High-frequency quenching virtual simulation teaching system combined with GRU neural network

Abstract

The invention relates to a high-frequency quenching virtual simulation teaching system combined with a GRU neural network, and belongs to the technical field of artificial intelligence and virtual simulation. The problems that an existing high-frequency quenching teaching system is low in parameter prediction precision, and teaching is lack of individuation are solved. The core module comprises a hierarchical user authorization module for realizing authority management; the intelligent teaching auxiliary module provides a knowledge base and learning behavior analysis; the virtual workshop supports scene self-defined layout and equipment interaction; the improved GRU neural network module adopts a GRU and self-attention mechanism fusion framework, bidirectional time sequence correlation prediction of heat treatment process parameters and material performance is achieved, and prediction interpretability is provided through SHAP value analysis. And the multi-dimensional teaching support module dynamically generates personalized exercises and assessment schemes in combination with immersive experiments and intelligent evaluation. The intelligent question answering module supports offline professional question analysis based on a localized knowledge graph.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and virtual simulation technology, specifically relating to a high-frequency quenching virtual simulation teaching system that combines a GRU neural network. Background Technology

[0002] High-frequency quenching, as a core process for surface strengthening of metallic materials, is crucial for the mechanical manufacturing field due to its parameter control and performance prediction. Traditional teaching relies on physical experiments, which has limitations such as high equipment costs, high operational risks, and long experimental cycles. Furthermore, the correlation between the microstructure and macroscopic properties of materials is difficult to present intuitively. While existing virtual simulation systems can partially replace physical operations, parameter prediction methods based on empirical formulas cannot capture multivariate coupling and time-varying characteristics, resulting in insufficient prediction accuracy. Fixed experimental procedures lack intelligent diagnostics, and interactive design is limited to two-dimensional interfaces, weakening immersive understanding.

[0003] Current technical bottlenecks include: difficulty in modeling the complex nonlinear mapping between process parameters and performance; insufficient personalized teaching support, making it impossible to dynamically adjust strategies based on student learning; rigid virtual scene interaction, lacking multimodal sensory feedback; and experimental result analysis limited to single indicators, making it difficult for students to build systematic understanding. Furthermore, existing systems rely on network environments and cannot provide intelligent services offline. Summary of the Invention

[0004] The purpose of this invention is to propose a high-frequency quenching virtual simulation teaching system that combines GRU neural networks, thereby solving the problems of insufficient accuracy in process parameter prediction, weak personalized teaching support, and limited immersion and interactivity in existing high-frequency quenching teaching systems.

[0005] To achieve the above objectives, the present invention provides a high-frequency quenching virtual simulation teaching system incorporating a GRU neural network, comprising a hierarchical user login authorization module, an intelligent teaching assistance module, a visual teaching resource module, an intelligent assessment and examination management module, an intelligent question bank management module, a virtual workshop scene management module, a virtual human operation guidance system, a multi-dimensional teaching support module, a heat treatment performance prediction module based on an improved GRU neural network, and an intelligent teaching Q&A module.

[0006] The hierarchical user login authorization module is used to verify user logins, dividing users into two levels: Level 1 users and Level 2 users, and assigning different system application permissions to users at different levels. The hierarchical user login authorization module includes a user information storage unit, a login verification unit, and a permission allocation unit. The user information storage unit stores user login information and user level information. The login verification unit receives login information input by the user and compares it with the stored information; if a match is found, the user is allowed to log in to the system. After successful login, the permission allocation unit assigns corresponding permissions according to the user level. Level 1 users are granted application permissions to all modules in the system, while Level 2 users are granted application permissions to all modules except the intelligent assessment and examination management module and the intelligent question bank management module.

[0007] The intelligent teaching assistance module is used to assist users in learning experiment-related knowledge and finding experiment-related materials. The intelligent teaching assistance module includes a knowledge base management submodule and a learning behavior analysis submodule. The knowledge base management submodule is used to provide users with experiment-related materials, including literature, experimental manuals, and books on annealing, normalizing, quenching, and tempering. The learning behavior analysis submodule is used by users to view their personal learning status, which includes system usage time, existing knowledge gaps, and completion status of assignments and tests.

[0008] The visualized teaching resource module provides users with learning resource display and interactive experience. It includes pre-defined knowledge graphs and knowledge maps, which visually represent various knowledge points related to high-frequency quenching and related heat treatment processes, along with their interrelationships. Users can perform the following operations on any knowledge point within the knowledge graph and knowledge map:

[0009] Favorites feature: Users have the right to favorite knowledge points that they are interested in. The system will record the favorite information so that users can quickly find and view the knowledge points the next time they log in to the system.

[0010] Highlighter marking: Users have the right to use a highlighter to mark specific knowledge points. The marked knowledge points will be presented in a prominent way, making it easy for users to quickly identify key content when reviewing later.

[0011] Add annotations: Users have the authority to add their own annotations to any knowledge point. The system will save the annotations added by users and associate them with the corresponding knowledge points.

[0012] The intelligent assessment and examination management module provides comprehensive assessment and examination management services for Level 1 users. This module includes a learning statistics submodule and an examination monitoring submodule. The learning statistics submodule displays the overall learning status of all users to administrators using intuitive methods, including histograms, pie charts, and heatmaps. The overall learning status includes the learning progress, test scores, mastery rates of different knowledge points, and homework completion rates of all users. The examination monitoring submodule monitors the examination behavior of Level 2 users during examinations. If a Level 2 user switches windows or other application windows appear on the screen during the examination, the system records this information.

[0013] The intelligent question bank management module is used to generate and manage test questions and exercises related to high-frequency quenching virtual simulation teaching, and interacts with the learning statistics submodule in the intelligent assessment and examination management module. The intelligent question bank management module includes a manual question creation submodule and an AI dynamic question creation submodule. The manual question creation submodule allows first-level users to create questions independently; the AI ​​dynamic question creation submodule automatically creates questions based on the knowledge points of high-frequency quenching and related heat treatment processes, as well as the learning data of second-level users.

