Intelligent teaching evaluation system and device

Abstract

The invention provides an intelligent teaching evaluation system and device, and the system comprises an obtaining module which is used for obtaining multi-source training data; wherein the multi-source training data comprises VR motion capture data, multi-scene practical operation data, man-machine collaborative auditing data, voice question and answer data and theoretical learning data; the interaction calculation module is used for determining a plurality of VR interaction features according to the VR motion capture data, the multi-scene practical operation data and the man-machine cooperation auditing data; and the AI evaluation module is used for determining an evaluation result according to the plurality of VR interaction features, the voice question and answer data, the theoretical learning data and an AI evaluation algorithm. Through deep analysis and fusion of multi-source data, the method shows unique and remarkable advantages in nuclear emergency medical rescue teaching evaluation. Specifically, on the basis of VR motion capture data and multi-scene practical operation data, interaction features during nuclear emergency medical rescue in a VR virtual scene are extracted, and advantages and deficiencies of students in nuclear emergency medical rescue skill learning are accurately positioned.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to an intelligent teaching evaluation system and device. Background Technology

[0002] In the field of nuclear emergency medical rescue, as the complexity and specialization of rescue missions continue to increase, the limitations of traditional teaching evaluation systems are becoming increasingly apparent. Current teaching evaluations mostly focus solely on the final assessment results, neglecting the dynamic monitoring of the learners' learning process. They fail to promptly capture the difficulties encountered by learners at each stage of knowledge absorption and skills training, and are also unable to provide personalized guidance tailored to the learners' pace and characteristics. In terms of theoretical knowledge assessment, traditional written examinations are frequently used, emphasizing memorization of knowledge points and lacking a comprehensive examination of the complex concepts, principles, application, and knowledge transfer abilities within the nuclear emergency medical rescue theoretical system. This leads to a disconnect between learners' theory and practice.

[0003] Furthermore, with the application of immersive virtual training models in nuclear emergency medical rescue teaching, existing evaluation methods lack quantitative assessment indicators and scientific evaluation methods for virtual scenarios, making it impossible to accurately measure trainees' actual emergency response capabilities in highly simulated environments. Summary of the Invention

[0004] In view of this, the present invention provides an intelligent teaching evaluation system and device, which aims to solve the problem that existing evaluation methods lack quantitative evaluation indicators and scientific evaluation methods for virtual scenarios, and therefore cannot accurately measure students' actual emergency response capabilities in highly simulated environments.

[0005] A first aspect of this invention provides an intelligent teaching evaluation system, comprising:

[0006] The acquisition module is used to acquire multi-source training data, which includes VR motion capture data, multi-scene practical data, human-computer collaborative review data, voice question and answer data, and theoretical learning data.

[0007] The interactive computing module is used to determine multiple VR interactive features based on VR motion capture data, multi-scene operation data, and human-computer collaborative review data;

[0008] The AI ​​evaluation module is used to determine the evaluation results based on multiple VR interaction features, voice question and answer data, theoretical learning data, and AI evaluation algorithms.

[0009] In one possible implementation, the interactive computing module is used for:

[0010] Based on VR motion capture data and multi-scenario practical data, the first interaction feature was determined;

[0011] Based on the human-machine collaborative review data and the pre-established knowledge-behavior mapping matrix, the second interaction feature is determined; the knowledge-behavior mapping matrix is ​​the correlation matrix between knowledge weaknesses and practical errors.

[0012] Based on the first interaction feature and the second interaction feature, multiple VR interaction features are determined.

[0013] In one possible implementation, the interactive computing module is used for:

[0014] Based on VR motion capture data, multi-scenario practical data, and intelligent skills training indicator system, multiple VR interaction features are determined; the intelligent skills training indicator system includes primary indicators and at least one secondary indicator corresponding to each primary indicator.

[0015] In one possible implementation, the interactive computing module is used for:

[0016] Based on spatiotemporal convolutional networks and dynamic time warping algorithms, features are extracted from VR motion capture data and multi-scene practical data according to multiple secondary indicators to determine the standardization of rescue actions, proficiency in equipment operation, continuity of medical operations, scene response speed, dynamic adaptability, and multi-task coordination ability in each scenario.

[0017] Based on the standardization of rescue actions, proficiency in equipment operation, and consistency of medical procedures, operational skill characteristics are determined as a primary indicator value.

[0018] Based on scenario response speed, dynamic adaptability, and multi-task coordination capabilities, emergency decision-making characteristics are determined as another primary indicator value.

[0019] The first interaction feature is determined based on the primary indicator values ​​for each scenario.

[0020] In one possible implementation, the interactive computing module is used for:

[0021] Based on the data from the human-machine collaborative review, identify the current knowledge gaps;

[0022] Based on the current knowledge gaps and the knowledge-behavior mapping matrix, identify practical errors;

[0023] Based on practical errors, the second interactive feature is determined.

[0024] In one possible implementation, multiple VR interaction features include comprehensive skill level features, learning ability features, knowledge practice transformation features, and team collaboration features; the interaction computing module is used for:

[0025] The first and second interaction features are processed using a weighted fusion algorithm based on an attention mechanism to determine the comprehensive skill level features;

[0026] The first and second interaction features are processed using the temporal difference learning algorithm and the hidden Markov model to determine the learning ability features;

[0027] The first and second interaction features are processed using the association rule mining algorithm to determine the knowledge practice transformation features;

[0028] The team collaboration characteristics are determined by processing the first and second interaction features using social network analysis algorithms.

[0029] In one possible implementation, the AI ​​evaluation module is used for:

[0030] The evaluation results are determined by weighting multiple VR interaction features, voice question-and-answer data, theoretical learning data, and spatiotemporal-semantic joint embedding networks.

[0031] In one possible implementation, the spatiotemporal-semantic joint embedding network is optimized through a comprehensive loss function L, which includes: an evaluation prediction loss term, a cross-modal alignment loss term, and a scene adaptation loss term.

[0032] The evaluation prediction loss term is used to measure the difference between the evaluation results predicted by the model and the actual evaluation labels.

[0033] A cross-modal alignment loss term is used to constrain the distance between VR interaction features and semantic features in the joint embedding space; the semantic features are determined based on speech question-answering data and theoretical learning data.

