the question-answer state features, specific language patterns (such as the frequency of repeating specific words or phrases) also show strong correlations. Through Pearson correlation coefficient analysis, these features are given higher weights. These weighted features are converted into feature vectors and fused into a target state feature vector.

[0034] S104, obtaining a model parameter range for each initial intelligent evaluation decision model, and performing parameter optimization analysis on the initial model parameters according to the model parameter range using a particle swarm algorithm to obtain target model parameters for each initial intelligent evaluation decision model;

[0035] Specifically, the initial model parameters are screened to obtain the corresponding model parameter set. This involves determining the key parameters in each initial intelligent evaluation decision model, such as the regularization parameter and the order of the polynomial kernel function in the support vector machine (SVM) model, the number of features and the maximum tree depth in the random forest model, and the number of boosting stages and the degree of contribution in the gradient boosting decision tree (GBDT) model. These parameters determine the complexity, fitting ability and prediction performance of the model. The parameter value range of these model parameter sets is calibrated. This process is to determine the effective value range of each parameter. For example, for the regularization parameter of the support vector machine model, a range from 0.1 to 100 is set; for the maximum depth of the random forest, a range from 1 to 50 is set. The determination of these ranges is based on the understanding of the model and the knowledge of the problem domain. After the parameter range is determined, random initial values ​​are generated to form a random initial value set. Each element in this set represents a set of model parameter configurations. In order to optimize these parameters, a particle swarm algorithm is used to construct a particle population. The particle swarm algorithm is an optimization algorithm based on swarm intelligence. It simulates the collective behavior of a flock of birds or a school of fish to find the optimal solution. In this algorithm, each "particle" represents a specific set of parameter configurations, and the entire "swarm" jointly explores the parameter space to find the most optimized set of parameters. The fitness calculation of each particle is based on the impact of the parameter configuration it represents on the model performance. The higher the fitness of the particle, the better the prediction effect of the model can be achieved by the corresponding parameter configuration. The calculation of particle fitness is usually based on the prediction accuracy or other performance indicators of the model. Through iterative calculation, the particle continuously adjusts its position (i.e., parameter configuration) according to its own fitness and the fitness information of other particles in the group to gradually approach the optimal solution. This process continues until the preset conditions are met, such as reaching a certain number of iterations or the fitness is no longer significantly improved. After this, the generated optimal solution represents the best parameter configuration. These optimal parameter configurations are then used to classify and adjust the parameters of each initial intelligent evaluation decision model to obtain the target model parameters of each model.

[0036] S105, performing parameter execution strategy analysis on multiple initial intelligent evaluation decision models through target model parameters to obtain multiple parameter execution strategies, and performing parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models;

[0037] Specifically, policy gradient analysis is performed on multiple initial intelligent evaluation decision models through target model parameters to determine the policy gradient corresponding to each initial intelligent evaluation decision model. Policy gradient is a machine learning technique mainly used in the field of reinforcement learning. It maximizes the expected reward by optimizing a policy. In this context, policy refers to the way a model makes decisions, while gradient refers to the direction of moving in the policy space to improve performance. For each initial intelligent evaluation decision model, such as support vector machine (SVM), random forest, and gradient boosted decision tree (GBDT), policy gradient analysis is used to determine the best policy direction corresponding to each model, which helps guide the optimization of model parameters. Policy execution is performed on the reward feedback parameters corresponding to each initial intelligent evaluation decision model using policy gradient. In reinforcement learning, reward feedback refers to the positive or negative feedback that a model receives based on the quality of its behavior. By performing policy execution on these reward feedback parameters, the policy execution parameters of each model can be obtained, which will guide the model on how to adjust its decision-making process to obtain higher rewards, that is, more accurate prediction results. Based on the policy update parameters of each model, policy analysis is performed on the initial action execution policy to generate multiple parameter execution policies. This stage is to apply the previously determined policy execution parameters to the decision-making process of the model to generate specific operation steps. These parameter execution strategies are key to the model optimization process because they directly affect the decision-making method and performance of the model. The model parameters are updated through these parameter execution strategies to obtain multiple target intelligent evaluation decision models. At this stage, all previous analyses and adjustments will be applied to the actual model parameters to optimize the initial intelligent evaluation decision model. This process involves fine-tuning the parameters of each model to ensure that each model performs best in its specific task. For example, the system uses three different initial intelligent evaluation decision models: SVM, random forest, and GBDT. Through policy gradient analysis, the system found that for SVM, adjusting the regularization parameter and kernel function type can significantly improve model performance; for random forest, adjusting the number of trees and maximum depth can optimize the prediction accuracy of the model; for GBDT, adjusting the learning rate and the number of trees is the key. Based on these findings, the corresponding reward feedback parameters are adjusted to ensure that each model update is in the direction of increasing rewards, that is, improving prediction accuracy. These policy execution parameters are used to further guide the optimization of the model decision process. Through this method, the multiple parameter execution strategies finally generated will be used to update the parameters of the initial intelligent evaluation decision model, making it a more accurate and effective target intelligent evaluation decision model.

