Intelligent manufacturing practice teaching system and method based on ai large model

Abstract

This invention provides an intelligent manufacturing practice teaching system and method based on an AI large-scale model, relating to the fields of smart education and intelligent manufacturing training technology. The system includes a capability index generation module, an intelligent manufacturing comprehensive practice platform module, a practice teaching management module, and an intelligent central module. A structured capability index map is generated and dynamically updated through a first AI large-scale model. Intelligent task decomposition, standardized training module encapsulation, and automatic generation of training case processes are achieved through a second AI large-scale model. By constructing a three-layer feedback optimization closed loop, the system drives the automatic, collaborative, and continuous evolution of the capability map, training cases, and teaching strategies, thereby achieving dynamic optimization and precise adaptation of the intelligent manufacturing practice teaching system.

Description

Technical Field

[0001] This invention relates to the fields of smart education and intelligent manufacturing training technology, and in particular to an intelligent manufacturing practice teaching system and method based on an AI large model, which is applicable to comprehensive practical teaching and talent training in intelligent manufacturing, automation and other related majors in higher engineering education. Background Technology

[0002] As the global manufacturing industry undergoes intelligent transformation and upgrading, cultivating intelligent manufacturing talents with interdisciplinary knowledge and innovative capabilities has become a core task of engineering education. Currently, universities and institutions both domestically and internationally have conducted numerous explorations in the development of intelligent manufacturing practical teaching platforms and AI-enabled teaching.

[0003] Abroad, RWTH Aachen University in Germany has collaborated with enterprises to develop an Industry 4.0 practical system based on a streaming data platform, achieving real-time processing of high-frequency data; Pennsylvania State University and other institutions in the United States are using simulators and machine learning tools to cultivate students' decision-making abilities; MIT, Stanford University, and others have experimented with using large models such as GPT-4 as virtual assistants in experimental teaching, providing code debugging and experimental guidance. Domestically, universities such as Zhejiang University and Nanjing Institute of Technology have built hierarchical experimental platforms and curriculum systems around intelligent manufacturing; Beijing University of Technology and Huazhong University of Science and Technology have also begun exploring the introduction of knowledge graphs and large models into curriculum platforms and experimental teaching. However, existing solutions still have significant shortcomings, including: The intelligent manufacturing practical teaching platform lacks the ability to automatically convert natural language project requirements into executable training task chains. When setting up practical projects, teachers usually need to spend a lot of time manually breaking down real industry needs or cutting-edge technology topics into experimental steps that conform to teaching principles. This process is not only inefficient, but also relies on the teacher's personal experience, making it difficult to unify and connect the competency development objectives between different courses.

[0004] The construction and updating of the competency indicator system lacks an automated, data-driven, dynamic calibration mechanism. In existing technologies, the competency map of talent training programs is usually a static document that is periodically revised by an expert committee, which cannot respond to industrial technological changes in real time. When new technologies such as "industrial multimodal large models" and "edge-side real-time inference" emerge rapidly, traditional competency maps cannot incorporate these emerging competency points in a timely manner, resulting in training objectives lagging behind industry needs.

[0005] Practical teaching systems lack a multi-layered, closed-loop optimization mechanism oriented towards teaching effectiveness. Existing teaching platforms mostly operate on a one-way "build-use" model, failing to systematically and data-driven optimize teaching strategies, practical training cases, and even overall competency development goals based on student learning behavior data.

[0006] Therefore, there is an urgent need for a smart manufacturing practical teaching solution that can clearly define the talent training goals for the new era, provide integrated practical training, and possess the ability to continuously evolve itself. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent manufacturing practice teaching system and method based on AI large model, so as to solve one or more technical problems existing in the prior art.

[0008] On one hand, this invention provides an intelligent manufacturing practice teaching system based on an AI large-scale model, including a capability index generation module, an intelligent manufacturing comprehensive practice platform module, a practice teaching management module, and an intelligent central module. The capability index generation module is used to collect multi-source heterogeneous data and extract structured capability triples, and then generate and dynamically update a structured capability index map based on the triples using a first AI large-scale model. The intelligent manufacturing comprehensive practice platform module includes a physical and virtual integrated practical environment, a second AI large-scale model, a training case library, and a case assembly engine. The second AI large-scale model is used to parse the input natural language project requirements into a structured project task tree, and then... Standardized training modules are generated and stored in the training case library; the case assembly engine is used to respond to requests from the intelligent central module, dynamically retrieve and assemble training modules from the training case library to generate a complete training project process; the practical teaching management module is used to collect multimodal practical data of students in the practical environment to construct a student practical data model; the intelligent central module is communicatively connected to the capability index generation module, the intelligent manufacturing comprehensive practice platform module, and the practical teaching management module, respectively, to integrate data from each module, and to make collaborative decisions through multiple dedicated intelligent agents deployed within it, generating personalized teaching strategies, training case optimization instructions, and capability map evolution schemes.

[0009] On the other hand, the present invention provides a smart manufacturing practice teaching method employing the system described in one aspect, the method comprising: S1, Capability Index Map Generation and Update: Through the capability index generation module, based on industry dynamic data, technology development data and internal teaching feedback data, the first AI big model is used to generate and dynamically update the intelligent manufacturing talent capability index map. S2, Training Project Generation and Push: Through the intelligent manufacturing comprehensive practice platform module, based on teaching needs or input natural language project descriptions, the second AI model generates project task trees and standardized training modules, and then dynamically assembles them from the training case library into a complete training project process through the case assembly engine before pushing them to the user terminal. S3, Teaching process data collection and processing: Students practice in a hands-on environment. The practical teaching management module collects multimodal practical data in real time, processes it, constructs structured student practical data based on a multi-layer entity and relationship model of "student-project-task-action", and sends it to the intelligent central module. S4, Ability Assessment and Optimization Trigger: The intelligent central module uses the assessment and profiling intelligent agent to analyze the student's ability achievement and, in conjunction with the global state view output by the context awareness unit, determines whether to trigger the system optimization mechanism. S5, Multi-level Optimized Execution: The meta-controller within the intelligent central module, based on the judgment results, initiates corresponding intelligent agents to perform optimization according to pre-set priorities and quantified judgment criteria. S51 optimizes the inner layer by pushing appropriate tutoring tips or learning resources to student terminals through a personalized teaching strategy intelligent agent, thus enabling personalized intervention. S52 optimizes towards the middle layer by optimizing the agent through practical training cases and generating optimization adjustment instructions for the design parameters of specific practical training cases based on the near-end policy optimization algorithm. S53, optimize outwards, evolve the agent through the capability graph, and when the external technology popularity or industry demand index is detected to exceed the preset threshold, generate revision suggestions for the capability index graph.

[0010] By adopting the above technical solution, the present invention has at least the following beneficial effects: This invention addresses several technical problems in existing intelligent manufacturing practical teaching systems by introducing a large AI model and multiple intelligent agents. Through the first large AI model, the system continuously and automatically mines and identifies emerging competency points from dynamic information sources such as massive industry recruitment data and technical literature, and structurally integrates them into a competency indicator map. Combined with a time-decay weighted algorithm, the system dynamically calculates and adjusts the weights of each competency indicator, ensuring the map reflects industry demand in real time and that talent development goals are always in sync with cutting-edge technology development, thus solving the problem of lagging talent development goals. Through the second large AI model, the system automatically and accurately parses and decomposes project requirements described in natural language into a structured project task tree composed of standardized training modules, replacing the inefficient process of manual decomposition by teachers. This achieves intelligent and standardized transformation from cutting-edge industry topics to executable teaching practices, greatly improving teaching preparation efficiency while ensuring that the competency development goals of practical projects across different courses have a unified internal logic and coherence. By coordinating multiple dedicated intelligent agents through a meta-controller to construct a three-layer continuous optimization mechanism of individual-case-system, the practical teaching system achieves self-adaptation and continuous evolution. This solves the problem of the single-item model of traditional platform construction and use, making the system of this invention an intelligent teaching system with organic collaboration of modules, data-driven, and capable of self-evolution, providing a holistic solution for cultivating intelligent manufacturing talents who can adapt to rapid technological changes. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall system framework provided in an embodiment of the present invention; Figure 2 This is an example of a capability index map generated by the first large AI model provided in an embodiment of the present invention; Figure 3 This is an example of a project task tree generated by the second large AI model provided in an embodiment of the present invention; Figure 4 This is an example of a three-layer continuous evolution closed-loop mechanism driven by a large-model intelligent hub provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the implementation process of a video-based ore particle size analysis training exercise provided in an embodiment of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The present invention will be further explained below with reference to specific embodiments.

[0014] Example 1 like Figure 1 As shown, this embodiment provides an intelligent manufacturing practice teaching system based on an AI large-scale model. The system includes a capability index generation module, an intelligent manufacturing comprehensive practice platform module, a practice teaching management module, and an intelligent hub module. Specifically: the capability index generation module collects multi-source heterogeneous data and extracts structured capability triples, and generates and dynamically updates a structured capability index map based on the triples using a first AI large-scale model; the intelligent manufacturing comprehensive practice platform module includes a physical and virtual integrated practical environment, a second AI large-scale model, a training case library, and a case assembly engine; the second AI large-scale model is used to parse the input natural language project requirements into a structured project task tree, and based on the project... The task tree generates standardized training modules, which are stored in the training case library. The case assembly engine responds to requests from the intelligent central module, dynamically retrieves and assembles training modules from the training case library to generate a complete training project process. The practical teaching management module collects multimodal practical data from students in the practical environment to construct a student practical data model and extract quantitative features. The intelligent central module communicates with the capability index generation module, the intelligent manufacturing integrated practice platform module, and the practical teaching management module, respectively, to integrate data from each module and make collaborative decisions through multiple dedicated intelligent agents to generate personalized teaching strategies, training case optimization instructions, and capability map evolution schemes.

[0015] Furthermore, the capability indicator generation module is the foundation for setting and dynamically calibrating teaching objectives in this invention. Its core task is to construct a structured capability indicator system that can respond to rapid changes in industry and technology and ensure its continuous evolution. Its implementation is a closed-loop process driven by data, empowered by a large model, and involving human-machine collaboration. This module includes a data acquisition and preprocessing unit, and a unit for publishing and updating the first AI large model and capability map.

[0016] Furthermore, the data acquisition and preprocessing unit is used to collect multi-source heterogeneous data and uses a joint training model based on BERT sequence labeling and relation classification to extract capability point-technical requirement-performance index triples from the multi-source heterogeneous data.

[0017] This module first establishes extensive data connection channels, with data sources including: (a) Industry dynamic data: Through API interfaces and web crawlers, it regularly scrapes job descriptions (JDs) related to "intelligent manufacturing," "industrial AI," and "digital twins" from major recruitment websites (such as LinkedIn and Zhaopin.com), as well as authoritative industry reports (such as Deloitte and McKinsey's Manufacturing Outlook), industry white papers, and technical blogs of well-known manufacturing companies (such as Siemens and Huawei). (b) Technology development data: It continuously monitors the latest research paper abstracts and keywords related to industrial large models, multimodal learning, embodied intelligence, and generative design in academic databases such as ArXiv, IEEE Xplore, and CNKI; and synchronously tracks the popularity and updates of relevant open-source projects on GitHub (such as industrial inspection and predictive maintenance models). (c) Internal teaching feedback data: It receives anonymized aggregated data from the practical teaching management module, including the ability achievement scores, high-frequency error points, project innovation scores, and tracking survey data on graduate employment positions and salaries (after anonymization).

