Papermaking employee virtual practical training method and system combined with large model

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a papermaking employee virtual practical training method and system combined with a large model. According to the papermaking employee virtual practical training method and system combined with the large model, immersive interaction with a high-fidelity virtual papermaking production line is achieved through data collection, virtual simulation and large model interaction and reasoning. According to the method, a large-scale language model is introduced, so that the method has strong natural language understanding, deep reasoning and multi-modal information fusion capabilities, complex and changeable virtual environment states can be analyzed in real time, and intelligent diagnosis and guidance are performed in combination with natural language questions of employees; and even non-preset or multi-factor coupled fault scenes which cannot be processed by a traditional rule system can be processed, man-machine dialogue interaction similar to experienced experts is realized, and the intelligent level of the system on solving complex problems is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and relates to a virtual training method and system for paper mill employees that combines large models. Background Technology

[0002] The development of virtual reality and simulation technologies has brought revolutionary breakthroughs to skills training in the industrial sector. Existing virtual training systems, by constructing highly realistic 3D virtual environments, have to some extent solved the time, space, and safety limitations faced by traditional training. These systems typically include a physical space data acquisition module, capable of acquiring static equipment parameters, environmental data, and operating procedures from the production site, thus providing basic data support for the construction of virtual scenarios. Based on this, the virtual space module uses 3D modeling software to accurately reproduce the equipment form and operating logic of core processes in a paper production line, such as pulping, papermaking, and drying. Through programming, it achieves an operational experience highly consistent with the physical entities, enabling employees to operate virtual equipment in an immersive way using VR devices.

[0003] Meanwhile, the data center module is responsible for aggregating, processing, analyzing, and storing this data, supporting the simulation and feedback of training scenarios, and recording employee operational behaviors. Through this hybrid virtual-real simulation approach, existing virtual training systems allow employees to repeatedly practice equipment operation and familiarize themselves with process flows in a safe and controllable environment, and to a certain extent simulate common production anomaly scenarios, thereby improving their basic operational skills and initial ability to identify specific faults. This technical solution demonstrates significant advantages in improving training efficiency and reducing training costs, and has become a mainstream trend in the current industrial skills training field.

[0004] However, with the continuous development of related technologies and the increasingly stringent requirements of the paper industry for intelligent, flexible, and risk-managed production, some inherent characteristics of the aforementioned traditional virtual training systems at the principle level have gradually revealed their deep-seated limitations in addressing new challenges. Although existing systems can provide a high-fidelity visual and operational experience and simulate preset fault scenarios, their core intelligent interaction and problem diagnosis mechanisms still mainly rely on pre-written rules, fixed decision trees, or limited knowledge bases. This means that when employees encounter complex anomalies, ambiguous fault signals, or problems requiring deep reasoning in the virtual environment, existing systems struggle to provide dynamic, adaptive, and logically profound intelligent guidance.

[0005] For example, malfunctions in paper production lines are often the result of multiple coupled factors and chain reactions, rather than a single event. A slight fluctuation in a parameter can trigger a series of complex follow-up problems. The diagnostic process requires experienced experts to conduct multi-dimensional analysis, trace causal chains, and even use natural language to ask questions, verify hypotheses, and retrieve knowledge. Traditional virtual systems, when faced with problems that require highly abstract thinking, associative reasoning, and the integration of massive amounts of knowledge, often only provide rigid prompts or pre-set solutions, failing to simulate the flexible, heuristic, and conversational learning process between experts and trainees.

[0006] Even worse, its feedback mechanism lacks insight into employees' thought processes, failing to provide personalized and precise feedback and guidance based on employees' questioning intentions, knowledge gaps, or operational habits. This makes it difficult for employees to develop a deep understanding beyond simply knowing what to do, and to apply their knowledge to other situations when solving complex problems. Furthermore, with the continuous optimization of papermaking processes and the introduction of new equipment, the cost of updating and maintaining the knowledge base is high and significantly lagging, making the disconnect between virtual training content and the actual production frontier increasingly serious.

[0007] Therefore, how to overcome the bottlenecks of existing virtual training systems in terms of intelligent interaction with complex problems, deep knowledge reasoning, dynamic fault diagnosis, and personalized learning guidance, and achieve a more adaptive, real-time, and intelligent immersive training experience, has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0008] To achieve the above-mentioned objectives, this invention provides a virtual training method and system for paper mill employees that incorporates a large model. This system addresses the limitations of existing virtual training systems in areas such as intelligent interaction with complex problems, deep knowledge reasoning, dynamic fault diagnosis, and personalized learning guidance. It also enhances the professional skills and emergency response capabilities of paper mill employees while reducing training and time costs.

[0009] The virtual training system for paper mill employees proposed in this invention, which incorporates a large-scale model, includes: The physical space data acquisition module is set up at the paper production site to acquire static parameters, dynamic operating data, environmental data, and on-site operational behavior data of paper production equipment in real time, and transmit the data to the data center module. The virtual space simulation module is used to construct an interactive three-dimensional virtual environment that is highly consistent with the actual paper production line based on the data provided by the physical space data acquisition module. It can also dynamically simulate the equipment operation logic, material transfer process, energy conversion and fault evolution process through the physical simulation engine. The data center module is used to aggregate, process, analyze, and store the data acquired by the physical space data acquisition module and the simulation data generated by the virtual space simulation module. The core of the large-scale model interaction and reasoning module is to deploy a large-scale language model optimized for the paper industry, which is used to realize natural language interaction with employees, provide deep knowledge reasoning, dynamic fault diagnosis and personalized learning guidance, and interact with the data center module and virtual space simulation module. The training service platform module serves as the user interface and management center for the entire system. It provides training managers, instructors, and trainees with functions such as training scenario management, employee file management, performance evaluation and feedback, knowledge base updates and maintenance, remote collaboration and guidance, and system configuration and log management. It also interacts with the data center module.

[0010] Preferably, the physical space data acquisition module acquires data by deploying multiple sensor arrays.

[0011] Preferably, the virtual space simulation module is used to construct the three-dimensional virtual environment, which includes equipment models of the pulping workshop, papermaking workshop, and winding and finished product packaging areas; The physical simulation engine is based on simulation algorithms using Newtonian mechanics, fluid mechanics, or thermodynamics principles, achieving a simulation accuracy within an error range of 0.1%. It can simulate the dewatering process of pulp in the wire section, the fiber arrangement direction, and the formation of wet paper sheets, as well as the heat transfer after steam is introduced into the drying cylinder in the drying section, the paper sheet moisture evaporation curve, and the dynamic balance of temperature and humidity in the drying cylinder.

[0012] Preferably, the data center module uses a distributed storage architecture and a combination of relational and non-relational databases to store massive amounts of data.

[0013] Preferably, the large-scale language model in the large-scale model interaction and inference module is based on the Transformer architecture and contains at least 10 billion parameters, and is trained and optimized through the following steps: Domain knowledge pre-training uses massive amounts of text data from the paper industry for pre-training, aiming to enable the large model to master the professional terminology, concepts, processes, equipment principles, fault types and their typical symptoms, solutions, and industry standards of the paper industry. Instruction fine-tuning and alignment involves fine-tuning a pre-trained large model using a large number of high-quality, task-specific instruction-response pairs. The aim is to improve the large model's ability to understand user intent, generate accurate answers following instructions, and enhance the reliability and security of the generated content. Aligning reinforcement learning with human feedback involves introducing human experts to evaluate and rank the output of large models, using human preferences as reward signals, and further optimizing large models through reinforcement learning algorithms to make the generated content more in line with human cognitive habits, problem-solving logic, and best practices in the professional field.

[0014] Preferably, the large model interaction and reasoning module includes: Natural Language Understanding and Intent Recognition is used to receive natural language queries from employees via voice or text input, and to use deep learning models to perform lexical analysis, syntactic analysis, semantic analysis, and sentiment analysis on the input text to accurately identify the employee's query intent and key entity information; Knowledge retrieval and graph fusion are used to perform efficient knowledge retrieval by combining the knowledge graph of the papermaking industry constructed by it. The knowledge retrieval adopts a vector database combined with an inverted index, and obtains the most relevant knowledge fragments as context information by semantic similarity matching and keyword matching and inputs them into the large model. Multimodal information fusion and context awareness are used to acquire the current virtual environment status data fed back by the virtual space simulation module in real time through the data center module, and to fuse the structured data with the natural language queries of employees to form a comprehensive perception of the current training context.