[0038] S106. Perform model weight analysis on multiple target intelligent evaluation decision models to obtain model weight data for each target intelligent evaluation decision model, and perform model integration through the model weight data to obtain an intelligent evaluation decision model set.

[0039] Specifically, the performance indicators of each target intelligent evaluation decision model are analyzed separately. The purpose is to evaluate the performance of each model in a specific task, including key indicators such as its accuracy, recall, precision and F1 score. For example, the analysis results of the performance indicators provide a quantitative description of the effect of each model, which is helpful for the subsequent weight allocation. Model weight analysis is performed on each model based on the performance indicator data. The weight of each model in the final model set is determined according to its performance indicator data. The weight of the model reflects its relative importance in the overall integrated model. Generally, better performing models are given higher weights because they are more reliable in prediction. The target intelligent evaluation decision model is integrated according to the model weight data using the stacking algorithm to obtain the intelligent evaluation decision model set. Stacking is an integrated learning technique that improves the overall prediction performance by combining multiple different models. In this process, the output of each model is provided as new input data to a "meta-learner", which makes the final prediction decision based on these inputs. The advantage of this method is that it can capture complex patterns that are ignored by a single model and make up for the weaknesses of each model by integrating the strengths of different models.

[0040] In an embodiment of the present invention, the multimodal information of motion state data and question-and-answer data is combined, and the information is integrated through natural language processing and state space tracking models to provide more comprehensive user evaluation data. Through Pearson correlation coefficient analysis and feature association processing, the method can effectively extract and select features with key influences, thereby improving the accuracy and interpretability of the evaluation model. The particle swarm algorithm is used to optimize the model parameters, and the model can be better adjusted to adapt to different types of evaluation tasks, thereby improving the generalization ability and performance of the model. Through policy gradient analysis and parameter execution strategy, the method can optimize the strategy of the evaluation decision model according to the target model parameters, so that it can better adapt to different evaluation conditions and requirements. Using the stacking algorithm for model integration, combining multiple intelligent evaluation decision models together can improve the overall evaluation performance and reduce the uncertainty of model selection. Since the model parameters and strategies can be automatically adjusted according to the user's motion and question-and-answer data, a more personalized intelligent evaluation can be achieved to better meet the needs of users. Through the state space tracking model and strategy analysis, the method can monitor and adjust the user's evaluation situation in real time to adapt to the user's changing needs and performance.

[0041] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0042] (1) Perform motion and speed detection on the target user to obtain motion data and speed data, and perform vector initialization on the motion data and speed data to obtain an initialization state vector, and initialize the covariance matrix of process noise and measurement noise;

[0043] (2) Construct the state transfer equation and measurement equation of the state space tracking model, and estimate the time-varying measurement noise through variational Bayesian estimation to obtain state prediction data;

[0044] (3) updating the initialized state vector according to the state prediction data to obtain an updated state vector;

[0045] (4) Performing motion state data analysis on the updated state vector to generate a target motion data set, and performing motion state feature matching on the target motion data set according to a preset evaluation motion feature knowledge base to obtain multiple motion state features.

[0046] Specifically, the target user is subjected to motion and speed detection, which can be achieved through various sensors, such as accelerometers and gyroscopes in smart devices. These sensors can capture the user's motion data, such as position, speed, and acceleration, which provide raw information about the user's motion state. Vector initialization is performed on these data to convert the raw data into a format more suitable for analysis. The initialization state vector is a mathematical representation used to describe the user's current motion state. In addition, it also involves the initialization of the covariance matrix of process noise and measurement noise. This can help quantify and manage the noise that is inevitably generated during data collection, which comes from sensor errors or external environmental factors. The state transition equation and measurement equation of the state space tracking model are constructed. These equations are the core of the model, which describe the changes in the user's motion state over time (state transition equation) and how to infer the motion state from sensor data (measurement equation). The establishment of these equations provides a theoretical basis for subsequent data processing. In order to deal with time-varying measurement noise, variational Bayesian estimation is used for noise estimation. Variational Bayesian estimation is a probabilistic method that can effectively estimate and update the noise level, thereby improving the accuracy of state prediction. The advantage of this approach is that it can learn from the initial noise estimate and continuously adjust the noise estimate based on new data to more accurately reflect the actual situation. After the state prediction data is available, the initial state vector is updated. Based on the latest measurement data, the understanding of the user's motion state is adjusted and updated. The updated state vector reflects the latest motion characteristics and lays the foundation for the next step of analysis. The updated state vector is analyzed for motion state data to generate target motion data sets. These data sets contain detailed descriptions of the user's motion state. According to the preset evaluation motion feature knowledge base, the target motion data set is matched with motion state features.