[0018] All collected raw data (mostly unstructured text, charts, and code) enters the preprocessing pipeline. First, it undergoes cleaning (deduplication, noise reduction, and encoding correction). Then, a BERT-based pre-trained model is used to extract key entities and relations, resulting in structured "skill point-technical requirement-performance indicator" triples. To ensure the accuracy of triple extraction, a joint training model based on sequence labeling and relation classification is employed. Specifically, BERT is used as the encoder. In its entity recognition (NER) task, the labeled entity types include "skill point" (SKILL), "technical tool" (TOOL), and "performance indicator" (METRIC). In the relation extraction (RE) task, relation categories such as "requirements" (REQUIRES) and "achieved performance" (ACHIEVES) are defined. By constructing training data with (subject, relation, object) triples, a multi-task learning model is trained, enabling it to simultaneously predict entities and their relationships within sentences. This effectively improves the precision and recall of the joint extraction, achieving an F1 score of over 85%, thus ensuring the usability of the generated triples. The triples are stored in graph databases, such as Neo4j, as raw material for subsequent analysis.

[0019] For example, from a job description for a "Senior Industrial Vision Algorithm Engineer," the system might extract entities such as: (SKILL: "Defect Detection"), (TOOL: "PyTorch"), (METRIC: "mAP>95%)," and relations such as: (Defect Detection, REQUIRES, PyTorch), (Defect Detection, ACHIEVES, mAP>95%). From an academic paper, it might extract (SKILL: "Visual Transformer Defect Detection"), (METRIC: "mAP 78.9%)," and (ACHIEVES).

[0020] Furthermore, the first large AI model is used to generate a capability index map using the capability point triples as input. The generation method includes: S101. Clustering and Abstraction: The semantic vector model is used to convert capability triples into semantic vectors, and unsupervised clustering algorithms are used to group semantically similar capability points into capability clusters.

[0021] The first AI model performs deep semantic analysis on tens of thousands of capability point triples stored in a graph database after preprocessing. To achieve deep semantic analysis, the Sentence-BERT model is used to convert the "capability point" text in all triples into high-dimensional semantic vectors. This model is based on BERT and trained through a Siamese network structure, enabling it to map semantically similar sentences to similar positions in the vector space.

[0022] Unsupervised clustering algorithms can be used to automatically group scattered capability points into capability clusters, such as the DBSCAN clustering algorithm combined with semantic vectors. The tool can be implemented using DBSCAN from the Scikit-learn library. Example configuration of key parameters: neighborhood distance threshold set to 0.3 (eps=0.3), minimum number of samples required to form core points set to 5 (…). The distance metric used is cosine distance to measure semantic similarity (metric='cosine'). This method can understand different expressions such as "proficient in TensorFlow", "skilled in PyTorch", and "able to complete projects using Paddle", and group these descriptions with similar distances into the same capability cluster in the semantic vector space, such as the capability cluster of "deep learning framework application".

[0023] S102. Generate a hierarchical architecture: Use a hierarchical text classification model to automatically predict the level to which the capability cluster belongs, and use a fine-tuned BERT text classifier to classify capability clusters within the same level into a preset dimension.

[0024] like Figure 2The example diagram showing the capability map generated by the first AI large model provided in the embodiment of the present invention illustrates the hierarchical and dimensional relationships of the capability index map, with three capability levels from shallow to deep (basic literacy level, professional depth level, and intelligent innovation level), as well as the dimensions contained in each level. A dimension is a set of specific capability index items. For example, the basic literacy level may include dimensions such as engineering mathematics, programming basics, and circuit and control principles.

[0025] The hierarchical structure is primarily achieved using pre-trained hierarchical text classification models, such as HierarchicalAttention Networks (HAN). This model captures multi-layered semantic information from the text. In practice, representative text for each capability cluster (such as the central sentence describing all capability points within the cluster) is input into the model. Based on its trained hierarchical prior knowledge (e.g., basic concepts typically use general vocabulary, while innovative technologies often contain specific terminology), the model automatically predicts the most likely level to which each capability cluster belongs.

[0026] After determining the hierarchical affiliation, a finely tuned BERT text classifier is used to perform secondary classification of capability clusters within the same level, categorizing them into preset dimensions, such as... Figure 2 Applications of industrial vision large-scale models, multimodal data fusion analysis, generative control algorithm design, and digital master line technology.

[0027] For example, the model categorizes "cross-modal data fusion innovation," "generation of process optimization schemes based on large models," and "industrial metaverse conceptual design" into the intelligent innovation layer. Then, within each layer, the model further summarizes these capability clusters under specific dimensions.

[0028] S103. Generation of Capability Indicator Definitions and Descriptions: Based on a text generation model, using the capability indicator name, its level, and its dimension as input, and combining retrieval-enhanced generation technology, a structured indicator definition description containing core knowledge points, typical behavioral manifestations, and related antecedent capability indicators is automatically generated. The specific generation mechanism is as follows: Employing a Transformer-based text generation model (such as T5 or BART), and learning with "zero-shot" or "few-shot" methods, the model uses the name of the capability indicator, its level, and its dimension as input prompts to generate structured descriptive text. The model has learned the standard narrative pattern of "concept-definition-key points-performance" from massive amounts of technical documents and educational materials, thus automatically generating professional descriptions that conform to standards.

[0029] Core knowledge points are obtained through retrieval-enhanced generation techniques (such as RAG): based on the name of the capability indicator, relevant paragraphs are retrieved from the preprocessed technical literature database, and then the generation model extracts and summarizes them.

[0030] Typical behavioral performance and associated prior ability indicators are automatically inferred and generated by analyzing other ability points and relationships (such as "required to master" and "prerequisites") that frequently co-occur with the ability point in the graph database.

[0031] For example, the following structured content can be generated for the capability indicator definition of "intelligent innovation layer - generative control algorithm design - generative conceptual design": Key knowledge points: It is necessary to understand the basic principles of Generative Adversarial Networks (GANs) and diffusion models; Typical performance: Able to use tools such as Stable Diffusion to generate and optimize product appearance sketches based on natural language descriptions; Related leading capability indicators: rely on "professional depth - digital image processing" capabilities.

[0032] The results of the hierarchical and dimensional summarization generated in step S102 and the capability index definitions generated in step S103 will serve as the target for case design in the intelligent manufacturing integrated practice platform module, and also as the benchmark for capability assessment in the practical teaching management module. For example, the specific capability indicators under the "intelligent application" dimension will directly correspond to the practical training case design involving the application of large AI models in the platform.

[0033] S104. Dynamic Weight Calculation: Combining multi-dimensional data, a time-decay weighted algorithm is used to dynamically calculate the weight of each capability indicator in the current period's map. The weight calculation formula includes: , in, Indicators of capability In time The weights; Indicators of capability A collection of related data entries; Represents a set of data entries A data entry in the database. Represents data entries Time; This represents the decay factor, with a value between 0 and 1. For example, 0.95 indicates a 5% decay per month. It is an indicator In data entries The normalized frequency of occurrence in; Represents data entries The significance score is calculated based on a combination of factors, including the authority of the data source and the salary of the relevant job position. Indicates the significance weighting coefficient; This represents the normalized denominator, ensuring that the sum of the weights is 1, and taking the sum of the corresponding weights or coefficients respectively.

[0034] Multi-dimensional data includes the frequency of occurrence of competency indicators, associated salary levels, the age of technical documents, and citation trends in academic papers. The core idea of ​​the time decay weighted algorithm is that recent, highly significant data contributes more to the weight. For example, if "large model hinting engineering" experiences explosive growth in both recent job postings and top conferences, its weight will automatically increase.

[0035] Data entries typically refer to an official news item, an academic journal article, an industry report, or a job posting. The time granularity of data entries is usually calculated on a monthly basis. Data entry sets are selected from relevant data entries according to a preset time period, such as 20 months or 24 months.

[0036] Furthermore, the quantitative calculation methods for data item saliency scores include:

[0037] in, Represents data entries Significance score in the data authority dimension. Represents data entries The weighting of significance scores in the salary-salary correlation dimension. Represents data entries Significance score for the technological innovation dimension; , , These represent the weighting coefficients for the significance scores of each dimension. Preferably, in this example, they are set separately. It equals 0.5. It equals 0.3. It equals 0.2.

[0038] Furthermore, different quantification methods were used for each dimension of the significance score. (Significance score for the data authority dimension) Based on the type of data source, an authority rating mapping table is established for quantitative assignment. For example, Table 1 shows an authority dimension rating mapping table.

[0039] Table 1: Example of Authority Dimension Rating Mapping

[0040] Significance score of salary correlation dimension For recruitment data items that include salary information, the quantification score is based on the quantile position of the salary level within the market data for the same position. For recruitment data items that include salary information, the quantification score is based on the quantile position of the salary level within the market data for the same position. For non-recruitment data items that do not contain salary information, the default value of 0.5 is used, as shown in Table 2.

[0041] Table 2: Example of Salary-Related Dimensional Scoring Mapping

[0042] Significance score of technological innovation dimension The scores are automatically generated by a pre-trained academic innovation assessment model based on the BERT architecture, fine-tuned on tens of thousands of manually annotated "technical text - innovation level" datasets, outputting a continuous value between 0 and 1. The score is based on the text's innovation level. For example, see Table 3.

[0043] Table 3: Example of Scoring Mapping for Innovation Dimension

[0044] Dynamic weights directly influence the decision-making of the intelligent agents evolving the capability graph in the intelligent central module. High-weight emerging indicators will trigger updates to the teaching system first, driving the iteration of courses and capability standards. At the same time, weights also serve as the basis for the allocation of case resources in the intelligent manufacturing comprehensive practice platform module. Training cases corresponding to high-weight indicators will receive more recommendation priority and development resources, ensuring that teaching focus is aligned with industry dynamics and improving the timeliness and effectiveness of talent training.

[0045] The first AI big data model automatically extracts competency points from massive recruitment data, technical literature and teaching feedback, and organizes them into a structured hierarchical indicator system. At the same time, it dynamically adjusts the weight of each competency indicator based on a time decay weighting algorithm to achieve continuous automatic evolution of the competency map.

[0046] Furthermore, the capability graph publishing and updating unit is used to push the capability indicator graph generated by the first AI large model to the review interface for manual review, and publish the reviewed graph in a standardized format.

[0047] The initial version of the capability indicator map generated by the first AI big model will be pushed to a review committee composed of subject matter experts, enterprise engineers, and teaching administrators. The system provides a visual review interface, displaying the relationship network and weight change trends between indicators, and providing data traceability. For example, clicking on the "Digital Mainline Technology Application" indicator will show which industry reports and papers it originates from. The committee can adjust, merge, or supplement the indicators on this interface. The approved map will be released as the official version in a standardized format (such as JSON-LD). This version, as an authoritatively reviewed teaching guide, will be synchronized to the big model intelligent hub and the intelligent manufacturing comprehensive practice platform module, becoming the authoritative benchmark for subsequent intelligent decision-making and case design. The capability map will be continuously and dynamically updated after its establishment. The triggering conditions for the update process include: Scheduled trigger: A full update process will be automatically initiated once per semester.