[0014] The virtual workshop scene management module allows users to create personalized virtual workshop scenes according to experimental requirements and preferences. This module includes a scene and equipment switching submodule and a custom scene creation submodule. The scene and equipment switching submodule allows users to select virtual scene styles according to their preferences and choose experimental equipment based on experimental requirements. Virtual scene styles include preset scenes of different styles or customizable wall, table, chair, and floor designs. Experimental equipment includes a forging air hammer, a trolley-type box furnace, a pit-type heat treatment furnace, a gear hanger, a high-frequency induction hardening heating device, and an impact toughness testing machine. The system includes an electronic universal testing machine and a Vickers hardness tester. The custom scene creation submodule provides a visual editing interface in a blank virtual workshop, allowing users to design scene layouts independently through drag-and-drop operations. It supports position adjustment and orientation rotation of tables and chairs, forging air hammers, trolley box-type heating furnaces, pit-type heat treatment furnaces, gear hangers, high-frequency induction quenching heating equipment, Vickers hardness testers, machining centers, electronic universal testing machines, water tanks, oil tanks, and sample storage cabinets. It also has scene saving and loading functions, allowing users to save custom scene configurations as templates and reuse them in subsequent experiments. The system automatically detects the rationality of the layout based on the physical properties of the equipment and prompts potential conflicts.

[0015] The virtual human operation guidance system is used to equip each experimental device with an NPC in a preset virtual workshop scene. When a user approaches or operates the corresponding experimental device, the NPC will provide a detailed explanation of the usage methods and precautions for the experimental device through voice and corresponding action demonstrations, based on pre-stored information about the usage methods and precautions of the corresponding experimental device. The virtual human operation guidance system works in conjunction with the intelligent teaching and Q&A module. Users have the authority to ask questions to the NPC in text. After receiving the question, the NPC transmits it to the intelligent teaching and Q&A module, which analyzes and processes the question and sends the answer back to the NPC, which then presents it to the user.

[0016] The multi-dimensional teaching support module includes a pre-learning knowledge learning sub-module, a coursework task sub-module, an immersive learning sub-module, and an intelligent assessment sub-module. The pre-learning knowledge learning sub-module stores videos, images, and data related to the experiment and interacts with the AI ​​dynamic question generation sub-module and the learning behavior analysis sub-module. By acquiring data such as user learning time and completion status of pre-learning exercises recorded by the learning behavior analysis sub-module, the AI ​​dynamic question generation sub-module generates personalized pre-learning exercises based on the user's knowledge gaps and pushes them to the pre-learning knowledge learning sub-module. After the user completes the exercises, the system sends back data such as the accuracy rate and answering time to the learning behavior analysis sub-module. The analysis submodule performs statistical analysis; the homework task submodule supports first-level users in assigning homework, exercises, and exams, and it interfaces with the intelligent assessment and exam management module to obtain specific content. After second-level users complete tasks through this module, the system automatically synchronizes the performance data to the learning behavior analysis submodule for generating individual learning reports and the learning situation statistics submodule for analyzing group learning situations; in the immersive learning submodule, users operate according to the following process: select mold → select material → forging → preheat treatment → rough machining → five process steps of free combination quenching, high frequency quenching, high temperature tempering, medium temperature tempering, and low temperature tempering → fine machining → performance testing. During the experiment, the system provides operation guidance through highlighted display areas, text box prompts, audio command broadcasts, and NPC interaction next to the virtual human operation guidance system. Simultaneously, the system transmits the user-input heat treatment route, temperature and time parameters, part type, and material data to the heat treatment performance prediction module based on an improved GRU neural network, receiving the hardness value and tensile test curve output by this module as the experimental results. The intelligent assessment submodule, based on customized questions generated by the AI ​​dynamic question-setting submodule, evaluates user capabilities in two ways: first, a theoretical assessment, testing the user's mastery of the experimental operation process and related knowledge and generating an assessment result; second, a practical assessment, simulating the real experimental process through the immersive learning submodule, recording user operation steps without prompts and generating operation judgment results. During the assessment, the examination monitoring submodule is simultaneously triggered to record user behavior, and the final assessment results and operation judgment results are fed back to the learning behavior analysis submodule and the learning statistics submodule for personalized learning suggestion generation and teaching quality evaluation. The completion status of pre-experiment exercises and the number of experimental rounds completed in the immersive learning submodule are displayed through the intelligent teaching assistance module.

[0017] The heat treatment performance prediction module based on the improved GRU neural network is used to receive the user-selected heat treatment route, temperature and time parameters of each process, part type and material data transmitted by the immersive learning submodule in the multidimensional teaching support module. The module performs feature extraction and time-series relationship analysis on the input data through a pre-trained improved GRU neural network model. Combined with historical experimental data and material performance database, it predicts and outputs the hardness value and tensile test curve of the material after heat treatment.

[0018] The intelligent teaching and Q&A module serves as a local intelligent assistant to provide teaching support. Users can trigger Q&A interactions by using a virtual human to operate the module's NPC next to the device, or they can directly access it through an independent entry point on any interface of the system. The intelligent teaching and Q&A module integrates a knowledge base in the field of high-frequency quenching and heat treatment processes, allowing users to ask professional questions about experimental operations, process parameters, and equipment usage in text form. The system automatically analyzes the questions and generates answers that include graphic explanations, experimental operation specifications, or experimental data comparisons.