[0047] In a specific embodiment, if Figure 2 As shown, the process of executing step S102 may specifically include the following steps:

[0048] S201, obtaining a target question-answering dataset of a target evaluation user;

[0049] S202, converting the target question-answering data set into question-answering text by using a natural language processing model to obtain question-answering text data, and converting the target question-answering data set into mel-spectrogram to obtain audio mel-spectrogram;

[0050] S203, performing feature vectorization processing on the question and answer text data to obtain a high-dimensional space word vector, and performing semantic relationship analysis on the high-dimensional space word vector according to a preset evaluation feature word frequency knowledge base to obtain multiple semantic relationship features;

[0051] S204, performing speech feature matching on the audio mel spectrum according to a preset evaluation feature speech segment knowledge base to obtain a plurality of speech change features;

[0052] S205, performing feature association processing on the multiple semantic relationship features and the multiple voice change features to obtain a feature association relationship;

[0053] S206: Perform feature enhancement on multiple semantic relationship features and multiple voice change features according to feature association relationships to obtain multiple question-answering state features.

[0054] Specifically, obtain the question-and-answer dataset of the target evaluation user, which comes from structured interviews, daily conversation records or other interactive communications. The question-and-answer dataset contains the user's language usage patterns and communication behaviors. The question-and-answer dataset is processed using natural language processing (NLP) technology. This includes two main steps: conversion of question-and-answer text and mel-spectrogram conversion of audio data. Question-and-answer text conversion is to convert spoken or written language into text data that can be processed by computers, which involves conversion from speech to text, or conversion of unstructured text into a structured format. Audio mel-spectrogram conversion converts speech data into mel-spectrogram, which is a technology for capturing and analyzing speech features, and is particularly suitable for processing speech features such as intonation, rhythm and speech speed. After the question-and-answer text data is processed, feature vectorization is performed to convert the text into a word vector in a high-dimensional space. Each word in the text is converted into a vector in a high-dimensional space, for example, using word embedding technology such as Word2Vec or GloVe. These word vectors in high-dimensional space can better represent the semantic characteristics of words. According to the preset evaluation feature word frequency knowledge base, these word vectors are subjected to semantic relationship analysis. This knowledge base contains specific language patterns related to the assessment, such as the frequency of repeating specific words or phrases. By analyzing the relationship between the high-dimensional space word vector and these feature words, multiple semantic relationship features can be extracted. In terms of audio data, the audio Mel spectrogram is matched with speech features according to the preset assessment feature speech segment knowledge base. This knowledge base contains specific speech patterns related to the assessment, such as specific pitch changes, rhythm or speaking speed. By matching the audio Mel spectrogram with these features, multiple speech change features can be obtained. Feature association processing is performed on the extracted semantic relationship features and speech change features in order to understand the relationship between these features. For example, certain specific semantic expressions are accompanied by specific speech patterns. By understanding these associations, these features can be analyzed at a deeper level. Feature enhancement is performed on the semantic relationship features and speech change features based on the feature association relationships. The identified association relationships are used to strengthen individual features to obtain more accurate and comprehensive question and answer status features.

[0055] In a specific embodiment, if Figure 3 As shown, the process of executing step S103 may specifically include the following steps:

[0056] S301, performing Pearson correlation coefficient analysis on multiple motion state features and multiple question-answer state features to obtain a target Pearson correlation coefficient;

[0057] S302, determining first feature weights of multiple motion state features according to the target Pearson correlation coefficient, and determining second feature weights of multiple question-answering state features according to the target Pearson correlation coefficient;

[0058] S303, performing feature weighting and vector conversion on multiple motion state features according to the first feature weight to obtain a motion state feature vector, and performing feature weighting and vector conversion on multiple question and answer state features according to the second feature weight to obtain a question and answer state feature vector;

[0059] S304, performing vector fusion on the motion state feature vector and the question-answer state feature vector to obtain a target state feature vector;

[0060] S305, inputting the target state feature vector into a plurality of initial intelligent evaluation decision models, wherein the plurality of initial intelligent evaluation decision models include: a support vector machine model, a random forest model, and a gradient boosting decision tree model;

[0061] S306. Perform decision analysis on the target state feature vector through multiple initial intelligent evaluation decision models to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model.