[0015] Preferably, the large model interaction and reasoning module further includes: Deep reasoning and problem diagnosis are used to diagnose complex problems based on context awareness and retrieved knowledge. The reasoning process includes causal chain tracing, differential diagnosis, hypothesis testing, trend prediction and risk assessment, and solution optimization and suggestions. The prediction model adopts a time-series prediction algorithm. Personalized learning guidance and adaptive feedback are used to continuously monitor employees' operational behavior, decision-making process, questioning patterns and error types in the virtual environment. By constructing employee learning profiles, highly personalized real-time error correction, knowledge supplementation, heuristic questioning, adaptive difficulty adjustment, and emotion perception and motivation are provided. The personalized learning guidance is achieved by combining reinforcement learning and expert systems. Natural language generation and multimodal output are used to generate natural, fluent and accurate natural language responses based on reasoning results and learning guidance needs, and output them in speech form through speech synthesis technology. At the same time, in the virtual environment, they are presented in multiple forms such as text box prompts, oral expressions of virtual expert figures or device parameter highlighting.

[0016] Preferably, the functions of the training service platform module include: The training scenario management is used by instructors to create, edit, import and export different training scenarios, including setting specific process parameters, preset fault types, defining initial equipment status and configuring learning objectives. The scenario creation tool supports both drag-and-drop interface and script programming. Employee file management is used to record and manage each employee's basic information, training history, learning progress, skills assessment results, and personalized learning profiles generated by a large model; Performance evaluation and feedback are used to monitor employees' operational performance in virtual training in real time, automatically record key indicators such as operation steps, decision-making time, problem-solving efficiency, resource consumption, and safety compliance, and provide detailed performance reports and targeted improvement suggestions generated by the large model. The performance evaluation indicator system follows the relevant skill standards of the paper industry. The knowledge base is updated and maintained, allowing lecturers and experts to perform CRUD operations on the paper industry knowledge base on which the large model relies. This includes uploading new equipment information, updating process flows, supplementing fault cases, and revising operating procedures. The knowledge base supports version control and access management. Remote collaboration and guidance are used to support multiple employees to conduct collaborative training in the same virtual environment, or for instructors to remotely access the virtual scene to provide real-time guidance to employees, including voice calls, screen sharing, and virtual pointer indication.

[0017] A virtual training method for paper mill employees that incorporates a large-scale model includes the following steps: The data acquisition and preprocessing step (S1) involves real-time acquisition of static equipment parameters, dynamic operating data, environmental data, and employee on-site operational behavior data from the paper production site through the physical space data acquisition module. The data is preprocessed by cleaning, noise reduction, and feature extraction, and then transmitted to the data center module for storage. At the same time, a knowledge base for the paper industry is built and updated to provide data support for the training and inference of large-scale language models. The virtual environment construction and dynamic simulation step (S2) involves constructing a high-fidelity three-dimensional virtual paper production line based on the static equipment parameters and dynamic operating data stored in the data center module. It also integrates a physical simulation engine to simulate the dynamic processes of material flow, energy flow, and information flow, as well as the equipment operating logic and fault evolution trends. The virtual environment is presented through visualization rendering technology and supports access to multi-channel immersive interactive devices. The large-scale language model training and optimization step (S3) involves training and optimizing the large-scale language model in the data center module. This includes using massive amounts of text data from the paper industry to pre-train the base model with domain knowledge, further training it using manually labeled instruction-response pairs through instruction fine-tuning technology, and introducing paper industry experts to evaluate and rank the quality of the content generated by the large model through reinforcement learning and alignment with human feedback for optimization. The model training process is continuously iterated. Immersive virtual training and multimodal interaction steps (S4): Employees enter the virtual paper production line constructed by the virtual space simulation module through the multimodal interaction interface for immersive training, and directly operate virtual equipment, observe instrument readings, adjust process parameters, and deal with simulated production anomalies. The virtual space simulation module feeds back the employees' operating behavior, virtual equipment status changes, and simulation results to the data center module in real time. The large-scale model intelligent interaction and deep reasoning step (S5) involves the large-scale model interaction and reasoning module receiving natural language queries from employees in virtual training in real time, as well as virtual environment status data and employee operation data provided by the data center module. It performs natural language understanding and intent recognition, and then combines the paper industry knowledge graph and real-time contextual data to perform deep reasoning. Personalized feedback and adaptive learning steps (S6): The large model interaction and reasoning module performs real-time personalized analysis based on employees' questions, operational behaviors, decision-making paths and error patterns, and provides customized and accurate error correction, principle explanation, heuristic question guidance, knowledge point expansion and dynamic adjustment of learning paths through natural language generation technology. The performance evaluation and training management steps (S7) involve the training service platform module generating detailed performance evaluation reports and improvement suggestions based on employees' performance in virtual training. Training managers can use the platform to view employee learning progress, manage training scenarios, update the knowledge base, and optimize the entire training system.

[0018] Preferably, the large model intelligent interaction and deep reasoning step (S5) specifically includes: Context awareness and association are used to comprehensively analyze employee questions, current virtual device status, historical operation records, and simulated fault progress to accurately understand the contextual information of the current training scenario; Multi-factor fault diagnosis is used to address fault phenomena reported by employees. The large model uses its trained knowledge and reasoning ability to analyze possible causes from multiple dimensions, trace the causal chain of the fault, eliminate irrelevant factors, and form a preliminary diagnostic conclusion. Dynamic problem analysis and solution generation are used to dynamically analyze complex problems based on employee follow-up questions or changes in training scenarios, and generate multi-step, executable solutions. Knowledge integration allows for the integration of disparate knowledge of papermaking processes, equipment principles, and troubleshooting experience into a large-scale model, providing comprehensive cross-disciplinary and multi-disciplinary guidance in the process of solving complex problems.

[0019] Through the above technical solution, the present invention achieves the following beneficial effects: Overcoming the bottleneck of intelligent interaction in complex problems: Existing virtual training systems struggle to handle complex anomalies and ambiguous fault signals that are not pre-set. This invention introduces a large-scale language model, giving it powerful natural language understanding, deep reasoning, and multimodal information fusion capabilities. It can analyze complex and ever-changing virtual environment states in real time and provide intelligent diagnosis and guidance based on employees' natural language questions. It can even handle non-pre-set or multi-factor coupled fault scenarios that traditional rule-based systems cannot handle, achieving human-computer dialogue-like interaction similar to that of experienced experts, significantly improving the system's intelligence level in solving complex problems.

[0020] Achieving deep knowledge reasoning and dynamic fault diagnosis: The large-scale model of this invention has been pre-trained and finely optimized with massive amounts of paper industry data, mastering rich domain knowledge and reasoning logic. It can perform causal chain tracing, differential diagnosis, hypothesis testing, and trend prediction, simulating expert-level thinking processes in a virtual environment. This guides employees to conduct in-depth fault analysis and decision-making, rather than simply providing preset solutions, thereby cultivating employees' deep understanding of the "why" behind problems and their ability to solve problems by analogy.

[0021] Personalized adaptive learning guidance is provided: The large-scale model interaction and reasoning module of this invention can monitor employees' operational behaviors, questioning patterns, and error types in real time, and provide highly personalized and accurate feedback based on employees' learning profiles. This feedback includes immediate error correction, knowledge point supplementation, heuristic questioning, and adaptive difficulty adjustment, effectively overcoming the shortcomings of rigid and untargeted feedback in traditional systems, greatly improving learning efficiency and effectiveness, and ensuring that each employee receives guidance most suitable for their learning progress.