[0019] The beneficial effects of this invention are as follows: This invention significantly improves teaching efficiency through algorithmic innovation and scenario reconstruction. The improved recurrent neural network model, supported by historical experimental data, breaks through the prediction accuracy bottleneck of traditional empirical formulas, achieving precise mapping between heat treatment parameters and material properties. The introduction of knowledge graph technology enables dynamic organization of teaching content; the system can generate reinforcement exercises for weak knowledge points in real time based on student behavior data, effectively solving the pain point of the traditional "one-size-fits-all" teaching approach. Multimodal interactive design combined with physical phenomenon simulation significantly enhances the realism of equipment operation, greatly improving students' understanding of complex processes. A localized intelligent Q&A module provides professional question analysis. Through interpretability analysis, students can intuitively understand the influence weight of process parameters on material properties, establishing a systematic understanding of "parameter-structure-performance," significantly improving both experimental operation standardization and knowledge mastery depth. This invention provides a complete solution for high-frequency quenching teaching that combines scientific rigor, safety, and engaging learning. This invention uses a deep learning model to predict the bidirectional time-series correlation between process parameters and performance, and combines a knowledge graph to establish a dynamic learning analysis mechanism to support personalized teaching guidance. It also employs multimodal interaction technology to enhance immersion and helps students understand the influence of parameters through multi-dimensional performance prediction and interpretability analysis. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall framework of a high-frequency quenching virtual simulation teaching system incorporating a GRU neural network, according to the present invention.

[0021] Figure 2 This is a flowchart of the experiment.

[0022] Figure 3 Here is a structural diagram of the intelligent teaching Q&A module;

[0023] Figure 4 Structure diagram of the hierarchical user login authorization module;

[0024] Figure 5 This is a structural diagram of the intelligent question bank management module;

[0025] Figure 6 This is a diagram of the virtual workshop structure. Detailed Implementation

[0026] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0027] See Figure 1 The present invention provides a high-frequency quenching virtual simulation teaching system incorporating a GRU neural network, comprising a hierarchical user login authorization module, an intelligent teaching assistance module, a visual teaching resource module, an intelligent assessment and examination management module, an intelligent question bank management module, a virtual workshop scene management module, a virtual human operation guidance system, a multi-dimensional teaching support module, a heat treatment performance prediction module based on an improved GRU neural network, and an intelligent teaching Q&A module.

[0028] The hierarchical user login authorization module is used to verify user logins, dividing users into two levels: Level 1 users and Level 2 users, and assigning different system application permissions to users at different levels. The module includes a user information storage unit, a login verification unit, and a permission allocation unit. The user information storage unit stores user login information and user level information. The login verification unit receives login information input by the user and compares it with the stored information; if a match is found, the user is allowed to log in. After successful login, the permission allocation unit assigns corresponding permissions based on the user level. Level 1 users are granted application permissions to all modules in the system, while Level 2 users are granted application permissions to all modules except the intelligent assessment and examination management module and the intelligent question bank management module. Level 1 users are administrators, and Level 2 users are student users.

[0029] The intelligent teaching assistance module is used to assist users in learning experiment-related knowledge and finding experiment-related materials. The module includes a knowledge base management submodule and a learning behavior analysis submodule. The knowledge base management submodule provides users with experiment-related materials, including literature, experimental manuals, and books on annealing, normalizing, quenching, and tempering. The learning behavior analysis submodule allows users to independently view their personal learning progress, including system usage time, existing knowledge gaps, completion status of assignments and tests, completion status of pre-experiment exercises in the immersive learning submodule, and the number of rounds of experiments completed in the immersive learning submodule.

[0030] The visualized teaching resource module provides users with an intuitive and convenient display and interactive experience of learning resources. It includes pre-defined knowledge graphs and knowledge maps, which visually represent various knowledge points related to high-frequency quenching and related heat treatment processes, along with their interrelationships. Users can perform the following operations on any knowledge point within the knowledge graph and knowledge map:

[0031] Favorites feature: Users can favorite knowledge points that they are interested in. The system will record this information so that users can quickly find and view the knowledge points the next time they log in to the system.

[0032] Highlighter marking: Users can use a highlighter to mark specific knowledge points. The marked knowledge points will be presented in a prominent way, making it easy for users to quickly identify key content when reviewing later.

[0033] Add annotations: Users can add their own annotations for any knowledge point, and the system will save the annotations added by users and associate them with the corresponding knowledge points;

[0034] The intelligent assessment and examination management module provides comprehensive assessment and examination management services for Level 1 users. This module includes a learning statistics submodule and an examination monitoring submodule. The learning statistics submodule displays the overall learning status of all users to administrators using intuitive methods, including histograms, pie charts, and heatmaps. The overall learning status includes the learning progress, test scores, mastery rates of different knowledge points, and homework completion rates of all users. The examination monitoring submodule monitors the examination behavior of Level 2 users during examinations. If a Level 2 user switches windows or other application windows appear on the screen during the examination, the system records this information.

[0035] The intelligent question bank management module is used to generate and manage test questions and exercises related to high-frequency quenching virtual simulation teaching. It interacts with the learning statistics submodule in the intelligent assessment and examination management module. The intelligent question bank management module includes a manual question-generating submodule and an AI dynamic question-generating submodule. The manual question-generating submodule allows first-level users to generate questions independently. It can obtain common knowledge gaps among second-level users (ordinary students) based on information provided by the learning statistics submodule, providing a reference for question generation. The AI ​​dynamic question-generating submodule automatically generates questions based on knowledge points of high-frequency quenching and related heat treatment processes, as well as learning data of second-level users, automatically generating test questions or exercises that meet teaching needs and assessment standards.