[0062] Specifically, a Pearson correlation coefficient analysis is performed on multiple motion state features (such as the frequency and amplitude of hand movements) and multiple question-answering state features (such as the pattern and complexity of language use) to obtain a target Pearson correlation coefficient. Through this analysis, the linear relationship between the motion state features and the question-answering state features can be quantified. The purpose is to determine the degree of association between these different types of features, so as to assign appropriate weights to them in subsequent steps. The Pearson correlation coefficient is a statistical measure used to evaluate the strength and direction of the linear association between two variables. The first feature weights of multiple motion state features and the second feature weights of multiple question-answering state features are determined according to the obtained target Pearson correlation coefficient. These features are weighted and vectorized. For the motion state features, feature weighting and vectorization are performed according to the first feature weight to obtain a motion state feature vector. This conversion process involves converting the motion state data into a numerical vector that can reflect its importance. Similarly, for the question-answering state features, similar processing is performed according to the second feature weight to obtain the question-answering state feature vector, and the original data is converted into a format that is more suitable for processing by the machine learning model. Vector fusion is performed on the motion state feature vector and the question-answering state feature vector. The two different types of feature vectors are merged into a unified target state feature vector to ensure that both motion and language features can be taken into account in subsequent model training and analysis. This fused feature vector provides a comprehensive perspective that can more accurately reflect the user's overall state. This fused target state feature vector is input into multiple initial intelligent assessment decision models. These models include support vector machines (SVM), random forests, and gradient boosting decision trees (GBDT). Each model has its own unique processing methods and advantages. For example, SVM is good at processing high-dimensional data, random forests are robust when processing a large number of features, and GBDT can effectively handle nonlinear relationships. Through these models, decision analysis of the target state feature vector can be performed to obtain the prediction data and initial model parameters of each initial intelligent assessment decision model. These data and parameters are the basis for evaluating model performance and further adjustments.

[0063] In a specific embodiment, if Figure 4 As shown, the process of executing step S104 may specifically include the following steps:

[0064] S401, performing model parameter screening on the initial model parameters to obtain a corresponding model parameter set, wherein the model parameter set includes: a regularization parameter and an order of a polynomial kernel function in a support vector machine model, a feature number and a maximum depth of a random forest model, and a boosting stage number and a contribution degree of a gradient boosting decision tree model;

[0065] S402, calibrating the parameter value range of the model parameter set to obtain a parameter value range set;

[0066] S403, generating random initial values ​​for the model parameter set through the parameter value range set to obtain a corresponding random initial value set, and constructing a particle population for the random initial value set through a particle swarm algorithm to obtain a particle population;

[0067] S404, performing particle fitness calculation on the particle population to obtain a particle fitness set corresponding to the particle population, and at the same time, performing iterative calculation on the particle fitness set until a preset condition is met, thereby generating an optimal solution corresponding to the particle population;

[0068] S405. Classify the initial model parameters through the optimal solution to obtain the target model parameters of each initial intelligent evaluation decision model.

[0069] Specifically, the initial model parameters are screened to determine which parameters have a greater impact on the performance of model prediction. This includes selecting the regularization parameter and the order of the polynomial kernel function in the support vector machine (SVM) model, the number of features and the maximum depth in the random forest (RF) model, and the number of boosting stages and the degree of contribution in the gradient boosted decision tree (GBDT) model. For example, the regularization parameter of the SVM can affect the strictness of the model's fit to the data, and the order of the polynomial kernel function determines how the model processes data in high-dimensional space; the number of features and the maximum depth of the random forest affect the complexity and fitting ability of the model; and the number of boosting stages and the degree of contribution of the GBDT are directly related to how the model gradually builds a decision tree to improve performance. The numerical range of these screened parameters is calibrated to determine the effective value range of each parameter. The setting of this range needs to be based on the understanding of the model and the knowledge of the specific problem domain. For example, the regularization parameter of SVM is set in a range from 0.01 to 100 to explore the degree of regularization from extremely loose to extremely strict; the maximum depth of random forest varies from 1 to 50 to test the impact of different tree depths on model performance; the number of boosting stages of GBDT is set between 10 and 500 to observe how different numbers of boosting stages improve the predictive ability of the model. After determining the numerical range of the parameters, random initial values ​​are generated to generate a set of parameter configurations based on the numerical range of each parameter. This set of random initial values ​​provides a starting point for the subsequent optimization algorithm. The particle swarm algorithm is used to construct a particle population for this random initial value set. The particle swarm algorithm is an optimization method based on swarm intelligence, which simulates the behavior of a flock of birds or a school of fish to find the optimal solution. In this algorithm, each "particle" represents a specific set of parameter configurations, and the entire "swarm" jointly explores the parameter space to find the most optimized parameter combination. The fitness calculation of each particle is based on the impact of the parameter configuration it represents on the model's predictive performance. The higher the fitness of the particle, the better the prediction effect of the model can be achieved by its corresponding parameter configuration. The calculation of fitness is usually based on the prediction accuracy of the model or other relevant performance indicators. Through iterative calculation, particles continuously adjust their positions (i.e. parameter configurations) according to their own fitness and the fitness information of other particles in the group to gradually approach the optimal solution. This process continues until the preset conditions are met, such as reaching a certain number of iterations or the fitness is no longer significantly improved. After this, the generated optimal solution represents the best parameter configuration. These optimal parameter configurations are then used to classify and adjust the parameters of each initial intelligent evaluation decision model to obtain the target model parameters of each model.

