A work-integrated learning intelligent teaching system for rail transit intelligent operation and maintenance

Abstract

This invention relates to the field of intelligent teaching technology and provides an integrated engineering-learning intelligent teaching system for intelligent operation and maintenance of rail transit. In this system, a teaching data platform module aggregates and processes teaching data; a teaching twin module constructs a three-layer twin model and generates multi-dimensional teaching tasks; an integrated engineering-learning teaching engine module automatically arranges the optimal sequence of teaching activities based on student competency profiles and teaching tasks through reinforcement learning; a metaverse scene construction module constructs immersive simulation training scenarios based on this sequence; a competency evolution assessment module analyzes student training data and generates a competency increment map; and a teaching feedback optimization module feeds this map back to the task generator to drive a new round of teaching. This invention achieves dynamic generation, intelligent arrangement, precise assessment, and closed-loop optimization of teaching in the field of intelligent operation and maintenance of rail transit, solving problems such as insufficient intelligence, mismatched teaching strategies, mismatch between engineering and learning, and a single method of evaluating teaching quality in rail transit teaching.

Description

Technical Field

[0001] This invention relates to the field of intelligent teaching technology, and more specifically, to an integrated engineering-learning intelligent teaching system for intelligent operation and maintenance of rail transit. Background Technology

[0002] Currently, the rail transit industry is undergoing a comprehensive transformation from traditional operation and maintenance (O&M) to "smart O&M" characterized by unmanned operation, intelligent monitoring, and predictive maintenance. This industry upgrade places unprecedentedly high demands on the composite capabilities of technical and skilled personnel, urgently requiring vocational education systems to cultivate high-quality talent who are both theoretically sound and practically capable. Against this backdrop, the work-integrated learning model has become the core path for vocational schools to cultivate such talent. This model emphasizes the deep integration of real-world work processes with theoretical learning; however, in the teaching practice of this specific high-end field of smart O&M in rail transit, its implementation faces a series of technical bottlenecks that urgently need to be addressed.

[0003] Traditional teaching models in this field suffer from the following shortcomings: First, the teaching scenarios are severely disconnected from real-world working conditions. Core rail transit equipment (such as train traction systems and signal interlocking equipment) is expensive, structurally complex, and operates in a high-risk environment, making it difficult to conduct in-depth skills training with high failure rates and risks in physical training. Existing virtual simulation software is mostly limited to pre-set, isolated operational procedures, unable to dynamically simulate the evolution of hidden equipment failures, and lacks realistic reproduction of multi-system coupled failures and the resulting chain of operational events. This results in students mastering individual operations but lacking systemic thinking and emergency decision-making abilities to cope with complex real-world conditions. Second, instructional design relies heavily on subjective experience, lacking personalization and adaptability. Teachers primarily design instruction based on personal experience and general syllabi, making it difficult to accurately grasp the significant differences in students' knowledge base, skill gaps, and learning pace, thus failing to achieve personalized teaching tailored to each individual. There is often a mismatch between teaching objectives, activities, and students' actual abilities, leading to low teaching efficiency and making it difficult to form an effective closed loop between "work" and "learning." Third, the teaching evaluation dimensions are singular and feedback is delayed. Existing evaluation systems primarily focus on the correctness of final operational results or theoretical exam scores, constituting summative assessment. They severely lack multi-dimensional and quantitative tracking of trainees' process-oriented abilities, collaborative communication skills, and innovative thinking demonstrated throughout the entire work process—from task comprehension and planning to decision-making, implementation, monitoring and control, evaluation and feedback, and transfer and innovation. Evaluation results are typically delayed, failing to provide timely and accurate data support for real-time adjustments and optimization of the teaching process.

[0004] In existing technologies, such as the patent with publication number CN120690080A, "Intelligent Operation and Maintenance Training System for Rail Transit Based on Cloud Computing," a method is provided to optimize the scheduling of training projects and the allocation of cloud computing resources by quantitatively analyzing student performance. However, this system is essentially a scheduling platform focused on resource management and load optimization. Its core lies in prioritizing and allocating resources to an existing, static training project library, without addressing fundamental teaching issues such as the dynamic generation of teaching scenarios, the intelligent arrangement of personalized teaching strategies, and the multi-dimensional process assessment of complex professional skills. Other related technical solutions also focus on single aspects such as online learning management, fixed process simulation, or simple behavioral analysis, failing to build a complete intelligent teaching system that can deeply integrate domain knowledge, dynamic twin environments, intelligent decision engines, and closed-loop evaluation feedback.

[0005] Therefore, there is an urgent need in this field for an innovative technical solution to systematically solve the above problems and realize the leap from concept to high-quality implementation of integrated engineering and learning teaching in the field of intelligent operation and maintenance of rail transit. Summary of the Invention

[0006] To address the technical problems mentioned in the background section of the present invention, this invention provides an integrated engineering and learning intelligent teaching system for intelligent operation and maintenance of rail transit, the specific technical solution of which is as follows:

[0007] An integrated engineering-education intelligent teaching system for intelligent operation and maintenance of rail transit includes:

[0008] The teaching data platform module is used to aggregate and process heterogeneous data for rail transit operation and maintenance teaching. The heterogeneous data includes static standard data, dynamic teaching process data, and external teaching resource data.

[0009] The teaching twin module is used to dynamically construct a three-layer twin model containing equipment mechanism, system business, and teaching tasks based on the heterogeneous data, and generate a multi-dimensional composite teaching task set;

[0010] The engineering-integrated teaching engine module has a built-in adaptive task generator based on a hybrid reinforcement learning algorithm, which is used to automatically arrange the optimal sequence of teaching activities according to the student's ability profile and the multi-dimensional composite teaching task set.

[0011] The metaverse scene construction module is used to call corresponding virtual materials based on the optimal teaching activity sequence to construct a simulation scene for students to conduct interactive virtual operation training;

[0012] The capability evolution assessment module is used to collect training data of trainees in the simulation scenario and compare it with the standard operation model in the teaching twin module to generate a capability increment map.

[0013] The teaching feedback optimization module is used to feed the capability increment map back to the adaptive task generator, driving the generation of a new round of optimal teaching activity sequence.

[0014] Furthermore, the teaching data platform module specifically includes:

[0015] A static standard library is used to store and associate the rail transit operation and maintenance capability standards, curriculum standard map and industry specifications that integrate the four dimensions of "job, course, competition and certificate". The curriculum standard map establishes a semantic mapping relationship between course knowledge points, skill points and items in the capability standards through ontology modeling.

[0016] The dynamic process library is used to collect and structured store multimodal data throughout the teaching process in real time. The multimodal data includes the operation sequence logs of students in virtual operation training, the dialogue text generated by interaction with the intelligent teaching assistant, the role-playing and communication records in collaborative tasks, and the key node annotation data of the teacher based on the standard operation model for the students' operation process. The standard operation model is the operation and maintenance specification model preset in the teaching twin module.

[0017] External resource library, used to index and associate parameterized 3D model library of rail transit equipment, fault case knowledge graph, maintenance procedure video clips and safety warning education materials.

[0018] Furthermore, the three-layer twin model in the instructional twin module specifically includes:

[0019] The equipment-level mechanism twin layer is used to simulate faults and predict conditions based on the physical simulation model and historical operation data of key rail transit equipment, and to generate a set of potential hidden fault modes.

[0020] The system-level business twin layer is connected to the device-level mechanism twin layer. It is used to associate and deduce the set of hidden fault modes with the device topology relationship and operation scheduling rules based on the operation and maintenance procedure knowledge graph, and generate a virtual business event chain that affects operational safety.

[0021] The teaching-level task twin layer, connected to the system-level business twin layer, is used to receive the virtual business event chain and, in conjunction with the teaching objectives in the static standard library, deconstruct the event chain into a multi-dimensional composite teaching task set including fault handling, scheduling and coordination, and safety briefing through task decomposition and reorganization algorithms.

[0022] Furthermore, the engineering-integrated learning engine module specifically includes:

[0023] The trainee capability profile generator is used to output a three-dimensional capability vector of "safety-efficiency-cost" based on trainees' historical training data and real-time eye movement, operation sequence, and voice command multimodal information through graph neural networks.