[0022] Reducing Knowledge Base Maintenance Costs and Lag: Traditional virtual training systems suffer from high costs and frequent delays in updating and maintaining their knowledge bases. This invention's large-scale model, through continuous learning, reinforcement learning, and alignment with human feedback mechanisms, can more flexibly and efficiently absorb the latest industry knowledge, expert experience, and failure cases, maintaining the real-time and cutting-edge nature of the knowledge base. This reduces the frequency and cost of manual maintenance, ensuring that training content is synchronized with the forefront of actual production.

[0023] Enhancing the immersion and realism of practical training: This invention combines a high-fidelity virtual space simulation module with precise equipment modeling, physics engine simulation, and multimodal interactive interfaces (such as VR / AR devices) to provide employees with an immersive practical training experience. The intelligent interaction of the large model further enhances the realism of the virtual environment, allowing employees to interact with real experts as if facing virtual challenges, greatly improving the participation and effectiveness of the training.

[0024] Optimizing the utilization and efficiency of training resources: This invention breaks through the limitations of traditional training in terms of time, space, and safety. Employees can practice high-risk operations repeatedly anytime and anywhere without occupying expensive production equipment or bearing actual operational risks. Through intelligent guidance from a large model, the training process is more efficient, significantly shortening the training cycle from novice to skilled worker, reducing training costs for paper manufacturing companies and improving the quality and efficiency of talent development. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the structure of the virtual training system for paper mill employees that incorporates a large model, as described in this invention. Figure 2 This is a schematic diagram illustrating the data interaction between modules in the virtual training system of this invention; Figure 3 This is a flowchart illustrating the virtual training method for paper mill employees that incorporates a large model, as described in this invention. Figure 4 for Figure 3 A detailed flowchart illustrating the large-scale language model training and optimization process in step S3; Figure 5 for Figure 3 A detailed flowchart illustrating the intelligent interaction and deep reasoning process of the S5 large model in the middle step; Figure 6 This is a schematic diagram illustrating a scenario where employees of this invention undergo immersive training in a virtual environment. Detailed Implementation

[0026] This invention provides a virtual training method and system for paper mill employees that integrates a large-scale model. This system and method aim to overcome the limitations of existing virtual training systems in areas such as intelligent interaction with complex problems, deep knowledge reasoning, dynamic fault diagnosis, and personalized learning guidance by deeply integrating large-scale language models with high-fidelity virtual simulation technology. Ultimately, it aims to improve the professional skills and emergency response capabilities of paper mill employees while significantly reducing training costs and time. This invention constructs a highly realistic and interactive virtual paper mill production environment and integrates a domain-specific large-scale model with powerful natural language processing and reasoning capabilities, enabling employees to engage in intelligent question answering, complex problem diagnosis, and personalized learning guidance during immersive virtual training.

[0027] Combination Figure 1-6 The structural diagram shown illustrates that the virtual training system for paper mill employees proposed in this invention, which incorporates a large model, mainly includes: a physical space data acquisition module, a virtual space simulation module, a data center module, a large model interaction and reasoning module, and a training service platform module. Data exchange between these modules is achieved through standardized communication interfaces, ensuring the real-time performance, collaborative nature, and high efficiency of the entire system.

[0028] Example 1: The physical space data acquisition module's core responsibility is to acquire various key data from the paper production site in real time and accurately, providing a solid data foundation for subsequent virtual simulation, data analysis, and large-scale model training. This module is typically deployed at various key workstations and equipment nodes in an actual paper production line, acquiring data through a high-precision sensor array. The static parameters that this module can acquire include, but are not limited to: the pulper model is HD-S3000, rated power 500kW, design speed 1500rpm; its detailed design drawings' dimensional data; the material properties of major structural components; the precise three-dimensional coordinates of the installation location; the calibration parameters of all integrated sensors; and the layout topology data of the entire production line, including each workshop.

[0029] Furthermore, the physical space data acquisition module also acquires dynamic operational data in real time, covering all stages of the papermaking process, including pulping, papermaking, drying, and winding. This data includes, but is not limited to: real-time temperature, pressure, flow rate, liquid level, rotation speed, current, voltage, vibration frequency, noise intensity, fiber concentration, pulp brightness, paper moisture content, basis weight, and density, as well as a series of real-time process variables such as equipment start-up and shutdown status, valve opening, pump speed, and motor load rate.

[0030] In addition, this module also collects environmental data, including workshop ambient temperature, humidity, dust concentration, light intensity, and harmful gas content. Simultaneously, to analyze the characteristics of employees' on-site operational behavior, the module deploys high-precision vision sensors and inertial measurement units to collect employees' operational trajectories, operation durations, operation frequencies, key operation points, and operation results.

[0031] The physical space data acquisition module achieves comprehensive data acquisition by deploying a high-precision, high-reliability sensor array. This array includes: Temperature sensors: Pt100 type resistance temperature detectors and K type thermocouples are used to measure the temperature of key components such as pulp, paper sheet and drying cylinder.

[0032] Pressure sensor: A diffused silicon pressure transmitter is selected to monitor pipeline pressure, vacuum level and hydraulic system pressure.

[0033] Flow meters: mainly electromagnetic flow meters and vortex flow meters.

[0034] Level gauges: Ultrasonic level gauges and hydrostatic level gauges are used to monitor the liquid level in storage tanks and vessels.

[0035] Speed ​​sensor: Employs Hall effect sensor and photoelectric encoder, installed in various rotating equipment, such as pumps, fans, rollers, etc.

[0036] Vibration sensor: A piezoelectric accelerometer is selected to monitor the vibration status of equipment such as bearings and motors, and to detect equipment failures at an early stage.

[0037] Current / voltage transformer: A 0.2-class high-precision transformer is used in conjunction with a smart meter to monitor motor load, power consumption and grid stability in real time.

[0038] Visual sensor: Employs industrial-grade high-definition cameras, combined with posture recognition and target detection algorithms, to achieve non-contact monitoring of employee operating behavior and equipment surface condition.

[0039] Environmental parameter sensors: including high-precision temperature and humidity sensors and various gas sensors, used to monitor workshop environmental safety and health indicators.

[0040] The physical space data acquisition module transmits the collected raw data to the data center module via an industrial communication protocol. The transmission medium primarily uses industrial-grade optical fiber or Category 5e / Category 6 shielded twisted-pair cable to ensure a transmission rate of no less than 100Mbps. In multi-point deployment and high-concurrency data stream scenarios, the peak transmission rate can reach 1Gbps. To ensure the real-time performance and integrity of data transmission, the system deploys time-sensitive networking technology and employs cyclic redundancy check (CRC) and sequence number mechanisms to verify the integrity of data packets. Simultaneously, to ensure data security, all transmitted data is encrypted using the AES-256 standard and end-to-end encrypted communication is performed using the TLS / SSL protocol to prevent data theft or tampering during transmission.

[0041] Example 2: The virtual space simulation module's main function is to construct a three-dimensional virtual environment that is highly consistent with the actual paper production line in terms of geometry, physical characteristics, motion logic, and interaction methods, and is capable of real-time interaction, based on the detailed data provided by the physical space data acquisition module. This virtual environment meticulously reproduces all core equipment models, auxiliary facilities, piping systems, electrical instruments, and control panels in the pulping workshop, papermaking workshop, and winding and finished product packaging areas.

[0042] When constructing the virtual environment, the virtual space simulation module employs professional 3D modeling software and advanced modeling techniques. Specifically, for the external structure and complex surfaces of the equipment, NURBS surface modeling technology is primarily used to ensure surface smoothness and geometric accuracy. For the internal mechanical structures, pipes, valves, and other components of the equipment, solid modeling and parametric modeling techniques are employed to facilitate precise dimensional control and subsequent parameter modifications. After all geometric models are constructed, topology optimization and polygon reduction are performed to improve rendering efficiency while maintaining visual appeal.

[0043] The generation of model materials and textures strictly follows the physically based rendering process. By creating base color maps, metallic maps, roughness maps, normal maps, and ambient occlusion maps, the virtual model can present optical reflection, diffuse reflection, and specular characteristics that are highly realistic to real devices under different lighting conditions.