[0070] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0071] (1) Performing policy gradient analysis on multiple initial intelligent evaluation decision models through target model parameters to determine the policy gradient corresponding to each initial intelligent evaluation decision model;

[0072] (2) The reward feedback parameters corresponding to each initial intelligent evaluation decision model are calculated through the policy gradient to obtain the policy execution parameters corresponding to each initial intelligent evaluation decision model;

[0073] (3) performing a strategy analysis on the initial action execution strategy corresponding to each initial intelligent evaluation decision model based on the strategy update parameters corresponding to each initial intelligent evaluation decision model, and generating multiple parameter execution strategies;

[0074] (4) Parameters are updated through multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models.

[0075] Specifically, policy gradient analysis is performed on multiple initial intelligent evaluation decision models through the target model parameters to determine the policy gradient corresponding to each initial intelligent evaluation decision model. Policy gradient is an important concept in the field of reinforcement learning, which involves optimizing a decision strategy to maximize a certain form of cumulative reward. Each initial intelligent evaluation decision model (such as support vector machine, random forest, gradient boosting decision tree) needs to determine a policy gradient suitable for itself. The purpose of policy gradient analysis is to find out how to adjust the parameters of each model to improve its performance, that is, how to move in the parameter space of the model to find a parameter combination that can produce more accurate prediction results. These policy gradients are used to adjust the reward feedback parameters of each model. In reinforcement learning, reward feedback is positive or negative feedback given based on the results of model behavior. By adjusting the reward feedback parameters, the learning process of the model can be changed to make it more inclined to take actions that lead to positive results. Such adjustments are based on the results of policy gradient analysis and are intended to improve the performance of the model in a specific task. Based on the policy update parameters of each model, the action execution strategy of the initial intelligent evaluation decision model is analyzed. This phase involves applying policy execution parameters to the model's decision-making process to generate specific operational steps that reflect how to adjust the model's behavior based on the optimized policy. For example, this involves changing certain key parameters of the model when processing data, or adjusting the way the model reacts when faced with different types of data. By executing these parameter execution strategies, the model's parameters are updated to obtain multiple optimized target intelligent evaluation decision models. This step applies all previous analysis and adjustments to the actual model parameters. The parameter adjustments for each model are designed to make it perform better in a specific task.

[0076] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0077] (1) Analyze the performance indicators of multiple target intelligent evaluation decision models respectively to obtain the performance indicator data of each target intelligent evaluation decision model;

[0078] (2) Performing model weight analysis on multiple target intelligent evaluation decision models based on performance indicator data to obtain model weight data for each target intelligent evaluation decision model;

[0079] (3) Using the stacking algorithm, multiple target intelligent evaluation decision models are integrated according to the model weight data to obtain an intelligent evaluation decision model set.

[0080] Specifically, the performance indicators of each target intelligent evaluation decision model are analyzed, including key indicators such as accuracy, recall, precision, F1 score, etc. of the evaluation model. These indicators are helpful to evaluate the advantages and limitations of the model, because different models show their own strengths and weaknesses in different performance indicators. Model weight analysis is performed on these target intelligent evaluation decision models based on performance indicator data. The purpose is to determine the relative importance of each model in the final model integration. Generally, models with better performance are given higher weights. This is because they are more reliable in prediction and therefore should account for a larger proportion in the final model integration. The stacking algorithm is used to integrate multiple target intelligent evaluation decision models according to model weight data to obtain an intelligent evaluation decision model set. Stacking is an ensemble learning technique that improves overall prediction performance by combining multiple different models. In the stacking process, the output of each model is provided as new input data to a "meta-learner", which makes the final prediction decision based on these inputs. The advantage of this method is that it can capture complex patterns that are ignored by a single model and make up for the weaknesses of each model by integrating the strengths of different models.

[0081] The above describes the intelligent evaluation method for logical thinking training in the embodiment of the present invention. The following describes the intelligent evaluation device for logical thinking training in the embodiment of the present invention. Figure 5 In one embodiment of the present invention, an intelligent evaluation device for logical thinking training includes:

[0082] The motion tracking module 501 is used to track the motion state of the target evaluation user through the state space tracking model to obtain a target motion data set, and extract motion state features from the target motion data set to obtain multiple motion state features;

[0083] A feature extraction module 502 is used to obtain a target question and answer data set of the target evaluation user, and extract question and answer state features from the target question and answer data set through a natural language processing model to obtain multiple question and answer state features;

[0084] A decision analysis module 503 is used to input the multiple motion state features and the multiple question-answer state features into multiple initial intelligent evaluation decision models for decision analysis, and obtain prediction data and initial model parameters of each initial intelligent evaluation decision model;

[0085] The parameter optimization module 504 is used to obtain the model parameter range of each initial intelligent evaluation decision model, and perform parameter optimization analysis on the initial model parameters according to the model parameter range by using a particle swarm algorithm to obtain the target model parameters of each initial intelligent evaluation decision model;

[0086] A parameter updating module 505 is used to perform parameter execution strategy analysis on the multiple initial intelligent evaluation decision models through the target model parameters to obtain multiple parameter execution strategies, and to perform parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models;

[0087] The model integration module 506 is used to perform model weight analysis on the multiple target intelligent evaluation decision models to obtain model weight data of each target intelligent evaluation decision model, and perform model integration through the model weight data to obtain an intelligent evaluation decision model set.