[0024] The course difficulty adjuster monitors the variance of the reward function. If the variance exceeds the threshold, the difficulty of subsequent tasks will be dynamically adjusted down or up according to the "zone of proximal development" principle.

[0025] The knowledge graph gap diagnosis interface is used to compare the four-dimensional knowledge graph of "job, course, competition and certification" with the three-dimensional ability vector of trainees to generate gap vectors;

[0026] An adaptive task generator, with a built-in reinforcement learning hybrid algorithm unit, is used to output a preliminary teaching activity sequence through an Actor-Critic architecture. The state is the concatenation of the three-dimensional capability vector and the embedding vector of the teaching task set. The action space is the task difficulty, task type, and resource combination. The reward function is the expected capability increment.

[0027] Furthermore, after the adaptive task generator outputs the preliminary teaching activity sequence, it also includes rearranging the content and replacing tasks based on the gap vector to generate a revised teaching activity sequence, and outputting the optimal teaching activity sequence based on the task difficulty adjustment requirements fed back by the course difficulty adjuster.

[0028] Furthermore, the metaverse scene construction module is specifically used for:

[0029] Receive the optimal teaching activity sequence and analyze its requirements for virtual environment, device status, fault phenomena and interactive objects;

[0030] Based on the analysis results, the system intelligently retrieves and calls matching parametric 3D models, material maps, animation scripts, and physics engine configuration parameters from the external resource library.

[0031] Based on a real-time rendering game engine, the called materials are instantiated, spatially laid out, and logically bound in real time according to the teaching logic, generating an interactive immersive simulation scene that is synchronized with the optimal teaching activity sequence.

[0032] The immersive simulation scenario allows trainees to conduct detailed operation training from a first-person perspective through their personal terminals, while also supporting global situation simulation and role-playing from a third-person perspective in a collaborative sandbox.

[0033] Furthermore, the capability evolution assessment module specifically includes:

[0034] The data acquisition unit is used to extract the trajectory data of trainees' operations, the time series of step completion, the tool selection sequence, and the communication content with other trainees from the background logs of virtual operation training.

[0035] The multi-dimensional comparative analysis unit is used to compare the trajectory data, time series, and tool selection sequence with the preset optimal path, standard working hours, and standard toolset in the standard operation model, respectively, and calculate the operation standardization score, efficiency score, and decision accuracy score; perform natural language processing on the communication content, analyze its collaborative intent and problem-solving logic, and generate collaborative communication and innovative thinking scores.

[0036] The graph generation unit is used to calculate the changes in the trainees' ability levels in each ability dimension relative to the previous assessment based on the collaborative communication and innovative thinking scores of the multi-dimensional comparative analysis unit and the project response theory. It generates a visual ability increment graph in the form of a combination of radar chart and ability node growth chart. The trainees' ability dimensions include six dimensions: "task understanding, planning, decision implementation, inspection and control, evaluation and feedback, and transfer and innovation".

[0037] Furthermore, the instructional feedback optimization module specifically includes:

[0038] The strategy adjustment unit is used to analyze the capability increment map and identify the common weaknesses of the trainee group and the specific capability shortcomings of individuals.

[0039] The resource rematching unit is used to rematch the corresponding resource list from the external resource library and static standard library of the teaching data platform module based on the common weak ability dimensions of the identified student group and the specific ability shortcomings of the individual. The resource list includes cases, micro-lessons, simulation faults or extended reading materials.

[0040] The task regeneration instruction unit is used to package and send the updated list of common weak abilities of the student group and individual specific ability shortcomings, as well as the rematched resource list, to the adaptive task generator of the integrated work-study teaching engine module. This triggers the generator to prioritize the generation of a new round of optimal teaching activity sequence in the next round of strategy iteration, targeting the common weak abilities and individual specific ability shortcomings. The new round of optimal teaching activity sequence includes reinforcement training tasks and personalized learning path suggestions.

[0041] Furthermore, it also includes a cross-platform collaboration and resource sharing module, which is communicatively connected to the teaching data platform module and the work-integrated learning engine module, specifically including:

[0042] The collaborative interaction unit adopts the WebRTC real-time communication protocol to realize voice / video interaction between students and tutors, synchronized operation screens, and collaborative text annotation between students, and supports multi-role permission hierarchical management;

[0043] The learning data synchronization unit, based on a distributed database with a cloud-edge collaborative architecture, synchronizes students' learning progress, training records, and ability assessment results to the cloud in real time, ensuring data consistency after multiple terminal logins.

[0044] The intelligent resource push unit, based on the collaborative filtering algorithm and combined with the weak points in the student's ability profile, selects and matches customized learning resources from the external resource library of the teaching data platform, and pushes them to the student's terminal in real time through the message push interface.

[0045] Furthermore, all modules of the engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit described in any of the above embodiments are deployed on the same cloud-edge collaborative architecture, which includes:

[0046] The central cloud is used to store historical operational big data, heterogeneous data, and cross-school teaching resources;

[0047] Edge cloud, deployed in the training base, is used to support real-time computing and low-latency rendering of the teaching twin;

[0048] Terminal edge nodes are embedded in personal immersive operating terminals and AR glasses to perform posture tracking and gesture recognition locally;

[0049] The unified time-series database adopts the time-series structured multi-label TSMT data model to label and store real operation and maintenance events, teaching task events, and student behavior events.

[0050] This invention offers the following advantages: It deeply integrates domain knowledge graphs, dynamic digital twins, reinforcement learning decision engines, and immersive virtual simulation technology to construct a self-iterable closed-loop intelligent teaching ecosystem. This system achieves a breakthrough by transforming teaching scenarios from pre-set static cases to dynamically generated scenarios based on real data and mechanistic models. It can simulate complex hidden fault chains and system-level business events in intelligent operation and maintenance of rail transit, providing trainees with a high-fidelity training environment that closely approximates real-world work pressure and complexity. Simultaneously, through reinforcement learning algorithms, based on trainees' multi-dimensional ability profiles and real-time twin events, the system intelligently arranges and generates personalized teaching task sequences that deeply integrate technical operations and process collaboration, achieving a fundamental leap from "resource scheduling" to "strategy generation," resulting in precise "one-person-one-policy" adaptive teaching. At the evaluation level, the system collects trainees' full-process, multi-modal behavioral data in the immersive environment and, based on... By conducting multi-dimensional and process-oriented in-depth analysis based on project response theory, a capability increment map that quantifies trainees' growth trajectory in core professional competency dimensions is generated, realizing the evolution from single-outcome evaluation to value-added and developmental evaluation. Finally, the above evaluation results are fed back to the system's task generation engine and twin model in real time, driving continuous dynamic optimization of teaching strategies and training content, forming a strong closed loop with self-learning and self-evolution capabilities. This fundamentally solves systemic problems in traditional vocational education such as distorted teaching scenarios, insufficient personalization, single and lagging evaluation methods, and the disconnect between theory and practice, providing a quantifiable, traceable, and optimizable solution for cultivating compound high-end technical and skilled talents who can meet the needs of industrial upgrading. Attached Figure Description

[0051] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a block diagram of an integrated engineering and science intelligent teaching system for intelligent operation and maintenance of rail transit, provided as an embodiment of the present invention.

[0053] Figure 2 This is a schematic diagram of the virtual operation score of the trainee and the evaluation radar of each capability dimension generated by the map generation unit in the capability evolution evaluation module provided in an embodiment of the present invention. Detailed Implementation

[0054] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an integrated engineering and learning intelligent teaching system for intelligent operation and maintenance of rail transit proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0055] Example

[0056] The following description, in conjunction with the accompanying drawings, details a specific scheme for an integrated engineering and learning intelligent teaching system for intelligent operation and maintenance of rail transit provided by this invention.

[0057] The specific scenario addressed by this invention is: addressing the core dilemma faced in implementing integrated work-study advanced skills training in the field of intelligent operation and maintenance of rail transit, namely, how to reproduce complex fault handling scenarios in real operation and maintenance in a safe and controllable environment with high fidelity, and to achieve process-oriented training and accurate assessment of trainees' composite abilities such as systematic diagnosis, emergency decision-making, and collaborative operation.