[0036] The virtual workshop scene management module allows users to create personalized virtual workshop scenes according to experimental requirements and preferences. This module includes a scene and equipment switching submodule and a custom scene creation submodule. The scene and equipment switching submodule allows users to select virtual scene styles according to their preferences and choose experimental equipment based on experimental requirements. Virtual scene styles include preset scenes of different styles or customizable wall, table, chair, and floor designs. Experimental equipment includes forging air hammers, trolley-type box furnaces, pit-type heat treatment furnaces, gear hangers, high-frequency induction hardening heating equipment, impact toughness testing machines, and electric... The system includes a universal testing machine, a Vickers hardness tester, and a machining center. The custom scene creation submodule provides a visual editing interface within a blank virtual workshop, allowing users to design scene layouts through drag-and-drop operations. It supports position adjustments and orientation rotations for tables and chairs, forging air hammers, trolley-type heating furnaces, pit-type heat treatment furnaces, gear hangers, high-frequency induction hardening heating equipment, Vickers hardness testers, machining centers, electronic universal testing machines, water tanks, oil tanks, and sample storage cabinets. It also features scene saving and loading functions, allowing users to save custom scene configurations as templates for reuse in subsequent experiments. The system automatically detects the layout rationality based on the physical properties of the equipment and alerts to potential conflicts.

[0037] The virtual human operation guidance system is used to equip each experimental device with an NPC in a preset virtual workshop scene. When a user approaches or operates the corresponding experimental device, the NPC will provide a detailed explanation of the usage methods and precautions of the corresponding experimental device (such as forging air hammer, trolley box furnace, pit heat treatment furnace, gear hanger, high frequency induction quenching heating equipment, Vickers hardness tester, machining center, electronic universal testing machine, water tank, oil tank) to the user through voice explanation and corresponding action demonstrations. The virtual human operation guidance system works in conjunction with the intelligent teaching and Q&A module. Users can ask questions to the NPC in text. After receiving the question, the NPC will transmit it to the intelligent teaching and Q&A module, which will analyze and process the question and send the answer back to the NPC, which will then present it to the user.

[0038] The multi-dimensional teaching support module includes a pre-learning knowledge learning sub-module, a coursework task sub-module, an immersive learning sub-module, and an intelligent assessment sub-module. The pre-learning knowledge learning sub-module stores videos, images, and data related to the experiment and interacts with the AI ​​dynamic question-generating sub-module and the learning behavior analysis sub-module. By acquiring data such as user learning time and pre-learning task completion recorded by the learning behavior analysis sub-module, the AI ​​dynamic question-generating sub-module generates personalized pre-learning tasks based on user knowledge gaps and pushes them to the pre-learning knowledge learning sub-module. After the user completes the tasks, the system sends data such as task accuracy and completion time back to the learning behavior analysis sub-module for statistical analysis. The task assignment submodule supports first-level users in assigning homework, exercises, and exams. It interfaces with the intelligent assessment and exam management module to obtain specific content. After second-level users complete tasks through this module, the system automatically synchronizes the performance data to the learning behavior analysis submodule for generating individual learning reports and the learning statistics submodule for analyzing group learning. In the immersive learning submodule, users operate according to the following process: select mold → select material → forging → pre-heat treatment (normalizing / annealing route selection) → rough machining → five process steps of free combination quenching, high-frequency quenching, high-temperature tempering, medium-temperature tempering, and low-temperature tempering (the system detects the rationality of the process in real time and prompts errors) → finish machining → performance testing. During the experiment, the system provides operation guidance through highlighted display areas, text box prompts, audio command broadcasts, and NPC interaction next to the device in the virtual human operation guidance system. Simultaneously, the system transmits the user-input heat treatment route, temperature and time parameters, part type, and material data to the heat treatment performance prediction module based on an improved GRU neural network, receiving the hardness value and tensile test curve output by this module as the experimental results. The intelligent assessment submodule, based on customized questions generated by the AI ​​dynamic question-setting submodule, evaluates user capabilities in two ways: first, a theoretical assessment, testing the user's mastery of the experimental operation process and related knowledge and generating an assessment result; second, a practical assessment, simulating the real experimental process through the immersive learning submodule, recording user operation steps without prompts and generating operation judgment results. During the assessment, the examination monitoring submodule is simultaneously triggered to record user behavior, and finally, the assessment results and operation judgment results are fed back to the learning behavior analysis submodule and the learning statistics submodule for personalized learning suggestion generation and teaching quality evaluation.

[0039] The heat treatment performance prediction module based on the improved GRU neural network is used to receive the user-selected heat treatment route (including process combinations such as quenching, high-frequency quenching, and tempering), temperature and time parameters (temperature value and duration), part type and material data transmitted by the immersive learning submodule in the multidimensional teaching support module. The module performs feature extraction and time series relationship analysis on the input data through a pre-trained improved GRU neural network model, and combines historical experimental data and material performance database to predict and output the hardness value and tensile test curve of the material after heat treatment.

[0040] The intelligent teaching and Q&A module serves as a localized intelligent assistant to provide teaching support. Users can trigger Q&A interactions through a virtual human-guided NPC next to the device, or access it directly from any system interface via a dedicated entry point. This module integrates a knowledge base in the field of high-frequency quenching and heat treatment processes, allowing users to submit professional questions about experimental operations, process parameters, and equipment usage in text form. The system automatically parses the questions and generates answers including graphic explanations, experimental operation specifications, or experimental data comparisons. The module has multi-turn dialogue capabilities, continuously providing guided answers based on user follow-up questions. Level 1 users (administrators) can manage the knowledge base content through this module, including uploading and updating professional documents, editing frequently asked questions and answers, and experimental specifications. The module adopts a localized deployment architecture, supporting offline operation, and achieves intelligent response functionality in offline environments through document parsing, semantic retrieval, and a lightweight inference technology stack.

[0041] See Figure 2 The immersive learning submodule serves as the core module of the virtual simulation experiment. Users can operate according to the following process: select mold → select material → forging → pre-heat treatment (normalizing / annealing route selection) → rough machining → free combination of quenching, high-frequency quenching, high-temperature tempering, medium-temperature tempering, low-temperature tempering, and other process steps (the system detects the rationality of the process in real time and prompts errors) → finish machining → performance testing. During the experiment, the system guides the operation through highlighted display areas, text box prompts, audio command broadcasts, and device-side NPC interaction combined with the virtual human operation guidance module. At the same time, the heat treatment route, temperature and time parameters, part type, and material data input by the user are transmitted to the heat treatment performance prediction module based on an improved GRU neural network, and the hardness value and tensile test curve output by this module are received as experimental results.