[0088] Optionally, the motion tracking module 501 is specifically used to: perform motion and speed detection on the target evaluation user to obtain motion data and speed data, and perform vector initialization on the motion data and speed data to obtain an initialized state vector, and initialize the covariance matrix of process noise and measurement noise; construct the state transfer equation and measurement equation of the state space tracking model, and perform time-varying measurement noise estimation through variational Bayesian estimation to obtain state prediction data; perform state update on the initialized state vector according to the state prediction data to obtain an updated state vector; perform motion state data analysis on the updated state vector to generate a target motion data set, and perform motion state feature matching on the target motion data set according to a preset evaluation motion feature knowledge base to obtain multiple motion state features.

[0089] Optionally, the feature extraction module 502 is specifically used to: obtain a target question and answer data set of the target evaluation user; perform question and answer text conversion on the target question and answer data set through a natural language processing model to obtain question and answer text data, and perform Mel-spectrogram conversion on the target question and answer data set to obtain an audio Mel-spectrogram; perform feature vectorization processing on the question and answer text data to obtain a high-dimensional space word vector, and perform semantic relationship analysis on the high-dimensional space word vector according to a preset evaluation feature word frequency knowledge base to obtain multiple semantic relationship features; perform speech feature matching on the audio Mel-spectrogram according to a preset evaluation feature speech segment knowledge base to obtain multiple speech change features; perform feature association processing on the multiple semantic relationship features and the multiple speech change features to obtain feature association relationships; perform feature enhancement on the multiple semantic relationship features and the multiple speech change features according to the feature association relationships to obtain multiple question and answer status features.

[0090] Optionally, the decision analysis module 503 is specifically used to: perform Pearson correlation coefficient analysis on the multiple motion state features and the multiple question-and-answer state features to obtain a target Pearson correlation coefficient; determine a first feature weight of the multiple motion state features according to the target Pearson correlation coefficient, and determine a second feature weight of the multiple question-and-answer state features according to the target Pearson correlation coefficient; perform feature weighting and vector conversion on the multiple motion state features according to the first feature weight to obtain a motion state feature vector, and perform feature weighting and vector conversion on the multiple question-and-answer state features according to the second feature weight to obtain a question-and-answer state feature vector; perform vector fusion on the motion state feature vector and the question-and-answer state feature vector to obtain a target state feature vector; input the target state feature vector into multiple initial intelligent evaluation decision models, the multiple initial intelligent evaluation decision models including: a support vector machine model, a random forest model, and a gradient boosting decision tree model; perform decision analysis on the target state feature vector through the multiple initial intelligent evaluation decision models to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model.

[0091] Optionally, the parameter optimization module 504 is specifically used to: perform model parameter screening on the initial model parameters to obtain a corresponding model parameter set, wherein the model parameter set includes: the regularization parameter and the order of the polynomial kernel function in the support vector machine model, the number of features and the maximum depth of the random forest model, and the number of boosting stages and the degree of contribution of the gradient boosting decision tree model; perform parameter value range calibration on the model parameter set to obtain a parameter value range set; generate random initial values ​​for the model parameter set through the parameter value range set to obtain a corresponding random initial value set, and construct a particle population for the random initial value set through the particle swarm algorithm to obtain a particle population; calculate the particle fitness of the particle population to obtain a particle fitness set corresponding to the particle population, and at the same time, iteratively calculate the particle fitness set until the preset conditions are met, and generate the optimal solution corresponding to the particle population; perform parameter classification on the initial model parameters through the optimal solution to obtain the target model parameters of each initial intelligent evaluation decision model.

[0092] Optionally, the parameter updating module 505 is specifically used to: perform policy gradient analysis on the multiple initial intelligent evaluation decision models through the target model parameters to determine the policy gradient corresponding to each initial intelligent evaluation decision model; perform policy execution parameters on the reward feedback parameters corresponding to each initial intelligent evaluation decision model through the policy gradient to obtain the policy execution parameters corresponding to each initial intelligent evaluation decision model; perform policy analysis on the initial action execution strategy corresponding to each initial intelligent evaluation decision model based on the policy update parameters corresponding to each initial intelligent evaluation decision model to generate multiple parameter execution strategies; perform parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models.

[0093] Optionally, the model integration module 506 is specifically used to: perform performance indicator analysis on the multiple target intelligent evaluation decision models respectively to obtain performance indicator data of each target intelligent evaluation decision model; perform model weight analysis on the multiple target intelligent evaluation decision models according to the performance indicator data to obtain model weight data of each target intelligent evaluation decision model; use a stacking algorithm to integrate the multiple target intelligent evaluation decision models according to the model weight data to obtain an intelligent evaluation decision model set.