[0058] Please see Figure 1 The diagram illustrates a block diagram of an integrated engineering and education intelligent teaching system for intelligent operation and maintenance of rail transit, according to an embodiment of the present invention. The system includes:

[0059] Teaching Data Platform Module 101, Teaching Twin Module 102, Work-Study Integrated Teaching Engine Module 103, Metaverse Scene Construction Module 104, Ability Evolution Assessment Module 105, Teaching Feedback Optimization Module 106.

[0060] The teaching data platform module 101 is used to aggregate and process heterogeneous data in rail transit operation and maintenance teaching. This heterogeneous data includes static standard data, dynamic teaching process data, and external teaching resource data. Specifically, the teaching data platform module includes: a static standard library, used to store and associate the four-dimensional integrated rail transit operation and maintenance capability standards, curriculum standard maps, and industry specifications, where the curriculum standard map establishes semantic mapping relationships between course knowledge points, skill points, and items in the capability standards through ontology modeling; a dynamic process library, used to collect and structure and store multimodal data throughout the teaching process in real time, including student operation sequence logs during virtual operation training, dialogue text generated from interactions with intelligent teaching assistants, role-playing and communication records in collaborative tasks, and key node annotation data of the student operation process based on a standard operation model, where the standard operation model is a pre-set operation and maintenance specification model in the teaching twin module; and an external resource library, used to index and associate parameterized rail transit equipment 3D model libraries, fault case knowledge graphs, maintenance procedure video clips, and safety warning education materials.

[0061] To realize the functions of the teaching data platform module 101, in its specific implementation, this module constructs a unified data governance framework integrating a static standard library, a dynamic process library, and an external resource library. The static standard library uses ontology modeling technology to transform the competency standard items, course knowledge points, and industry standard provisions corresponding to "job training, course certification, and competition certification" into a semantically related knowledge graph. For example, for the competency standard of "train traction system fault diagnosis," the graph establishes a mapping relationship with the knowledge point of "Traction Inverter Working Principle" in the course "Urban Rail Transit Vehicle Maintenance," the safety testing procedures in the local standard "DB11 / T2390-2025 Technical Specifications for Maintenance of Urban Rail Transit Suburban Express Vehicles," and relevant competition questions from the "Rail Transit Intelligent Operation and Maintenance Skills Competition," forming a structured description of competency requirements.

[0062] The dynamic process library, deployed in the virtual training environment, captures and structures the multimodal data streams generated during the teaching process in real time through a data acquisition interface. When trainees operate in a virtual train maintenance scenario, the system not only records their operation steps and tool usage order, but also simultaneously collects their voice dialogue text with the intelligent teaching assistant, the command text sent through the communication interface in collaborative tasks, and the standard compliance annotations added by teachers to key operation nodes in the playback recording through the management terminal. This data is timestamped and associated with specific teaching tasks, equipment twin instances, and trainee identities for storage.

[0063] The external resource library manages parametric 3D models and multimedia materials. For example, a 3D model of a subway bogie is associated with the geometric parameters, material properties, and physical simulation parameters of its detachable components. An operation video of "abnormal wear detection of pantograph carbon sliding plate" is cut into step segments and linked to the corresponding nodes in the fault case knowledge graph.

[0064] The core processing logic of the teaching data platform module lies in the association and fusion of the three types of heterogeneous data mentioned above. When the engineering-integrated teaching engine module requests the generation of a teaching task on "intermittent faults in train network control systems," the data platform determines the knowledge points that the task needs to cover, such as "fault code analysis" and "network topology analysis," based on the capability map in the static standard library. Simultaneously, it queries the target student's historical operation records in the dynamic process library to identify weaknesses in their "logical judgment" dimension. Furthermore, it retrieves matching fault case 3D models, network topology diagrams, and relevant technical manual excerpts from external resource libraries. This cross-database semantic retrieval and association matching ensures that the subsequently generated teaching tasks accurately align with job competency requirements, adapt to individual student learning situations, and are equipped with high-fidelity practical training resources.

[0065] In this embodiment, static standard data ensures the standardization of teaching objectives, dynamic process data ensures the objectivity of learning analysis, and external resource data guarantees the richness of teaching scenarios. These three types of data support each other in data processing logic: standard data defines the dimensions and objectives required for analyzing dynamic data; the analysis results of dynamic data drive the personalized selection and retrieval of resource data; and the associativity of resource data enables the instructions generated based on standard and dynamic data to be concretely presented. This interaction between data collectively supports the subsequent dynamic construction of complex teaching scenarios by the teaching twin module and the intelligent generation of personalized teaching strategies by the work-integrated learning engine module.

[0066] The structured heterogeneous data aggregated by the teaching data platform module provides a solid data foundation for the intelligent evolution of the teaching process. However, simple data retrieval cannot automatically generate dynamic training scenarios that match the complexity of real-world operations and are adapted to the development of individual learners' abilities. Therefore, the system introduces the teaching twin module 102. As the core engine connecting the underlying data and the top-level teaching applications, this module is responsible for dynamically integrating and constructing an interactive, predictable, and trainable virtual teaching environment based on the standardized knowledge, real-time process information, and parameterized resources from the data platform, according to the physical laws of the equipment, the system's operational logic, and the teaching objectives.

[0067] The teaching twin module 102 is used to dynamically construct a three-layer twin model containing equipment mechanisms, system business, and teaching tasks based on the heterogeneous data, generating a multi-dimensional composite teaching task set. Specifically, the three-layer twin model in the teaching twin module includes: an equipment-level mechanism twin layer, used to simulate faults and predict states based on the physical simulation model and historical operating data of key rail transit equipment, generating a set of potential latent fault modes; a system-level business twin layer, connected to the equipment-level mechanism twin layer, used to correlate and deduce the latent fault mode set with equipment topology relationships and operation scheduling rules based on the operation and maintenance procedure knowledge graph, generating a virtual business event chain affecting operational safety; and a teaching-level task twin layer, connected to the system-level business twin layer, used to receive the virtual business event chain and, in conjunction with the teaching objectives in the static standard library, deconstruct the event chain into a multi-dimensional composite teaching task set including fault handling, scheduling coordination, and safety briefing through task decomposition and recombination algorithms.

[0068] Specifically, the instructional twin module 102, through the construction of a three-layered twin model coupling equipment mechanism, system business, and teaching tasks, realizes the transformation from data to dynamic teaching scenarios. The equipment-level mechanism twin layer, based on historical and real-time monitoring data from the dynamic process library, drives the physical simulation models of key equipment such as traction motors and signal relays. This model can simulate the state evolution of equipment under specific operating conditions and potential degradation conditions. When the simulation parameters and measured data trends show a continuous deviation, a latent fault mode, such as "the vibration amplitude of the traction motor bearing increases exponentially but does not exceed the threshold," is marked and generated. This fault mode is then passed to the system-level business twin layer as a basic event.

[0069] The system-level business twin layer embeds an operation and maintenance procedure knowledge graph, which defines the topological connections between devices and the operational scheduling logic. Upon receiving a set of latent fault modes from the device layer, this layer performs correlation deduction based on the graph. For example, a latent fault such as "abnormal current waveform of a turnout switch machine" can, based on the interlocking relationships and train operation rules defined in the knowledge graph, deduce a series of virtual business event chains that may lead to "route arrangement failure," "inaccurate train platform alignment," or even "decrease in section throughput capacity." This deduction process simulates the cascading amplification effect of a single device anomaly in the operation system.

[0070] The teaching-level task twin layer receives the aforementioned virtual business event chain and deconstructs the tasks by combining them with the teaching objectives corresponding to the "job training, competition, and certification" from the static standard library. This layer uses task decomposition and reorganization algorithms to transform the abstract event chain into a set of teaching tasks containing specific actions, roles, and resource requirements. For the aforementioned "turnout anomaly causing route failure" event chain, the algorithm automatically generates a multi-dimensional composite teaching task set based on the capability requirements for "emergency fault handling," "train operation adjustment," and "safety risk management" in the teaching objective library. This task set may include: a turnout on-site electrical inspection and data measurement task performed by one trainee; a route rearrangement and train number adjustment task performed by another trainee on the simulated dispatch terminal; and a fault information notification and safety protection setup task completed collaboratively by both trainees. The task set clearly defines the execution order, logical dependencies, required 3D models and procedural documents, and corresponding capability assessment dimensions for each sub-task.