[0044] The virtual space simulation module can accurately reproduce the internal structure, external shape, control panel layout, and pipe connection method of each device.

[0045] Furthermore, the virtual space simulation module integrates a powerful physical simulation engine to dynamically simulate equipment operation logic, material transfer processes, energy conversion, and fault evolution. This physical simulation engine is developed based on various physical principles such as Newtonian mechanics, fluid mechanics, and thermodynamics, and employs simulation algorithms such as finite element analysis and computational fluid dynamics. The simulation accuracy has been rigorously verified, achieving an error range of 0.1%, ensuring a high degree of consistency between the simulation results and the actual production process.

[0046] For example, in the wire section, the module can precisely simulate the dewatering process of the pulp on the forming wire, including stages such as free dewatering and vacuum dewatering, as well as the fiber alignment direction under the action of water flow, and calculate the formation rate, basis weight, and moisture content of the wet paper sheet in real time. In the drying section, the module can simulate the heat transfer path after steam enters the drying cylinder, accurately calculate the moisture evaporation curve of the paper sheet under different drying cylinder temperatures, steam pressures, and machine speeds, as well as the dynamic balance relationship between the surface temperature of the drying cylinder and the moisture content of the paper sheet and the ambient humidity. It can even simulate the impact of local overheating or condensate accumulation on the drying effect.

[0047] The virtual space simulation module also integrates a multi-channel virtual interaction interface, allowing employees to perform immersive operations using commercially available virtual reality headsets and high-precision hand tracking controllers. These operations include, but are not limited to: precisely opening and closing pipeline valves, starting / stopping various equipment, adjusting operating parameters, checking virtual instrument readings, performing equipment inspections, and simulating fault diagnosis. The virtual space simulation module receives real-time updates of physical world data from the data center module and intelligent commands from the large model interaction and inference module via a high-speed data bus.

[0048] Meanwhile, this module feeds back the sequence of employee actions in the virtual environment, changes in the status of virtual devices, and simulation results to the data center module in real time, ensuring bidirectional synchronization and consistency between virtual and real data. The data transmission protocol uses TCP / IP, with a maximum packet size of 64KB, and a heartbeat mechanism maintains the connection and ensures stable data flow.

[0049] Example 3: The data center module is the core data hub of the entire system. Its function is to efficiently aggregate, process, analyze, and store the massive amounts of raw data acquired by the physical space data acquisition module and the simulation data generated by the virtual space simulation module. To cope with the diversity, massive volume, and high concurrency characteristics of data in the paper industry, the data center module adopts a hybrid storage architecture. For massive amounts of unstructured and semi-structured data, a distributed storage architecture is used, specifically, for example, Hadoop HDFS or Ceph.

[0050] For structured data, such as equipment parameters, employee files, training records, and knowledge graph relational data, a combination of relational and non-relational databases will be used. The data storage capacity is designed to be no less than 100TB and has linear scalability, with a data throughput designed to be no less than 1GB / s to ensure support for high-concurrency data read and write operations.

[0051] The data processing flow includes, but is not limited to, the following steps: Data cleaning and preprocessing: This involves comprehensive and meticulous cleaning and preprocessing of the raw data to ensure data quality and consistency. Specific operations include: removing duplicate data points, filtering sensor noise, detecting and correcting outliers, performing data format conversion, time series alignment, and imputing missing values.

[0052] Data feature engineering involves extracting high-dimensional features related to papermaking processes, equipment operating status, and failure modes from cleaned raw data to improve the effectiveness of subsequent model training. For example, Fourier transform is used to convert vibration signals from the time domain to the frequency domain, extracting spectral features of key frequency components; wavelet analysis is used to extract transient signal features, suitable for detecting early equipment damage; sliding window techniques are used to calculate statistical features such as mean, variance, skewness, and kurtosis to capture dynamic changes in data distribution; and it also includes extracting multi-sensor fusion features, equipment operating condition features, and trend features based on historical data.

[0053] Data fusion: Deeply integrate data from different sources and modalities to construct a unified knowledge graph and multimodal dataset. This includes associating and fusing real-time sensor data with historical production records, maintenance logs, expert experience documents, employee operation logs, fault case libraries, industry standards and specifications, etc. The constructed knowledge graph is stored in a graph database and contains ontology relationships between equipment, components, faults, causes, symptoms, and solutions, as well as causal relationships between process flow, parameters, and effects, providing structured background knowledge for large models.

[0054] Real-time data analysis: Utilizing stream processing technology, the system performs real-time monitoring, trend analysis, anomaly detection, and early warning of incoming sensor data and virtual operation data. For example, when the system detects that the whiteness of the slurry is continuously below a set threshold at multiple points or that the local temperature of the drying cylinder is abnormally high, it can trigger an alarm within 100 milliseconds and display the abnormal trend in real time through a data visualization interface, supporting rapid response from training managers and employees.

[0055] Offline data analysis and model training: In-depth mining of aggregated and stored historical data is used to support the training of fault prediction models, production efficiency optimization models, and employee performance evaluation models. The model training employs a distributed computing framework to accelerate the model training process on large-scale datasets.

[0056] The data center module securely and efficiently interacts with the virtual space simulation module, the large model interaction and inference module, and the training service platform module through standard API interfaces. All APIs undergo strict authentication and authorization mechanisms to ensure the security of data access.

[0057] The large-scale model interaction and reasoning module is the core component of this invention. It innovatively introduces a large-scale language model to achieve natural language interaction with employees, provide deep knowledge reasoning, dynamic fault diagnosis, and personalized learning guidance. The core of this module lies in deploying a large-scale language model specifically optimized for the paper industry. This large model is based on an advanced Transformer architecture, and its training and optimization process involves the following refined steps: Domain knowledge pre-training: This is the foundation for building industry-specific large-scale models. We use massive amounts of high-quality text data from the paper industry to pre-train the base model. Internal enterprise data: strictly anonymized historical production logs, expert diagnostic records, training materials, Q&A responses, fault reports, and maintenance records.

[0058] The goal of the pre-training is to enable the large model to fully grasp the professional terminology, core concepts, complex processes, working principles of various equipment, common fault types and their typical symptoms, corresponding solutions, and strict industry standards and safety requirements of the papermaking industry. In this way, the large model can develop a deep and structured understanding of the papermaking field.

[0059] Instruction Fine-tuning and Alignment: Building upon domain knowledge pre-training, the large model undergoes instruction fine-tuning. This stage utilizes a large number of high-quality, task-specific instruction-response pairs for training to improve the model's ability to understand user intent, generate accurate answers following instructions, and enhance the reliability and security of generated content. Instruction design covers various typical scenarios, such as: Fault Diagnosis: The user commands fluctuations in steam pressure in the drying section and abnormal paper moisture content. Please analyze possible causes and provide troubleshooting steps. Large Model Response: Steam pressure fluctuations may be caused by steam valve malfunction, poor condensate drainage, or partial blockage in the steam pipeline. First, check if the steam trap is working properly and if condensate is draining. Second, check if the steam regulating valve opening is stable; try manual adjustment and observe the response. Finally, measure the pressure changes upstream and downstream of the steam pipeline to determine if there is a partial blockage and check if the steam filter is clean.

[0060] Operational Instructions: How to correctly start the pulper using user commands? Large Model Generation Response: Before starting the pulper, please ensure that the slurry level has reached the safe starting range, check that all safety guards are correctly in place, and confirm that there are no foreign objects inside the pulper or in the feeding area. Then, strictly follow the equipment operating procedures, opening the water supply valves sequentially to add water to the pulper to the preset level. After the level stabilizes, open the slurry feed valve. After a small amount of slurry enters, start the main motor of the pulper and slowly adjust it to the required speed, while observing whether the current and vibration are normal.