[0094] Through the collaboration of the above components, combined with the multimodal information of motion state data and question-and-answer data, the information is integrated through natural language processing and state space tracking models to provide more comprehensive user evaluation data. Through Pearson correlation coefficient analysis and feature association processing, this method can effectively extract and select features with key influence, thereby improving the accuracy and interpretability of the evaluation model. The particle swarm algorithm is used to optimize the model parameters, which can better adjust the model to adapt to different types of evaluation tasks and improve the generalization ability and performance of the model. Through policy gradient analysis and parameter execution strategy, this method can optimize the strategy of the evaluation decision model according to the target model parameters, so that it can better adapt to different evaluation conditions and needs. Using the stacking algorithm for model integration, combining multiple intelligent evaluation decision models together can improve the overall evaluation performance and reduce the uncertainty of model selection. Since the model parameters and strategies can be automatically adjusted according to the user's motion and question-and-answer data, more personalized intelligent evaluation can be achieved to better meet the needs of users. Through the state space tracking model and strategy analysis, this method can monitor and adjust the user's evaluation situation in real time to adapt to the user's changing needs and performance.

[0095] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0096] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0097] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent evaluation method for logical thinking training, characterized in that: The intelligent evaluation method for logical thinking training includes: The motion state of the target evaluation user is tracked through the state space tracking model to obtain a target motion data set, and the motion state feature of the target motion data set is extracted to obtain multiple motion state features; Obtain a target question and answer data set of the target evaluation user, and extract question and answer state features of the target question and answer data set through a natural language processing model to obtain multiple question and answer state features; Inputting the plurality of motion state features and the plurality of question-answer state features into a plurality of initial intelligent evaluation decision models for decision analysis, respectively, to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model; Obtaining a model parameter range for each initial intelligent evaluation decision model, and performing parameter optimization analysis on the initial model parameters according to the model parameter range using a particle swarm algorithm to obtain target model parameters for each initial intelligent evaluation decision model; Performing parameter execution strategy analysis on the multiple initial intelligent evaluation decision models through the target model parameters to obtain multiple parameter execution strategies, and updating parameters through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models; Model weight analysis is performed on the multiple target intelligent evaluation decision models to obtain model weight data for each target intelligent evaluation decision model, and model integration is performed through the model weight data to obtain an intelligent evaluation decision model set.

2. The intelligent evaluation method for logical thinking training according to claim 1 is characterized in that: The state space tracking model is used to track the motion state of the target evaluation user to obtain a target motion data set, and motion state feature extraction is performed on the target motion data set to obtain multiple motion state features, including: Performing motion and speed detection on the target evaluation user to obtain motion data and speed data, performing vector initialization on the motion data and speed data to obtain an initialization state vector, and initializing the covariance matrix of process noise and measurement noise; Construct the state transfer equation and measurement equation of the state space tracking model, and estimate the time-varying measurement noise through variational Bayesian estimation to obtain the state prediction data; According to the state prediction data, the initialization state vector is updated to obtain an updated state vector; The updated state vector is subjected to motion state data analysis to generate a target motion data set, and motion state feature matching is performed on the target motion data set according to a preset evaluation motion feature knowledge base to obtain a plurality of motion state features.

3. The intelligent evaluation method for logical thinking training according to claim 1 is characterized in that: The target question and answer data set of the target evaluation user is obtained, and question and answer state feature extraction is performed on the target question and answer data set through a natural language processing model to obtain multiple question and answer state features, including: Obtain a target question-answering dataset of the target evaluation user; Performing question-answer text conversion on the target question-answer data set through a natural language processing model to obtain question-answer text data, and performing mel-spectrogram conversion on the target question-answer data set to obtain audio mel-spectrogram; Performing feature vectorization processing on the question and answer text data to obtain a high-dimensional space word vector, and performing semantic relationship analysis on the high-dimensional space word vector according to a preset evaluation feature word frequency knowledge base to obtain multiple semantic relationship features; According to a preset evaluation feature speech segment knowledge base, speech feature matching is performed on the audio mel spectrum to obtain a plurality of speech change features; Performing feature association processing on the multiple semantic relationship features and the multiple voice change features to obtain a feature association relationship; The multiple semantic relationship features and the multiple voice change features are enhanced according to the feature association relationship to obtain multiple question and answer state features.