[0034] The scenario adaptation loss term is used to ensure the model's generalization ability under different emergency scenarios.

[0035] A second aspect of the present invention provides an intelligent teaching evaluation device, characterized in that it includes a VR device, a motion capture system, and a server equipped with the intelligent teaching evaluation system as described in the first aspect.

[0036] In one possible implementation, the VR device includes VR glasses and a VR backpack; the motion capture system includes a motion capture suit with markers and inertial sensors, and a camera for capturing the markers.

[0037] The intelligent teaching evaluation system and device provided in this invention includes: an acquisition module for acquiring multi-source training data, including VR motion capture data, multi-scenario practical data, human-computer collaborative review data, voice question-and-answer data, and theoretical learning data; an interaction calculation module for determining multiple VR interaction features based on the VR motion capture data, multi-scenario practical data, and human-computer collaborative review data; and an AI evaluation module for determining the evaluation result based on the multiple VR interaction features, voice question-and-answer data, theoretical learning data, and AI evaluation algorithm. This invention demonstrates unique and significant advantages in nuclear emergency medical rescue teaching evaluation through in-depth analysis and fusion of multi-source data. Specifically, based on VR motion capture data and multi-scenario practical data, it accurately extracts the interaction features during nuclear emergency medical rescue in VR virtual scenarios, precisely identifies the strengths and weaknesses of trainees in learning nuclear emergency medical rescue skills, provides detailed evidence for subsequent personalized teaching and targeted training, and effectively improves teaching efficiency and trainees' comprehensive emergency rescue capabilities. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is an application scenario diagram of the intelligent teaching evaluation system provided in the embodiments of the present invention;

[0040] Figure 2 This is a schematic diagram of the structure of the intelligent teaching evaluation system provided in an embodiment of the present invention. Detailed Implementation

[0041] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0042] Figure 1 This is an application scenario diagram of the intelligent teaching evaluation system provided in an embodiment of the present invention. The method of the present invention can be applied to this scenario, including but not limited to. Figure 1 As shown, the scene includes: VR device 11, motion capture system 12, teaching terminal 13, and server 14.

[0043] The VR device 11 includes VR glasses and a VR backpack; the motion capture system 12 includes a motion capture suit with markers and inertial sensors, and a camera for capturing the markers.

[0044] During theoretical learning, trainees or instructors can log in to the relevant interface of the server through the teaching terminal to learn theoretical knowledge and take exams, and upload relevant data to the intelligent teaching evaluation system of server 14.

[0045] During nuclear emergency medical rescue simulation training, trainees wear motion capture suits, VR glasses, and VR backpacks. Data collected by inertial sensors on the motion capture suits and data from marker points captured by cameras are reported to server 14 in real time. Server 14 processes this data using appropriate algorithms to capture the trainees' actual movements, which are then displayed on the VR glasses via the VR backpack. The VR backpack or VR glasses are equipped with a language / text input device, allowing users to ask questions to the question-and-answer database system. The input device transmits the user's question to the server via the VR backpack. The question-and-answer database system on the server analyzes the question using the VR-based nuclear emergency medical rescue question-and-answer method described in this invention, derives the standard answer, and provides feedback to the trainees. Throughout this process, the server records relevant data from the nuclear emergency medical rescue simulation training in real time and transmits it to the intelligent teaching evaluation system.

[0046] The main training tasks of nuclear emergency medical rescue simulation exercises include individual training, group training, and comprehensive drills. Individual training includes personal protective equipment testing and the use of instruments and equipment. Group training includes command groups, reconnaissance and defense groups, triage groups, and decontamination groups. Comprehensive drills include casualty triage and on-site medical treatment.

[0047] Figure 2 This is a schematic diagram of the intelligent teaching evaluation system provided in an embodiment of the present invention. Figure 2 As shown, the intelligent teaching evaluation system includes:

[0048] The acquisition module 210 is used to acquire multi-source training data; among which, the multi-source training data includes VR motion capture data, multi-scene practical data, human-computer collaborative review data, voice question and answer data, and theoretical learning data;

[0049] In this embodiment of the invention, high-precision motion capture equipment is used to collect real-time action sequence data of trainees in a virtual nuclear emergency medical rescue scenario. For example, in a simulated casualty treatment scenario, the system can accurately record the angle, force, speed, and duration of a series of actions performed by the trainee, from approaching the casualty, assessing the injury, performing CPR, to transporting the casualty. It can also capture subtle operational issues such as incorrect order of donning protective equipment or improper first aid techniques. This data not only reflects the trainee's operational proficiency but also, by comparing it with a standard action library, quantitatively assesses the standardization of their actions, providing crucial evidence for subsequent analysis of the trainee's practical skill deficiencies.

[0050] In this embodiment of the invention, for complex scenarios that may arise in nuclear emergency medical rescue, such as scenarios with multiple types of casualties and overlapping injuries, group casualty treatment in contaminated areas, scenarios with limited rescue resources, and scenarios with dynamic environmental changes, the system records the emergency response methods of trainees in the face of sudden events, as well as their priority judgment in the case of multiple tasks running in parallel. The multi-scenario practical data can realistically reproduce the complexity of rescue work and reflect the trainees' comprehensive adaptability in a near-real-world environment.

[0051] In this embodiment of the invention, to ensure the accuracy and reliability of the data, the acquisition module incorporates a human-machine collaborative review mechanism. Experts can use their experience to deeply interpret complex rescue decision-making scenarios that are difficult for AI to judge, integrating their professional knowledge into the data. This human-machine combined approach effectively filters data noise, improves data quality, and provides accurate input for subsequent interactive calculations and evaluations.

[0052] Meanwhile, this invention establishes a dual review mechanism of "intelligent automatic grading of objective questions + intelligent human-machine collaborative review of subjective questions," significantly improving the quality and efficiency of data collection for human-machine collaborative review and laying a solid data foundation for the intelligent teaching evaluation system. Regarding objective question review, the system utilizes natural language processing and image recognition technology to intelligently and automatically grade objective question types such as multiple choice, true / false, and fill-in-the-blank questions in students' theoretical learning tests and online assignments. Taking the nuclear emergency medical rescue theory assessment as an example, the system can quickly identify the protective measures answers filled in by students, accurately compare them with a preset standard answer database, instantly complete the grading and record the answer status, and simultaneously analyze the students' mastery of knowledge points such as emergency regulations and protective principles, forming a detailed knowledge point mastery map.