[0061] Principle Inquiry: User command: Please explain the dewatering principle of the press section. Large Model Generation Response: Dewatering in the press section mainly relies on mechanical squeezing to expel a large amount of water from the fiber web of the wet paper sheet. Its core principle is to utilize the linear pressure applied between one or more pairs of pressure rollers, subjecting the wet paper sheet to immense pressure as it passes through the narrow roller gap. Under the action of the pressure gradient, water is discharged from the capillary channels inside the paper sheet to the outer surface of the pressure rollers or the press felt. Simultaneously, the coatings on the pressure roller surfaces and the porous press felt have excellent water absorption and pressure transmission properties, effectively absorbing and guiding the transfer of moisture, thereby significantly reducing the paper sheet's moisture content and preparing it for subsequent drying.

[0062] Emergency Handling: How to handle a paper machine paper breakage as instructed by the user? Large Model Response: When a paper breakage occurs on the paper machine, the primary task is to ensure personnel safety and prevent further equipment damage. Immediately press the paper machine emergency stop button to cut off all power supply. Quickly and systematically remove the broken paper to prevent it from tangling in the equipment. Inspect the cause of the paper breakage, such as pulp quality issues, poor wire forming, press roller malfunction, or dryer cylinder sticking. If there are no other safety risks and the fault has been eliminated, prepare to re-feed the paper. During paper feeding, remain highly vigilant, strictly adhere to safety operating procedures, ensure the paper feeding path is clear, and monitor paper formation and transmission status.

[0063] By fine-tuning the instructions, the large model can more accurately understand the user's complex intentions and generate professional responses that conform to the expected format and logic.

[0064] Aligning Reinforcement Learning with Human Feedback: To further improve the quality, accuracy, and security of content generated by the large model, reinforcement learning incorporates feedback from human experts. In this phase, paper industry experts evaluate and rank the various outputs generated by the large model during the fine-tuning stage, providing detailed feedback. These human preferences serve as reward signals, further optimizing the large model through reinforcement learning algorithms. RLHF aims to ensure that the content generated by the large model is not only accurate but also more aligned with human cognitive habits, problem-solving logic, best practices in the professional field, and ethical and safety guidelines. For example, for the same question, the large model will learn to generate more structured, easier-to-understand, and safety-conscious answers.

[0065] Example 4: The large model interaction and inference module has the following core functions during operation: Natural Language Understanding and Intent Recognition: This module receives natural language queries from employees via voice or text input. It employs deep learning models, such as BERT or RoBERTa-based semantic analyzers, to perform lexical, syntactic, semantic, and sentiment analysis on the input text. Through these analyses, the system can accurately identify the employee's query intent, such as for problem diagnosis, operational guidance, principle queries, solution comparisons, safety inquiries, or fault prediction, and extract key entity information from the query, such as equipment name, fault symptoms, parameter values, and process steps.

[0066] Knowledge Retrieval and Graph Fusion (RAG Mechanism): Upon recognizing the employee's query intent, the module performs efficient knowledge retrieval by combining it with a pre-constructed knowledge graph of the paper industry. This knowledge graph contains ontology relationships between equipment, components, faults, causes, symptoms, and solutions, as well as causal relationships between process flow, parameters, and impacts, forming a complex knowledge network. Knowledge retrieval employs a vector database combined with an inverted index, using semantic similarity matching and keyword matching to retrieve the most relevant knowledge fragments to the current query from a massive knowledge base. These retrieved knowledge fragments are then used as enhanced contextual information and input into the large model along with the employee's original query, greatly enhancing the accuracy, relevance, and timeliness of the generated response. This Retrieval Enhanced Generation (RAG) mechanism effectively compensates for the timeliness deficiencies of knowledge within the large model and reduces the risk of illusions.

[0067] Multimodal Information Fusion and Context Awareness: The large-scale model interaction and reasoning module acquires real-time virtual environment status data from the virtual space simulation module via the data center module. This data is structured and includes, but is not limited to: real-time operating parameters of virtual devices, alarm information, the progress of simulated faults, the sequence of employee operational behaviors in the virtual environment, and visual feedback in the virtual scene. The module deeply integrates this structured data with the employee's natural language queries to form a comprehensive perception of the current training context. This fusion employs a multimodal encoder for feature extraction and fusion, enabling the large model to obtain complementary clues from different modal information, thereby more accurately understanding the context and making inferences.

[0068] Deep Reasoning and Problem Diagnosis: Based on context awareness and retrieved knowledge, the large model utilizes its powerful reasoning capabilities to diagnose complex problems. This reasoning process is not a simple information matching, but a deep logical deduction simulating expert thinking, including but not limited to: Causal chain tracing: Based on observed symptoms, large models can reverse-engineer possible chains of causes for failure, gradually delving from the symptoms to the root cause of the problem.

[0069] Differential diagnosis: Among multiple possible causes of failure, by comparing subtle differences in symptoms, changing trends of relevant parameters, and reviewing similar historical failure cases, the impossible is gradually eliminated, and finally the most likely cause of failure is focused.

[0070] Hypothesis validation: For specific fault hypotheses initially diagnosed by the large model, it can guide employees to perform validation operations in a virtual environment and dynamically adjust or confirm the fault hypotheses based on the feedback from these operations.

[0071] Trend Prediction and Risk Assessment: Combining historical operating data, fault development models, and real-time simulation data, this system predicts potential fault development trends in the current equipment state and their impact on production, assesses relevant risk levels, and assists employees in making risk decisions. The prediction model employs advanced time-series prediction algorithms, such as Long Short-Term Memory networks, gated recurrent units, or Transformer-based prediction models.

[0072] Solution Optimization and Suggestions: For the diagnosed problems, the large model can generate multiple possible solutions and comprehensively analyze the advantages and disadvantages, implementation costs, impact on production efficiency, and safety considerations of each solution to help employees choose the best response strategy. For example, for a motor overheating problem, the large model may provide multiple solutions such as checking the cooling fan, cleaning the vents, reducing the load, and checking the bearings, and analyze the applicability of each one.

[0073] Personalized learning guidance and adaptive feedback: This module continuously monitors employees' operational behavior, decision-making processes, questioning patterns, and error types in the virtual environment. By dynamically building and updating employee learning profiles, the large model can provide highly personalized and accurate feedback and guidance. This feedback includes, but is not limited to: Instant error correction: Provides real-time, specific, and principled corrections for employee erroneous operations. For example, when an employee attempts to inspect high-voltage electrical equipment without disconnecting the power, the system will immediately warn of the error: "Please disconnect the power first! Always ensure safety before inspecting electrical equipment, otherwise there may be a risk of electric shock," and explain the underlying safety procedures.

[0074] Additional information: When the large model detects that employees lack sufficient understanding of a certain concept or principle, it will proactively provide relevant background knowledge, graphic materials, or links to further reading materials.

[0075] Heuristic questioning: Instead of giving direct answers, it stimulates employees' independent thinking through questioning and guiding questions, helping them to understand not only what but also why, and cultivating their ability to solve problems independently.

[0076] Adaptive Difficulty Adjustment: Based on employees' learning progress and knowledge acquisition, the large model collaborates with the training service platform module to dynamically adjust the difficulty of subsequent training scenarios, the complexity of faults, and the challenge of the situations. For example, if employees perform well in handling simple valve jamming faults, the system will automatically introduce more complex fault scenarios with multiple coupled factors, such as pumping difficulties caused by abnormal slurry viscosity accompanied by increased vibration.

[0077] Emotional perception and motivation: By analyzing employees' voice tone, the big data model can determine their learning emotions and provide appropriate encouragement or adjust the interaction method to make it more approachable and improve the learning experience.

[0078] The personalized learning guidance is achieved by combining reinforcement learning with expert systems. Through continuous interaction and learning profile updates, the accuracy and effectiveness of the guidance are constantly improved.

[0079] Natural Language Generation and Multimodal Output: Based on deep reasoning results and personalized learning guidance needs, the module generates natural, fluent, and accurate natural language responses. These responses are output in speech form using speech synthesis technology, such as deep neural network models, to generate speech that closely resembles human intonation, emotion, and rhythm. Simultaneously, the response content is presented multimodally in the virtual environment through various forms such as text box prompts, virtual expert avatars, or highlighted device parameters, ensuring the clarity and effectiveness of information delivery and improving learning outcomes.