[0071] The three-layer twin model features tightly coupled operational logic and progressively advancing data flow. The physical simulation of the equipment mechanism layer provides the system business layer with raw input conforming to engineering principles, avoiding the pitfalls of business simulations detached from physical reality. The rule-based simulation of the system business layer places equipment failures under realistic operational constraints, generating complex scenarios with job-specific realism. The algorithmic deconstruction of the teaching task layer transforms these complex scenarios into trainable and assessable specific teaching activities, directly linked to curriculum standards. This addresses the technical problems of virtual training scenarios being singular and static, disconnected from the dynamic complexity of real operation and maintenance, and insufficient matching between teaching tasks and comprehensive job competency requirements. By simulating the dynamic process of a failure from its latent occurrence to its systemic impact, and deconstructing it into composite tasks according to teaching principles, trainees can undergo systematic competency training covering the entire work process and involving multi-position collaboration in a highly realistic virtual environment, effectively enhancing the depth and effectiveness of integrated work-industry learning.

[0072] The multi-dimensional, composite teaching task set generated by the instructional twin module 102 provides a rich and structured content foundation for conducting work-integrated learning training. However, effectively transforming these tasks into a sequence of teaching activities that adapts to the specific ability levels of individual students and achieves optimal training results within limited teaching resources still requires a core engine with intelligent decision-making capabilities. To this end, the system introduces the work-integrated learning engine module 103. This module uses the task set output by the instructional twin module as "raw materials" and the student's ability profile as a "personalized blueprint," performing dynamic planning and optimization through a built-in reinforcement learning hybrid algorithm. Its function is to intelligently arrange a training path that progresses from easy to difficult, step by step, and highly adaptable to the student's real-time ability weaknesses by combining a static task list with dynamic learning status. This transforms pre-set standardized teaching tasks into dynamic, personalized teaching activity execution plans.

[0073] The integrated learning and work-integrated teaching engine module 103 incorporates an adaptive task generator based on a hybrid reinforcement learning algorithm. This generator automatically arranges the optimal sequence of teaching activities based on the student's ability profile and the multi-dimensional composite teaching task set. Specifically, the integrated learning and work-integrated teaching engine module includes: a student ability profile generator, which outputs a three-dimensional ability vector of "safety-efficiency-cost" using a graph neural network based on the student's historical training data and real-time eye-tracking, operation sequences, and voice command multimodal information; a course difficulty adjuster, which monitors the variance of the reward function and dynamically adjusts the difficulty of subsequent tasks according to the "zone of proximal development" principle if the variance exceeds a threshold; a knowledge graph gap diagnosis interface, which compares the four-dimensional knowledge graph of "job, course, competition, and certification" with the student's three-dimensional ability vector to generate a gap vector; and an adaptive task generator, which incorporates a hybrid reinforcement learning algorithm unit. This unit uses the concatenation of the three-dimensional ability vector and the embedded vector of the teaching task set as the state, task difficulty, task type, and resource combination as the action space, and the expected ability increment as the reward function, to output a preliminary sequence of teaching activities through an Actor-Critic architecture. After the adaptive task generator outputs the initial teaching activity sequence, it also includes rearranging the content and replacing tasks based on the gap vector to generate a revised teaching activity sequence, and outputting the optimal teaching activity sequence based on the task difficulty adjustment requirements fed back by the course difficulty adjuster.

[0074] More specifically, the work-integrated learning engine module 103, through the progressive collaborative work of its internal components, realizes a complete intelligent decision-making process from student status perception to the generation of personalized teaching sequences. The student competency profile generator continuously receives and processes multimodal data streams from practical training scenarios, such as the tool operation sequence of students in a virtual train braking system maintenance task, the time spent on each step, the content of discussions with the voice assistant about the causes of failures, and the duration of gaze at key dashboards recorded by VR glasses. A graph neural network is trained to model these temporal and correlated data, outputting a quantified three-dimensional competency vector, such as [Safety: 0.82, Efficiency: 0.58, Cost: 0.70], which objectively characterizes the student's real-time status in terms of procedure compliance, operational proficiency, and resource utilization.

[0075] The knowledge graph gap diagnosis interface compares this three-dimensional vector with the four-dimensional knowledge graph of "job skills, courses, competitions, and certifications" stored in the teaching data platform. This graph pre-defines target values ​​for each ability dimension and their associated knowledge and skill nodes. By calculating the difference, the interface generates a structured gap vector. For example, it clearly identifies a significant deficiency in the skill point of "brake cylinder piston stroke adjustment," with a quantified gap value of 0.35. This vector indicates the specific learning objectives that need to be strengthened.

[0076] The adaptive task generator's built-in reinforcement learning hybrid algorithm unit initiates the first round of decision-making, concatenating the three-dimensional capability vector with the embedding vector extracted from the teaching task set to form the initial state input. The algorithm defines the action space based on task difficulty level, task type (e.g., diagnosis, operation, collaboration), and resource combination (e.g., whether AR guidance is enabled), and constructs a reward function aimed at maximizing the incremental increase in predicted capability. Through iterative exploration of the Actor-Critic architecture, under the condition of satisfying basic teaching constraints, the algorithm outputs a preliminary sequence of teaching activities, such as [Task A: Learning about braking principles, Task B: Basic brake pad replacement practice].

[0077] Subsequently, the generator initiates an optimization and correction phase. Based on the gap vector provided by the knowledge graph gap diagnosis interface, the system rearranges the content and replaces tasks in the initial sequence. For the aforementioned "brake cylinder adjustment" gap, the algorithm may replace the "basic brake pad replacement drill" task in the sequence with a more suitable "brake cylinder stroke measurement and adjustment" specialized training task, generating a revised teaching activity sequence. During this process, the course difficulty adjuster runs in parallel, monitoring the variance of the reward function during reinforcement learning training. If the variance is detected to continuously exceed a preset threshold, indicating a systematic deviation between the current task difficulty and the suitability for the learner group, a global difficulty adjustment parameter is generated based on the "zone of proximal development" principle, for example, "uniformly reducing the baseline difficulty coefficient of all subsequent practical tasks by 10%." The adaptive task generator receives this adjustment requirement, performs final difficulty calibration on the revised sequence, and outputs the optimal teaching activity sequence.

[0078] In this module, the student competency profile generator and knowledge graph gap diagnosis interface provide quantifiable and accurate data on student status and learning objectives. The reinforcement learning algorithm unit provides the core capability for sequence optimization under complex constraints, while the course difficulty adjuster ensures the macro-adaptability of the optimization process and the rationality of teaching. These technical features are functionally mutually supportive and interactive, collectively forming a technical means that can dynamically respond to individual learning differences and intelligently generate personalized teaching paths. This solution effectively solves the technical problems in the background technology, such as instructional design relying on experience and difficulty in dynamically adapting to individual student competency gaps and learning paces. Through data-driven intelligent decision-making, it achieves a shift from general teaching arrangements to precise "one-person-one-policy" teaching, significantly improving the relevance and efficiency of intelligent operation and maintenance skills training for rail transit.

[0079] The optimal sequence of teaching activities output by the engineering-integrated learning engine module 103 clearly defines "what to train" and "in what logical order to train." However, to transform these abstract teaching instructions into a concrete training environment that students can perceive, interact with, and operate, an execution interface capable of achieving virtual-real integration and scenario construction is needed. Therefore, the system introduces the metaverse scene construction module 104.