[0053] For subjective questions such as case analysis and rescue plan design, this module initiates a human-machine collaborative intelligent review process. First, the AI ​​algorithm performs a preliminary analysis of the subjective question answers, extracting key information and semantic features. Through text similarity calculation and sentiment analysis, it provides preliminary scores and evaluation suggestions, such as pointing out logical flaws or oversights in the trainees' casualty transfer plans. Subsequently, a team of experts in nuclear emergency medical rescue, combining industry standards and practical rescue experience, reviews and corrects the key content and scoring suggestions marked by the AI. Experts can use their extensive on-site rescue experience to provide in-depth interpretation and supplementary evaluation of the trainees' answers regarding emergency decision-making and resource allocation, integrating professional knowledge and practical wisdom into the review process. This deeply integrated human-machine review model leverages AI's advantages in efficient data processing and rapid feature extraction while compensating for its limitations in professional depth and complex situation judgment. It effectively filters data noise, accurately identifies trainees' strengths and weaknesses in knowledge application and emergency decision-making, and provides high-quality, in-depth review data for subsequent interactive calculations and AI evaluation, comprehensively improving the accuracy and professionalism of teaching evaluation.

[0054] In theoretical knowledge assessment and emergency decision-making communication scenarios, the audio Q&A content of trainees is collected to determine the depth of trainees' understanding of nuclear emergency medical rescue knowledge, thereby comprehensively evaluating the trainees' knowledge reserves, language communication skills and emergency decision-making thinking.

[0055] In this embodiment of the invention, theoretical learning data of trainees is acquired, including course viewing time, e-textbook reading progress, online test scores, and homework completion status. This data reflects the trainees' engagement and mastery of theoretical knowledge related to nuclear emergency medical rescue, such as protective principles, emergency medicine theory, and emergency management regulations. By analyzing the theoretical learning data, weaknesses in specific knowledge areas can be identified, complementing the practical data and providing comprehensive information for constructing a complete learning profile of the trainees.

[0056] The interactive computing module 220 is used to determine multiple VR interactive features based on VR motion capture data, multi-scene operation data, and human-computer collaborative review data;

[0057] In this embodiment of the invention, the interactive computing module 220 serves as the "analysis center" of the intelligent teaching evaluation system. Through in-depth analysis and fusion of VR motion capture data, multi-scenario practical data, and human-machine collaborative review data, it accurately extracts multiple key VR interactive features. Based on VR motion capture data, it uses spatiotemporal convolutional networks and dynamic time warping algorithms to quantitatively analyze the details of trainees' actions in virtual rescue scenarios, accurately locating practical ability characteristics such as the standardization of rescue actions and proficiency in equipment operation. Combined with multi-scenario practical data, from an emergency decision-making perspective, it mines emergency decision-making characteristics such as trainees' scene response speed and dynamic adaptability in complex rescue environments. Utilizing human-machine collaborative review data and relying on a knowledge-behavior mapping matrix, it establishes a correlation between trainees' knowledge weaknesses and practical errors, analyzing the characteristics of knowledge-practice transformation. Finally, through algorithms such as attention mechanisms and temporal differential learning, it organically integrates the features extracted from multi-source data to generate VR interactive features across multiple dimensions, including comprehensive skill level, learning ability, and teamwork. This provides a comprehensive and in-depth analytical basis for the AI ​​evaluation module, helping to achieve accurate assessment of trainees' core emergency medical rescue capabilities.

[0058] AI evaluation module 230 is used to determine the evaluation results based on multiple VR interaction features, voice question and answer data, theoretical learning data and AI evaluation algorithms.

[0059] In this embodiment of the invention, the AI ​​evaluation module 230 serves as the "decision brain" of the intelligent teaching evaluation system. By integrating multi-source data and advanced AI algorithms, it achieves a precise assessment of trainees' nuclear emergency medical rescue capabilities. This module first performs feature engineering on VR interaction features, voice question-and-answer data, and theoretical learning data to construct a multimodal feature vector space, comprehensively characterizing trainees' performance in dimensions such as knowledge mastery, practical skills, and emergency decision-making. Subsequently, it dynamically adjusts the evaluation weights for each dimension; for example, in simulated scenarios, it increases the weight of emergency response speed and operational standardization. Simultaneously, it establishes a semantic association between theoretical knowledge and practical behavior, achieving a quantitative assessment of trainees' knowledge transfer capabilities. Finally, it captures the trainees' psychological state under high-pressure rescue scenarios through a sentiment analysis model and combines this with fuzzy comprehensive evaluation to generate multi-dimensional evaluation results containing capability profiles and improvement suggestions. This provides data-driven decision support for personalized teaching and promotes the transformation of nuclear emergency medical rescue talent training models towards intelligence and precision.

[0060] In some embodiments, the interactive computing module 220 is configured to: determine a first interactive feature based on VR motion capture data and multi-scene practical data; determine a second interactive feature based on human-computer collaborative review data and a pre-established knowledge-behavior mapping matrix; the knowledge-behavior mapping matrix is ​​a correlation matrix between knowledge weaknesses and practical errors; and determine multiple VR interactive features based on the first and second interactive features.

[0061] In some embodiments, the interactive computing module 220 is used to: determine multiple VR interactive features based on VR motion capture data, multi-scene practical data, and an intelligent skills training indicator system; wherein the intelligent skills training indicator system includes primary indicators and at least one secondary indicator corresponding to each primary indicator.