4. The intelligent evaluation method for logical thinking training according to claim 1 is characterized in that: The step of inputting the plurality of motion state features and the plurality of question-answer state features into a plurality of initial intelligent evaluation decision models for decision analysis to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model includes: Performing a Pearson correlation coefficient analysis on the multiple motion state features and the multiple question-answer state features to obtain a target Pearson correlation coefficient; Determining first feature weights of the plurality of motion state features according to the target Pearson correlation coefficient, and determining second feature weights of the plurality of question-answer state features according to the target Pearson correlation coefficient; Performing feature weighting and vector conversion on the multiple motion state features according to the first feature weight to obtain a motion state feature vector, and performing feature weighting and vector conversion on the multiple question and answer state features according to the second feature weight to obtain a question and answer state feature vector; Performing vector fusion on the motion state feature vector and the question-answer state feature vector to obtain a target state feature vector; Inputting the target state feature vector into a plurality of initial intelligent evaluation decision models, wherein the plurality of initial intelligent evaluation decision models include: a support vector machine model, a random forest model, and a gradient boosting decision tree model; The target state feature vector is subjected to decision analysis through the multiple initial intelligent evaluation decision models to obtain prediction data and initial model parameters of each initial intelligent evaluation decision model.

5. The intelligent evaluation method for logical thinking training according to claim 4 is characterized in that: The step of obtaining a model parameter range for each initial intelligent evaluation decision model, and performing parameter optimization analysis on the initial model parameters according to the model parameter range by using a particle swarm algorithm to obtain target model parameters for each initial intelligent evaluation decision model includes: Performing model parameter screening on the initial model parameters to obtain a corresponding model parameter set, wherein the model parameter set includes: a regularization parameter and an order of a polynomial kernel function in a support vector machine model, a feature number and a maximum depth of a random forest model, and a boosting stage number and a contribution degree of a gradient boosting decision tree model; Calibrate the parameter value range of the model parameter set to obtain a parameter value range set; Generating random initial values ​​for the model parameter set through the parameter value range set to obtain a corresponding random initial value set, and constructing a particle population for the random initial value set through a particle swarm algorithm to obtain a particle population; Performing particle fitness calculation on the particle population to obtain a particle fitness set corresponding to the particle population, and performing iterative calculation on the particle fitness set until a preset condition is met, thereby generating an optimal solution corresponding to the particle population; The initial model parameters are classified by the optimal solution to obtain the target model parameters of each initial intelligent evaluation decision model.

6. The intelligent evaluation method for logical thinking training according to claim 1 is characterized in that: The step of performing parameter execution strategy analysis on the multiple initial intelligent evaluation decision models through the target model parameters to obtain multiple parameter execution strategies, and performing parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models includes: Performing a policy gradient analysis on the multiple initial intelligent evaluation decision models through the target model parameters to determine the policy gradient corresponding to each initial intelligent evaluation decision model; The reward feedback parameters corresponding to each initial intelligent evaluation decision model are measured by the policy gradient to obtain the policy execution parameters corresponding to each initial intelligent evaluation decision model; Performing strategy analysis on the initial action execution strategy corresponding to each initial intelligent evaluation decision model based on the strategy update parameters corresponding to each initial intelligent evaluation decision model, and generating multiple parameter execution strategies; Parameters are updated through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models.

7. The intelligent evaluation method for logical thinking training according to claim 1 is characterized in that: The performing of model weight analysis on the multiple target intelligent evaluation decision models to obtain model weight data of each target intelligent evaluation decision model, and performing model integration through the model weight data to obtain an intelligent evaluation decision model set, includes: Performing performance index analysis on the multiple target intelligent evaluation decision models respectively to obtain performance index data of each target intelligent evaluation decision model; Performing model weight analysis on the multiple target intelligent evaluation decision models according to the performance indicator data to obtain model weight data of each target intelligent evaluation decision model; A stacking algorithm is used to integrate the multiple target intelligent evaluation decision models according to the model weight data to obtain an intelligent evaluation decision model set.

8. An intelligent evaluation device for logical thinking training, characterized in that: The intelligent evaluation device for logical thinking training comprises: The motion tracking module is used to track the motion state of the target evaluation user through the state space tracking model to obtain a target motion data set, and extract motion state features from the target motion data set to obtain multiple motion state features; A feature extraction module is used to obtain a target question and answer data set of the target evaluation user, and extract question and answer state features of the target question and answer data set through a natural language processing model to obtain multiple question and answer state features; A decision analysis module, used to input the multiple motion state features and the multiple question-answer state features into multiple initial intelligent evaluation decision models for decision analysis, and obtain prediction data and initial model parameters of each initial intelligent evaluation decision model; A parameter optimization module is used to obtain the model parameter range of each initial intelligent evaluation decision model, and perform parameter optimization analysis on the initial model parameters according to the model parameter range through a particle swarm algorithm to obtain the target model parameters of each initial intelligent evaluation decision model; A parameter updating module, used to perform parameter execution strategy analysis on the multiple initial intelligent evaluation decision models through the target model parameters to obtain multiple parameter execution strategies, and perform parameter update through the multiple parameter execution strategies to obtain multiple target intelligent evaluation decision models; The model integration module is used to perform model weight analysis on the multiple target intelligent evaluation decision models to obtain model weight data for each target intelligent evaluation decision model, and perform model integration through the model weight data to obtain an intelligent evaluation decision model set.

Images