[0080] The metaverse scene construction module 104 is used to construct a simulation scene for students to conduct interactive virtual operation training by calling corresponding virtual materials according to the optimal teaching activity sequence. Specifically, the metaverse scene construction module 104 is used to: receive the optimal teaching activity sequence and analyze its requirements for virtual environment, equipment status, fault phenomena, and interactive objects; based on the analysis results, intelligently retrieve and call matching parametric 3D models, material textures, animation scripts, and physics engine configuration parameters from the external resource library; based on a real-time rendering game engine, instantiate, spatially arrange, and logically bind the called materials in real time according to the teaching logic to generate an interactive immersive simulation scene synchronized with the optimal teaching activity sequence; wherein, the immersive simulation scene supports students to conduct fine operation training from a first-person perspective through their personal terminals, and also supports global situational simulation and role-playing from a third-person perspective in a collaborative sandbox.

[0081] Specifically, the metaverse scene construction module 104 transforms abstract teaching logic into a concrete, interactive, immersive training environment based on the dynamic instructions of the optimal teaching activity sequence. After receiving the optimal teaching activity sequence data from the engineering-integrated teaching engine module, this module first analyzes the specific requirements of each teaching step in the sequence for the constituent elements of the virtual environment. These requirements include the type of target training site, the rail transit equipment objects involved and their initial states, the characteristics of the fault phenomena to be simulated, and the interactive objects or roles expected to appear in this step. For example, for a teaching task step of "detection and replacement of abnormal wear of carbon sliding plate of train pantograph", the analysis results will clearly define the requirements as: a virtual roof working environment containing the contact wire and pantograph, a 3D model of the pantograph with the status parameter showing "carbon sliding plate thickness is below the threshold and there is uneven wear", particle effects simulating fault phenomena such as "sparks" or "abnormal noises", and virtual toolkits and safety protection equipment available to trainees.

[0082] Based on the parsed requirement description, this module performs intelligent retrieval and matching from an external resource library. The parametric 3D models in the resource library are associated with metadata tags, such as equipment name, moving parts, physical attributes, and configurable state variables; the fault case knowledge graph links to the animation scripts, sound files, and physics engine parameters required to represent specific faults. The system uses a semantic matching algorithm to compare requirement keywords such as "pantograph" and "carbon contactor wear" with the resource metadata, retrieving highly matching basic models, corresponding wear texture maps, disassembly and assembly animation sequences, and a set of physical parameters simulating the dynamic contact between the pantograph and the overhead contact line.

[0083] The retrieved materials are input into a real-time rendering game engine for scene composition. The engine instantiates and spatially arranges the materials in real time according to the teaching logic. For example, it precisely positions the pantograph model at designated coordinates on the roof of a virtual train, places tool models in an easily accessible tool cart, and sets a series of logical trigger points in the scene according to the pre-set work processes in the teaching sequence. These trigger points are bound to the student's actions on their personal terminal (such as viewpoint movement and virtual controller grabbing). When the student's virtual gaze focuses on a specific part of the carbon skateboard or performs a disassembly action, the scene will trigger interactive responses such as displaying precise measurement readings, playing disassembly animations, or providing force feedback. The scene constructed by this module has multimodal interactive capabilities, supporting students to perform precise first-person perspective operations through VR headsets and controllers, such as using a virtual measuring instrument to measure dimensions or using a wrench to install and remove bolts. Simultaneously, the scene also supports providing multiple students with a third-person global overview view in collaborative sandbox mode for task division and collaboration, train dispatching simulations, or safety monitoring role-playing. The scenario state is strictly synchronized with the teaching sequence, and the operation result of the previous step serves as the initial condition for the scenario in the next step.

[0084] By retrieving, assembling, and binding parameterized resources in real time based on dynamic teaching instructions, this module enables on-demand generation and high configurability of teaching scenarios, avoiding the limitations of traditional simulation software with its single scenario and fixed process. The tight binding of teaching logic with the state and interactive behavior of virtual objects allows abstract knowledge and skills requirements to be directly perceived and trained in a concrete operational environment. The high-fidelity, highly interactive, and immersive scenarios generated by this module not only provide students with a practical platform for "learning by doing," but more importantly, they provide a native environment for the subsequent competency evolution assessment module to collect multi-dimensional behavioral data from students throughout the entire process, making it possible to quantitatively analyze operational standardization, decision-making logic, and collaborative processes.

[0085] The immersive simulation scenario generated by the metaverse scenario construction module 104 provides trainees with a concrete operational space to execute the "optimal sequence of teaching activities," continuously generating a large amount of raw, real-time behavioral data in the process. However, this data is merely a record of the training process and requires systematic processing and analysis to be transformed into a scientific assessment of trainees' ability development and training effectiveness. Therefore, the system introduces the ability evolution assessment module 105. This module, acting as the "analysis engine" for training effectiveness, receives and processes data streams from the virtual scenario backend in real time. Based on the standard operating procedure model and multi-dimensional evaluation system pre-set in the teaching twin module, it quantitatively analyzes and compares the trainees' operational processes and interactive performance in the simulation environment. Its role is to aggregate and elevate the trainees' discrete, temporal behavioral traces in virtual training into a structured, measurable ability evolution map, thereby providing objective and accurate data feedback for the closed-loop optimization of teaching strategies.

[0086] The capability evolution assessment module 105 is used to collect the trainees' training data in the simulation scenario and compare it with the standard operation model in the teaching twin module to generate a capability increment map. The capability evolution assessment module 105 specifically includes: a data acquisition unit, used to extract the trajectory data, step completion time series, tool selection series, and communication content with other trainees from the background logs of virtual operation training; a multi-dimensional comparative analysis unit, used to compare the trajectory data, time series, and tool selection series with the preset optimal path, standard working hours, and standard toolset in the standard operation model, respectively, and calculate the operation standardization score, efficiency score, and decision accuracy score; perform natural language processing on the communication content, analyze its collaborative intent and problem-solving logic, and generate collaborative communication and innovative thinking scores; and a graph generation unit, used to calculate the trainee's capability level change in each capability dimension relative to the previous assessment based on the collaborative communication and innovative thinking scores of the multi-dimensional comparative analysis unit and the project response theory, and generate a visualized capability increment graph in the form of a combination of radar chart and capability node growth graph, wherein the trainee's capability dimensions include six dimensions: "task understanding, planning, decision implementation, inspection and control, evaluation and feedback, and transfer innovation".

[0087] Specifically, the capability evolution assessment module 105 collects and analyzes trainees' behavioral data in virtual scenarios to achieve a quantitative assessment and growth tracking of their comprehensive professional abilities. The data acquisition unit continuously extracts the raw data streams generated by trainees during practical training from the background operation logs of the metaverse scenario construction module. This data includes the coordinate sequence of the trainee's virtual avatar's movement and operation trajectory in three-dimensional space, the timestamp sequence of completing each teaching sub-step, the sequence of selecting and using different tools from the virtual toolbox, and the text content of communication with other trainees or intelligent characters through text or voice channels in collaborative tasks.

[0088] After receiving the above data, the multi-dimensional comparative analysis unit compares it item by item with the predefined standard operation model in the teaching twin module. The standard operation model includes the optimal operation path for a specific teaching task, the standard working time range for each step, the recommended tool selection sequence, and key safety regulations. For example, for the task of "replacing train brake pads," the unit calculates the spatial similarity between the trainee's actual movement trajectory and the optimal path (such as the shortest safe route from the tool cart to the brake cylinder) to obtain an operation standardization score; it compares the actual completion time with the standard working time to calculate an efficiency score; and it performs a matching degree analysis between the trainee's actual tool selection sequence (such as whether to use a torque wrench to pre-loosen the bolts before using a special puller) and the standard toolset and usage order to obtain a decision accuracy score. At the same time, the unit performs natural language processing on the collected communication content, and generates scores for two dimensions: collaborative communication and innovative thinking by analyzing the intent classification of the statements, information completeness, terminology accuracy, and whether the proposed solutions go beyond the conventional process. These processes transform discrete behavioral records into initial scores on six preset capability dimensions: "task understanding, planning, decision implementation, inspection and control, evaluation and feedback, and transfer innovation."