[0062] In this embodiment of the invention, the intelligent skills training index system serves as an evaluation framework, including primary indicators such as operational skills and emergency decision-making, as well as detailed secondary indicators such as the standardization of rescue actions, proficiency in equipment operation, and scene response speed. The interactive computing module 220 first preprocesses the VR motion capture data, extracting the spatial posture and temporal sequence features of the trainee's movements through a spatiotemporal convolutional network. Combined with environmental parameters and operation records from multi-scenario practical data, the raw data is mapped to the secondary dimensions of the index system. For example, in a simulated casualty rescue scenario, the system accurately calculates the compression frequency, depth, and gesture angle of the trainee's cardiopulmonary resuscitation based on the VR motion capture data, compares it with standard operating procedures, and quantifies the secondary indicator value of the standardization of rescue actions. By analyzing the smoothness and completion time of the trainee's operation of the radiation detection equipment, the proficiency index of equipment operation is determined. Based on the trainee's decision-making and adjustment speed in the event of sudden dose changes, the scene response speed is evaluated.

[0063] At the primary indicator construction level, the module adopts a weighted aggregation strategy to integrate secondary indicators. Taking operational skills as an example, related secondary indicators such as the standardization of rescue actions, proficiency in equipment operation, and continuity of medical operations are assigned different weights according to their importance in nuclear emergency rescue, and the quantitative value of operational skill level is calculated by weighted summation. For the emergency decision-making primary indicator, secondary indicators such as scenario response speed, dynamic adaptability, and multi-task coordination ability are comprehensively evaluated to reflect the trainees' decision-making ability in complex rescue environments.

[0064] In some embodiments, the interactive computing module 220 is configured to: extract features from VR motion capture data and multi-scene practical data according to multiple second indicators based on spatiotemporal convolutional networks and dynamic time warping algorithms; determine the standardization of rescue actions, proficiency in equipment operation, continuity of medical operations, scene response speed, dynamic adaptability, and multi-task coordination ability in each scene; determine operational skill features based on the standardization of rescue actions, proficiency in equipment operation, and continuity of medical operations, as a primary indicator value; determine emergency decision-making features based on scene response speed, dynamic adaptability, and multi-task coordination ability, as another primary indicator value; and determine a first interactive feature based on the primary indicator values ​​in each scene.

[0065] In this embodiment of the invention, VR motion capture data and multi-scene practical data are first preprocessed. VR motion capture data includes the trainee's 3D skeletal information, joint movement trajectories, and touch data of operating equipment in a virtual emergency medical rescue scenario. Multi-scene practical data covers environmental parameters and task execution records from the actual simulated scenario. The system removes noise interference from the data through filtering algorithms and unifies the data format through normalization, laying the foundation for subsequent analysis.

[0066] Spatiotemporal convolutional networks, as a core analytical tool, possess a unique structure capable of simultaneously processing data features in both spatial and temporal dimensions. In the spatial dimension, they extract spatial features such as trainee posture and equipment operation position through convolutional layers, for example, identifying the bending angle of a trainee's hand and the degree of body tilt when wearing protective clothing. In the temporal dimension, they capture the sequence and duration of actions through recurrent layers, such as recording the time interval from when a trainee discovers an injured person to when they begin administering first aid.

[0067] The dynamic time warping algorithm solves the evaluation challenge caused by differences in the operating speed of different trainees. In nuclear emergency medical rescue practice, trainees may take different times to complete the same task due to differences in personal habits and proficiency. By finding the optimal matching path between two time sequences (the trainee's operating sequence and the standard operating sequence), the time axis is flexibly aligned, ensuring fair comparison of operating sequences at different speeds. For example, even if trainee A completes the patient transfer 10 seconds faster than the standard procedure, but omits certain steps, the difference from the standard procedure can still be accurately identified.

[0068] The system compares the trainee's action sequence frame-by-frame with a standard action library established in the field of nuclear emergency medical rescue. Taking cardiopulmonary resuscitation (CPR) as an example, the standard action library clearly stipulates that the compression rate should be 100-120 times per minute, the compression depth should reach 5-6 centimeters, and the arm should be perpendicular to the victim's chest. A spatiotemporal convolutional network extracts parameters such as the trainee's compression rate, depth, and arm angle for each compression and compares them with the standard parameters. If the trainee's compression rate is only 80 times per minute, the system will deduct corresponding points according to the preset scoring rules, ultimately obtaining a quantitative value for the standardization of the rescue action.

[0069] For various specialized equipment used in nuclear emergency medical rescue, such as detectors and defibrillators, the system records the entire process data of trainees, from equipment startup and parameter setting to actual operation completion. It analyzes the completeness of steps, operation time, and number of errors during the operation. For example, if a trainee misses a calibration step or exceeds the standard testing time by 30%, it will be considered as insufficient proficiency and points will be deducted accordingly. Simultaneously, the system will also assess the trend of trainees' proficiency improvement by comparing historical data.

[0070] A state transition diagram of the medical operation process is constructed, breaking down the operation process in nuclear emergency medical rescue into multiple key nodes, such as injury assessment, hemostasis and bandaging, and patient transfer. The transition time and operational continuity between each node are analyzed. If a trainee completes hemostasis and bandaging but then delays patient transfer without a reasonable explanation, it indicates a problem with the continuity of their operation, and the system will deduct points based on the delay time and the importance of the operation.

[0071] Various emergency situations are simulated in virtual or real-world scenarios, such as sudden increases in concentration or warnings of secondary building collapse. When these events are triggered, the system records the time interval between when trainees receive the information and when they take an effective response (such as activating emergency protection procedures or adjusting rescue routes). The system simulates dynamically changing rescue environments and observes whether trainees can adjust their rescue strategies and procedures in a timely manner after environmental changes. The system evaluates the effectiveness of trainees' decision-making and adjustment timelines by analyzing the effectiveness of their strategies.

[0072] The system designs complex rescue scenarios involving multiple tasks, such as simultaneously treating the wounded, monitoring, and reporting information. It analyzes trainees' judgment of task priorities, the rationality of resource (time, manpower, equipment) allocation, and the smoothness of task switching. The system establishes a task priority evaluation model, assigning weights to each task based on its urgency and importance. If trainees focus excessively on information reporting while neglecting the treatment of seriously injured personnel, points will be deducted based on the difference in task weights.