[0042] The knowledge base management submodule uses MongoDB and relational databases for data storage, storing unstructured data such as full-text experimental papers and structured data such as metadata. For retrieval, it integrates the Elasticsearch full-text search engine with Jieba word segmentation technology to build a text index, enabling intelligent retrieval that supports Boolean logic combinations. In terms of personalized recommendations, it calculates the semantic matching degree between keywords highlighting user weaknesses and the provided text using a cosine similarity algorithm. This is combined with data popularity, update time, user history, and user feedback data to establish a multi-dimensional recommendation strategy for generating accurate recommendations. Finally, this submodule interacts with the learning behavior analysis submodule through a data interface, triggering the aforementioned recommendation strategy based on the latter's feedback on weaknesses, and providing users with relevant materials.

[0043] See Figure 3 The intelligent teaching and Q&A module uses the PyMuPDF library to achieve localized parsing and text extraction of PDF documents. It uses regular expressions to match and filter special symbols, redundant spaces, and unstructured content, completing text cleaning and sentence segmentation. The Sentence-BERT model is loaded and run locally to convert the cleaned text into 384-dimensional semantic vectors. The generated vector data is saved to the local disk using ChromaDB's file storage mode. An IVFFlat index is built using the FAISS library for efficient retrieval. Based on the spaCy offline model, the module parses the subject-verb-object structure of user questions, extracts core intent keywords, and calculates the matching degree between the question vector and knowledge block vectors in the database using a cosine similarity algorithm, returning the five knowledge blocks with the highest similarity. The module then calls the LLaMA-7B quantization model for local inference, concatenating the retrieval results with the question vector and inputting the result into the model to generate an answer. The Jinja2 template engine is used to match the answer content with the cited sources using regular expressions. The system matches the original document chapters of knowledge blocks using Dass and renders them in Markdown format. It integrates the Neo4j embedded database to construct a knowledge graph, extracts relationships using the Stanza library to establish semantic connections between entities, and combines custom terminology replacement rules to achieve domain adaptation. It uses the OpenBLAS library to optimize matrix operations, and the Watchdog library to monitor local folders for incremental updates, triggering the Dask parallel processing framework to parse, encode, and index new documents. Finally, it packages the system into a standalone executable file using PyInstaller, storing model parameters and data in a local SQLite database, enabling a fully network-free workflow. Through an integrated technology stack encompassing document processing, vector storage, question-and-answer generation, citation annotation, domain enhancement, performance optimization, and local deployment, the system completes the entire process from data import to intelligent response in a network-free environment, supporting three core functions: multi-format document parsing, semantic retrieval, and lightweight inference.

[0044] See Figure 4 The hierarchical user login authorization module is used to verify user logins and divide users into two levels: Level 1 users and Level 2 users. Different system application permissions are assigned to users at different levels. The hierarchical user login authorization module includes a user information storage unit for storing user login information and user level information; a login verification unit for receiving user-input login information and comparing it with the stored information. If they match, the user is allowed to log in to the system; and a permission allocation unit for assigning corresponding permissions according to the user level after successful login. Level 1 users are granted application permissions to all modules in the system, while Level 2 users are granted application permissions to all modules except the intelligent assessment and examination management module and the intelligent question bank management module.

[0045] The manual question-setting submodule is used for first-level users, i.e. administrators, to independently create questions. Based on the knowledge gaps commonly found by second-level users, i.e. ordinary students, provided by the learning statistics submodule, administrators can enter the question content, associate it with preset knowledge points, difficulty level, question type, standard answer and explanation on the visual question-setting interface. The test data is stored in a structured relational database.

[0046] See Figure 5 The AI-powered dynamic question generation submodule, based on pre-set knowledge points and data from the learning progress statistics submodule regarding users' knowledge mastery and weak areas, achieves intelligent question generation through the following mechanisms: It determines the weight of weak knowledge points based on learning progress data, prioritizing questions of that type; it filters and sorts questions in the pre-set question bank based on difficulty coefficients, adjusting the question distribution according to the principle of low difficulty emphasizing basic concepts and high difficulty emphasizing comprehensive application; and it generates a combination of multiple-choice and short-answer questions in proportion to meet question type distribution requirements through a combined query algorithm. The module employs association rule mining technology to analyze historical question generation data, optimizes the matching strategy between knowledge points and question types, and uses clustering algorithms to identify similarities between questions, avoiding duplicate question selection. The generated test paper uses a set coverage algorithm to treat all preset knowledge points as the entire set, checking whether the subsets of knowledge points in the test paper cover key knowledge points. If any important knowledge points are not covered, they are marked as anomalies. An entropy weighting method is used to assign weights to each knowledge point, calculating the entropy value of the knowledge point distribution in the test paper to assess its uniformity and completeness. If the entropy value is too low, it indicates that there may be missing knowledge points. Box plot analysis is used to create a dataset of test paper question difficulty coefficients and draw a box plot. The upper and lower quartiles, median, and outliers are used to determine whether the difficulty gradient is reasonable; the presence of outliers indicates that the question difficulty does not match the overall difficulty. Linear regression is used to fit the assumption that the test paper difficulty changes linearly with the question order. The question difficulty and sequence number are fitted; if the goodness of fit is low, it indicates that the difficulty gradient does not conform to a linear law, and there may be difficulty jumps. After these anomaly detection algorithms are verified, the test paper is output to the intelligent assessment module.