[0089] The graph generation unit is responsible for integrating scores from various dimensions and calculating ability increments. This unit applies the Item Response Theory (OPT) model, using scores from each dimension obtained within the current assessment period along with the student's historical assessment data as input. The OPT model estimates the student's potential ability value in each dimension, eliminating the impact of differences in task difficulty on the score. By comparing the ability value calculated in this assessment with that in the previous assessment, the unit calculates the student's "change in ability level" in each dimension, i.e., the ability increment. For example, if a student's ability value in the "decision implementation" dimension increases from 60% to 80%, the increment is recorded as 20%. Finally, the system visualizes the student's current absolute ability values ​​in the six dimensions using a radar chart and visually displays the incremental changes in each dimension compared to the previous assessment using a node growth chart. The combination of these two elements constitutes a complete graph of the student's ability increment. (See attached image) Figure 2 The diagram showing the virtual operation scores and evaluation radar diagrams for each ability dimension generated by the graph generation unit in the ability evolution assessment module visually presents the students' virtual operation performance: On the left, in the "Student Virtual Operation Score" section, scores for "Operational Efficiency" and "Decision Accuracy" are above 80 points, while the "Innovative Thinking" score is only 25 points, reflecting strong practical skills but insufficient collaborative innovation. On the right, in the "Student Ability Dimensions" section, scores for "Planning" and "Inspection and Control" are above 80%, while "Proactive Innovation" is only 20%, indicating that students' early planning abilities are up to standard, but collaborative innovation is weak. This graph can provide direct evidence for subsequent collaborative teaching tasks.

[0090] This module directly addresses the problems in the background technology of the intelligent operation and maintenance teaching evaluation system for rail transit, such as its singular approach, lack of process data collection, and delayed evaluation results that fail to reflect competency growth. By directly collecting native behavioral data from the entire process and multiple modalities in an immersive virtual environment, this module evaluates the trainees' operational process rather than just the final result; through refined comparison with standard operating procedures, it transforms subjective experience judgments into objective quantitative scores; and in particular, by introducing Project Response Theory to calibrate and compare intertemporal competency values, it shifts the evaluation focus from static "competency level" to dynamic "competency growth," achieving value-added evaluation.

[0091] The competency evolution assessment module 105 generates a competency increment map, providing a refined quantitative diagnosis of the teaching process's effectiveness. However, the assessment results themselves are not the end point; their core value lies in driving the self-optimization and iteration of the teaching system. To ensure that assessment conclusions can be transformed into practical strategies to improve the relevance and effectiveness of subsequent teaching, the system introduces a teaching feedback optimization module 106. This module, acting as a hub connecting "assessment" and "re-decision-making," receives and analyzes the common shortcomings and individual deficiencies revealed by the competency increment map, thereby generating specific teaching strategy adjustment instructions and resource reallocation suggestions. Its role is to transform the data insights generated in the assessment phase into actionable teaching interventions and feed them back to the upstream work-integrated learning engine module, thus initiating a new cycle of more targeted and personalized instructional design. The operation of the teaching feedback optimization module marks the completion of a full closed loop from "teaching implementation - effectiveness assessment" to "assessment feedback - strategy optimization," achieving adaptive evolution of the teaching process.

[0092] The teaching feedback optimization module 106 is used to feed back the capability increment map to the adaptive task generator, driving the generation of a new round of optimal teaching activity sequence. Specifically, the teaching feedback optimization module 106 includes: a strategy adjustment unit, used to analyze the capability increment map and identify common weaknesses in the learning group and specific weaknesses in individuals; a resource re-matching unit, used to re-match corresponding resource lists from the external resource library and static standard library of the teaching data platform module based on the identified common weaknesses in the learning group and specific weaknesses in individuals, the resource lists including case studies, micro-lessons, simulation faults, or extended reading materials; and a task regeneration instruction unit, used to package and send the updated common weaknesses in the learning group, specific weaknesses in individuals, and the re-matched resource list to the adaptive task generator of the integrated engineering and learning teaching engine module, triggering it to prioritize the generation of a new round of optimal teaching activity sequence targeting the common weaknesses and specific weaknesses in individuals in the new round of strategy iteration, the new round of optimal teaching activity sequence including reinforcement training tasks and personalized learning path suggestions.

[0093] The teaching feedback optimization module 106, based on the capability increment map generated by the capability evolution assessment module, drives the closed-loop optimization of teaching strategies and the dynamic adaptation of learning resources. After receiving the map data, the strategy adjustment unit first performs clustering and statistical analysis on the capability increments in the six dimensions of "task understanding, planning, decision implementation, inspection and control, evaluation and feedback, and transfer and innovation". This unit identifies dimensions that generally show negative growth or slow growth in a specific student group and marks them as common weak capability dimensions; at the same time, it also identifies capability shortcomings of individual students that significantly deviate from the group average in certain dimensions and marks them as individual-specific capability shortcomings. For example, the analysis may find that the current student group generally scores low in the "inspection and control" dimension, while a certain student has a prominent weakness in the "decision implementation" dimension.

[0094] The resource rematching unit adjusts the weak link identifiers output by the previous unit according to the strategy, initiating a new round of resource association retrieval. This unit accesses the static standard library and external resource library in the teaching data platform module, and retrieves learning materials with a higher degree of matching with the target weak ability dimension based on semantic association and preset mapping rules. For the common weak dimension "inspection and control", the system may match a series of micro-lecture videos and 3D interactive cases on "standardized inspection process of rail transit equipment" and "hidden defect characteristics that are easily overlooked in fault diagnosis" from the external resource library; for individual shortcomings in "decision implementation", it may select complex fault handling simulation tasks that emphasize "multiple solution comparison and risk assessment" from the fault case knowledge graph associated with the static standard library. The resource retrieval process not only considers the relevance of the content theme, but also evaluates whether the difficulty gradient of the materials and the teaching format are suitable for targeted reinforcement training.

[0095] The task regeneration instruction unit encapsulates the updated learning data into a structured instruction package. This package contains a list of identified common weaknesses, a description of each student's specific skill gaps, and a list of alternative resources recommended by the resource rematching unit for each type of weakness. This instruction package is sent to the adaptive task generator in the integrated engineering and learning teaching engine module. The adaptive task generator uses this instruction package as a crucial input for the next round of policy iteration, quantifying the weakness information within it as a component of the reinforcement learning state vector. When generating a new optimal sequence of teaching activities, the algorithm, while satisfying the original constraints, prioritizes combinations of teaching tasks that can effectively train the identified weaknesses and tends to utilize the relevant resources recommended in the instruction package to construct these tasks. For example, for a group weak in "inspection and control," the generator might orchestrate a virtual inspection task sequence containing multiple hidden defects, requiring students to check each item according to a standard procedure; for individuals with weaknesses in "decision implementation," a decision-making task requiring them to weigh multiple repair options, each with its own advantages and disadvantages, might be inserted into their individual learning path.

[0096] This technology directly addresses the problems of delayed feedback of evaluation results, lack of data support for adjusting teaching strategies, and the inability to form a continuous optimization loop in the background technologies. By automatically parsing the evaluation map into actionable strategy adjustment instructions and resource matching suggestions, and directly feeding them back to the task generation engine, the system achieves an automated link from learning outcome diagnosis to the implementation of teaching intervention. The analytical function of the strategy adjustment unit enables teaching optimization to take into account both group patterns and individual differences; the resource rematching unit ensures that intervention measures have highly relevant content support; and the task regeneration instruction unit completes the transformation from optimization intentions to specific teaching actions.

[0097] In addition, the system also includes a cross-platform collaboration and resource sharing module, which is connected to the teaching data platform module 101 and the work-study integrated teaching engine module 103. Specifically, it includes: a collaborative interaction unit, which adopts the WebRTC real-time communication protocol to realize voice / video interaction, operation screen synchronization, and text collaborative annotation between students and tutors, and between students, and supports multi-role permission hierarchical management; a learning data synchronization unit, which is based on a distributed database of cloud-edge collaborative architecture to synchronize students' learning progress, training records, and ability assessment results to the cloud in real time to ensure data consistency after multiple terminal logins; and an intelligent resource push unit, which is based on a collaborative filtering algorithm and combines the weak points in the student's ability profile to select matching customized learning resources from the external resource library of the teaching data platform and push them to the student's terminal in real time through the message push interface.