[0048] Event Trigger: When the number of key papers for a certain technology increases beyond a threshold (e.g., 50%) in a short period of time, or when a new job title (e.g., "Industrial Big Model Prompt Engineer") appears frequently, a local update analysis of the capability cluster related to that technology will be initiated immediately.

[0049] Feedback trigger: When a systemic signal is received from the practical teaching management module indicating that several outstanding students have shown strong interest and innovation in a certain emerging technology project, and that the technology is not fully reflected in the existing map, an update will also be triggered.

[0050] The update process is similar to the initial construction mechanism, but it will focus on analyzing the differences between the old and new versions and automatically generate update logs to ensure the transparency and interpretability of the graph evolution process.

[0051] Furthermore, the first large AI model is prepared using a two-stage method of knowledge distillation and domain fine-tuning, specifically including: S201, using industry and technology documents, trains a general teacher model for understanding industry technologies; S202 utilizes active learning and crowdsourced annotation technologies to acquire large-scale educational data labeled according to a predetermined competency indicator system. The specific process involves: first, education experts define a three-tiered architecture of basic literacy, professional depth, and intelligent innovation, along with preliminary dimensions (such as soft skills and hard skills) for each tier, and label a small number of high-quality samples (approximately 1000). These seed samples are then used to train an initial classifier to predict unlabeled data. Data with low prediction confidence or significant discrepancies (i.e., "difficult samples" where the model is uncertain) are submitted to crowdsourced annotators for labeling. This iterative process yields large-scale labeled data at a relatively low cost. During the annotation process, each data point is labeled not only with its broad competency category (layer) but also with its fine-grained dimensions, ensuring the hierarchical and dimensional nature of the annotation definition.

[0052] S203 uses labeled data (text, hierarchical labels, dimension labels) to supervise the training of the student model, enabling it to initially learn data classification. The student model employs a lightweight base model with a small parameter size (such as LLaMA-7B) and is trained in a supervised manner using the standard cross-entropy loss function.

[0053] S204, The same unlabeled industry technology text is simultaneously input into the teacher model and the student model. By minimizing the KL divergence between the prediction results of the student model and the prediction results of the teacher model, as well as the classification cross-entropy loss of the student model on the labeled data, the student model is jointly optimized to obtain the first AI large model.

[0054] Specifically, the student model is trained using knowledge distillation techniques to mimic the predictive logic of the teacher model, learning to abstract, classify, and categorize competency requirements from complex descriptions. The total loss function used includes: ; ; in, This represents the classification cross-entropy loss of the student model on labeled data; The distillation loss is represented by the KL divergence, which measures the probability distribution of the teacher model output. With student model The difference in output probability distributions prompts the student model to mimic the prediction logic of the teacher model; balancing parameters This is a hyperparameter, with a value between 0 and 1, used to balance the two parts of the loss.

[0055] The performance of the student model trained using distillation loss is evaluated on an independent test set using classification accuracy and macro-F1 score. The goal is to ensure that the final student model achieves a predetermined macro-F1 score and parameter count on the educational domain's ability classification task, resulting in a lightweight domain-specific large model for this module, i.e., the first AI large model. More specifically, the student model is required to have a macro-F1 score of 0.88 or higher, while the number of parameters is required to be 1 / 10 of the teacher model, achieving a balance between lightweight design and high performance.

[0056] Furthermore, the intelligent manufacturing integrated practice platform module serves as the concrete entity for capacity building. Its core function is to use a second large AI model to select practical training cases based on teaching needs, or to intelligently disassemble and dynamically assemble existing case combinations from the training case library, and to update the knowledge points and / or technologies of the training cases according to the map update requirements. Its implementation is not a simple equipment integration, but a flexible training environment based on AI models and data-driven approaches.

[0057] Specifically, the practical environment includes a physics teaching unit, a virtual simulation unit, and simulation middleware for synchronizing and interacting data between the two, supporting three training modes: pure physics, pure virtual, and virtual-real linkage.

[0058] More specifically, the physics teaching platform is a physical model of an industrial system capable of performing industry-standard tasks, possessing high-fidelity mechanisms and full-element digital mapping. Taking the conveyor belt system as an example, on the 1:30 physical model, each motor, roller, and idler is equipped with miniature sensors (vibration, temperature, current) of the same type or principle as those in real industrial applications. The data interfaces, communication protocols (such as Modbus TCP, OPCUA), installation, and calibration procedures of the laser scanner, near-infrared spectrometer, industrial camera, and other detection equipment are completely consistent with those in the industrial field. The control system, composed of a PLC and edge computing gateway (such as one equipped with an NVIDIA Jetson series), also uses industry-standard programming software and configuration environment. Any state changes of the physics teaching unit (such as motor speed, belt misalignment, and material flow rate) are collected and digitized in real time with high precision.

[0059] The virtual simulation unit comprises a digital twin corresponding to the physics teaching unit, built on platforms such as Unity3D or Unreal Engine. It features a 3D visualized digital model with a physics engine and process mechanism model. For example, the digital twin of a belt conveyor incorporates a dynamic model to simulate tension changes under different loads; the digital twin of a silo incorporates a particle flow model to simulate material accumulation and flow.

[0060] The virtual simulation unit and the physics teaching unit interact synchronously in real time through a high-speed network. The real data from the physics teaching unit drives the display and simulation of the virtual unit. Conversely, the virtual simulation unit can run "hypothesis analysis", such as simulating an extreme fault (like a drum jamming), and inject the simulation results back into the control system of the physics unit for safe fault simulation teaching.

[0061] The virtual simulation unit and the physics teaching unit are synchronized and interact through a simulation middleware (such as ROS2 or based on the IEEE HLA standard) to achieve collaborative work. This allows for flexible configuration of teaching scenarios: purely virtual training, purely physical training, and virtual-physical integrated training. Purely virtual training includes rapid algorithm verification and scheme design on digital twins; purely physical training includes hardware wiring, debugging, and troubleshooting on physical devices; virtual-physical integrated training is the most common mode, where students design and verify control strategies in a virtual environment, then deploy them to physical devices with one click, and the results are fed back to the virtual environment for comparative analysis.

[0062] Furthermore, the second AI model includes a semantic understanding subunit, a task decomposition subunit, and a training module encapsulation subunit.

[0063] The semantic understanding subunit is used to parse natural language project requirements into a structured semantic representation that includes core objectives, constraints, and technical fields, using a pre-trained semantic understanding model finely tuned by the engineering domain.

[0064] Sources of input requirements: The project requirement texts (case materials) input into the semantic understanding subunit are usually one of the following three categories: (a) real problem descriptions from partner companies, such as "improving the grinding efficiency of a ball mill in a mineral processing plant"; (b) solutions from cutting-edge technology papers, such as "anomaly detection of industrial time series data based on Transformer"; (c) innovative ideas proposed by teachers or students, such as "designing an AR-based equipment inspection assistance system".

[0065] Model construction and processing: The semantic understanding subunit adopts a Transformer-based bidirectional encoder architecture, preferably the RoBERTa-base model. Three parallel classification heads are added on top of the Encoder to handle different semantic parsing tasks respectively. Core target classification: Using sequence labeling, output the start and end positions of the target phrase in the input text; Constraint extraction: A multi-label classification method is used to determine whether the input text contains predefined constraint types, such as "real-time", "accuracy", and "cost". Technical field identification: A multi-label classification method is used to determine whether the input text belongs to a predefined field category, such as "machine vision", "control system" and "data analysis".

[0066] For example, such as: Input: Design a system to detect surface tears in conveyor belts, requiring real-time alarms and a detection accuracy of no less than 99%. Output (structured semantics): Core objective: "Detect belt surface tear", Constraints: "Real-time", "Accuracy ≥ 99%", Technical fields: "Machine vision", "Embedded systems", "Real-time computing".

[0067] The task decomposition subunit has a built-in pre-built engineering task decomposition knowledge graph, which is used to decompose the structured semantic representation (core objectives, constraints, technical fields) output by the semantic understanding subunit and map it to generate a project task tree composed of basic action nodes. Its core is to use the pre-built engineering task decomposition knowledge graph to perform semantic-graph matching and subgraph retrieval through graph neural networks, and to complete the structured decomposition using a tree sequence generation model.

[0068] More specifically, the pre-built engineering task decomposition knowledge graph, described in the OWL ontology language, contains hundreds of basic engineering actions and their data flow and logical sequence relationships. The graph is stored in the form of triples, defining various basic engineering actions (e.g., action: acquire images, input: camera ID, output: image matrix; action: train classification model, input: labeled dataset, output: model file) and their logical relationships (e.g., "precedes", "requires").

[0069] The specific steps for generating the project task tree include: First, using a Graph Neural Network (GNN), the structured semantic representation of the project requirements is matched and reasoned against a pre-built knowledge graph of engineering task decomposition. This retrieves a series of integrated engineering action nodes related to the project requirements and the relationships between these nodes, forming a candidate basic action node subgraph. Then, based on the project requirement semantics and the candidate basic action node subgraph, a tree-based sequence generation model (such as Tree-LSTM or Graph Transformer) is used to generate the structured project task tree.

[0070] Although the relationships between the basic action nodes retrieved by the graph neural network are related, they are usually loose and not yet structured according to the task flow required by the project. The sequence generation model combines the semantics of the project requirements (vector representations of objectives, constraints, etc.) with the vector representations of the knowledge graph. Through a multi-head attention mechanism, it gradually generates tree nodes (actions) and establishes parent-child / sibling relationships, decomposing and mapping the semantics of the project requirements into a directed acyclic graph composed of basic action nodes, i.e., a structured project task tree.

[0071] Furthermore, the task decomposition subunit adopts a Transformer-based encoder-decoder architecture, with the decoder employing a tree-structured attention mechanism. Specifically, the encoder uses RoBERTa-large (pre-trained with domain adaptation) to encode the input requirements; the decoder, based on a Tree-LSTM structure, generates a tree node (basic action) and its parent-child relationships with existing nodes at each time step. Internally, the decoder maintains a node stack, simultaneously monitoring the encoder output and the states of generated parent nodes through an attention mechanism. Tree structure constraints are implemented through a masking mechanism to ensure that the generated graph is a directed acyclic graph.

[0072] like Figure 3As shown, taking the "Belt Tear Detection System" as an example, a schematic diagram of a project task tree is given. The root node of the project task tree is "Complete the tear detection system". The first-level sub-nodes include "1. Deployment of image acquisition subsystem", "2. Development of tear detection algorithm", and "3. Implementation of real-time alarm logic". Each sub-node can be further decomposed.

[0073] The training module encapsulation subunit is used to encapsulate each basic action node on the project task tree, generating a standardized training module that includes teaching objectives, prerequisite knowledge, input and output specifications, implementation paths, and evaluation criteria.

[0074] Specifically, based on the type of the basic action node, the node's position in the project task tree, and the parent and child nodes associated with the context, a training module containing complete teaching metadata corresponding to the node is automatically generated through a retrieval and filling mechanism.

[0075] by Figure 3 The example of the node "2.2 Model Selection and Training (YOLO / UNet)" in the project task tree illustrates the pedagogical encapsulation of the project task tree node, as shown in Table 4.