[0080] The training service platform module serves as the user interface and management center of the entire virtual training system, providing comprehensive and visual functional support for training managers, instructors, and trainees. The module's functions include, but are not limited to: Training Scenario Management: Instructors can create, edit, import, and export different training scenarios through this module. When creating a scenario, instructors can set specific initial values ​​for process parameters, preset multiple fault types, define clear learning objectives, and configure initial equipment states. The scenario creation tool supports an intuitive drag-and-drop graphical user interface and more advanced scripting methods, allowing instructors to flexibly define complex logic and event triggers.

[0081] Employee file management: Records and manages detailed basic information, training history, learning progress, skills assessment results, and personalized learning profiles generated in real time by a large model for each employee.

[0082] Performance Evaluation and Feedback: This module monitors employee performance in virtual training in real time, automatically recording key indicators such as operational steps, decision-making time, problem-solving efficiency, resource consumption, and safety compliance. It provides detailed performance reports, including quantitative data visualization charts and targeted improvement suggestions generated from large-scale models. The performance evaluation indicator system strictly adheres to relevant skill standards in the paper industry and the company's internal SOPs.

[0083] Knowledge Base Updates and Maintenance: Instructors and experts can use this module to perform CRUD operations on the paper industry knowledge base upon which the large model relies. This includes uploading new equipment information, updating process flow diagrams and descriptions, supplementing the latest fault cases and their solutions, revising operating procedures, and adding new industry standards. The knowledge base supports version control and fine-grained access control, ensuring that only authorized personnel can make modifications and guaranteeing the accuracy and authority of the knowledge.

[0084] Remote Collaboration and Guidance: Supports collaborative training among multiple employees in the same virtual environment, simulating teamwork to solve complex problems. For example, one employee can operate while another observes and records parameters. Simultaneously, instructors can remotely access the virtual environment to provide real-time guidance to employees, including providing instructions via voice calls, displaying shared content on the screen within the virtual environment, using virtual pointers to indicate key equipment or operational points, and even directly taking over some virtual equipment for demonstration operations.

[0085] System configuration and log management: Allows for parameter configuration, user permission management, and recording and querying of system operation logs for the entire virtual training system, ensuring stable operation and traceability. Log management includes detailed operation logs, system event logs, and fault logs, facilitating system administrators in maintenance and troubleshooting.

[0086] Combination Figure 3 The flowchart shown illustrates a virtual training method for paper mill employees that combines a large model, proposed in this invention. The method includes the following steps: Step S1: Data acquisition and preprocessing.

[0087] This step utilizes the physical space data acquisition module to collect various types of data from the paper production site in real time, comprehensively, and from multiple dimensions. This includes, but is not limited to, static parameters and dynamic operating data of the equipment, as well as workshop environmental data and employee on-site operational behavior data.

[0088] The collected raw data first undergoes rigorous data cleaning, including noise reduction, deduplication, and outlier detection and repair, to ensure data quality and consistency. Subsequently, feature engineering is performed to extract features strongly correlated with equipment status, process flow, and failure modes from the raw data. This preprocessed data is then transmitted to the data center module for long-term storage via industrial communication protocols and encrypted transmission.

[0089] Simultaneously, a knowledge base for the paper industry is continuously built and updated within the data center module. This knowledge base is a dynamically evolving comprehensive information repository containing massive amounts of structured and unstructured data. This data, after cleaning, annotation, and indexing, forms a high-quality corpus and knowledge graph, providing robust data support for the subsequent training and inference of large-scale language models. The knowledge base's update mechanism ensures the real-time nature and authoritativeness of its content.

[0090] Step S2: Virtual environment construction and dynamic simulation.

[0091] This step is completed in the virtual space simulation module, which constructs a high-fidelity, immersive 3D virtual paper production line based on the static equipment parameters and dynamic operating data stored in the data center module. The construction process is a multi-stage engineering workflow: First, using professional 3D modeling software, the geometry, dimensions, appearance details, and internal structure of all key equipment in the paper production line are accurately reproduced based on the design drawings and parameters of the actual equipment. For example, the pulper's cylinder and impeller, the refiner's blade gaps, the forming wire in the wire section, the press rollers and their covering layers, and the drying cylinder arrangement and steam pipes in the drying section are all modeled with millimeter-level precision and given realistic material properties and texture maps.

[0092] Secondly, an advanced physical simulation engine is integrated to achieve dynamic simulation of the papermaking process. This includes, but is not limited to: pulp transport in pipes, uniform distribution in the headbox, dewatering and forming process in the wire section, mechanical dewatering of the wet paper sheet in the press section, and heat transfer, moisture evaporation curves, and dynamic balance of temperature and humidity in the drying cylinder in the drying section. The physical engine also simulates the operating logic of the equipment and the evolution trends of various faults.

[0093] Secondly, based on the actual production process, the initial operating parameters and various controllable variables of the virtual environment are set to ensure that the parameter range and operating characteristics of the simulation environment are highly consistent with those of the real production scenario.

[0094] Finally, the virtual environment is presented using high-performance visualization rendering technology, providing employees with a realistic visual experience. This environment supports multi-channel immersive interactive device access, such as VR headsets, hand tracking controllers, and force feedback devices, enabling employees to operate and perceive as if they were actually there. The virtual space simulation module feeds back employee actions, virtual device status changes, and simulation results to the data center module in real time, ensuring data synchronization and sharing.

[0095] Step S3: Large-scale language model training and optimization.

[0096] Combination Figure 4 The detailed flowchart shown illustrates that this step involves training and optimizing a large-scale language model on the high-performance computing cluster of the data center module to equip it with expertise in the paper industry and powerful reasoning capabilities. The training process is continuously iterated, incorporating new domain knowledge and expert experience to maintain the model's advanced nature.

[0097] The specific training process includes: Phase 1: Domain Knowledge Pre-training. First, a large-scale pre-training of a general-purpose base model is conducted using the massive amount of textual data on the paper industry domain constructed in step S1. This data covers all aspects of the papermaking process, including but not limited to equipment structure principles, process parameters, fault diagnosis procedures, safe operating procedures, historical maintenance records, and expert experience. Through this phase of pre-training, the large model can comprehensively master the professional terminology, core concepts, deep knowledge structure, and language patterns of the papermaking industry, establishing a broad understanding of this domain.

[0098] Phase Two: Instruction Fine-Tuning. Building upon pre-training, the large model undergoes further refinement training using instruction fine-tuning techniques. This phase utilizes a large dataset of high-quality, human-annotated instruction-response pairs. These pairs cover various specific tasks and problem types that employees in the paper industry might encounter, such as fault diagnosis, operational guidance, principle lookup, and emergency handling. Through instruction fine-tuning, the large model is trained to understand and precisely follow user instructions, generating accurate, professional responses that meet specific task requirements, significantly improving its ability to solve concrete problems.

[0099] Phase Three: Alignment of Reinforcement Learning with Human Feedback. To further optimize the large-scale model's reasoning logic, answer accuracy, naturalness and professionalism of language expression, and to ensure the security and reliability of its output, paper industry experts were brought in for human feedback alignment. The expert team evaluated and ranked the various answers generated by the large-scale model during the instruction fine-tuning phase, providing detailed feedback information. This human preference data was transformed into reward signals, which were used to iteratively optimize the parameters of the large-scale model through reinforcement learning algorithms. The purpose of aligning reinforcement learning with human feedback is to enable the large-scale model to generate answers that better align with human values, logical thinking habits, and industry best practices, thereby making the model's output more practical and credible.

[0100] The model training process is a continuous iterative and feedback-driven optimization process. As new domain knowledge, expert experience, failure cases, and production data are continuously added to the knowledge base, the large model will undergo incremental training periodically or as needed to maintain the real-time and cutting-edge nature of its knowledge base, ensuring that it can always provide the most advanced and accurate intelligent guidance.

[0101] Step S4: Immersive virtual training and multimodal interaction.

[0102] Employees access a high-fidelity virtual paper production line constructed by the virtual space simulation module through the multimodal interaction interface. Employees can undergo immersive practical training in this virtual environment.