[0073] After acquiring the data for the aforementioned secondary indicators, the interactive calculation module 220 uses a weighted summation method to determine two key primary indicator values: the standardization of comprehensive rescue actions, the proficiency in equipment operation, and the continuity of medical operations—three secondary indicators. Different weights are assigned to these secondary indicators based on their importance in nuclear emergency medical rescue. A quantitative value for the operational skill characteristics is then calculated. This value comprehensively reflects the trainee's ability level at the basic operational level.

[0074] The three secondary indicators of scenario response speed, dynamic adaptability and multi-task coordination capability are weighted and calculated to obtain the numerical value of emergency decision-making characteristics, which reflects the trainees' decision-making ability in complex rescue scenarios.

[0075] The system aggregates trainees' operational skills and emergency decision-making characteristics across various emergency rescue scenarios to form the first interactive feature. This feature is presented in a multi-dimensional matrix, with each row representing a scenario and each column corresponding to different primary indicator values ​​and related secondary indicator details. This structured data representation not only visually demonstrates trainees' performance in different scenarios but also facilitates subsequent integration and analysis with the second interactive feature and other data. For example, by combining human-machine collaborative review data and the second interactive feature determined by the knowledge-behavior mapping matrix, the correlation between trainees' performance and knowledge acquisition can be further analyzed. Simultaneously, the system can utilize machine learning algorithms to analyze the first interactive feature, uncovering potential patterns in trainees' skill development and providing a basis for developing personalized teaching plans. If a trainee scores low in operational skills but demonstrates strong emergency decision-making ability in a particular scenario, targeted training in that scenario can be strengthened to improve their overall rescue capabilities.

[0076] In some embodiments, the interactive computing module 220 is configured to: determine the current knowledge gap based on human-machine collaborative review data; determine the practical error based on the current knowledge gap and the knowledge-behavior mapping matrix; and determine the second interactive feature based on the practical error.

[0077] In this embodiment of the invention, the human-machine collaborative review data includes the trainees' original answers in theoretical assessments and practical operations, preliminary AI scores, and expert correction opinions. The interactive computing module 220 first performs semantic analysis on this unstructured data, using natural language processing technology to extract key concepts and logical relationships from the trainees' answers. In the nuclear emergency medical rescue theoretical assessment, the system identifies fundamental deviations in the trainees' answers to "contamination treatment principles" and discovers gaps in their understanding of "the mechanism of action of eliminating drugs" through knowledge graph matching. Simultaneously, combined with key error tags marked by experts, such as "incorrect selection of protection level," the system constructs a heatmap of trainees' knowledge mastery, quantitatively assesses the mastery of each knowledge point, and finally determines the current set of knowledge weaknesses through threshold screening, such as {biological effects under pollution, principles for selecting emergency protective equipment}.

[0078] The knowledge-behavior mapping matrix, serving as a bridge connecting theoretical knowledge and practical operation, is pre-constructed using domain expert experience and historical data. Rows in the matrix represent knowledge domains (e.g., protection, first aid), columns correspond to specific operational error types (e.g., inadequate protective measures, incorrect equipment operation sequence), and matrix element values ​​represent the correlation strength between the two. When the system determines that a trainee has a knowledge weakness of "dosage calculation error," querying the mapping matrix reveals a correlation of 0.85 with "improper placement of personal dosimeter" and 0.62 with "non-standard decontamination process." The interactive calculation module 220 uses a Bayesian inference algorithm, combined with current scenario characteristics (e.g., accident type, contamination level), to calculate the probability of each operational error. In high-risk scenarios, the system predicts a 73% probability of the trainee "missing the airtightness check of protective equipment," and, through evidence theory, integrates multi-source correlation information to ultimately determine the list of operational errors with the highest confidence, such as {incorrect protective equipment wearing sequence, omission of radioactive material detection process}.

[0079] In this embodiment of the invention, a risk matrix defined by experts is used, combined with the standard operating procedures for nuclear emergency medical rescue, to classify the risk of each practical error. "Failure to record personal dosimeter readings" is classified as a medium-risk error, while "failure to wear airtight protective clothing in a high-risk area" is classified as a high-risk error, with weights of 0.35 and 0.75 respectively. The number of occurrences of the same type of practical error in multiple training sessions is statistically analyzed to calculate the error recurrence rate. If a trainee misses the equipment surface decontamination process after working in a contaminated area three times consecutively, the system marks this error as a "persistent error" and reinforces its representation in the feature vector. A causal graph model is used to analyze the transmission path between knowledge gaps and practical errors. It was found that the trainee's "insufficient understanding of the concept of half-life" led to "setting the sampling time interval too long," ultimately causing the practical error of "distorted detection results." The system encodes this causal chain as a dimension of the feature vector, fully recording the impact mechanism of knowledge deficiencies on practical operations.

[0080] In some embodiments, multiple VR interaction features include comprehensive skill level features, learning ability features, knowledge practice transformation features, and team collaboration features; the interaction calculation module 220 is used to: process the first interaction feature and the second interaction feature according to an attention-based weighted fusion algorithm to determine the comprehensive skill level feature; process the first interaction feature and the second interaction feature according to a temporal difference learning algorithm and a hidden Markov model to determine the learning ability feature; process the first interaction feature and the second interaction feature according to an association rule mining algorithm to determine the knowledge practice transformation feature; and process the first interaction feature and the second interaction feature according to a social network analysis algorithm to determine the team collaboration feature.

[0081] In this embodiment of the invention, the weighted fusion algorithm based on the attention mechanism assigns dynamic weights to different features by simulating human attention allocation patterns. In a nuclear emergency medical rescue scenario, the system first encodes the operational skill features (such as the standardization of rescue actions and proficiency in equipment operation) and emergency decision-making features (scenario response speed and dynamic adaptability) from the first interactive feature, and the risk of practical errors caused by knowledge weaknesses from the second interactive feature. When simulating high-risk scenarios, the algorithm automatically identifies scenario keywords through a convolutional neural network, increasing the weight of protective operation-related features to 60%, while reducing the weight of routine first aid operations to 25%, and allocating the remaining 15% to other features. For example, if a trainee makes a mistake in the airtightness check of protective equipment (second interactive feature), even if their cardiopulmonary resuscitation operation standardization score (first interactive feature) is high, the weighting of the protective error will significantly reduce their overall score during the comprehensive evaluation.