[0047] The heat treatment performance prediction module based on the improved GRU neural network adopts a fusion architecture of gated GRU units and self-attention mechanism, and its specific structure is as follows:

[0048] The heat treatment performance prediction module based on the improved GRU neural network comprises an input layer, a bidirectional GRU layer, a self-attention layer, and a multi-task output layer. The input layer encodes the heat treatment route, temperature parameters, time parameters, part type, and material data into a fixed-length feature vector, with the input dimension determined based on the data preprocessing results. The bidirectional GRU layer consists of two GRU units, each containing 128 neurons, capturing the bidirectional dependencies between parameters through forward and backward propagation. The self-attention layer employs a 4-head attention mechanism to calculate the correlation weight matrix between parameters of different processes, and the output is concatenated with the features of the bidirectional GRU layer before being input into the feedforward neural network. The multi-task output layer outputs hardness value prediction (mean square error loss) and tensile curve parameters (dynamic time warping loss) through two branches, with the hardness prediction branch containing a fully connected layer of 64 neurons and the tensile curve prediction branch containing a fully connected layer of 32 neurons.

[0049] The training process of the heat treatment performance prediction module based on the GRU neural network includes the following steps:

[0050] First, historical experimental data covering different materials, part types, and heat treatment process combinations were collected to construct a training dataset containing material, process parameters, hardness values, and tensile curves. After data standardization, the dataset was divided into training and validation sets in an 8:2 ratio. During training, the AdamW optimizer was used with an initial learning rate of 0.001, and a cosine annealing strategy was used to dynamically adjust the learning rate. The network was trained for 3000 epochs, and model parameters were saved every 500 epochs. The loss function included the mean squared error (MSE) of hardness prediction and the dynamic time warping (DTW) loss of curve prediction, while an attention regularization term was introduced to prevent overfitting. During the validation phase, the model performance was evaluated by comparing the similarity between the predicted curve and the true curve and the hardness prediction error.

[0051] The different materials are specifically 40Cr, 20CrMnTi, 35CrMo and No. 45 steel.

[0052] The specific types of parts are cylindrical gears, bevel gears, helical gears, and sprockets.

[0053] The heat treatment performance prediction module based on the improved GRU neural network is interpretable and can provide explanatory information for the prediction results. Specifically, during the model training process, the SHAP (SHapley Additive ex Planations) value feature importance analysis algorithm is used to calculate the contribution of each input feature, such as heat treatment route, temperature and time parameters, part type, and material data, to the prediction results such as hardness value and tensile test curve. This feature importance information is then displayed to the user so that the user can understand which factors have a greater impact on material properties.

[0054] The visualized teaching resource module supports multimodal interactive functions of knowledge graphs, specifically including:

[0055] Users can zoom and rotate the knowledge graph, and the system updates the graph layout in real time while maintaining a clear display of the relationships between knowledge points. When a user selects a specific knowledge point, the system automatically retrieves the corresponding 3D process animation, experimental data charts, and historical exam error sets, displaying them synchronously through a floating window. Furthermore, the system supports users creating knowledge point association paths, generating personalized learning roadmaps by dragging and dropping knowledge point nodes.

[0056] The immersive learning submodule combines lighting effects, spark effects, and smoke simulation during experimental operations. When users perform high-temperature processes such as quenching, the system simulates flickering light and shadow effects under high temperatures; when using equipment such as air hammers for forging operations, spark effects are used to enhance the realism of the operation; and when gas-related processes are involved, corresponding smoke effects are simulated, along with sound prompts such as the sound of equipment operation and gas flow, comprehensively enhancing the user's immersive experience.