[0103] During the training, employees can directly operate virtual equipment, observe virtual instrument readings, adjust process parameters, and respond to simulated production anomalies and malfunctions preset by the system or dynamically injected into a large model. For example, employees may need to diagnose and handle simulated malfunctions such as sudden drops in steam pressure in the drying section or abnormal vibrations in the pressing section.

[0104] The virtual space simulation module captures all employee actions, changes in virtual device status, and simulation results in real time within the virtual environment. This detailed data is transmitted to the data center module in real time and efficiently, providing foundational data for subsequent large-scale intelligent interaction and performance evaluation.

[0105] Step S5: Large-scale intelligent interaction and deep reasoning.

[0106] Combination Figure 5 The detailed flowchart shown illustrates the core role of the large-scale model interaction and reasoning module in this step. It receives queries submitted by employees in real-time using natural language during the immersive virtual training process, while also receiving current virtual environment status data and historical employee operation data provided by the data center module.

[0107] The large model first performs accurate natural language understanding and intent recognition, conducting lexical, syntactic, and semantic analysis on the employee's questions to analyze the employee's questioning intent and key entity information in the questions.

[0108] Subsequently, combining its internally constructed knowledge graph of the paper industry and real-time contextual data, the large model performs multi-dimensional and in-depth reasoning, including but not limited to: Context awareness and correlation: The large model comprehensively analyzes the content of employees' questions, the real-time operating status of virtual devices in the simulation environment, employees' historical operation records, and the progress of simulated faults.

[0109] Multi-factor fault diagnosis: For fault phenomena reported by employees, the large model utilizes its rich knowledge and powerful reasoning capabilities acquired during training in step S3 to analyze potential causes from multiple underlying dimensions. It can trace the causal chain of the fault and gradually eliminate irrelevant factors, forming preliminary, progressively layered diagnostic conclusions.

[0110] Dynamic Problem Analysis and Solution Generation: The large-scale model not only provides initial fault diagnosis but also dynamically analyzes complex problems based on employee inquiries or changes in training scenarios. It can generate multi-step, actionable solutions. For example, for a pump malfunction, the large-scale model can guide employees through a series of troubleshooting steps: first, check if the pump's power supply is normal; then, measure the pressure difference between the pump inlet and outlet; and finally, check for abnormal vibrations or noises in the pump body. Based on feedback at each step, subsequent guidance will be provided, and the guidance content will be dynamically adjusted based on the feedback at each step.

[0111] Knowledge Integration: The large-scale model can deeply integrate and synthesize information scattered across different knowledge domains. In solving complex problems, it can provide comprehensive guidance across fields and disciplines. For example, a problem of insufficient paper strength may involve multiple aspects such as the beating control in the pulping process, the forming conditions in the wire section, the dewatering efficiency in the press section, and the sizing amount in the sizing machine. The large-scale model can comprehensively analyze these aspects and provide integrated suggestions.

[0112] Step S6: Personalized feedback and adaptive learning.

[0113] In this step, the large-scale model interaction and reasoning module acts as an intelligent tutor during the employee's learning process. It performs real-time and detailed personalized analysis based on the employee's questioning patterns, operational behavior sequences, decision-making path selections, and error types encountered during training. Through natural language generation technology, the large-scale model provides customized and inspiring feedback and guidance, aiming to maximize employee learning effectiveness and efficiency. This feedback includes, but is not limited to: Precise Error Correction and Explanation of Principles: For employee errors or inappropriate decisions, the large model immediately provides specific and precise corrective suggestions. More importantly, it delves into the underlying physical or technological principles, helping employees understand why their actions are wrong and why proper operation is necessary. For example, if an employee misoperates while cleaning press rolls, the large model will point out the error and explain: Cleaning press rolls requires stopping the machine, disconnecting the power, and ensuring the roll surface cools down, as the high-speed rotating rolls and high-temperature steam can cause burns or entrapment hazards.

[0114] Heuristic questioning: The large model does not directly provide the final answer, but instead guides employees to think independently, analyze problems, and explore solutions by posing a series of thought-provoking questions, thereby cultivating their ability to solve problems independently and their critical thinking.

[0115] Knowledge expansion and deep learning: When the large model discovers through interaction that employees have knowledge gaps or insufficient understanding in a specific field or concept, it will proactively expand relevant knowledge points or recommend more in-depth learning resources to promote employees' in-depth understanding of papermaking technology and improve their knowledge system.

[0116] Dynamic adjustment of learning paths: The training service platform module combines the assessment results of employees' learning progress, knowledge mastery, and skill bottlenecks from a large model to dynamically adjust the difficulty, complexity, and content of subsequent practical training scenarios. For example, if an employee performs poorly in handling pump malfunctions, the system will automatically add more different types of pump malfunction training scenarios or provide more detailed explanations of pump equipment principles and repair videos. Conversely, if an employee has already mastered a skill, the system will recommend more challenging comprehensive tasks.

[0117] The large model interaction and inference module continuously records every interaction data, learning feedback content, and update status of the employee's learning profile, and transmits it to the data center module for further system optimization and model iteration.

[0118] Step S7: Performance evaluation and training management.

[0119] This step is handled by the training service platform module. It generates a comprehensive and detailed performance evaluation report based on all employee performance data from the virtual training. These evaluation indicators include, but are not limited to: operational accuracy, problem-solving efficiency, safety compliance, resource consumption, and decision-making ability.

[0120] The performance evaluation report includes not only quantitative analytical data but also personalized qualitative evaluations and targeted improvement suggestions generated by a large model. For example, the report might indicate that while employees generally followed the correct procedures when handling paper breaks in the wire section, they spent too much time identifying the cause of the break, and therefore recommended strengthening their training in fiber clogging identification.

[0121] Training managers and instructors can use the training service platform module to comprehensively view all employees' learning progress, individual performance reports, manage and create new training scenarios, update and maintain the knowledge base, and conduct remote collaborative guidance. These functions enable training managers to have macro-level control and continuous optimization of the entire training system, ensuring the effectiveness and timeliness of training content, thereby continuously improving the overall skill level and productivity of employees in the paper manufacturing enterprise.

[0122] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "inclusion" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0123] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A virtual training system for paper mill employees that incorporates a large-scale model, characterized in that: include: The physical space data acquisition module is set up at the paper production site to acquire static parameters, dynamic operating data, environmental data, and on-site operational behavior data of paper production equipment in real time, and transmit the data to the data center module. The virtual space simulation module is used to construct an interactive three-dimensional virtual environment that is highly consistent with the actual paper production line based on the data provided by the physical space data acquisition module. It can also dynamically simulate the equipment operation logic, material transfer process, energy conversion and fault evolution process through the physical simulation engine. The data center module is used to aggregate, process, analyze, and store the data acquired by the physical space data acquisition module and the simulation data generated by the virtual space simulation module. The core of the large-scale model interaction and reasoning module is to deploy a large-scale language model optimized for the paper industry, which is used to realize natural language interaction with employees, provide deep knowledge reasoning, dynamic fault diagnosis and personalized learning guidance, and interact with the data center module and virtual space simulation module. The training service platform module serves as the user interface and management center for the entire system. It provides training managers, instructors, and trainees with functions such as training scenario management, employee file management, performance evaluation and feedback, knowledge base updates and maintenance, remote collaboration and guidance, and system configuration and log management. It also interacts with the data center module.

2. The virtual training system for paper mill employees combining a large model as described in claim 1, characterized in that, The physical space data acquisition module acquires data by deploying multiple sensor arrays.

3. The virtual training system for paper mill employees combining a large model as described in claim 1, characterized in that, The virtual space simulation module is used to construct the three-dimensional virtual environment, which includes equipment models of the pulping workshop, papermaking workshop, and winding and finished product packaging areas. The physical simulation engine is based on simulation algorithms using Newtonian mechanics, fluid mechanics, or thermodynamics principles, achieving a simulation accuracy within an error range of 0.1%. It can simulate the dewatering process of pulp in the wire section, the fiber arrangement direction, and the formation of wet paper sheets, as well as the heat transfer after steam is introduced into the drying cylinder in the drying section, the paper sheet moisture evaporation curve, and the dynamic balance of temperature and humidity in the drying cylinder.