[0082] A multi-head attention mechanism is employed to process features across different dimensions in parallel, with each attention head focusing on a specific type of feature. For example, the first attention head specifically handles the correlation between operational skills and knowledge gaps, while the second head focuses on the mapping between emergency decisions and practical errors. These are then fused using a weighted summation formula to ultimately generate a comprehensive skill level feature vector with 128 dimensions. Each dimension corresponds to a specific skill element in emergency rescue, such as the accuracy of monitoring equipment operation and the deviation rate of team command execution.

[0083] The temporal difference learning algorithm models trainees' multi-round training data by constructing a Q-learning framework. The system records the first and second interaction features of each simulated rescue operation. Temporal difference errors are calculated. For example, if a trainee's protective measures fail due to a dosage calculation error in the third training session, the algorithm lowers the Q-value of the related skill and focuses on improving that skill in subsequent training. Hidden Markov models are used to model the hidden states of trainees' skill development. Three hidden states are defined, and state transition matrices and observation probability matrices are constructed. For example, if a trainee correctly executes the decontamination process (observation sequence) twice consecutively in a radioactive contaminated area, the Viterbi algorithm is used to infer their most probable state. The final generated learning ability features include six quantitative indicators, such as learning rate (skill improvement per round) and state stability (number of rounds maintaining the same ability state).

[0084] The association rule mining algorithm employs the FP-Growth algorithm to perform frequent itemset mining on the combination data of the first and second interaction features. The system processes student knowledge acquisition and practical behavior data in a transactional manner. By setting a minimum support threshold of 0.3 and a minimum confidence threshold of 0.7, strong association rules in the form of "Mastering dosage calculation methods → Correctly selecting protection levels (support 35%, confidence 82%)" are discovered. Based on the mined association rules, a knowledge-practice transformation evaluation model is constructed. For each rule, an improvement index is calculated to assess the degree of influence of knowledge acquisition on practical behavior.

[0085] Social network analysis algorithms abstract the team rescue process as a directed weighted graph. Data on team collaboration behavior (such as the number of information transmissions and command response times) from the first interaction feature and knowledge complementarity (such as the complementarity of knowledge gaps among members) from the second interaction feature are collected. For example, if trainee A sends monitoring data to trainee B 5 times, the weight of the edge (A, B) is set to 5. If trainee A is skilled in equipment operation and trainee B is familiar with emergency decision-making, their knowledge complementarity is high, so the edge weight coefficient is increased. A dynamically evolving collaborative network is constructed by continuously collecting training data from multiple rounds. Network topology indicators are calculated to evaluate team collaboration capabilities. Regarding node centrality, degree centrality reflects the degree of collaborative activity of members (e.g., trainee C collaborates with 5 members in a rescue, with a degree centrality of 5), while betweenness centrality measures its criticality in information transmission (e.g., trainee D acts as an information hub connecting multiple sub-teams, resulting in a high betweenness centrality). The degree of team division of labor is evaluated by combining clustering coefficients; for example, an overall team clustering coefficient of 0.7 indicates close collaboration among members. The final team collaboration features include eight evaluation dimensions such as leadership index (the proportion of influence of members with high betweenness centrality) and information entropy (the degree of dispersion of team communication), which fully characterize the role positioning and collaborative effectiveness of trainees in team rescue.

[0086] In some embodiments, the AI ​​evaluation module 230 is used to: perform weighted evaluation based on multiple VR interaction features, voice question-and-answer data, theoretical learning data, and spatiotemporal-semantic joint embedding network, and determine the evaluation result.

[0087] In this embodiment of the invention, a spatiotemporal attention mechanism is designed for encoding VR interaction features such as operational skills and emergency decision-making. In the simulated scenario, the trainee's action sequence (spatiotemporal dimension) and knowledge gaps (semantic dimension) are jointly modeled. For example, when an error in the sequence of actions (operational skill feature) is detected in a trainee wearing protective clothing, and this error is related to a knowledge blind spot in the "diffusion path" (knowledge practice transformation feature), the network extracts the temporal features of the actions through spatiotemporal convolutional layers, and simultaneously uses a graph attention network to capture knowledge associations, generating a spatiotemporal-semantic feature vector containing 128 dimensions.

[0088] In this embodiment of the invention, the speech signal is first converted into a text sequence, semantic representation is extracted, and temporal features during the question-and-answer process are captured. In the emergency decision-making simulation question-and-answer session, the system analyzes not only the correctness of the content (semantic dimension) of the trainee's answer to "how to deal with a sudden increase in dosage" but also the fluency of the answer (temporal dimension). For example, if the trainee's answer is accurate but has obvious pauses, the system will lower their score in the emergency communication ability dimension.

[0089] A knowledge graph-enhanced network is constructed to process theoretical learning data. Students' course viewing records and online test scores are transformed into nodes and edges in the knowledge graph, and a graph neural network is used to capture the relationships between knowledge points. For example, if a student scores low in a certain chapter and that knowledge point has a strong correlation with "casualty classification standards," the system predicts that the student may have weaknesses in related practical skills and adjusts the weight accordingly in the evaluation. A spatiotemporal-semantic joint embedding network generates adaptive weights through a meta-learning mechanism. During the training phase, the network learns the importance distribution of each modality in different scenarios. In high-risk scenarios, the weight of protective operation standardization in VR interaction features automatically increases to 0.6, while the weight of protective theoretical knowledge in theoretical learning data is 0.3, and the weight of voice question-and-answer data is 0.1. The feature vectors of each modality are input into a fully connected layer for scoring mapping, generating preliminary scores in 12 dimensions, including operational skills, emergency decision-making, knowledge mastery, and communication skills. A weighted summation formula is used to calculate the final evaluation results.

[0090] Based on the weighted scoring results, a multi-dimensional competency profile is generated, including radar charts and heatmaps. The radar chart displays the trainee's competency distribution across dimensions such as operational skills, emergency decision-making, and knowledge mastery, while the heatmap presents the degree of mastery of each knowledge point. For example, if a trainee scores low in the "Radioactive Material Detection" dimension, and the corresponding "Energy Spectrum Analysis Principle" node in the knowledge graph is displayed in red (not mastered), the system will automatically link it to the "Detection Equipment Operation Error" record in the VR interaction features, forming a complete competency diagnostic chain.