[0098] Specifically, the collaborative interaction unit deploys services based on the WebRTC protocol. When trainees or instructors access the system through different terminals, this unit establishes point-to-point real-time audio and video communication channels and data channels. In collaborative training tasks, such as conducting a simulation of "emergency handling of interlocking faults in depots," trainees playing the roles of dispatchers, on-site maintenance personnel, and safety supervisors can join the same virtual collaborative space through their respective personal terminals. This unit not only transmits audio and video streams but also synchronizes thumbnail views of key operation screens on each terminal or annotation information on the shared whiteboard, and supports access control, such as allowing only instructors to annotate on the shared screen or freezing the operations of a specific trainee. This low-latency interaction supports instant communication and collaborative decision-making among distributed roles.

[0099] The learning data synchronization unit interacts with the distributed database deployed in the cloud-edge collaborative architecture, responsible for maintaining the consistency of global data. When a student completes a virtual training session on their local edge terminal, this unit synchronizes incremental process data, such as operation logs and provisional assessment scores, to the central cloud database. Simultaneously, when the student logs in from another terminal, the unit retrieves their latest learning progress, complete competency profile, and historical training records from the cloud, ensuring a continuous learning experience. This synchronization mechanism allows students to seamlessly switch between immersive terminals in the training lab and desktop terminals on their personal computers, while the teaching progress and assessment data remain synchronized.

[0100] The intelligent resource recommendation unit, as an extension of personalized learning support, relies on student ability profile data from the integrated engineering and learning teaching engine module. This unit employs a collaborative filtering algorithm, which not only analyzes the weaknesses identified in the target student's ability profile but also references learning resources commonly accessed or highly rated by other student groups with similar ability profile characteristics. For example, if a target student has a low ability score in the "Traction System Fault Diagnosis" dimension, and the algorithm finds that other students with similar weaknesses generally show improved abilities after learning a specific micro-course on "Torsion Inverter IGBT Module Fault Principles and Troubleshooting," then this unit will mark this micro-course resource as highly relevant from the external resource library. The recommendation unit then sends this resource link and related recommendation reasons to the student's currently logged-in terminal interface in real time via an integrated message push interface. The timing of the push can be combined with the teaching process, such as triggering it immediately after a student completes a related task but the evaluation shows flaws in their decision-making logic.

[0101] This addresses the problems in the background technology, such as teaching being limited to fixed locations and single devices, isolated learning process data, and a lack of targeted guidance for after-class extension learning. The collaborative interaction unit, through a standardized real-time communication protocol, virtually aggregates physically dispersed students and instructors into the same teaching context, enabling remote training of complex operational processes requiring multi-role cooperation, breaking through the time and space limitations of traditional practical training. The learning data synchronization unit, through cloud-edge data synchronization, achieves continuous recording of the learning process and multi-terminal adaptation, providing a technological foundation for ubiquitous learning. The intelligent resource recommendation unit utilizes a collaborative filtering algorithm to combine group learning behavior patterns with individual skill gaps, achieving more socially intelligent resource recommendations and promoting self-directed and in-depth learning after class.

[0102] Furthermore, all modules are deployed on the same cloud-edge collaborative architecture, which includes: a central cloud for storing historical operation and maintenance big data, heterogeneous data, and cross-school teaching resources; an edge cloud deployed in the training base for supporting real-time computing and low-latency rendering of the teaching twin; terminal edge nodes embedded with personal immersive operation terminals and AR glasses for local posture tracking and gesture recognition; and a unified time-series database that uses a time-series structured multi-label TSMT (a storage engine structure designed specifically for time-series databases) data model to label and store real operation and maintenance events, teaching task events, and student behavior events.

[0103] Specifically, all functional modules of the system are deployed on the same cloud-edge collaborative architecture. This architecture uses layered processing based on data characteristics and computing needs to optimize overall performance. The central cloud, as a data aggregation and macro-management node, centrally stores historical operation and maintenance big data generated from the long-term operation of rail transit lines, heterogeneous data from the refined and integrated teaching data platform module, and standardized teaching resources accumulated from different colleges or enterprises. This data is characterized by its large volume, relatively low update frequency, but high global sharing value. The edge cloud is deployed within the training bases of specific vocational colleges or enterprise training centers. Its physical location is close to the users, and it is responsible for handling tasks with extremely high requirements for real-time computing and response latency. The iterative calculation of equipment-level physical simulation in the teaching twin module and the high frame rate rendering of complex 3D models in the metaverse scene construction module are all completed on the edge cloud to ensure the smoothness and real-time interaction of students in the virtual environment and avoid dizziness or asynchronous operation caused by network latency. The terminal edge node refers to the personal immersive operation terminal or AR glasses device directly used by the student. Its built-in sensors and processors are used to track and recognize the student's head posture and hand movements locally in real time, and upload the processed simplified action command stream instead of the original video stream to the edge cloud, thereby greatly reducing the uplink data transmission bandwidth requirements and further reducing end-to-end interaction latency.

[0104] The unified time-series database, serving as the core of data organization across the cloud, edge, and cloud layers, employs a time-series structured multi-label data model to store various events generated by the system. This model assigns a timestamp as the main label to each data record, along with multiple structured dimension labels, such as event type, associated device ID, involved student ID, and associated teaching task ID. For example, a record might be labeled "Timestamp T, Event Type: Student Operation Behavior, Device: Train B Braking Unit, Student: A, Task: Brake Pad Replacement Training," containing specific operational parameters of student A at that moment. This storage method enables the unified ingestion, aligned storage, and efficient correlation querying of heterogeneous but time-correlated data streams, including sensor data from real-world operations, task scheduling instructions generated by the teaching engine, and behavioral logs generated by students on their terminals. When teaching retrospective analysis or causal inference is required, the system can quickly retrieve and reconstruct a complete event sequence within a specific time period, from changes in real device status to the triggering of virtual tasks and student operational responses, based on the timestamp and multi-dimensional labels.

[0105] This cloud-edge collaborative architecture and its unified time-series data model address the issues of concentrated system computational load, high response latency, and difficulty in fusing and analyzing multi-source heterogeneous data in the background technologies. By offloading computationally intensive and real-time-critical tasks to the edge, it effectively ensures the quality of immersive training experiences. The unified time-series data model provides a reliable and efficient data foundation for refined analysis and intelligent decision-making throughout the teaching process by standardizing and associating cross-layer and multi-source events. This architecture and data model are not isolated; they are tightly coupled with and mutually supportive of the functional requirements of upper-layer application modules: the real-time computing needs of upper-layer modules define the necessity of edge cloud deployment; their access to historical data and cross-scenario knowledge relies on the storage of the central cloud; and the process data they generate depends on the unified time-series model for standardized governance and value extraction.

[0106] The embodiments of this invention have achieved significant technical effects through the implementation of the above-described technical solutions. The system successfully constructs a complete technical chain from data aggregation, dynamic twin modeling, intelligent task orchestration, immersive training, multi-dimensional capability assessment to closed-loop feedback optimization, realizing full-link intelligentization of the teaching process. Compared with existing technologies, this system not only solves specific problems such as static pre-setting of teaching scenarios, fragmented resources, and a single evaluation method, but more importantly, it achieves dynamic and precise integration of "work" and "learning" in teaching activities through deep data and model-driven approaches. Students can undergo systematic training targeting their individual skill gaps and covering the entire work process in a highly realistic virtual operation and maintenance environment, resulting in quantifiable and traceable continuous improvement in their comprehensive professional abilities. Simultaneously, the burden of instructional design for teachers is reduced, and the formulation and adjustment of teaching strategies gain objective data support. This system provides a complete technical solution with good operability and scalability for the large-scale, personalized, and efficient training of high-end technical and skilled personnel in the field of intelligent operation and maintenance of rail transit, effectively promoting the substantial implementation and innovative development of the work-integrated learning teaching model.