[0076] Table 4: Example of instructional encapsulation of project task tree nodes

[0077] The packaged standardized training modules are stored in the training case library, along with the aforementioned structured teaching metadata.

[0078] More specifically, the training module encapsulation subunit adopts a retrieval augmented generation (RAG) architecture, which includes two core components: a retrieval unit and a generator. The retrieval unit is based on a dense vector retrieval model using Sentence-BERT, which encodes basic action node descriptions (such as "model selection and training (YOLO / UNet)") into vectors and retrieves Top-K similar historical encapsulation examples from a pre-built teaching description library. The generator is based on a BART-large text generation model, which takes as input the concatenation of basic action node features (type, position, context parent / child node name) and the retrieved K examples, and outputs structured teaching metadata (JSON format).

[0079] The second AI model automatically deconstructs the project requirements described in natural language into standardized, independently decomposable and recombinable training modules, and ensures that the decomposition results are precisely aligned with the preset competency development goals.

[0080] Furthermore, the construction of the second large AI model adopts a multi-stage joint training strategy. The overall architecture is based on an Encoder-Decoder framework, and a tree-structured decoder is introduced at the Decoder end to generate a hierarchical task tree. The specific construction and training steps are as follows: S401 uses large-scale general-domain code-text pair data (such as CodeSearchNet) to pre-train the base model, enabling the model to acquire basic code understanding and generation capabilities. Preferably, the base model adopts the CodeT5 or CodeBERT architecture with approximately 220 million parameters.

[0081] S402 uses text data from industrial automation, machine vision, and intelligent manufacturing to perform domain-adaptive pre-training; it collects approximately 500GB of text data, including open-source project code, technical documents, and experimental manuals from industrial automation, machine vision, and intelligent manufacturing, and uses Masked Language Modeling (MLM) as the objective to continue pre-training, so that the model becomes familiar with engineering domain terminology and task patterns.

[0082] S403 uses paired data of "project requirement text - structured semantic labels" to perform supervised fine-tuning of the semantic understanding subunit of the model; approximately 100,000 pieces of labeled data of "project requirement text - structured semantic labels" are constructed to perform supervised fine-tuning of the Encoder part of the pre-trained model, enabling it to accurately extract core objectives, constraints, and technical fields. The loss function is multi-label classification cross-entropy loss.

[0083] S404 uses paired "project requirement-project task tree" data and trains the model's task decomposition ability through teacher-mandated training and reinforcement learning fine-tuning based on a teaching simulation environment. It constructs a paired dataset of approximately 50,000 "project requirement-project task tree" data points, with each tree stored in JSON format. A two-stage training method is employed. Phase 1: The model is trained using the Teacher Forcing method to generate tree structures. The loss function is a weighted sum of the cross-entropy loss of tree node prediction and the binary cross-entropy loss of tree structure relationship prediction.

[0084] The second stage involves introducing reinforcement learning fine-tuning, evaluating the task tree generated by the model in a simulated teaching environment, using the "expected improvement in students' abilities on the task tree" as a reward signal, and optimizing the model parameters using the policy gradient method to make the generated task tree more in line with teaching principles.

[0085] S405 uses paired data of "basic action nodes - instructional module descriptions" and combines retrieval-enhanced generation technology to train the model's ability to encapsulate training modules. It constructs approximately 200,000 paired data entries of "basic action nodes - instructional module descriptions" and employs the Retrieval-Enhanced Generation (RAG) framework. Specifically, a vector database containing thousands of high-quality instructional descriptions is pre-built. For each basic action node, its similarity to existing descriptions in the database is first calculated using Sentence-BERT. The top-3 similar descriptions are retrieved as prompts, and then input into the fine-tuned generative model to generate the final encapsulated result. The generative model adopts the BART architecture and undergoes supervised fine-tuning on approximately 50,000 data entries of "node features + retrieval results → encapsulated descriptions".

[0086] In addition, the second AI model also uses a large number of historical successful project cases (requirement documents - task decomposition trees - module descriptions) as training data to train the model's ability to generate task trees from requirements in a supervised learning manner, and fine-tunes it through reinforcement learning so that the generated task trees can obtain better feedback on students' ability improvement in subsequent simulated teaching assessments.

[0087] Furthermore, the training case library is used to store standardized training modules generated by the second AI model. This training case library is not a static, pre-defined list by the teacher, but a dynamically generated case factory. Its core is to use the second AI model to deconstruct complex engineering tasks, perform pedagogical transformations, and dynamically generate standardized training modules in batches, along with logical relationship graphs describing the dependencies between modules.

[0088] Furthermore, the training case library also includes teaching resource data for personalized instruction and metadata and parameters for the case optimization agent. The teaching resource data includes basic guidance content and tiered support resources. Basic guidance content includes training content, requirements, and steps. Tiered support resources include prompts, code libraries, charts and documents, and video tutorials, providing personalized guidance for students of different ability levels and training cases of varying difficulty. Metadata and parameters are crucial for dynamic case assembly and adaptive adjustment. They store the control parameters for each training module or case, including design parameter vectors and performance indicators and labels. The design parameter vectors include difficulty coefficients, resource support, and the amount of prompts. The case assembly engine adjusts these parameters based on student profiles (e.g., increasing resource support and prompts for struggling students), while the training case optimization agent optimizes these parameters through reinforcement learning. Performance indicators and labels include target ability indicators and historical performance data. The target ability indicators for the training case are stored, while the historical performance data collection stores the average time spent by students completing the case, completion rate, and innovation score, which are used by the training case optimization agent to calculate the reward function.

[0089] Furthermore, the case assembly engine is used to retrieve matching training modules from the training case library based on personalized teaching requests from the intelligent central module, and dynamically assemble them into a complete training project process according to logical relationships. The assembly can adaptively adjust the difficulty coefficient, resource support, or prompt information of the assembled cases according to the student's ability profile.

[0090] More specifically, the case assembly engine is an application interface or functional module of the second AI large model. It calls the sequence generation model of the second AI large model and retrieves and sorts cases from the case library according to the input conditions.

[0091] The intelligent hub module can send requests to this module, such as when the assessment and profiling agent or the personalized teaching strategy agent in the intelligent hub module needs to generate / adjust practical training tasks for students or groups. The request includes a list of target ability indicators and the current ability profile of the student from the practical teaching management module.

[0092] Furthermore, the case assembly engine can adaptively adjust based on students' ability assessment profiles: for students with weak foundations, the assembled cases may include more detailed step-by-step prompts and more reference code; for students with strong abilities, cases may omit intermediate steps or add open-ended challenge tasks (such as "Please try to achieve detection with the same accuracy using no more than three sensors"). This assembly is real-time and dynamic, ensuring the personalization and tiered nature of practical projects.

[0093] Furthermore, the intelligent manufacturing integrated practice platform module also includes a scheduling unit for managing the allocation and conflict resolution of physical equipment and virtual resources in the practical environment. When multiple student groups simultaneously apply to use physical equipment (such as robotic arms), the system will intelligently schedule based on project priority, appointment time, and equipment status, or suggest using virtual equipment for parallel development. The runtime environment (software dependencies, toolchain) of all training cases can be containerized (Docker), enabling one-click deployment and environment isolation, ensuring the stability and reproducibility of the teaching process.

[0094] Furthermore, the practical teaching management module includes a multimodal data acquisition layer, a data fusion and standardization processing layer, and a real-time feedback and intervention interface layer. Its core purpose is to collect and structure multi-dimensional, full-process, and fine-grained data on students' practical activities, providing high-quality raw materials for subsequent analysis.

[0095] The multimodal data acquisition layer is used to collect video data of students' operational behaviors, digital operation log data, and interaction and achievement data during practical activities. The multimodal data acquisition layer consists of a series of hardware and software acquisition points, ensuring that data is seamlessly and naturally integrated into the teaching process. Specifically, the collected data includes: Operational Behavior Acquisition: An intelligent vision system with edge computing capabilities is deployed in the physical experimental area to analyze student operational behaviors in real time using pre-trained computer vision models (such as YOLO and OpenPose). For example: it identifies whether students wear insulated gloves before operating the PLC (safety regulations); identifies the contact positions of multimeter probes and compares them with standard circuit diagrams to automatically determine wiring correctness; it records the frequency and duration of students switching from teach pendant operation to code debugging when debugging the robotic arm (operational habits). These behavioral video streams are processed in real time at the edge, and only structured behavioral events and keyframe images are uploaded. Structured behavioral events are recorded using text, such as: "Student A, at 10:05, completed the wiring of the PLC power module, taking 2 minutes and 13 seconds, and was visually verified as correct." Keyframe images are image data for the corresponding time period.

[0096] Digital operation log collection: Deep integration with all development and simulation tools. Install a custom plugin in programming IDEs (such as VSCode) to anonymously record students' code editing process: frequency of code completion usage, number of debugging attempts, error types (syntax errors, logic errors), and records of online documentation searches. In virtual simulation software, record every parameter modification, simulation run, and result viewing by the student. In project management tools (such as GitLab), fully record code commit history, branch merging, and issue discussion content to assess team collaboration and project management capabilities.

[0097] Interaction and Outcome Data Collection: All interactions within the teaching platform are recorded: the quality of students' questions and answers in the forum, the content of questions asked to the intelligent tutor (see Module 104) and the satisfaction with the feedback, and the answers and time taken in online quizzes. Final deliverables submitted by students, such as lab reports (text), design drawings (vector graphics), demonstration videos, and defense slides, are all collected by the system.

[0098] The raw data collected by the data acquisition layer comes in various formats, including time series, images, text, and event logs. The data fusion and standardization processing layer is used to associate and align the collected multimodal data based on a unified student practice data model, and to aggregate low-level operation events into high-level actions with pedagogical significance through a complex event processing engine, while extracting key features from the raw data.

[0099] Furthermore, the construction of student practice data models includes: S301 defines a multi-level entity and relationship model of "student-project-task-action" based on a graph database (such as Neo4j) to achieve data hierarchical alignment and association. The entity and relationship model definition includes: Entity type definitions: Student, Project, Task, Action.

[0100] Relation type definition: PARTICIPATES_IN(Student->Project) indicates that a student is participating in a project; CONTAINS(Project->Task) indicates the tasks contained in the project; COMPOSED_OF(Task->Action) indicates that the task consists of actions; PERFORMS (Student->Action) indicates that the student performs an action.

[0101] S302, the collected multimodal practice data is associated and aligned with the multi-layer entity and relationship model to form a complete data traceability chain.

[0102] Specifically, using student login ID, project ID, task ID, and timestamps accurate to milliseconds, all collected data (operation logs, code commits, video events) are used as attributes or relationships and attached to the corresponding Action nodes, or new Action nodes are created directly. For example, in the "Ore Particle Size Classification Project Based on Video Images" in Example 3, student Zhang San (Student node) participates in project P001 (Project node), which includes the "Algorithm Debugging" task T01 (Task node). When Zhang San executes the "Set Breakpoint" operation at 10:05:30.500, the system creates an Action node A001 of type "Debug_SetBreakpoint" and establishes the relationship (Zhang San)-[:PERFORMS]->(A001), and (T01)-[:COMPOSED_OF]->(A001), recording attributes such as timestamp, code file, and line number on the A001 node.

[0103] S303 uses semantic aggregation to aggregate low-level operational events into high-level actions with pedagogical significance.