[0091] Personalized improvement suggestions are generated using knowledge graphs and rule engines. When the system detects that a trainee's score in the "emergency decision-making speed" dimension is below the threshold and that there are weaknesses in the "dosage calculation" knowledge node, it automatically generates specific suggestions such as "strengthen rapid dosage estimation training" and "learn emergency decision tree methods." Simultaneously, by incorporating reinforcement learning mechanisms, the system recommends the most beneficial training modules to trainees, such as "rapid decision-making simulation in high-risk scenarios," achieving a seamless transition from evaluation results to teaching intervention.

[0092] In some embodiments, the spatiotemporal-semantic joint embedding network is optimized by a comprehensive loss function L, which includes: an evaluation prediction loss term, a cross-modal alignment loss term, and a scene adaptation loss term. The evaluation prediction loss term measures the difference between the evaluation result predicted by the model and the actual evaluation label. The cross-modal alignment loss term constrains the distance between VR interaction features and semantic features in the joint embedding space. The semantic features are determined based on voice question answering data and theoretical learning data. The scene adaptation loss term ensures the model's generalization ability in different emergency scenarios.

[0093] In this embodiment of the invention, the evaluation prediction loss term uses mean squared error (MSE) or cross-entropy loss function to measure the difference between the evaluation results predicted by the model and the actual evaluation labels. In the evaluation of nuclear emergency medical rescue teaching, the actual evaluation labels are comprehensively assessed by an expert team based on the trainees' actual performance in simulated rescue, covering multiple dimensions such as operational skills, emergency decision-making, and knowledge mastery. This error is fed back to the network. The network adjusts its parameters based on the error to optimize the prediction of trainees' abilities, making subsequent prediction results closer to the actual level and ensuring the reliability and accuracy of the evaluation results.

[0094] The cross-modal alignment loss term aims to constrain the distance between VR interaction features and semantic features in the joint embedding space, promoting the effective fusion of data from different modalities. Semantic features are obtained by processing speech question-and-answer data and theoretical learning data, for example, by semantically encoding the text after speech conversion and theoretical learning materials to extract feature vectors containing knowledge content and logical relationships; VR interaction features are extracted from VR motion capture data, multi-scene practical data, etc., through the interaction computing module.

[0095] During training, the cross-modal alignment loss term employs a contrastive learning approach to narrow the gap in feature representations of the same learner across different modalities, while widening the gap between features of different learners. For example, for learner A, their operational behavior features (VR interaction features) in VR simulated rescue and their descriptive features (semantic features) of rescue knowledge in voice question answering should be as close as possible in the joint embedding space. By minimizing the cross-modal alignment loss function, the network learns how to align information from different modalities, enabling the model to comprehensively utilize multimodal data for more accurate evaluation.

[0096] The scenario adaptation loss term aims to ensure the model's generalization ability across different emergency scenarios, enabling it to adapt to the diverse task requirements in nuclear emergency medical rescue. In nuclear emergency rescue scenarios, each scenario presents different skill requirements and operational priorities for rescue personnel.

[0097] The scene adaptation loss term employs a domain adversarial training method. This involves introducing a scene discriminator to distinguish the scene category of the input data, while simultaneously training the network to prevent the scene discriminator from accurately identifying the model. For example, during training, the scene discriminator attempts to determine which scene category the input VR interaction and semantic features originate from, while the spatiotemporal-semantic joint embedding network adjusts parameters to make data from different scenes as similar as possible in the feature space, thus confusing the scene discriminator's judgment. By minimizing the scene adaptation loss function, the network enhances its adaptability to different scenes, avoiding overfitting to specific scenarios, and thus enabling stable and accurate evaluation in various emergency scenarios.

[0098] The comprehensive loss function L is a weighted sum of the three loss terms mentioned above, and is optimized through methods such as cross-validation. During training, the network simultaneously optimizes these three loss terms, updating network parameters through the backpropagation algorithm to minimize the comprehensive loss function L. As training progresses, the network continuously improves in terms of evaluation accuracy, multimodal data fusion capability, and scenario generalization ability, ultimately achieving a precise and comprehensive evaluation of trainees' nuclear emergency medical rescue capabilities.

[0099] This invention can generate corresponding evaluation systems based on different training tasks. During the construction process, the system provides a visual operating interface, allowing instructors to flexibly adjust the weights of primary indicators such as operational skills, emergency decision-making, and knowledge acquisition, as well as their sub-indicators, according to different nuclear emergency medical rescue scenarios. For example, in a specific scenario, the weights of indicators such as the standardization of protective operations and the accuracy of monitoring can be appropriately increased to highlight the importance of key skills in that scenario.

[0100] This invention's system comprehensively tracks and monitors the teaching process and self-study status of both instructors and students. By deploying cameras, sensors, and other equipment in the teaching environment, combined with image recognition and behavioral analysis technologies, the system can monitor students' learning status in real time, such as their focus and whether they are engaging in any unauthorized activities. For instructors, the system can record their teaching methods, teaching duration, and interactions with students, providing data support for teaching quality assessment. Simultaneously, the system also has an anomaly warning function; when it detects prolonged inactivity or dangerous behavior by students, it promptly issues an alarm to alert relevant personnel.

[0101] This invention can monitor learners' learning status on both PC and VR platforms. On the PC, it uses a camera to monitor learners' facial expressions and body posture to determine their focus. On the VR platform, it utilizes eye-tracking technology and head movement monitoring to understand the learner's level of concentration in the virtual learning environment. For example, if a learner frequently shifts their gaze or remains still for extended periods during a VR course, the system will record this and remind the learner to pay attention to their learning status.

[0102] The system of this invention possesses highly efficient evaluation result feedback capabilities, ensuring that feedback speed is within one minute. After trainees complete the theoretical assessment, the system quickly grades and evaluates the results, providing feedback such as scores and error analysis to the trainees, allowing them to understand their learning outcomes promptly. The system supports multiple assessment methods, including true / false questions, multiple-choice questions, and other question types to meet different training needs and assessment scenarios. Simultaneously, the system can intelligently generate assessment questions based on teaching objectives and knowledge point distribution, ensuring that the assessment content is comprehensive and reasonable.