Claims

1. A high-frequency quenching virtual simulation teaching system combining GRU neural networks, characterized in that, It includes a tiered user login authorization module, an intelligent teaching assistance module, a visual teaching resource module, an intelligent assessment and examination management module, an intelligent question bank management module, a virtual workshop scene management module, a virtual human operation guidance system, a multi-dimensional teaching support module, a heat treatment performance prediction module based on an improved GRU neural network, and an intelligent teaching Q&A module; The hierarchical user login authorization module is used to verify user login, divide users into two levels: Level 1 users and Level 2 users, and assign different system application permissions to users of different levels; the hierarchical user login authorization module includes a user information storage unit, a login verification unit, and a permission allocation unit; The user information storage unit is used to store the user's login information and user level information; the login verification unit is used to receive the login information entered by the user and compare it with the stored information. If they match, the user is allowed to log in to the system; after the user successfully logs in, the permission allocation unit allocates corresponding permissions according to the user level. The first-level user is granted the application permission to use all modules in the system, and the second-level user is granted the application permission to use all modules except the intelligent assessment and examination management module and the intelligent question bank management module. The intelligent teaching assistance module is used to assist users in learning experiment-related knowledge and finding experiment-related materials. The intelligent teaching assistance module includes a knowledge base management submodule and a learning behavior analysis submodule. The knowledge base management submodule is used to provide users with experiment-related materials, including literature, experimental manuals, and books on annealing, normalizing, quenching, and tempering. The learning behavior analysis submodule is used by users to view their personal learning status, which includes system usage time, existing knowledge gaps, and completion status of assignments and tests. The visualized teaching resource module provides users with learning resource display and interactive experience. It includes pre-defined knowledge graphs and knowledge maps, which visually represent various knowledge points related to high-frequency quenching and related heat treatment processes, along with their interrelationships. Users can perform the following operations on any knowledge point within the knowledge graph and knowledge map: Favorites feature: Users have the right to favorite knowledge points that they are interested in. The system will record the favorite information so that users can quickly find and view the knowledge points the next time they log in to the system. Highlighter marking: Users have the right to use a highlighter to mark specific knowledge points. The marked knowledge points will be presented in a prominent way, making it easy for users to quickly identify key content when reviewing later. Add to Note: Users have the authority to add their own notes for any knowledge point. The system will save the notes added by users and associate them with the corresponding knowledge points. The intelligent assessment and examination management module provides comprehensive assessment and examination management services for Level 1 users. This module includes a learning statistics submodule and an examination monitoring submodule. The learning statistics submodule displays the overall learning status of all users to administrators using intuitive methods, including histograms, pie charts, and heatmaps. The overall learning status includes the learning progress, test scores, mastery rates of different knowledge points, and homework completion rates of all users. The examination monitoring submodule monitors the examination behavior of Level 2 users during examinations. If a Level 2 user switches windows or other application windows appear on the screen during the examination, the system records this information. The intelligent question bank management module is used to generate and manage test questions and exercises related to high-frequency quenching virtual simulation teaching, and interacts with the learning statistics submodule in the intelligent assessment and examination management module. The intelligent question bank management module includes a manual question creation submodule and an AI dynamic question creation submodule. The manual question creation submodule allows first-level users to create questions independently; the AI ​​dynamic question creation submodule automatically creates questions based on the knowledge points of high-frequency quenching and related heat treatment processes, as well as the learning data of second-level users. The virtual workshop scene management module allows users to create personalized virtual workshop scenes according to experimental requirements and preferences. This module includes a scene and equipment switching submodule and a custom scene creation submodule. The scene and equipment switching submodule allows users to select virtual scene styles according to their preferences and choose experimental equipment based on experimental requirements. Virtual scene styles include preset scenes of different styles, or users can customize different styles of walls, tables, chairs, and floors. Experimental equipment includes forging air hammers, trolley-type box furnaces, pit-type heat treatment furnaces, gear hangers, high-frequency induction hardening heating equipment, and impact toughness testing machines. Electronic universal testing machine, Vickers hardness tester, machining center; the custom scene creation submodule provides a visual editing interface in a blank virtual workshop, allowing users to design scene layouts independently through drag and drop operations. It supports position adjustment and orientation rotation of tables and chairs, forging air hammers, trolley box-type heating furnaces, pit-type heat treatment furnaces, gear hangers, high-frequency induction quenching heating equipment, Vickers hardness testers, machining centers, electronic universal testing machines, water tanks, oil tanks, and sample storage cabinets. It also has scene saving and loading functions, allowing users to save custom scene configurations as templates and reuse them in subsequent experiments. The system automatically detects the rationality of the layout based on the physical properties of the equipment and prompts potential conflicts. The virtual human operation guidance system is used to equip each experimental device with an NPC in a preset virtual workshop scene. When a user approaches or operates the corresponding experimental device, the NPC will provide a detailed explanation of the usage methods and precautions for the experimental device through voice and corresponding action demonstrations, based on pre-stored information about the usage methods and precautions of the corresponding experimental device. The virtual human operation guidance system works in conjunction with the intelligent teaching and Q&A module. Users have the authority to ask questions to the NPC in text. After receiving the question, the NPC transmits it to the intelligent teaching and Q&A module, which analyzes and processes the question and sends the answer back to the NPC, which then presents it to the user. The multi-dimensional teaching support module includes a pre-learning knowledge learning sub-module, a coursework task sub-module, an immersive learning sub-module, and an intelligent assessment sub-module. The pre-learning knowledge learning sub-module stores videos, images, and data related to the experiment and interacts with the AI ​​dynamic question generation sub-module and the learning behavior analysis sub-module. By acquiring data such as user learning time and completion status of pre-learning exercises recorded by the learning behavior analysis sub-module, the AI ​​dynamic question generation sub-module generates personalized pre-learning exercises based on the user's knowledge gaps and pushes them to the pre-learning knowledge learning sub-module. After the user completes the exercises, the system sends back data such as the accuracy rate and answering time to the learning behavior analysis sub-module. The analysis submodule performs statistical analysis; the homework task submodule supports first-level users in assigning homework, exercises, and exams, and it interfaces with the intelligent assessment and exam management module to obtain specific content. After second-level users complete tasks through this module, the system automatically synchronizes the performance data to the learning behavior analysis submodule for generating individual learning reports and the learning situation statistics submodule for analyzing group learning situations; in the immersive learning submodule, users operate according to the following process: select mold → select material → forging → preheat treatment → rough machining → five process steps of free combination quenching, high frequency quenching, high temperature tempering, medium temperature tempering, and low temperature tempering → fine machining → performance testing. During the experiment, the system provides operation guidance through highlighted display areas, text box prompts, audio command broadcasts, and NPC interaction next to the device in the virtual human operation guidance system. Simultaneously, the system transmits the user-input heat treatment route, temperature and time parameters, part type, and material data to the heat treatment performance prediction module based on an improved GRU neural network, receiving the hardness value and tensile test curve output by this module as the experimental results. The intelligent evaluation submodule, based on customized questions generated by the AI ​​dynamic question-setting submodule, assesses user capabilities in two ways: first, a theoretical assessment, testing the user's mastery of the experimental operation process and related knowledge and generating an assessment result; second, a practical assessment, simulating the real experimental process through an immersive learning submodule, recording the user's operation steps without prompts and generating an operation evaluation result. During the assessment process, the examination monitoring submodule is triggered simultaneously to record user behavior. Finally, the assessment results and operation evaluation results are fed back to the learning behavior analysis submodule and the learning statistics submodule for the generation of personalized learning suggestions and teaching quality evaluation. The completion status of pre-experiment exercises in the immersive learning submodule and the number of rounds of experiments completed in the immersive learning submodule are displayed through the learning status in the intelligent teaching assistance module; The heat treatment performance prediction module based on the improved GRU neural network is used to receive the user-selected heat treatment route, temperature and time parameters of each process, part type and material data transmitted by the immersive learning submodule in the multidimensional teaching support module. The module performs feature extraction and time-series relationship analysis on the input data through a pre-trained improved GRU neural network model. Combined with historical experimental data and material performance database, it predicts and outputs the hardness value and tensile test curve of the material after heat treatment. The intelligent teaching and Q&A module serves as a local intelligent assistant to provide teaching support. Users can trigger Q&A interactions by using a virtual human to operate the module's NPC next to the device, or they can access it directly from any interface of the system through a separate entry point. The intelligent teaching and Q&A module integrates a knowledge base in the field of high-frequency quenching and heat treatment processes, allowing users to ask professional questions about experimental operations, process parameters, and equipment usage in text form. The system automatically analyzes the questions and generates answers that include graphic explanations, experimental operation specifications, or experimental data comparisons.