[0104] Specifically, a rule engine (such as Drools) and a simple event processing engine (CEP) are used to aggregate low-level operation events (such as "clicked the 'Run' button" or "submitted a commit") into high-level instructional actions. The aggregation steps include: 1) Defining an event pattern, for example: "After the Debug_SetBreakpoint event, at least three consecutive Debug_StepOver events occur within 60 seconds, ultimately followed by a Code_Modification event." 2) When the CEP engine detects an event sequence matching this pattern, it triggers the rule, creating a new high-order Action node of type "Iterative Algorithm Debugging," and associating the original low-level event node as its child node; for example, a series of IDE events ("wrote a for loop," "set a breakpoint," "stepped through 5 times," "modified the loop condition") that match the preset pattern can be aggregated into an "Iterative Algorithm Debugging" action.

[0105] Furthermore, quantitative feature vectors are extracted from the raw data and aggregated teaching actions from multiple dimensions and stored as node attributes in the graph database. The dimensions of feature extraction include three aspects: code quality, operational behavior, and text interaction, specifically as follows: Code quality features: The cyclomatic complexity of the code was calculated using the radon library (Python); the number of lines of code was calculated using the cloc tool; and the semantic embedding vectors of the code were generated using a pre-trained CodeBERT model.

[0106] Operational behavior characteristics: From the video stream, the proportion of time the student's face is turned toward the device is calculated using OpenCV and the dlib library (attention level); the variance of the time interval between two consecutive valid operations is calculated (operational fluency).

[0107] Text interaction features: Sentiment analysis of forum posts was performed using TextBlob; keywords in the question text were extracted using TF-IDF and matched with a knowledge base to assess the depth of the question.

[0108] These features are quantized into numerical vectors and stored in the graph database as attributes of the corresponding Action or Student nodes.

[0109] The real-time feedback and intervention interface layer provides real-time student teaching practice data to the intelligent central module, while also offering low-latency local feedback. The processed structured data flows in real-time to the large-scale intelligent central module for decision-making, and also provides low-latency local feedback. For example, when the vision system detects a student performing a high-risk operation (such as operating on live electricity), this module can immediately control the device to cut off power and display a safety warning on the screen. When a student is detected stuck on the same programming error for a certain period, a prompt can be displayed in the IDE, suggesting the student review a knowledge card or request assistance from the intelligent tutor. This layer achieves on-site intelligentization of the teaching process.

[0110] Furthermore, the intelligent central module includes a unified data lake and context-aware unit, a multi-agent decision-making and optimization execution unit, and a meta-controller. As the system's "intelligent brain" and "evolutionary engine," the core of this module lies in constructing a three-layered nested, continuously optimized closed loop based on reinforcement learning and multi-agent collaboration. This integrates three different time scales and decision-making levels—individual teaching intervention (inner layer), group case optimization (middle layer), and overall knowledge system evolution (outer layer)—into a unified, data-driven, and self-iteratory intelligent framework. The meta-controller then coordinates globally based on the system's global state view, triggering and executing the optimization of each agent. For example... Figure 4 The diagram shows the three-layer continuous evolution closed-loop mechanism of the intelligent central driving module.

[0111] The unified data lake and context-aware unit are used to fuse multi-source data and output a global status view representing teaching load, learning outcomes, and alignment of internal and external trends. Interface data between the unified data lake and other modules includes capability indicator maps from the capability indicator generation module, platform status and case library metadata from the intelligent manufacturing integrated practice platform module, teaching process data from the practical teaching management module, and external data sources. The following provides an illustrative explanation of the definitions of each interface data point: Capability Indicator Map: Transmitted in JSON-LD format, containing fields such as {“indicator_id”: “I001”, “name”: “Industrial Vision Algorithm”, “weight”: 0.85, “hierachy”: “Specialty Depth”, ...}.

[0112] Platform status and case library metadata: transmitted in structured messages (such as Apache Avro format), containing information such as {"device_id":"RobotArm01","status":"busy", "next_available":"2023-10-27 14:30:00"} and {"case_id":"C203", "invoke_count":150, "avg_completion_rate": 0.72}.

[0113] Teaching process data: Accessed in the form of event streams (such as Apache Kafka Topic), with the event format being {"student_id": "S1001", "timestamp": "...", "action_type": "code_submit", "project_id": "P001", "data": {...}}.

[0114] External Data Sources: The intelligent hub module directly receives external industry and technology dynamic summary data streams (such as those via RSS subscriptions and API callbacks) after preliminary filtering and deduplication. Unlike the raw crawled data used by the capability indicator generation module, the external data used by this module focuses more on real-time alerts of macro trends and hot events (such as "Gartner releases 10 strategic technology trends for 2024" and "A major company open-sources an industrial vision basic model"), for rapid situation assessment at the context awareness layer, rather than for in-depth mining of capability indicators.

[0115] The context-aware agent is essentially a multimodal fusion model for time-series prediction and anomaly detection, used to continuously fuse and analyze multi-source heterogeneous data fed into a unified data lake. Specifically, a Transformer-based encoder maps different types of data (structured metrics, text summaries, time-series event counts) to a unified vector space, and utilizes a self-attention mechanism to analyze the correlations between different data sources, periodically outputting a quantized global state vector to construct a global view of the system's current state. The global state vector includes: a) Teaching load status, used to indicate the usage of platform resources, with quantitative indicators including the number of concurrent active projects and the average resource utilization rate; b) Learning performance status, used to indicate the achievement of teaching objectives and knowledge mastery. Quantitative indicators include the average achievement of students' abilities in each ability dimension and the deviation between the current average score and the historical average score of each practical training case.

[0116] c) Internal and external trend alignment, used to represent the degree of matching between teaching content and the development of cutting-edge industry technologies. Quantitative indicators include the correlation coefficient between high-weight indicators in the capability map and external real-time technology hot search terms.

[0117] The multi-agent decision-making and optimization execution unit includes an assessment and profiling agent for evaluating students' abilities and generating ability profiles, a personalized teaching strategy agent for generating personalized teaching intervention strategies, a practical training case optimization agent for optimizing practical training case design parameters, and an ability graph evolution agent for analyzing and generating ability graph update proposals. Each agent shares the underlying computing framework and builds and completes its own tasks based on its specific objective function and using different AI model bases.

[0118] Each agent executes its corresponding optimization strategy in an orderly manner according to the instructions and commands of the meta-controller. Among them, the evaluation and profiling agent is used to evaluate students' knowledge mastery, skill proficiency, and innovation ability based on the student practice data model, and generate student ability profiles; the personalized teaching strategy agent is a policy network based on deep reinforcement learning, used to generate and execute personalized teaching intervention strategies based on the student ability profiles; the practical training case optimization agent is an optimization engine based on reinforcement learning, used to analyze the group teaching effect data of practical training cases, and generate optimization adjustment instructions for case design parameters; the ability graph evolution agent is used to monitor internal and external trend differences, analyze the update needs of the ability indicator system, and generate graph update proposals with impact assessment.

[0119] Furthermore, the evaluation and profiling agent subscribes to the standardized "Student Practice Data Model" event stream output by the practical teaching management module to obtain in real time all Action nodes and their attributes (feature vectors) for individual students or groups in a specific project. For example, when a student completes the "model training" task and submits the results, the practical teaching management module generates an event containing attributes such as task ID, code quality score, model performance index (mAP), and time consumption. The evaluation agent receives this event. The agent extracts the input features required by each ability evaluation model from the event stream.

[0120] The assessment and profiling agent includes multiple ability assessment models, including a knowledge mastery assessment model, a skill proficiency assessment model, and an innovation ability assessment model, which are used to conduct multi-dimensional assessments of students' abilities based on the received event streams.

[0121] The knowledge mastery assessment model adopts the gradient boosting decision tree (GBDT, such as XGBoost) model. The input features include the test history score sequence of each knowledge point, the frequency and complexity of calling relevant APIs / libraries in the code, and the output is the probability distribution of the student's mastery level of each knowledge point. Configuration example: n_estimators=100, max_depth=6.

[0122] The skill proficiency assessment model uses a multilayer perceptron (MLP) neural network. The input features include task completion time (normalized), operation fluency score, number of debugging loops, and first-time debugging success rate. The output is a comprehensive proficiency score between 0 and 1.

[0123] The innovation capability assessment model combines semantic similarity calculation and rule-based reasoning. First, Sentence-BERT is used to vectorize student project proposals, and cosine similarity is calculated between these vectors and all historical proposal vectors in the case library to obtain a "proposal uniqueness" score (1 - highest similarity). Second, the breadth of the technology stack explicitly mentioned in the proposal is analyzed (e.g., simultaneous involvement of vision, control, and data analysis). Finally, a small text generation assessment model scores the "Innovation Summary" section of the project report (0-1 points), and the weighted average of these three scores yields the innovation capability score.

[0124] The assessment and profiling agent integrates the outputs of the above ability assessment models to generate student ability profile vectors. And compared with the target vector for this stage obtained from the capability index generation module. Compare and generate a difference vector. .

[0125] The personalized teaching strategy agent is used to perform inner-layer individual teaching strategy optimization. This agent is implemented based on a deep reinforcement learning (DRL) conditional policy network. The operation and training method of this agent includes: When the agent is triggered, its input state A fusion vector containing the following information: student ability gap vector The current error context's text embedding vector, the student's historical intervention records, and the effect feedback vector.

[0126] The agent's action space A set of predefined executable intervention strategies, such as the action space. Including but not limited to: {Push video_V001, Push code snippet_C005, Initiate guided dialogue_Q, Suggest switching task_T010}.

[0127] The conditional policy network Receive the input status and output the action space. The agent selects a probability distribution for each action and then selects and executes a specific intervention action based on this probability distribution. This agent is trained using the Actor-Critic reinforcement learning framework, where the Critic network is used to evaluate the response in a given state. Take a certain action The long-term expected value (Q value) is used as the reward signal, based on the student's subsequent short-term (such as the next task) improvement in ability. Through continuous interaction and learning with the environment (students), the agent learns to select the most effective intervention strategy for students with different characteristics under specific difficulties.

[0128] The training case optimization agent is used to execute the mid-level group case optimization strategy. Through the reinforcement learning engine, it continuously optimizes individual cases in the training case library during the teaching process, such as updating knowledge points or adding new technologies, in order to improve the effectiveness of the training cases and achieve the teaching objectives.

[0129] Furthermore, the training case optimization agent explores optimal combinations of training case design parameters in a virtual simulation environment using a proximal policy optimization algorithm. Taking the training case "Ore Granularity Analysis Based on Video Images" (see Example 3) as an example, when the agent detects that three consecutive cohorts of students have exceeded the average time limit and achieved a low completion rate on the "Model Parameter Tuning" subtask of this training case, the meta-controller triggers the training case optimization agent to call the reinforcement learning engine to optimize the training case module. Optimization methods include: S501, Define State The design parameter vector for the training case includes at least the difficulty coefficient, resource support, and amount of prompting information of the training case.

[0130] For example, ( , and (Values ​​are all in the range of 0-1) This indicates the difficulty level of the practical training case; a higher value indicates greater difficulty. This indicates resource support; a higher value indicates higher resource support. Indicates the amount of information provided. The larger the value, the more information needs to be provided. When students exceed the time limit for a task and the completion rate is low, these three parameters are dynamically adjusted, such as reducing the difficulty level, increasing resource support, and increasing the amount of information provided, so that more students can complete the task within the expected time.