[0103] After users complete the test, the system automatically grades the submitted papers, scoring the objective questions (single-choice, multiple-answer, true / false, and ranking). The system uses OCR technology or text matching algorithms to quickly and accurately determine the correctness of the student's answers and assigns scores according to preset scoring standards. After automatic grading, the system automatically generates evaluation information based on the answers and the difficulty of the questions. For example, if a student scores highly on an objective question related to a particular knowledge point, the system will evaluate their mastery of that knowledge point; if a student answers a difficult question incorrectly, the system will indicate that the knowledge point is difficult and suggest further study. After automatic grading and evaluation, the results await human review and confirmation to generate the final score and evaluation. During the human review process, graders can refer to the automatic grading results and evaluation suggestions provided by the system, combining their own professional knowledge to ensure the accuracy and reasonableness of the scoring. At the same time, graders can correct errors in the system's automatic grading and improve the evaluation results.

[0104] In this embodiment of the invention, the overall evaluation result of trainees is divided into four levels: Excellent, Good, Pass, and Fail, corresponding to 0.9-1.0, 0.7-0.9, 0.5-0.7, and 0.0-0.5, respectively. When calculating the overall evaluation result, the system comprehensively considers the trainee's performance in various stages, including theoretical assessment, skills training, and comprehensive practice, and performs a weighted calculation according to preset weights. For example, the theoretical assessment score accounts for 30% of the total score, the skills training score accounts for 50%, and the comprehensive practice score accounts for 20%. The trainee's final overall evaluation level is determined by weighted summation of these scores.

[0105] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0106] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An intelligent teaching evaluation system, characterized in that, include: The acquisition module is used to acquire multi-source training data; wherein, the multi-source training data includes VR motion capture data, multi-scene practical data, human-computer collaborative review data, voice question and answer data, and theoretical learning data; An interactive computing module is used to determine multiple VR interactive features based on the VR motion capture data, the multi-scene operation data, and the human-computer collaborative review data; The AI ​​evaluation module is used to determine the evaluation result based on multiple VR interaction features, the voice question and answer data, the theoretical learning data, and the AI ​​evaluation algorithm.

2. The intelligent teaching evaluation system according to claim 1, characterized in that, The interactive computing module is used for: Based on the VR motion capture data and the multi-scene practical data, the first interaction feature is determined; Based on the human-machine collaborative review data and the pre-established knowledge-behavior mapping matrix, the second interaction feature is determined; The knowledge-behavior mapping matrix is ​​the correlation matrix between knowledge gaps and practical errors. Based on the first interaction feature and the second interaction feature, multiple VR interaction features are determined.

3. The intelligent teaching evaluation system according to claim 2, characterized in that, The interactive computing module is used for: Based on the VR motion capture data, the multi-scene practical data, and the intelligent skills training indicator system, multiple VR interaction features are determined; The intelligent skills training indicator system includes primary indicators and at least one secondary indicator corresponding to each primary indicator.

4. The intelligent teaching evaluation system according to claim 3, characterized in that, The interactive computing module is used for: Based on spatiotemporal convolutional networks and dynamic time warping algorithms, features are extracted from the VR motion capture data and the multi-scene practical data according to multiple second indicators to determine the standardization of rescue actions, proficiency in equipment operation, continuity of medical operations, scene response speed, dynamic adaptability, and multi-task coordination ability in each scenario. Based on the standardization of rescue actions, proficiency in equipment operation, and consistency of medical procedures, operational skill characteristics are determined as a primary indicator value. Based on scenario response speed, dynamic adaptability, and multi-task coordination capabilities, emergency decision-making characteristics are determined as another primary indicator value. The first interaction feature is determined based on the primary indicator values ​​for each scenario.

5. The intelligent teaching evaluation system according to claim 2, characterized in that, The interactive computing module is used for: Based on the human-machine collaborative review data, the current knowledge gaps are identified; Based on the current knowledge gaps and the knowledge-behavior mapping matrix, identify practical errors; Based on the aforementioned operational errors, a second interactive feature is determined.

6. The intelligent teaching evaluation system according to claim 2, characterized in that, The multiple VR interaction features include comprehensive skill level features, learning ability features, knowledge practice transformation features, and team collaboration features; The interactive computing module is used for: The first interaction feature and the second interaction feature are processed by a weighted fusion algorithm based on an attention mechanism to determine the comprehensive skill level feature; The first interaction feature and the second interaction feature are processed using a temporal difference learning algorithm and a hidden Markov model to determine the learning ability feature; The first interaction feature and the second interaction feature are processed according to the association rule mining algorithm to determine the knowledge practice transformation feature; The first and second interaction features are processed using a social network analysis algorithm to determine the team collaboration features.

7. The intelligent teaching evaluation system according to claim 2, characterized in that, The AI ​​evaluation module is used for: The evaluation result is determined by weighting the evaluation based on multiple VR interaction features, the voice question-and-answer data, the theoretical learning data, and the spatiotemporal-semantic joint embedding network.

8. The intelligent teaching evaluation system according to claim 7, characterized in that, The spatiotemporal-semantic joint embedding network is optimized using a comprehensive loss function L, which includes: Evaluate the prediction loss term, the cross-modal alignment loss term, and the scene adaptation loss term; The evaluation prediction loss term is used to measure the difference between the evaluation result predicted by the model and the actual evaluation label. The cross-modal alignment loss term is used to constrain the distance between VR interaction features and semantic features in the joint embedding space; the semantic features are determined based on the speech question-answering data and the theoretical learning data. The scenario adaptation loss term is used to ensure the model's generalization ability under different emergency scenarios.

9. An intelligent teaching evaluation device, characterized in that, It includes VR devices, motion capture systems, teaching terminals, and servers equipped with the intelligent teaching evaluation system described in any one of claims 1 to 8.

10. The intelligent teaching evaluation device according to claim 9, characterized in that, The VR device includes VR glasses and a VR backpack; the motion capture system includes motion capture clothing with markers and inertial sensors, and a camera for capturing the markers.

Images