[0107] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0108] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An integrated engineering-learning intelligent teaching system for intelligent operation and maintenance of rail transit, characterized in that, include: The teaching data platform module is used to aggregate and process heterogeneous data for rail transit operation and maintenance teaching. The heterogeneous data includes static standard data, dynamic teaching process data, and external teaching resource data. The teaching twin module is used to dynamically construct a three-layer twin model containing equipment mechanisms, system business, and teaching tasks based on the heterogeneous data, generating a multi-dimensional composite teaching task set. Specifically, the equipment-level mechanism twin layer is used to perform fault simulation and state prediction based on the physical simulation model and historical operating data of key rail transit equipment, generating a set of potential latent fault modes. The system-level business twin layer, connected to the equipment-level mechanism twin layer, is used to correlate and deduce the set of latent fault modes with equipment topology relationships and operation scheduling rules based on the operation and maintenance procedure knowledge graph, generating a virtual business event chain affecting operational safety. The teaching-level task twin layer, connected to the system-level business twin layer, is used to receive the virtual business event chain and, in conjunction with the teaching objectives in the static standard library, decompose the event chain into a multi-dimensional composite teaching task set including fault handling, scheduling coordination, and safety briefing through task decomposition and recombination algorithms. The engineering-integrated teaching engine module has a built-in adaptive task generator based on a hybrid reinforcement learning algorithm, which is used to automatically arrange the optimal sequence of teaching activities according to the student's ability profile and the multi-dimensional composite teaching task set. The metaverse scene construction module is used to call corresponding virtual materials according to the optimal teaching activity sequence to construct a simulation scene for students to conduct interactive virtual operation training; wherein, the simulation scene constructed by the metaverse scene construction module keeps the state synchronized with the optimal teaching activity sequence, and the operation results of the students in the previous teaching step are instantiated as the initial conditions of the scene in the next step. The competency evolution assessment module is used to collect trainees’ training data in the simulation scenario and compare it with the standard operation model in the teaching twin module. Based on the project response theory, it calculates the changes in trainees’ competency levels in each competency dimension relative to the previous assessment and generates a visual competency increment map in the form of a combination of radar chart and competency node growth map. Trainees’ competency dimensions include six dimensions: “task understanding, planning, decision implementation, inspection and control, evaluation and feedback, and transfer and innovation”. The teaching feedback optimization module is used to feed back the incremental capability map to the adaptive task generator, driving the generation of a new round of optimal teaching activity sequence. The teaching feedback optimization module includes a task regeneration instruction unit, used to package and send the updated common weak capability dimensions of the student group and the individual specific capability shortcomings, along with a rematched resource list, to the adaptive task generator of the integrated work-study teaching engine module. This triggers the generator to prioritize the generation of a new round of optimal teaching activity sequence targeting the common weak capabilities and individual specific capability shortcomings in the next round of strategy iteration. The new round of optimal teaching activity sequence includes reinforcement training tasks and personalized learning path suggestions.

2. The engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to claim 1, characterized in that, The teaching data platform module specifically includes: A static standard library is used to store and associate the rail transit operation and maintenance capability standards, curriculum standard map and industry specifications that integrate "job, course, competition and certificate" in four dimensions. The curriculum standard map establishes a semantic mapping relationship between course knowledge points, skill points and items in the capability standards through ontology modeling. The dynamic process library is used to collect and structured store multimodal data throughout the teaching process in real time. The multimodal data includes the operation sequence logs of students in virtual operation training, the dialogue text generated by interaction with the intelligent teaching assistant, the role-playing and communication records in collaborative tasks, and the key node annotation data of the teacher based on the standard operation model for the students' operation process. The standard operation model is the operation and maintenance specification model preset in the teaching twin module. External resource library, used to index and associate parameterized 3D model library of rail transit equipment, fault case knowledge graph, maintenance procedure video clips and safety warning education materials.

3. The engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to claim 1, characterized in that, The integrated work-industry teaching engine module specifically includes: The trainee capability profile generator is used to output a three-dimensional capability vector of "safety-efficiency-cost" based on trainees' historical training data and real-time eye movement, operation sequence, and voice command multimodal information through graph neural networks. The course difficulty adjuster monitors the variance of the reward function. If the variance exceeds a threshold, the difficulty of subsequent tasks is dynamically adjusted downwards or upwards according to the "zone of proximal development" principle. The knowledge graph gap diagnosis interface is used to compare the four-dimensional knowledge graph of "job, course, competition and certification" with the three-dimensional ability vector of trainees to generate gap vectors; An adaptive task generator with a built-in reinforcement learning hybrid algorithm unit is used to output a preliminary teaching activity sequence through an Actor-Critic architecture. The state is the concatenation of the three-dimensional capability vector and the embedding vector of the teaching task set. The action space is the task difficulty, task type, and resource combination. The reward function is the expected capability increment.

4. The engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to claim 3, characterized in that, After the adaptive task generator outputs the initial teaching activity sequence, it also includes rearranging the content and replacing tasks based on the gap vector to generate a revised teaching activity sequence, and outputting the optimal teaching activity sequence based on the task difficulty adjustment requirements fed back by the course difficulty adjuster.

5. The engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to claim 2, characterized in that, The metaverse scene construction module is specifically used for: Receive the optimal teaching activity sequence and analyze its requirements for virtual environment, device status, fault phenomena and interactive objects; Based on the analysis results, the system intelligently retrieves and calls matching parametric 3D models, material maps, animation scripts, and physics engine configuration parameters from the external resource library. Based on a real-time rendering game engine, the called materials are instantiated, spatially laid out, and logically bound in real time according to the teaching logic, generating an interactive immersive simulation scene that is synchronized with the optimal teaching activity sequence. The immersive simulation scenario allows trainees to conduct detailed operational training from a first-person perspective through their personal terminals, while also supporting global situation simulation and role-playing from a third-person perspective in a collaborative sandbox.

6. The engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to claim 2, characterized in that, The capability evolution assessment module specifically includes: The data acquisition unit is used to extract the trajectory data of trainees' operations, the time series of step completion, the tool selection sequence, and the communication content with other trainees from the background logs of virtual operation training. The multi-dimensional comparative analysis unit is used to compare the trajectory data, time series, and tool selection sequence with the preset optimal path, standard working hours, and standard toolset in the standard operation model, respectively, and calculate the operation standardization score, efficiency score, and decision accuracy score; perform natural language processing on the communication content, analyze its collaborative intent and problem-solving logic, and generate collaborative communication and innovative thinking scores. The graph generation unit is used to generate scores based on the collaborative communication and innovative thinking scores of the multi-dimensional comparative analysis unit.

7. The engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to claim 2, characterized in that, The teaching feedback optimization module also includes: The strategy adjustment unit is used to analyze the capability increment map and identify the common weaknesses of the trainee group and the specific capability shortcomings of individuals. The resource rematching unit is used to rematch the corresponding resource list from the external resource library and static standard library of the teaching data platform module based on the common weak ability dimensions of the identified student group and the specific ability shortcomings of the individual. The resource list includes cases, micro-lessons, simulation faults or extended reading materials.

8. The engineering-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to claim 1, characterized in that, It also includes a cross-platform collaboration and resource sharing module, which is communicatively connected to the teaching data platform module and the work-integrated learning engine module, specifically including: The collaborative interaction unit adopts the WebRTC real-time communication protocol to realize voice / video interaction between students and tutors, synchronized operation screens, and collaborative text annotation between students, and supports multi-role permission hierarchical management; The learning data synchronization unit, based on a distributed database with a cloud-edge collaborative architecture, synchronizes students' learning progress, training records, and ability assessment results to the cloud in real time, ensuring data consistency after multiple terminal logins. The intelligent resource push unit, based on the collaborative filtering algorithm and combined with the weak points in the student's ability profile, selects and matches customized learning resources from the external resource library of the teaching data platform, and pushes them to the student's terminal in real time through the message push interface.

9. A work-integrated intelligent teaching system for intelligent operation and maintenance of rail transit according to any one of claims 1-8, characterized in that, All modules are deployed on the same cloud-edge collaborative architecture, which includes: The central cloud is used to store historical operational big data, heterogeneous data, and cross-school teaching resources; Edge cloud, deployed in the training base, is used to support real-time computing and low-latency rendering of the teaching twin; Terminal edge nodes are embedded in personal immersive operating terminals and AR glasses to perform posture tracking and gesture recognition locally; The unified time-series database adopts the time-series structured multi-label TSMT data model to label and store real operation and maintenance events, teaching task events, and student behavior events.

Images