[0131] S502, Define Action This refers to the adjustment amount of the design parameter vector; such as... This means that while slightly reducing the difficulty level, increasing resource support and prompts; This indicates the adjustment amount for the difficulty level of the practical training case. This represents the change in resource support. This indicates the amount of adjustment to the prompt message.

[0132] S503, In a virtual simulation environment, explore different actions and calculate a reward function based on the real student usage data in the next round after applying the action, and update its policy network. The reward function includes at least a weighted sum of changes in case completion rate, changes in average completion time, and changes in innovation score. Define reward function After designing the modified training case for application in a virtual student group (simulator), the following calculations are performed based on the performance data from the next round of real student usage: ,in, The change in case completion rate represents the change in the proportion of students who completed the case before and after optimization. This indicator reflects the feasibility of the case; a high completion rate means that the case is of moderate difficulty and has sufficient resources. The change in average completion time represents the change in the average time students spend completing the case before and after optimization; negative values ​​indicate this. (Reduced time) is a positive reward, reflecting the improved teaching efficiency of the case study; The average change in innovation score represents the change in the innovation score demonstrated by students in the case before and after optimization. The innovation score is given by the innovation ability assessment model. This indicator reflects the inspiration of the case. The optimized case should be able to inspire students to make more innovative attempts. , , They represent , and The weighting coefficients are dynamically adjusted according to the priority of the teaching objectives.

[0133] S504. After the strategy network converges, a structured optimization instruction containing updated design parameters and resources to be added is generated and sent to the intelligent manufacturing integrated practice platform module for execution.

[0134] The reinforcement learning engine employs the Proximal Policy Optimization (PPO) algorithm in its learning process, where the agent explores various policies in a simulated environment. The agent updates its policy network by maximizing the cumulative reward. Once the policy network converges or finds a better combination of parameters, the agent generates a structured optimization instruction, such as the instruction {"case_id":"C101","action":"update_params","new_params": {"h": 0.4},"add_resource": "tutorial_V002.mp4"}, which means increasing the information prompt from 0.3 to 0.4 and sending the corresponding learning video tutorial_V002.mp4. The instruction is sent to the case management interface of the intelligent manufacturing integrated practice platform module through an internal message queue.

[0135] The capability graph evolution agent is used to execute optimization strategies for the evolution of the outer-layer overall knowledge system, ensuring that talent development programs are synchronized with rapidly changing technological and industrial needs. This agent is activated periodically by the context-aware agent or when significant technological or industrial changes are detected. Once activated, the core processes executed by this agent include: S601, based on the technology of trigger judgment, performs in-depth analysis of external data sources and constructs a corresponding future demand map; S602 compares the constructed future demand map with the current system capability map and identifies the capability indicators that differ between the two. S603. For each competency indicator to be added or adjusted, the graph propagation and impact analysis algorithm is used to simulate and evaluate the impact of the competency indicator on the existing teaching system, and a quantitative impact assessment report is generated. S604. Based on the above analysis, a structured map update proposal containing an impact assessment is generated and submitted to the Teaching Management Committee for final decision-making. S605. After the proposal to update the capability map is approved, the updated capability map will be automatically synchronized to the capability indicator generation module, and the training case library of the intelligent manufacturing integrated platform module will be triggered to make corresponding supplementary or adjustment suggestions.

[0136] Furthermore, the intelligent agent identifies major technological or industrial shifts by monitoring specific indicators derived from external data analysis. These indicators include: Technology popularity metrics: the frequency growth rate of specific keywords (such as "industrial multimodal large model") in external data sources (papers, technology news), and the growth trend of the number of stars in related repositories on GitHub.

[0137] Industry demand indicators: the month-on-month changes in the number of relevant job postings and median salaries on recruitment websites. When the growth rate exceeds a preset threshold (e.g., a month-on-month increase of more than 50%) and continues for a certain period, it is considered a significant change.

[0138] Furthermore, the specific steps for simulation evaluation using graph propagation and impact analysis algorithms include: S701 maps the existing curriculum system and practical training cases into a directed acyclic dependency network graph (DAG). S702, Analyze all theoretical prerequisite knowledge nodes (dependency edges) of the proposed new capability indicator X, and locate these nodes in the dependency network graph; S703, simulate adding node X and its dependent edges with the previous node in the graph; S704 uses graph algorithms (such as a variant of PageRank) to calculate the impact of a new node on the critical path of all nodes in the network (i.e., the entire teaching plan); S705 estimates the potential additions of prerequisite knowledge modules, the number of courses that need to be adjusted, the additional class hours, and the number of supporting practical training cases that need to be developed, and forms a quantitative impact assessment report.

[0139] The meta-controller is used to coordinate the triggering and execution order of the multiple agents based on the system global state view output by the context-aware agent, and to manage the experience data pool for continuous training and fine-tuning of each agent.

[0140] The meta-controller coordinates the orderly execution of optimization strategies by each agent based on the global state view. It then tracks the optimization performance using data collected from the practical teaching module, forming a "state-action-reward-result" quadruple record data pair stored in the experience data pool. This experience data pool is used for continuous training and fine-tuning of each agent, especially the reinforcement learning engine (the meta-controller's lightweight policy network), forming a continuous optimization loop. This allows the overall decision-making capability of the large model's intelligent central module to continuously improve with system runtime, achieving closed-loop iterative optimization at the system level, evolving from individualized tutoring to group cases, and finally to a macro-level knowledge system. Specific implementation methods include: First, the priorities for optimization execution are pre-defined. Typically, personalized interventions at the inner layer are triggered first, resulting in the shortest feedback cycle. If the problem becomes widespread (group-oriented), optimization of the middle-layer training cases is initiated. If structural or trend mismatches are found, the evolution of the outer-layer optimization map is initiated. Therefore, the priorities are set from highest to lowest based on response speed: inner layer, middle layer, outer layer; and from highest to lowest based on the scope of impact: outer layer, middle layer, inner layer.

[0141] Secondly, pre-defined quantitative judgment criteria are established. The quantitative criterion for an individual problem to escalate into a general (group) problem is: when the gap vector output by the ability assessment agent... In the context of the learning process, if the proportion of students whose negative gap (target - current) in a certain dimension (such as "model parameter tuning") exceeds a threshold (e.g., 0.3) is greater than 30%, it is considered a common problem. The quantification criteria for common problems to escalate into structural and trend problems are provided by the context-aware agent. When the "internal and external trend alignment" in the global state vector shows that the popularity of a certain external technology trend is consistently higher than the corresponding indicator weight in the capability map for more than two update cycles, and the relevant capability dimension in the teaching effectiveness trend is stable or declining, it is considered a structural trend mismatch.

[0142] The meta-controller adopts a hybrid architecture that combines rule-based decision trees with lightweight policy networks. It receives a global state view and, upon detecting the aforementioned quantified judgment signals, schedules the corresponding agents to execute based on preset priorities and resource usage, and manages their execution order to avoid conflicts.

[0143] The experience data pool is a centralized experience replay buffer maintained by the system. After each agent makes a decision, it generates a standardized experience record, which is a four-tuple record. ,in, This indicates the state of the agent before the decision is executed. This refers to the specific actions or instructions performed by the intelligent agent; This represents the reward value, which is calculated by the system in subsequent periods based on the actual effects of the decision (such as the improvement in students' abilities and changes in task completion rates). It represents the new state after the agent performs an action; the generated experience records are uniformly stored in the experience data pool to form a shared knowledge base for learning across agents.

[0144] The system has an offline training pipeline that periodically (e.g., weekly) samples batches of data from the experience pool for continuous training and fine-tuning of each agent. More specifically, training steps using algorithms such as PPO or DQN are used to update the policy network and value network parameters of the personalized teaching strategy agents and training cases; new student performance data is used as annotations to incrementally train the GBDT and MLP evaluation models to maintain their predictive accuracy.

[0145] Update meta-controller strategy: Based on the success / failure results of historical scheduling decisions (based on the final system goal such as the overall capability improvement rate), fine-tune and update the decision rules or network parameters of the meta-controller through imitation learning or reinforcement learning.

[0146] In summary, this invention, through the deep integration and intelligent collaboration of the above four modules, constructs an intelligent manufacturing practice teaching ecosystem with dynamic and precise objectives, flexible content system, panoramic controllable process, and self-evolutionary capabilities.

[0147] Example 2 This embodiment provides a smart manufacturing practice teaching method using the system described in Embodiment 1. The method explains its application in teaching practice from a system perspective, involving three roles: student, teacher, and system administrator. Specific steps include: S1, Capability Index Map Generation and Update: Through the capability index generation module, based on industry dynamic data, technology development data, and internal teaching feedback data, the first AI big model is used to generate and dynamically update the intelligent manufacturing talent capability index map; S2, Training Project Generation and Push: Through the intelligent manufacturing comprehensive practice platform module, based on teaching needs or input natural language project descriptions, the second AI model generates project task trees and standardized training modules, and then dynamically assembles them from the training case library into a complete training project process through the case assembly engine before pushing them to the user terminal. S3, Teaching process data collection and processing: Students practice in a hands-on environment. The practical teaching management module collects multimodal practical data in real time, processes it, constructs structured student practical data based on a multi-layer entity and relationship model of "student-project-task-action", and sends it to the intelligent central module. The teaching process data includes one or more of the following: operational behavior, digital operation log interaction, and outcome data. S4, Ability Assessment and Optimization Trigger: The intelligent central module uses the assessment and profiling intelligent agent to analyze the student's ability achievement and, in conjunction with the global state view output by the context awareness unit, determines whether to trigger the system optimization mechanism. S5, Multi-level Optimized Execution: The meta-controller within the intelligent central module, based on the judgment results, initiates corresponding intelligent agents to perform optimization according to pre-set priorities and quantified judgment criteria. S51 optimizes the inner layer by pushing appropriate tutoring tips or learning resources to student terminals through a personalized teaching strategy intelligent agent, thus enabling personalized intervention. S52 optimizes towards the middle layer by optimizing the agent through practical training cases and generating optimization adjustment instructions for the design parameters of specific practical training cases based on the near-end policy optimization algorithm. S53, optimize outwards, evolve the agent through the capability graph, and when the external technology popularity or industry demand index is detected to exceed the preset threshold, generate revision suggestions for the capability index graph.

[0148] like Figure 4 The image shows an example of a three-layer continuous evolution closed-loop mechanism driven by a large-model intelligent hub, as provided in an embodiment of the present invention.

[0149] For students, the system provides case recommendations based on competency profiles, project-based practice environments, and real-time feedback. Students can view competency radar charts, project progress, and personalized learning suggestions.

[0150] For teachers, the system provides a teaching dashboard that displays class performance, a heatmap of case completion, and a list of high-risk students. Teachers can publish new cases using natural language descriptions, and the system automatically generates a project task tree for teachers to review before publication. Teachers can also manually intervene or adopt case optimization suggestions.

[0151] For system administrators, the responsibilities include configuring physical device parameters, managing user permissions and monitoring system status, reviewing capability map evolution proposals, and driving macro-level system updates.