2. The high-frequency quenching virtual simulation teaching system combined with GRU neural network according to claim 1, characterized in that, The knowledge base management submodule uses MongoDB and relational databases for data storage, storing unstructured data such as full-text experimental papers and structured data such as metadata. For retrieval, it integrates the Elasticsearch full-text search engine and Jieba word segmentation technology to build a text index, enabling intelligent retrieval that supports Boolean logic combinations. In terms of personalized recommendations, it calculates the semantic matching degree between keywords highlighting user weaknesses and the provided text using a cosine similarity algorithm. This is combined with data popularity, update time, user history, and user feedback data to establish a multi-dimensional recommendation strategy for generating accurate recommendations. Finally, this submodule interacts with the learning behavior analysis submodule through a data interface, triggering the aforementioned recommendation strategy based on the latter's feedback on weaknesses, and providing users with relevant materials.

3. The high-frequency quenching virtual simulation teaching system combined with GRU neural network according to claim 1, characterized in that, The intelligent teaching and Q&A module uses the PyMuPDF library to achieve localized parsing and text extraction of PDF documents. It uses regular expressions to match and filter special symbols, redundant spaces, and unstructured content, completing text cleaning and sentence segmentation. The Sentence-BERT model is loaded and run locally to convert the cleaned text into 384-dimensional semantic vectors. The generated vector data is saved to the local disk using ChromaDB's file storage mode. An IVFFlat index is built using the FAISS library to achieve efficient retrieval. Based on the spaCy offline model, the module parses the subject-verb-object structure of user questions, extracts core intent keywords, and calculates the matching degree between the question vector and the knowledge block vectors in the database using a cosine similarity algorithm, returning the five knowledge blocks with the highest similarity. The system utilizes the LLaMA-7B quantization model for local inference, concatenating search results with question vectors before inputting them into the model to generate answers. The Jinja2 template engine then uses regular expressions to match the answer content with the cited sources, matching them against the original document chapters of the knowledge block and rendering them in Markdown format. A knowledge graph is constructed using the Neo4j embedded database, and the Stanza library is used for relation extraction and semantic association between entities. Custom terminology replacement rules are combined to achieve domain adaptation. The OpenBLAS library is used to optimize matrix operations, and the Watchdog library monitors local folders for incremental updates, triggering the Dask parallel processing framework to parse, encode, and update the index of newly added documents. Finally, PyInstaller packages the system into a standalone executable file, storing model parameters and data in a local SQLite database, enabling a fully automated process without network dependencies.

4. The high-frequency quenching virtual simulation teaching system combined with GRU neural network according to claim 1, characterized in that, The heat treatment performance prediction module based on the improved GRU neural network adopts a fusion architecture of gated GRU units and self-attention mechanism, and its specific structure is as follows: The heat treatment performance prediction module based on the improved GRU neural network comprises an input layer, a bidirectional GRU layer, a self-attention layer, and a multi-task output layer. The input layer encodes the heat treatment route, temperature parameters, time parameters, part type, and material data into a fixed-length feature vector, with the input dimension determined based on the data preprocessing results. The bidirectional GRU layer consists of two GRU units, each containing 128 neurons, capturing the bidirectional dependencies between parameters through forward and backward propagation. The self-attention layer employs a 4-head attention mechanism to calculate the correlation weight matrix between parameters of different processes, and the output is concatenated with the features of the bidirectional GRU layer before being input into the feedforward neural network. The multi-task output layer outputs hardness prediction and tensile curve parameters through two branches, with the hardness prediction branch containing a fully connected layer of 64 neurons and the tensile curve prediction branch containing a fully connected layer of 32 neurons.

5. The high-frequency quenching virtual simulation teaching system combined with GRU neural network according to claim 1, characterized in that, The training process of the heat treatment performance prediction module based on the GRU neural network includes the following steps: First, historical experimental data covering different materials, part types, and heat treatment process combinations were collected to construct a training dataset containing material, process parameters, hardness values, and tensile curves. After data standardization, the dataset was divided into training and validation sets in an 8:2 ratio. During training, the AdamW optimizer was used with an initial learning rate of 0.001, and a cosine annealing strategy was used to dynamically adjust the learning rate. The network was trained for 3000 epochs, and model parameters were saved every 500 epochs. The loss function included the mean squared error (MSE) of hardness prediction and the dynamic time warping (DTW) loss of curve prediction, while an attention regularization term was introduced to prevent overfitting. During the validation phase, the model performance was evaluated by comparing the similarity between the predicted curve and the true curve and the hardness prediction error.

6. The high-frequency quenching virtual simulation teaching system combined with GRU neural network according to claim 5, characterized in that, The different materials are specifically 40Cr, 20CrMnTi, 35CrMo and No. 45 steel.

7. The high-frequency quenching virtual simulation teaching system combined with GRU neural network according to claim 5, characterized in that, The specific types of parts are cylindrical gears, bevel gears, helical gears, and sprockets.

8. A high-frequency quenching virtual simulation teaching system combining GRU neural networks according to claim 5, characterized in that, The heat treatment performance prediction module based on the improved GRU neural network uses the SHAP value feature importance analysis algorithm during model training to calculate the contribution of each input feature, such as heat treatment route, temperature and time parameters, part type, and material data, to the prediction results such as hardness value and tensile test curve, and displays this feature importance information to the user.

9. The high-frequency quenching virtual simulation teaching system combined with GRU neural network according to claim 1, characterized in that, The immersive learning submodule combines light and shadow effects, spark splash effects, and smoke simulation for joint demonstration during the experimental operation.

Images