4. The virtual training system for paper mill employees combining a large model as described in claim 1, characterized in that, The data center module uses a distributed storage architecture and a combination of relational and non-relational databases to store massive amounts of data.

5. The virtual training system for paper mill employees combining a large model according to claim 1, characterized in that, The large-scale language model in the large-scale model interaction and inference module is based on the Transformer architecture and contains at least 10 billion parameters, and is trained and optimized through the following steps: Domain knowledge pre-training uses massive amounts of text data from the paper industry for pre-training, aiming to enable the large model to master the professional terminology, concepts, processes, equipment principles, fault types and their typical symptoms, solutions, and industry standards of the paper industry. Instruction fine-tuning and alignment involves fine-tuning a pre-trained large model using a large number of high-quality, task-specific instruction-response pairs. The aim is to improve the large model's ability to understand user intent, generate accurate answers following instructions, and enhance the reliability and security of the generated content. Aligning reinforcement learning with human feedback involves introducing human experts to evaluate and rank the output of large models, using human preferences as reward signals, and further optimizing large models through reinforcement learning algorithms to make the generated content more in line with human cognitive habits, problem-solving logic, and best practices in the professional field.

6. The virtual training system for paper mill employees combined with a large model according to claim 5, characterized in that, The large model interaction and reasoning module includes: Natural Language Understanding and Intent Recognition is used to receive natural language queries from employees via voice or text input, and to use deep learning models to perform lexical analysis, syntactic analysis, semantic analysis, and sentiment analysis on the input text to accurately identify the employee's query intent and key entity information; Knowledge retrieval and graph fusion are used to perform efficient knowledge retrieval by combining the knowledge graph of the papermaking industry constructed by it. The knowledge retrieval adopts a vector database combined with an inverted index, and obtains the most relevant knowledge fragments as context information by semantic similarity matching and keyword matching and inputs them into the large model. Multimodal information fusion and context awareness are used to acquire the current virtual environment status data fed back by the virtual space simulation module in real time through the data center module, and to fuse the structured data with the natural language queries of employees to form a comprehensive perception of the current training context.

7. The virtual training system for paper mill employees combining a large model according to claim 6, characterized in that, The large model interaction and reasoning module also includes: Deep reasoning and problem diagnosis are used to diagnose complex problems based on context awareness and retrieved knowledge. The reasoning process includes causal chain tracing, differential diagnosis, hypothesis testing, trend prediction and risk assessment, and solution optimization and suggestions. The prediction model adopts a time-series prediction algorithm. Personalized learning guidance and adaptive feedback are used to continuously monitor employees' operational behavior, decision-making process, questioning patterns and error types in the virtual environment. By constructing employee learning profiles, highly personalized real-time error correction, knowledge supplementation, heuristic questioning, adaptive difficulty adjustment, and emotion perception and motivation are provided. The personalized learning guidance is achieved by combining reinforcement learning and expert systems. Natural language generation and multimodal output are used to generate natural, fluent and accurate natural language responses based on reasoning results and learning guidance needs, and output them in speech form through speech synthesis technology. At the same time, in the virtual environment, they are presented in multiple forms such as text box prompts, oral expressions of virtual expert figures or device parameter highlighting.

8. The virtual training system for paper mill employees combining a large model according to claim 1, characterized in that, The functions of the training service platform module include: The training scenario management is used by instructors to create, edit, import and export different training scenarios, including setting specific process parameters, preset fault types, defining initial equipment status and configuring learning objectives. The scenario creation tool supports both drag-and-drop interface and script programming. Employee file management is used to record and manage each employee's basic information, training history, learning progress, skills assessment results, and personalized learning profiles generated by a large model; Performance evaluation and feedback are used to monitor employees' operational performance in virtual training in real time, automatically record key indicators such as operation steps, decision-making time, problem-solving efficiency, resource consumption, and safety compliance, and provide detailed performance reports and targeted improvement suggestions generated by the large model. The performance evaluation indicator system follows the relevant skill standards of the paper industry. The knowledge base is updated and maintained, allowing lecturers and experts to perform CRUD operations on the paper industry knowledge base on which the large model relies. This includes uploading new equipment information, updating process flows, supplementing fault cases, and revising operating procedures. The knowledge base supports version control and access control. Remote collaboration and guidance are used to support multiple employees to conduct collaborative training in the same virtual environment, or for instructors to remotely access the virtual scene to provide real-time guidance to employees, including voice calls, screen sharing, and virtual pointer indication.

9. A virtual training method for paper mill employees combining large-scale models, characterized in that, Includes the following steps: The data acquisition and preprocessing step (S1) involves real-time acquisition of static equipment parameters, dynamic operating data, environmental data, and employee on-site operational behavior data from the paper production site through the physical space data acquisition module. The data is preprocessed by cleaning, noise reduction, and feature extraction, and then transmitted to the data center module for storage. At the same time, a knowledge base for the paper industry is built and updated to provide data support for the training and inference of large-scale language models. The virtual environment construction and dynamic simulation step (S2) involves constructing a high-fidelity three-dimensional virtual paper production line based on the static equipment parameters and dynamic operating data stored in the data center module. It also integrates a physical simulation engine to simulate the dynamic processes of material flow, energy flow, and information flow, as well as the equipment operating logic and fault evolution trends. The virtual environment is presented through visualization rendering technology and supports access to multi-channel immersive interactive devices. The large-scale language model training and optimization step (S3) involves training and optimizing the large-scale language model in the data center module. This includes using massive amounts of text data from the paper industry to pre-train the base model with domain knowledge, further training it using manually labeled instruction-response pairs through instruction fine-tuning technology, and introducing paper industry experts to evaluate and rank the quality of the content generated by the large model through reinforcement learning and alignment with human feedback for optimization. The model training process is continuously iterated. Immersive virtual training and multimodal interaction steps (S4): Employees enter the virtual paper production line constructed by the virtual space simulation module through the multimodal interaction interface for immersive training, and directly operate virtual equipment, observe instrument readings, adjust process parameters, and deal with simulated production anomalies. The virtual space simulation module feeds back the employees' operating behavior, virtual equipment status changes, and simulation results to the data center module in real time. The large-scale model intelligent interaction and deep reasoning step (S5) involves the large-scale model interaction and reasoning module receiving natural language queries from employees in virtual training in real time, as well as virtual environment status data and employee operation data provided by the data center module. It performs natural language understanding and intent recognition, and then combines the paper industry knowledge graph and real-time contextual data to perform deep reasoning. Personalized feedback and adaptive learning steps (S6): The large model interaction and reasoning module performs real-time personalized analysis based on employees' questions, operational behaviors, decision-making paths and error patterns, and provides customized and accurate error correction, principle explanation, heuristic question guidance, knowledge point expansion and dynamic adjustment of learning paths through natural language generation technology. The performance evaluation and training management steps (S7) involve the training service platform module generating detailed performance evaluation reports and improvement suggestions based on employees' performance in virtual training. Training managers can use the platform to view employee learning progress, manage training scenarios, update the knowledge base, and optimize the entire training system.

10. The method according to claim 9, characterized in that, The large-scale model intelligent interaction and deep reasoning step (S5) specifically includes: Context awareness and association are used to comprehensively analyze employee questions, current virtual device status, historical operation records, and simulated fault progress to accurately understand the contextual information of the current training scenario; Multi-factor fault diagnosis is used to address fault phenomena reported by employees. The large model uses its trained knowledge and reasoning ability to analyze possible causes from multiple dimensions, trace the causal chain of the fault, eliminate irrelevant factors, and form a preliminary diagnostic conclusion. Dynamic problem analysis and solution generation are used to dynamically analyze complex problems based on employee follow-up questions or changes in training scenarios, and generate multi-step, executable solutions. Knowledge integration allows for the integration of disparate knowledge of papermaking processes, equipment principles, and troubleshooting experience into a large-scale model, providing comprehensive cross-disciplinary and multi-disciplinary guidance in the process of solving complex problems.

Images