[0152] Example 3 like Figure 5 The diagram shows the complete implementation process of the specific training project "Ore Particle Size Analysis Based on Video Images". This embodiment uses a conveyor belt system as an application scenario to illustrate the usage of the system of the present invention and the collaborative working process of each intelligent agent in the large model intelligent hub.

[0153] Belt conveyor systems are a crucial link in mine production, and the particle size distribution of ore directly affects the efficiency of subsequent crushing and grinding processes. Traditional manual visual particle size assessment suffers from high subjectivity and the inability to perform real-time statistical analysis. This training project requires students to design an online ore particle size analysis and early warning system based on deep learning, enabling real-time acquisition of ore images on the belt conveyor system, particle size distribution calculation, and alarm for large ore pieces.

[0154] All modules of the practical teaching system are ready according to project requirements, including: System hardware components: industrial area scan camera (5 megapixels, with wide-angle lens), LED light source (adjustable brightness), industrial computer (equipped with NVIDIA RTX 3060 GPU), PLC controller (Siemens S7-1200), edge computing gateway (NVIDIA Jetson Xavier NX).

[0155] The system software consists of: an image acquisition module, a deep learning model training module, a real-time inference module, a host computer visualization module, and a communication module.

[0156] The first AI model has been trained according to the methods and steps provided in Example 1, and a structured capability index map of skill points in the field of conveyor belt transportation systems has been generated. Tables 5 and 6 show the multi-source heterogeneous data and extracted structured capability triples related to intelligent sensing conveyor belt transportation systems, respectively, and the generated structured capability index map is shown in Table 7.

[0157] The second AI model has been trained according to the methods and steps provided in Example 1, and the training case library stores the standardized training modules required for the project.

[0158] Table 5: Examples of Data Source Collection for the First Large AI Model

[0159] Table 6: Examples of Structured Capability Point Triplets

[0160] Table 7: Examples of Capability Indicator Definitions

[0161] 1. Teachers publish projects: Teachers input project requirements through the terminal: "Develop a deep learning-based online analysis and early warning system for ore particle size in belt conveyors".

[0162] 2. Modular Case Dynamic Assembly: After receiving project requirements, the second large AI model processes them, including: S801, Semantic Understanding: The semantic understanding subunit parses structured semantics. The core objective is "online analysis and early warning of ore particle size", and the constraints are "based on deep learning", "real-time performance", and "accuracy ≥ 90%". The technical fields are "machine vision", "embedded systems" and "real-time computing".

[0163] S802, Task Decomposition: The task decomposition subunit retrieves relevant basic actions ("industrial camera selection", "image acquisition programming", "dataset construction and annotation", "YOLO model training", "model quantization", "ONNX Runtime deployment", "PLC communication", etc.) by matching a pre-set knowledge graph with a graph neural network. The Tree-LSTM decoder generates a hierarchical project task tree. For example, the root node of the project task tree contains first-level nodes such as "1. Requirements analysis and solution design", "2. Hardware system construction", "3. Software algorithm development", "4. System integration and testing", and "5. Data analysis and reporting". Among them, "3. Software algorithm development" contains second-level nodes such as "3.1 Image annotation", "3.2 Model selection and training", and "3.3 Model optimization and lightweighting".

[0164] S803, Module Encapsulation: The training module encapsulation subunit encapsulates each node in the task tree in a pedagogical manner. Taking node "3.2 Model Selection and Training (YOLO / UNet)" as an example, the retrieval machine retrieves three historical encapsulations of similar modules from the pedagogical description database, and the generator combines the node features to generate the following encapsulation results, as shown in Table 8.

[0165] Table 8: Examples of instructional encapsulation of each node in the task tree by the encapsulation subunit of the practical training module

[0166] S804, Dynamic Assembly: The case assembly engine adaptively adjusts based on student ability profiles. Assume the current project group consists of 5 students, 3 of whom have strong programming skills (historical Python project scores ≥85), and 2 have extensive hardware experience. The difficulty adjuster calculates an average hint detail coefficient of 0.4, an average step completeness coefficient of 0.8, and a challenge task coefficient of 0.6. Therefore, the assembly engine decides: (1) Retain the complete task tree structure (high step integrity); (2) Attach the "Hyperparameter Tuning Guide" document to the "3.2 Model Selection and Training" module (with moderate level of detail). (3) Automatically add a challenge task at the end of the project: “Try to replace CNN with Swing Transformer and compare the difference between mAP and inference speed.”

[0167] S805, Project Push: Push practical training projects to students based on their ability profiles.

[0168] 3. Student Practice and Data Collection During the four-week project implementation process, students' operational behaviors were fully collected by the practical teaching management module.

[0169] Week 1: Student groups conducted camera selection simulations in a virtual simulation environment and completed hardware selection reports. The system recorded the number of group discussions (5 times), report submission time (day 5), and report plagiarism rate (12%). Examples of data storage for the Week 1 teaching process are shown in Tables 9-16, including project entities, student entities, the relationship between students and projects, and task and action data within the project.

[0170] Table 9: Project Entity Creation

[0171] Table 10: Student Entity Creation

[0172] Table 11: Student Participation in Project Relationship Building

[0173] Table 12: Creating a "Hardware Selection" Task

[0174] Table 13: Actions for Creating Group Discussions (Zhang San)

[0175] Table 14: Number of group discussions this week (automatically counted by the system)

[0176] Table 15: Action for Creating and Submitting a Hardware Selection Report (Zhang San)

[0177] Table 16: System-recorded report submission time (days since project start)

[0178] Week 2: Students installed cameras and debugged image acquisition on a physical tape model. The intelligent vision system detected that a student was not wearing protective glasses while adjusting the light source and immediately displayed a safety warning on the terminal. Examples of data storage for the Week 2 teaching process are shown in Tables 17-19.

[0179] Table 17: Creating a "Hardware Installation and Debugging" Task

[0180] Table 18: Actions for Creating Safety Violation Incidents (Li Si)

[0181] Table 19: Number of violations recorded by the system for this student

[0182] Week 3: Students train the model. The IDE plugin records code editing logs: Student A repeatedly adjusted the learning rate parameter in the training script (modified 8 times), debugged the loop 15 times, and finally ran it successfully. The system records GPU usage time (cumulative 23 hours) and the loss curve in the training log. Data storage for the teaching process in Week 3 is shown in Tables 20-23. Table 20: Creating the "Model Training" Task

[0183] Table 21: Code debugging action sequence for creating student A (Zhang San)

[0184] Table 22: Number of times student A's learning rate was modified according to system statistics:

[0185] Table 23: Creating Debug Loop Actions (Aggregated via CEP Engine)

[0186] Week 4: Students conduct model deployment tests, converting the trained model to ONNX format and deploying it on Jetson Xavier NX for real-time inference testing. The system records inference speed (average 18ms / frame), alarm accuracy (92%), project report submission time, and plagiarism rate. Data storage for the Week 4 teaching process is shown in Tables 25-27 (example). Table 25: Creating the "Model Deployment Test" Task

[0187] Table 26: Creating ONNX Conversion and Inference Test Actions

[0188] Table 27: Actions for Creating and Submitting the Final Project Report

[0189] Examples of data querying and usage are shown in Table 28-30: Table 28: Query all actions and timelines of student A throughout the entire project.

[0190] Table 29: Completion Status of Each Task in the Statistical Project

[0191] Table 30: Calculation of Student A's debugging efficiency in the "Model Training" task

[0192] Through the aforementioned student practice data model based on graph database, all teaching process data is uniformly stored as an attribute graph with a four-layer structure of "student-project-task-action", providing a standardized and directly queryable data foundation for the evaluation and profiling of intelligent agents by the intelligent central module.

[0193] In the teaching practice, the assessment and profiling agent subscribes to the event stream of the practical teaching management module and updates the student's ability profile in real time. Taking student A as an example: Knowledge mastery assessment: The XGBoost model outputs the mastery level probability based on its test scores (image processing 92 points, deep learning 85 points) and the frequency of PyTorch API calls in the code (high frequency): PyTorch [proficient: 0.85, average: 0.15].

[0194] Skill proficiency assessment: The MLP model outputs a proficiency score of 0.82 based on task completion time (the model training task time is 15% faster than average) and operation fluency (15 debugging loops, slightly higher than the average of 12, but the success rate of debugging on the first try is high).

[0195] Innovation capability assessment: Sentence-BERT calculates the similarity between Student A's project report and historical solutions (uniqueness score 0.75), the breadth of the technology stack (involving YOLOv5, ONNX, TensorRT, Flask), and the text generation assessment model scores the "Summary of Innovation Points" section at 0.8, resulting in a weighted innovation capability score of 0.78.

[0196] Final Student A's Ability Profile Vector , and the target vector of this stage of life Comparison yields the difference vector. The largest difference was found in the dimension of "model optimization and lightweighting" with a value of -0.25.

[0197] During student practice, the intelligent central module uses assessment and profiling agents to analyze students' ability achievement and, in conjunction with the global state view output by the context-aware unit, determines whether to trigger the system optimization mechanism.

[0198] 4. Real-time personalized tutoring When a student's individual problem is detected, the personalized teaching strategy agent is triggered to initiate inner-layer optimization and provide personalized tutoring. For example, the meta-controller detects that student A has a significant gap (<-0.2) in the "model optimization and lightweighting" dimension, and the student's historical intervention records show that he has a high acceptance of text-based tutorials, thus triggering the personalized teaching strategy agent.

[0199] The policy network input state s includes: the difference vector (Focusing on the "model optimization" dimension), current error context (students encountered an "operator not supported" error when trying to export from ONNX), historical intervention feedback (positive feedback was received when video V003 was pushed). The probability of each action output by the policy network is as follows: [Pushing video V012 (detailed explanation of ONNX operators): 0.65, Pushing code example C008 (model quantization code): 0.25, Initiating guided dialogue: 0.10].

[0200] The system selects the action with the highest probability, pushes a video link through the student's terminal, and includes a prompt: "We detected that you encountered an operator problem when exporting from ONNX. We suggest you first learn the operator types supported by ONNX (video V012), and then try our preset quantization script (link)." 5. Project Completion and Capability Assessment After all students complete the project, the assessment and profiling intelligent system will further evaluate the model's performance indicators, generate a student ability increment report, identify common problems in the project, and feed them back to the intelligent central module.

[0201] 6. Feedback Optimization Set The intelligent central module uses project feedback and the global state view provided by the context-aware intelligent agent to statistically analyze common problems and determine whether to trigger middle-layer and outer-layer optimizations.

[0202] For example, when the context-aware agent monitors and finds that in the past three months, the average time spent on the "lightweight deployment of the model" subtask (2.8 days) by all teams that completed the project was significantly higher than that of other subtasks (average 1.2 days), and the completion rate was only 65%, the meta-controller determines that this is a common problem and triggers the training case optimization agent, i.e., mid-level optimization.

[0203] The reinforcement learning engine state s_t=[d=0.7, r=0.5, h=0.3] (difficulty gradient 0.7, resource support 0.5, cue information 0.3). After 1000 explorations in a virtual environment using the PPO algorithm, it was found that the action a_t=[Δd=-0.1, Δr=+0.2, Δh=+0.1] yields the highest cumulative reward. The agent generation optimization instructions are shown in Table 31.

[0204] Table 31: Examples of Agent Generation Optimization Instructions

[0205] The instruction was sent to the teacher's terminal for review. After the teacher reviewed and confirmed its adoption, the optimized module took effect in the next round of projects. Subsequent monitoring data showed that the average time for this subtask decreased to 1.9 days after optimization, and the completion rate increased to 82%.

[0206] The capability graph evolution agent regularly analyzes external data sources and found that keywords such as "real-time edge inference" and "lightweight model deployment" have increased by 120% in the past two months, and the number of stars in related GitHub repositories has increased by 80%. However, the weight of the "embedded AI deployment" indicator in the current capability graph is only 0.35, which is significantly low. The agent judges this to be a major technological shift and initiates the impact assessment process. S901. Construct a future demand map and temporarily increase the weight of the "Embedded AI Deployment" indicator to 0.75; S902. Calculate the difference. If the cosine similarity between the current map and the future map is 0.62 (< threshold 0.85), trigger in-depth analysis. S903. Graph propagation analysis maps the existing curriculum system into a dependency network graph (85 nodes), simulating the addition of the "Embedded AI Deployment" indicator node and its dependency edges (depending on "C++ Programming", "Deep Learning Fundamentals", and "ONNX / TensorRT"). S904. Quantization Impact Assessment: It is necessary to add 1 prerequisite course module "Embedded Linux Basics", adjust 2 existing courses ("Deep Learning" adds a lightweight chapter, "Real-time Systems" adds an ONNX case study), and add 3 practical training cases ("YOLOv5 Model Quantization Practice", "TensorRT Deployment Optimization", "Jetson Platform Performance Tuning"), with an estimated total increase of 18 hours of class time. S905. Generate an update proposal and submit it to the Teaching Management Committee.

[0207] After the committee approves the updated capability map, it will be synchronized to the capability indicator generation module. At the same time, suggestions will be sent to the intelligent manufacturing comprehensive practice platform module, requiring priority to be given to the development of practical training cases related to "embedded AI deployment" to form a new closed loop for the construction of teaching resources.

[0208] As can be seen from this embodiment, the two AI models in this invention form a complete intelligent teaching support system in a real intelligent manufacturing teaching scenario (ore particle size analysis project in a conveyor belt system): The first AI big model: automatically learns from industry data and generates cutting-edge capability indicators such as "online analysis of ore particle size" to ensure that training objectives are in sync with industry needs; The second major AI model automatically deconstructs the natural language project requirements proposed by teachers into an executable training task tree and generates standardized teaching modules, thereby improving the efficiency of project implementation by more than 80%. Intelligent Hub: During project implementation, through the collaborative work of three layers of intelligent agents, continuous closed-loop optimization is achieved from individual tutoring (inner layer), case optimization (middle layer) to competency map evolution (outer layer), which improves teaching effectiveness (completion rate, innovation) by 20%-30% and shortens the update cycle of talent training programs from annual to monthly.

[0209] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A smart manufacturing practical teaching system based on an AI large-scale model, characterized in that, It includes a capability indicator generation module, an intelligent manufacturing comprehensive practice platform module, a practical teaching management module, and an intelligent central module, among which; The capability index generation module is used to collect multi-source heterogeneous data and extract structured capability triples. The first AI big model generates and dynamically updates the structured capability index map based on the triples. The intelligent manufacturing integrated practice platform module includes a physical and virtual integrated practical environment, a second AI large model, a training case library, and a case assembly engine. The second AI large model is used to parse the input natural language project requirements into a structured project task tree, and generate standardized training modules based on the project task tree and store them in the training case library. The case assembly engine is used to respond to requests from the intelligent central module, dynamically retrieve and assemble training modules from the training case library to generate a complete training project process. The practical teaching management module is used to collect multimodal practical data of students in the practical environment in order to construct a student practical data model and extract quantitative features; The intelligent central module communicates with the capability indicator generation module, the intelligent manufacturing integrated practice platform module, and the practical teaching management module, respectively. It is used to integrate data from each module and make collaborative decisions through multiple dedicated intelligent agents to generate personalized teaching strategies, practical training case optimization instructions, and capability map evolution schemes.

2. The practical teaching system according to claim 1, characterized in that, The capability index generation module uses a BERT-based sequence labeling and relation classification joint training model to extract capability point-technical requirement-performance index triples from multi-source heterogeneous data.

3. The practical teaching system according to claim 2, characterized in that, The The first AI large-scale model uses capability point triples as input to generate capability index maps. The methods include: S101, Clustering and Abstraction: The semantic vector model is used to convert the triplet of capability points into semantic vectors, and the semantically similar capability points are grouped into capability clusters by an unsupervised clustering algorithm; S102, Generate a hierarchical architecture: Use a hierarchical text classification model to automatically predict the level to which the capability cluster belongs, and use a fine-tuned BERT text classifier to classify the capability clusters in the same level into a preset dimension. S103, Generate capability indicator definition and description: Based on the text generation model, the capability indicator name, level and dimension are used as input. Combined with retrieval enhancement generation technology, a structured indicator definition and description containing core knowledge points, typical behavioral performance and related preliminary capability indicators are automatically generated. S104, Dynamic Weight Calculation: Combining multi-dimensional data, the weight of each capability indicator in the current map is dynamically calculated using a time-decay weighted algorithm. The weight calculation formula includes: in, Indicators of capability In time The weights; Indicators of capability A collection of related data entries; Represents data entries, Represents data entries Time; Indicates the attenuation factor; It is an indicator In data entries The normalized frequency of occurrence in; Represents data entries The significance score; Indicates the significance weighting coefficient; This represents the normalized denominator.

4. The practical teaching system according to claim 1, characterized in that, The first large AI model was prepared using a two-stage method of knowledge distillation and domain fine-tuning. The specific methods include: S201, using industry and technology documents, trains a general teacher model for understanding industry technologies; S202 uses active learning and crowdsourced annotation technology to obtain large-scale educational data labeled according to a predetermined capability indicator system. S203, using labeled data to supervise the training of student models, enabling student models to initially learn data classification; S204, the same unlabeled industry technology text is simultaneously input into the teacher model and the student model. By minimizing the KL divergence between the prediction results of the student model and the prediction results of the teacher model, as well as the classification cross-entropy loss of the student model on the labeled data, the student model is jointly optimized to obtain the first AI large model.

5. The practical teaching system according to claim 1, characterized in that, The second AI model includes a semantic understanding subunit, a task decomposition subunit, and a module encapsulation subunit; The semantic understanding subunit is used to parse natural language project requirements into a structured semantic representation that includes core objectives, constraints, and technical fields, using a pre-trained semantic understanding model finely tuned by the engineering domain. The task decomposition subunit has a built-in pre-built engineering task decomposition knowledge graph, which is used to decompose the structured semantic representation output by the semantic understanding subunit and map it to generate a project task tree composed of basic action nodes. The training module encapsulation subunit is used to encapsulate each basic action node on the project task tree in a teaching-oriented manner, generating a standardized training module that includes teaching objectives, prerequisite knowledge, input and output specifications, implementation paths, and evaluation criteria.

6. The practical teaching system according to claim 1, characterized in that, The method for constructing a student practice data model in the practical teaching management module includes: S301 defines a multi-level entity and relationship model of "student-project-task-action" based on a graph database; S302, associate and align the collected multimodal practice data with the multi-layer entity and relation model; S303 aggregates associated low-level operation events into high-level actions with pedagogical significance.

7. The practical teaching system according to claim 1, characterized in that, The second large AI model is constructed through a multi-stage joint training strategy, and the training steps include: S401 uses large-scale code-text pre-training on the base model, enabling the model to master the basic ability of code understanding and generation; S402 uses text data from fields related to industrial automation, machine vision, and intelligent manufacturing for domain-adaptive pre-training. S403, using paired data of "project requirement text - structured semantic tags" to perform supervised fine-tuning of the semantic understanding part of the model; S404 uses "project requirements - project task tree" paired data to train the model's task decomposition ability through teacher-mandated and reinforcement learning fine-tuning methods based on teaching simulation environments. S405 uses "basic action node - instructional module description" paired data, combined with retrieval enhancement generation technology, to train the model's practical module encapsulation capabilities.

8. The practical teaching system according to claim 1, characterized in that, The intelligent central module includes: A unified data lake and context-aware units are used to integrate multi-source data and output a global state view that represents the alignment of teaching load, learning outcomes, and internal and external trends. The multi-agent decision-making and optimization execution unit includes an assessment and profiling agent for evaluating students' abilities and generating ability profiles, a personalized teaching strategy agent for generating personalized teaching intervention strategies, a practical training case optimization agent for optimizing practical training case design parameters, and an ability graph evolution agent for analyzing and generating ability graph update proposals. The meta controller is used to coordinate the triggering and execution of the multiple smart agents based on the global state view.

9. The system according to claim 8, characterized in that, The training case optimization agent is trained and optimized using a proximal policy optimization algorithm, and its optimization methods include: S501, define the design parameter vector of the state as a training case, the design parameter vector shall include at least the difficulty coefficient, resource support degree and prompt information amount; S502, define the action as the adjustment amount of the design parameter vector; S503, In a virtual simulation environment, explore different actions and calculate a reward function based on the real student usage data in the next round after applying the action, and update its policy network. The reward function includes at least a weighted sum of changes in case completion rate, changes in average completion time, and changes in innovation score. S504. After the strategy network converges, a structured optimization instruction containing updated design parameters and resources to be added is generated and sent to the intelligent manufacturing integrated practice platform module for execution.

10. A practical teaching method for intelligent manufacturing using the system described in any one of claims 1-9, characterized in that, The methods include: S1, Capability Index Map Generation and Update: Through the capability index generation module, based on industry dynamic data, technology development data, and internal teaching feedback data, the first AI big model is used to generate and dynamically update the intelligent manufacturing talent capability index map; S2, Training Project Generation and Push: Through the intelligent manufacturing comprehensive practice platform module, based on teaching needs or input natural language project descriptions, the second AI model generates project task trees and standardized training modules, and then dynamically assembles them from the training case library into a complete training project process through the case assembly engine before pushing them to the user terminal. S3, Teaching process data collection and processing: Students practice in a hands-on environment. The practical teaching management module collects multimodal practical data in real time, processes it, constructs structured student practical data based on a multi-layer entity and relationship model of "student-project-task-action", and sends it to the intelligent central module. S4, Ability Assessment and Optimization Trigger: The intelligent central module uses the assessment and profiling intelligent agent to analyze the student's ability achievement and, in conjunction with the global state view output by the context awareness unit, determines whether to trigger the system optimization mechanism. S5, Multi-level Optimized Execution: The meta-controller within the intelligent central module, based on the judgment results, initiates corresponding intelligent agents to perform optimization according to pre-set priorities and quantified judgment criteria. S51 optimizes the inner layer by pushing appropriate tutoring tips or learning resources to student terminals through a personalized teaching strategy intelligent agent, thus enabling personalized intervention. S52 optimizes towards the middle layer by optimizing the agent through practical training cases and generating optimization adjustment instructions for the design parameters of specific practical training cases based on the near-end policy optimization algorithm. S53, optimize outwards, evolve the agent through the capability graph, and when the external technology popularity or industry demand index is detected to exceed the preset threshold, generate revision suggestions for the capability index graph.

Images